HANDBOOKS IN HEALTH

ECONOMIC EVALUATION

Decision Modelling for

Health Economic Evaluation

ANDREW BRIGGS

KARL CLAXTON

MARK SCULPHER
Decision modelling for health
economic evaluation
Handbooks in Health Economic Evaluation Series Decision modelling
Series Editors: Alastair Gray and Andrew Briggs

Forthcoming volumes in the series: for health economic

evaluation
Economic evaluation in 'clinical trials
Henry A Glick, lalpa A Doshi, Seema S Sonnad and Daniel Polsky

Andrew Briggs
Lindsay Chair in Health Policy & Economic Evaluation,
Section of Public Health & Health Policy,
University of Glasgow, UK

Karl Claxton
Professor of Economics, Department of Economics
and Related Studies and Centre for Health Economics,
University of York, UK

Mark Sculpher
Professor of Health Economics, Centre for Health
Economics, University of York, UK

OXFORD
UNIVERSITY PRESS
 For Eleanor, Zoe and Clare

OXFORD
UNIVERSITY I'RESS

Great Clarendon Street, Oxford OX2 6DP

Oxford University Press is a department of the University of Oxford.
It furthers the University's objective of exceUence in research, scholarship,
and education by publishing worldwide in
Oxford New York
Auckland Cape Town Dar es Salaam Hong Kong Karachi
Kuala Lumpur Madrid Melbourne Mexico City Nairobi
New Delhi Shanghai Taipei Toronto
With offices in
Argentina Austria Brazil Chile Czech Republic France Greece
Guatemala Hungary Italy Japan South Korea Poland Portugal
Singapore Switzerland Thailand Turkey Ukraine Vietnam

Oxford is a registered trade mark of Oxford University Press

in the UK and in certain other countries
Published in the United States
by Oxford University Press Inc., New York

© Oxford University Press 2006

The moral rights of the author have been asserted

Database right Oxford University Press (maker)
Reprinted 2007 (with corrections). 2011
All rights reserved. No part of this publication may be reproduced.
stored in a retrieval system, or transmitted, in any form or by any means,
without the prior permission in writing of Oxford University Press,
or as expressly pennitted by law, or under tenns agreed with the appropriate
reprographics rights organization. Enquiries concerning reproduction
outside the scope of the above should be sent to the Rights Department,
Oxford University Press, at the address above
You must not circulate this book in any other binding or cover
And you must impose this same condition on any acquirer

ISBN 978-0-19-852662-9

Printed and bound in Great Britain by

CPI Antony Rowe, Chippenham and Eastbourne
 Preface

Decision Modelling for Health Economic Evaluation

''All models are wrong .. some are useful"

George Box

This book reflects the increasing demand for technical details of the use of
decision models for health economic evaluation. The material presented in
this book has evolved from a number of sources, but the style of presentation
owes much to the three-day residential course on "Advanced Modelling
Methods for Health Economic Evaluation" that has been running since
September 2003, and which offered us the opportunity to tryout the material
on willing participants from a range of professional backgrounds. The book,
like the course, remains essentially practical, the aim has always been to
provide a help to analysts, perhaps struggling with modelling methods for the
first time, by demonstrating exactly how things might be done. For this
reason, many of the teaching examples) spreadsheet templates and solutions
are provided in the book and on the accompanying website. We are indebted
to a long list of course alumni who have helpfully ironed out many of the
typos) errors and gremlins that are an inevitable consequence of producing
such material.
In addition to the participants of the course itself, a number of people have
been instrumental in checking and commenting on previous drafts of mate­
rial at various stages of its development. Particular thanks must go to Elisabeth
Fenwick, who has been a constant presence from the inception of the very first
course and has commented from the very first draft of the course materials
right through to the final version of the book. We would also like to thank
Pelham Barton, lonathon Kamon, Karen Kuntz, and Gillian Sanders for
comments on specific chapters; Gemma Dunn and Garry Barton for checking
exercises and proofs as part of the final stages of development; and finally
Alastair Gray for comments and guidance in his role as series editor. Of
course) remaining errors are our own responsibility.
Finally, in seeing the final version of the manuscript published, we are
acutely aware of the fast-moving nature of the field. The opening quote was
made in the 1970s in relation to statistical modelling of data, but we believe it
applies just as well to decision modelling. The methods that we present in this
viii I PREFACE

book are just one way of looking at the world, it is not necessarily the correct
way and there may well be other equally valid approaches to handle some of
Series Preface
the issues we present as part of this book. Nevertheless, while future research
will undoubtedly offer improvements on the techniques we propose here, we
hope that the drawing together of this material in this form, if not entirely
correct, will at least prove usefuL
Economic evaluation in health care is a thriving international activity that is
AB,KC&MS increasingly used to allocate scarce health resources, and within which applied
May 2006 and methodological research, teaching, and publication are flourishing.
Several widely respected texts are already well-established in the market, so
what is the rationale for not just one more book, but for a series? We believe
that the books in the series Handbooks in Health Economic Evaluation share a
strong distinguishing feature, which is to cover as much as possible of this
broad field with a much stronger practical flavour than existing texts, using
plenty of illustrative material and worked examples. We hope that readers will
use this series not only for authoritative views on the current practice of
economic evaluation and likely future developments, but for practical and
detailed guidance on how to undertake an analysis. The books in the series are
textbooks, but first and foremost they are handbooks.
Our conviction that there is a place for the series has been nurtured by the
continuing success of two short courses we helped develop - Advanced
Methods of Cost-Effectiveness Analysis, and Advanced Modelling Methods for
Economic Evaluation. Advanced Methods was developed in Oxford in 1999
and has run several times a year ever since, in Oxford, Canberra and Hong Kong.
Advanced Modelling was developed in York and Oxford in 2002 and has also
run several times a year ever since, in Oxford, York, Glasgow and Toronto.
Both courses were explicitly designed to provide computer-based teaching
that would take participants through the theory but also the methods and
practical steps required to undertake a robust economic evaluation or
construct a decision-analytic model to current standards. The proof-of­
concept was the strong international demand for the courses - from academic
researchers, government agencies and the pharmaceutical industry and the
very positive feedback on their practical orientation.
So the original concept of the Handbooks series, as well as many of the
specific ideas and illustrative material, can be traced to these courses. The
Advanced Modelling course is in the phenotype of the first book in the series,
Decision Modelling for Health Economic Evaluation, which focuses on the role
and methods of decision analysis in economic evaluation. The Advanced
Methods course has been an equally important influence on Applied Methods
of Cast-Effectiveness, the third book in the series which sets out the key
x I SERIES PREFACE

elements of analysing costs and outcomes, calculating cost-effectiveness and

reporting results. The concept was then extended to cover several other Contents
important topic areas. First, the design, conduct and analysis of economic
evaluations alongside clinical trials has become a specialised area of activity
with distinctive methodological and practical issues, and its own debates and
controversies. It seemed worthy of a dedicated volume, hence the second book 1 Introduction 1
in the series, Economic Evaluation in Clinical Trials. Next, while the use of cost­
benefit analysis in health care has spawned a substantial literature, this is 2 Key aspects of decision modelling for economic evaluation 15
mostly theoretical, polemical, or-focused on specific issues such as willingness 3 Further developments in decision analytic models
to pay. We believe the fourth book in the series, Applied Methods of Cost­ for economic evaluation 45
Benefit Analysis in Health Care, fills an important gap in the literature by 4 Making decision models probabilistic 77
providing a comprehensive guide to the theory but also the practical conduct 5 Analysing and presenting simulation output from
of cost-benefit analysis, again with copious illustrative material and worked probabilistic models 121
examples. 6 Decision-making) uncertainty and the value of information 165
Each book in the series is an integrated text prepared by several contributing
authors, widely drawn from academic centres in the UK, the United States, 7 Efficient research design 201
Australia and elsewhere. Part of our role as editors has been to foster a consis­ 8 Future challenges for cost-effectiveness modelling
tent style, but not to try to impose any particular line: that would have been of health care interventions 225
unwelcome and also unwise amidst the diversity of an evolving field. News
and information about the series, as well as supplementary material for each
book, can be found at the series website: http://www.herc.ox.ac.uk/books

Alastair Gray Andrew Briggs

Oxford Glasgow
July 2006
 Chapter 1

Introduction

Economic evaluation is increasingly used to inform the decisions of various

health care systems about which health care interventions to fund from
available resources. This is particularly true of decisions about the coverage
or reimbursement of new pharmaceuticals. The first jurisdictions to
use economic evaluation in this way were the public health systems in
Australia and Ontario, Canada (Commonwealth Department of Health 1992;
Ministry of Health 1994); since then many others have developed similar
arrangements (Hjelmgren et al. 2001). In the UK, the National Institute
for Health and Clinical Excellence (NICE) has a wider purview in terms of
health technologies, and uses economic evaluation to inform decisions about
medical devices, diagnostic technologies and surgical procedures, as well as
pharmaceuticals (NICE 2004a). The ever-present need to allocate finite
resources between numerous competing interventions and programmes
means that economic evaluation methods are also used) to a greater or
lesser degree, at 'lower levels' within many health care systems (Hoffmann
et al. 2000).
The increasing use of economic evaluation for decision making has placed
some very clear requirements on researchers in terms of analytic methods
(Sculpher et al. 2005). These include the need to incorporate all appropriate
evidence into the analysis, to compare new technologies with the full range of
relevant alternative options and to reflect uncertainty in evidence in the
conclusions of the analysis. The need to satisfy these requirements provides
a strong rationale for decision analytic modelling as a framework for economic
evaluation. This book focuses on the role and methods of decision analysis in
economic evaluation. It moves beyond more introductory texts in terms of
modelling methods (Drummond et al. 2005), but seeks to ground this exposi­
tion in the needs of decision making in cOllectively-funded health care
systems. Through the use of a mixture of general principles, case studies and
exercises, the book aims at providing a thorough understanding of the latest
methods in this field, as well as insights into where these are likely to move
over the next few years.
2 I INTRODUCTION
COST�EFFECTJVENESS ANALYSIS I N H EALTH CARE I 3

In this introductory chapter, we seek to provide a brief overview of the main the basis of health-related quality of life (HRQL). There is also an implicit view
tenets of economic evaluation in health care) the needs of decision making that the current distribution of income is, if not optimal, then at least acceptable
and the rationale for decision analytic modelling. (Pauly 1995), and that the distributive impacts of health care programmes, and
the failure actually to pay compensation, are negligible.
1 . 1 . Defining economic evaluation Cost-benefit analysis is the form of economic evaluation that springs from
this theoretical paradigm, based on the concept of potential Pareto improve­
It is not the purpose of this book to provide a detailed introduction to ment (Sugden and Williams 1979). Cost-benefit analysis seeks to value the full
economic evaluation in health care in general. Good introductory texts are range of health and other consequences of a policy change and compare this
available (Sloan 1995; Gold (t al. 1996; Drummond et al. 2005 ) , and we with .resource costs as a form of compensation test. In health care) benefit
provide only an overview of the key aspects of these methods here. valuation has usually been in the form of contingent valuation or willingness
Economic evaluation in health care can be defined as the comparison to pay methods (Johanesson 1995; Pauly 1995). This involves seeking individuals'
of alternative options in terms of their costs and consequences (Drummond valuation of the consequences of health care programmes in terms of hypothet­
et al. 2005). Alternative options refer to the range of ways in which health care
ical monetary payment to be paid to obtain a benefit or to avoid a disbenefit
resources can be used to increase population health; for example) pharmaceutical (Gafni 1991; Diener et al. 1998).
and surgical interventions) screening and health promotion programmes. In A second disciplinary origin for economic evaluation in health care is in
this book terms like 'options') 'technologies» 'programmes) and 'interventions) operations research and management science. In general, this has taken the
are used interchangeably. Health care costs refer to the value of tangible form of constrained maximization: the maximization of a given objective
resources available to the health care system; for example, clinical and function subject to a set of constraints. However, it should be noted that this
other staff, capital equipment and buildings, and consumables such as drugs. view of economic evaluation is also consistent with the concept of social decision
Non-health service resources are also used to produce health care) such as the making which has been described in the economics literature as an alternative
time of patients and their families. Consequences represent all the effects of to standard welfare theory (Sugden and Williams 1979). There has also been
health care programmes other than those on resources. These generally focus some development in economics of the concept of (extra welfarism) as a
on changes in individuals' health) which can be positive or negative) but can normative framework for decision making (Culyer 1989). In essence, these
also include other effects that individuals may value, such as reassurance and non-welfarist perspectives take an exogenously defined societal objective and
information provision. budget constraint for health care. Cost-effectiveness analysis (CEA) is the
As is clear in the definition above) economic evaluation is strictly comparative. form of economic evaluation that has generally been used in health care to
It is not possible to establish the economic value of one configuration of apply these principles of resource allocation. It is possible to justify CEA
resources (e.g. the use of a particular medical intervention) unless its costs and within a welfare theoretic framework (Garber and Phelps 1997; Meltzer 1997;
consequences are compared with at least one alternative option. V\'einstein and Manning 1997), but generally it is the social decision making
view that has implicitly or explicitly provided the methodological foundations
1 . 2 . Alternative paradigms for economic evaluation of CEA in health.
It is possible to trace the disciplinary origins of economic evaluation back in
several directions. One direction relates to welfare economic theory (Ng 1983) 1 .3. Cost-effectiveness analysis in health care
which implies that health care programmes should be judged in the same way as
any other proposed change in resource allocation. That is, the only question is There is debate about the appropriate normative theory for economic evalua­
whether they represent a potential Pareto improvement in social welfare - could tion in health care. There can be little doubt, however, that some form of CEA
the gainers from a policy cbange compensate the losers and remain in a preferred predominates in terms of applied research in health (Pritchard 1998). In the
position compared with before the change? In the context of resource allocation context of health care, CEA would typically be characterized with a health­
in health care, therefore, welfare theory does not have has an interest solely in related objective function and constraints centred on a (narrow or broad)
whether policy changes improve health outcomes as measured, for example, on health care budget. There are many examples in the CEA literature which use
4 I INTRODUCTION T H E ROLE OF DECISION ANALYSIS IN ECONOMIC EVALUATION ! 5

measures of health specific to the disease or intervention under consideration. options generating constant returns to scale and being perfectly divisible
Examples of such measures are episode-free days (asthma) (Sculpher and (Birch and Gafni 1992).
Buxton 1993), true positive cases detected (breast cancer) (Bryan et al. 1995) Standard cost�effectiveness decision rules involve relating differences in
and percentage reduction in blood cholesterol (coronary heart disease) costs between options under comparison to differences in benefits. In the case
(Schulman et al. 1990). However, given the need in most health care systems of an option being dominant (costing less and generating greater effects than
to make resource allocation decisions across a whole range of disease areas, all the alternatives with which it is being compared), it is unequivocally cost­
CEA has increasingly been based on a single (,generic') measure of health. effective. However, if an option generates additional benefits but at extra cost
Although other measures have been suggested, the quality-adjusted life-year it can still be considered cost-effective. In such a situation, the option)s incre­
(QALY) is the most frequently used measure for this purpose. mental costs and effects are calculated and compared with those of other uses
The use of the QALY as the measure of effect in a CEA is often referred to as of health service resources. For example, if a new drug therapy for Aizheimer's
cost-utility analysis (Drummond et al. 2005). On the basis that health care disease is found to generate more QALYs than currently available treatment
programmes and interventions aim to impact on individuals) length of life options, but also to add to costs, then a decision to fund the new therapy will
and health-related quality of life, the QALY seeks to reflect these two aspects in involve opportunity costs falling on the health care system (i.e. the QALYs
a single measure. Various introductions to the QALY as used in CEA are available forgone from programmes or interventions which are removed or down­
(Torrance and Feeny 1989; Drummond et al. 2005), as well as more detailed scaled to fund the new drug). The analytical question is whether the QALYs
descriptions of the methods used to derive 'quality-weights' (often called generated from the new Alzheimer's therapy are greater than the opportunity
values or utilities) (Torrance 1986; Patrick and Erickson 1993) and of the costs. Given limited information on the costs and effects associated with the
assumptions underlying QALYs (Pliskin et al. 1980; Loomes and McKenzie full range of uses of health service resources, simplified decision rules have
1989). Despite its limitations, the QALY remains the only generic measure of centred on the calculation of an incremental cost-effectiveness ratio (ICER);
health that has been used in a large range of clinical areas. that is, the additional cost per extra unit of effect (e.g. QALY) from the more
In some introductory texts on economic evaluation in health) and in a effective treatment. When the lCER is compared with those of other interven­
number of applied studies, cost-minimization analysis is described as a separate tions. or with some notional threshold value which decision makers are
type of economic evaluation. This has been used under the assumption that (assumed to be) willing to pay for an additional unit of effect, the preferred
effects (on health and other possible attributes) do not differ between the option from those being evaluated can be established. A further concept in deci­
options under consideration. In such circumstances, the option with the lowest sion rules is (extended dominance). This can occur when three or more options
cost represents the greatest value for money. This is essentially a simplified are being compared and is present when an option has a higher lCER than a
CEA} but it has been criticized because) in many circumstances) the assumption more effective comparator. This concept is described further in the next chapter.
of equal effects is based on an erroneous interpretation of a statistical hypothesis
test and ignores the uncertainty which will invariably exist in the differential
effects between options (Briggs and O'Brien 200 1). 1 .4. The role of decision analysis in economic evaluation
As mentioned earlier, the origin of CEA is generally seen to be in constrained Decision analysis represents a set of analytic tools that are quite distinct from
optimization. In principle, the methods of identifying whether a particular cost-benefit analysis and CEA but can be seen as complementary to both.
technology is appropriate (i.e. cost-effective) would involve looking at each Decision analysis has been widely used in a range of disciplines including
and every use of resources and selecting those which maximize the health­ business analysis and engineering (Raiffa and Schlaifer 1959). In health care, it
related objective function subject to the budget constraint (Stinnett and Paltiel is an established framework to inform decision making under conditions of
1996). The information needed to implement these methods in full, however, uncertainty (Weinstein and Fineberg 1980; Sox et al. 1988; Hunink et al.
has so far made their practical use for decision making impossible. The result 2001). Decision analysis has been used more generally in health care evalua­
has been the use of simplified 'decision rules' for the identification of the most tion than in economic evaluation, in terms of informing clinical decisions at
cost-effective option from among those being compared (Johannesson and population and individual levels (McNeil et al. 1976; Schoenbaum et al. 1976;
Weinstein 1993). These are based on some simplifying assumptions, such as Gottlieb and Pauker 1981).
6 I INTRODUCTION T H E ROLE O F DECISION ANALYSIS I N ECONOMIC EVALUATION I 7

Basic concepts in decision modelling for CEA have been covered elsewhere decision modelling as a vehicle for evaluation. The strongest evidence for this
(Hunink et al. 2001; Drummond et al. 2005). Here we summarize some of the is probably the 2004 methods guidelines from th, National Institute for
key concepts and principles in the area. Clinical Excellence in the UK (NICE 2004b), but it is also apparent with other
decision makers.
1.4.1. What is decision modelling? In part, the increasing use of modelling in this context can be explained by
Decision analysis has been defined as a systematic approach to decision the required features of any economic evaluation seeking to inform decision
making under uncertainty (Raiffa 1968). In the context of economic evaluation, making. The key requirements of such studies are considered below.
a decision analytic model uses mathematical relationships to define
Synthesis
a series of possible consequences that would flow from a set of alternative
options being evaluated. Based on the inputs into the model, the likelihood of It is essential for economic evaluation studies to use aU relevant evidence. In
each consequence is expressed in terms of probabilities, and each consequence the context of parameters relating to the effectiveness of interventions, this is
has a cost and an outcome. It is thus possible to calculate the expected cost and consistent with a central tenet of evidence-based medicine (Sackett et al.
expected outcome of each option under evaluation. For a given option, the 1996). However, it should also apply to other parameters relevant to economic
expected cost (outcome) is the sum of the costs (outcomes) of each consequence evaluation, such as resource use and quality-of-life weights (utilities). Rarely
weighted by the probability of that consequence. will all relevant evidence come from a single source and, typically, it will have
A key purpose of decision modelling is to allow for the variability and to be drawn from a range of disparate sources. A framework is, therefore,
uncertainty associated with all decisions. The concepts of variability and needed within which to synthesize this range of evidence. This needs to
uncertainty are developed later in the book (particularly in Chapter 4). The provide a structure in which evidence can be brought to bear on the decision
way a decision model is structured will reflect the fact that the consequences problem. Hence it should provide a means to characterize the natural history
of options are variable. For example, apparently identical patients will respond of a given condition, the impact of alternative interventions and the costs and
differently to a given intervention. This might be characterized in terms, health effects contingent on clinical events. This framework will also include
for example, of dichotomous events such as 'response' and (no response' to the relationship between any intermediate clinical measure of effect and the
treatment. The model will be structured to reflect the fact that, for an individual ultimate measure of health gain required for CEA (Drummond et al. 2005).
patient, whether or not they respond will be unknown in advance. The like­ Consideration of all relevant comparators
lihood of a response will be expressed as a probability, which is a parameter to The cost-effectiveness of a given technology, programme or intervention can
the model. However, the estimation of this parameter is uncertain and this only be achieved in comparison with all alternative options that could feasibly
should also be allowed for in the model using sensitivity analysis. be used in practice. These alternatives could relate to different sequences
As a vehicle for economic evaluation, the 'decision' relates to a range of of treatments and/or stop-go decision rules on intervention. In most instances,
resource allocation questions. Examples of these include: Should a collectively­ a single study, such as a randomized trial, will not compare all alternatives
funded health system fund a new drug for Alzheimer's disease? What is the relevant to the economic evaluation. There will, therefore, be a need to bring
most cost-effective diagnostic strategy for suspected urinary tract infection in together data from several clinical studies using appropriate statistical synthesis
children? Would it represent a good use of our limited resources to undertake methods (Sutton and Abrams 2001; Ades et al. 2006). Again, the decision
additional research regarding one or more parameters in our decision model? model provides the framework to bring this synthesis to bear on the decision
1.4.2. The role of decision modelling problem.
for economic evaluation Appropriate time horizon
Decision analysis has had a controversial role in economic evaluation in For decision making, economic evaluation requires that studies adopt a time
health care (Sheldon 1996; Buxton er al. 1997). However, the growing use of horizon that is sufficiently long to reflect all the key differences between
economic evaluation to inform specific decision problems facing health care options in terms of costs and effects. For many interventions, this will effec­
decision makers (Hjelmgren et al. 2001) has seen an increased prominence for tively require a lifetime time horizon. This is particularly true of interventions
8 I INTRODUCTION SUM MARY AND STRUCTURE OF T H E BOOK I 9

with a potential mortality effect, where life expectancy calculations require full have been based on data from a single trial (www.york.ac.ukiinsticrd).
survival curves to be estimated. Economic evaluations based on a single Given the requirements of decision making described in the last section,
source of patient-level data (e.g. a randomized trial or observational study) however - in particular, the need to include all relevant evidence and compare
will rarely have follow-up which is sufficiently long to facilitate a lifetime time all appropriate options - the appropriate role for an economic study based on
horizon. Again, the decision model becomes the framework within which to a single triai, rather than a decision model, remains a source of debate.
structure the extrapolation of costs and effects over time. There are two
elements to this. The first relates to extending baseline costs and effects
beyond the primary data source, where this may relate to natural history or 1 .5 . Summary and structure of the book
one of the 'standard' therapies being evaluated (i.e. baseline effects). The This introductory chapter provides an overview of the key concepts behind
second element concerns the continuation of the treatment effect; that is, the economic evaluation ofhealth care in general, and the role of decision analysis
effectiveness of the interventions being evaluated relative to the baseline. in this context. It has described the alternative theoretical origins lying behind
economic evaluation - welfare economic theory and extra-welfarist
Uncertainty
approaches including social decision making. These alternative paradigms
A key requirement of economic evaluation for decision making is to indicate largely explain the existence of competing evaluation methods in economic
how uncertainty in the available evidence relating to a given policy problem evaluation - cost-benefit analysis and cost-effectiveness analysis - although
translates into decision uncertainty; that is, the probability that a given decision the latter predominates in the applied literature. The role of decision analysis
is the correct one. A key objective of this book is to show how probabilistic in economic evaluation can be seen as independent of either of these types of
decision models can fully account for this uncertainty, present this to decision evaluation although it can complement both approaches. Within the social
makers and translate this into information about the value and optimal design decision making paradigm, the requirements of decision making regarding
of additional research. resource allocation decisions emphasize the value of decision models. In
particular, the need to synthesize all relevant evidence and to compare all
1 .4.3. Models versus trials options over an appropriate time horizon necessitates the need for a decision
The importance of the randomized controlled tria! in generating evidence for analytical framework.
the evaluation of health care programmes and interventions has seen it In the remainder of this book, we develop further the methods of decision
develop a role as a vehicle for economic evaluation. That is, the trial provides analysis for economic evaluation. Chapter 2 focuses on the cohort model and,
the sole source of evidence on resource use and health effects that, together in particular, the Markov model, a widely used form of decision model in
with external valuation data (in the form of unit costs and utilities), forms the CEA. Chapter 3 considers a range of possible extensions to the standard
basis of the estimate of cost-effectiveness. It has been well recognized that Markov model including reflecting time dependency and the use of patient­
many trials exhibit weaknesses when used in this way, particularly 'regulatory level simulations. Chapter 4 describes the methods for handling uncertainty in
trials' designed to support licensing applications for new pharmaceuticals decision models, in particular, the use of probabilistic sensitivity analysis to
(Drummond and Davies 1991). As noted previously, this includes the limited reflect parameter uncertainty. Chapter 5 shows how decision uncertainty and
number of comparisons, short follow-up and a failure to collect all the heterogeneity in a model can be presented to decision makers using methods
evidence needed to address cost-effectiveness. As a result of these limitations, such as cost-effectiveness acceptability curves. Chapter 6 builds on the
there has been extensive consideration of the appropriate design and analysis concepts in earlier chapters to show how probabilistic decision models can be
of trials for economic evaluation - these have variously been termed prag­ used to quantify the cost of decision uncertainty and hence the value of addi­
matic or 'real world' studies (Freemantle and Drummond 1997; Coyle et al. tional information as a basis for research prioritization. Chapter 7 extends the
1998; Thompson and Barber 2000). value-of-information methods by looking at how probabilistic decision
A large proportion of economic evaluation studies could be described models can be used to identify the most efficient design of future research
as trial-based economic evaluations: since 1994, approximately 30 per cent of studies. Chapter 8 draws together the key conclusions from the preceding
published economic evaluations on the NBS Economic Evaluation Database chapters.
10 I INTRODUCTION REFERENCES 111
A website for the book has been set up to give the reader access to the Excel
1 ,6, Exercises templates that form the basis of the exercises, The web address is
One of the features of this book is the emphasis on practical exercises that are www.herc.ox.ac.uk/bookslmodelling.html and contains links to the exercise
designed to illustrate the application of many of the key issues covered in each templates labelled by exercise number together with the solution files.
chapter. These exercises are integrated into Chapters 2-6 and are based References
around the development of two example models, building in sophistication as
Ades, A. E., Sculpher, M. J., Sutton, A., Abrams, K., Cooper, N., Welton, N., et al. (2006)
the book progresses. The first example is a replication of a previously 'Bayesian methods for evidence synthesis in cost�effectiveness analysis',
published model of combination therapy for HIV/AIDS (Chancellor et al. PharmacoEconomics24: 1-19.
1 997), which, although somewhat out of date in terms of the treatment under Birch, S. and Gafni, A. (1992) 'Cost effectiveness/utility analyses: do current decision rules
evaluation, nevertheless represents a useful example that serves to illustrate a lead us to where we want to be?: Journal of Health Economics, 11: 279-296.
number of general modelling issues. The second example is a slightly simplified Briggs, A. H. and O'Brien, B. J.(2001) 'The death of cost�minimisation analysis?: Health
version of a model examining the cost-effectiveness of a new cemented hip Economics, 10: 179-184.

prosthesis compared with the standard prosthesis that has been used in the Briggs, A., Sculpher, M., Dawson, J., Fitzpatrick, R" Murray, D. and Ma1chau, H.(2004) 'Are
new cemented prostheses cost-effective? A comparison of the Spectron and the
NHS for many years (Briggs et al. 2004). The practical exercises are facilitated Charnley'. Applied Health Economics & Health Policy, 3(2): 78-89.
by a series of Microsoft ExceFM templates that set out the structure of the Bryan, S., Brown, J. and Warren, R.(1995) 'Mammography screening: an incremental cost­
exercise and a set of solutions for each exercise. In most cases the templates in effectiveness analysis of two-view versus one-view procedures in London', Journal of
one chapter are based on the solution for the preceding chapter, such that the Epidemiology and Community Health, 49: 70-78.
exercises build to a comprehensive evaluation of the decision problem. Buxton, M. L Drummond, M. F.. Van Hout, B. A., Prince, R. L., Sheldon, T, A" Szucs, T. and
It is worth being clear why we have chosen to use Excel as a platform for the Vray, M, (1997) 'Modelling in economic evaluation: an unavoidable fact ofUfe', Health
Economics,6: 217-227.
exercises and recommend it more widely for undertaldng cost-effectiveness
Chancellor, J, V., Hill, A. M., Sabin, C. A., Simpson, K. N. and Youle, M.(1997) 'Modelling
modelling. Firstly, Excel is the most popular example of spreadsheet software the cost effectiveness oflamivudine/zidovudine combination therapy in HIV infection',
(much of what we cover in the exercises is directly transferable to other PharmacoEconomics,12: 54-66.
spreadsheet packages). Although there are a number of good dedicated decision Commonwealth Department of Health, H. a. C. S. (1992) Guidelines for the pharmaceutical
analysis packages available, in our experience none is capable of all the industry on preparation of submissions to the Pharmaceutical Benefits Advisory
functions and presentation aspects of many full health economic models. Committee, Canberra, APGS.

Furthermore, where some functions are available (such as the ability to corre­ Coyle, D., Davies, 1. and Drummond, M. (1998) 'Trials and tribulations - emerging issues
in designing economic evaluations alongside clinical trials', International Journal of
late parameters) the application can become something of a 'black box' - it is Technology Assessment in Health Care, 14; 135-144.
always worth knowing how to implement such methods from first principles. Culyer, A. J,(1989) 'The normative economics of health care finance and provision', Oxford
There are also a number of popular 'add-ins' for Excel, such as Crystal Ball and Review of Economic Policy,5: 34-58.
@Risk. These programs add to the functionality of Excel, particularly in Diener, A., O'Brien) B. and Gafni, A, (1998) 'Health care contingent valuation studies: a
relation to simulation methods. While these are much less of a black box, in review and classification of the literature', Health Economics, 7: 313-326.
that they can be employed to complement existing spreadsheet models, there Drummond, M. F. and Davies, L.(1991) 'Economic analysis alongside clinical trials',
is a problem in that models built with these add-ins can only be used by other International Journal a/Technology Assessment in Health Care, 7: 561-573 .

people with the add-in software. This can severely limit the potential user base Drummond, M, E, Sculpher, M, J., Torrance, G. W., O'Brien, B. and Stoddart, G. L (2005)
Methods for the economic evaluation of health care programmes. Oxford, Oxford
for a model. For these reasons we have chosen to demonstrate how the basic University Press,
Excel package can be used for the complete modelling process. This has the Freemantle, N. and Drummond, M. (1997) 'Should clinical trials with concurrent
advantage that the exercises are concerned with implementing the methods economic analyses be blinded', Journal of the American Medical Association,277:63-64.
from first principles rather than coaching the reader in the use of software to Gafni, A. (1991) 'Willingness to pay as a measure of benefits', Medical Care,29: 1246--1252,
implement the methods. That said, some basic familiarity with Excel operations Garber, A. M, and Phelps, C. E.(1997) 'Economic foundations of cost-effectiveness analysis'.
is assumed. 16: 1-31.
12 I INTRODUCTION REFERENCES 13

Gold, M. R., Siegel, J. E., Russell, 1. B. and Weinstein, M. C. (1996) Cost-effectiveness in Sackett, D. 1., Rosenberg, W. M. c., Gray, J.A. M., Haynes, R. B. and Richardson, W. S.
health and medicine. New York, Oxford University Press. (1996) 'Evidence-based medicine: what it is and what it isn't', British Medical Journal,
Gottlieb, J. E. and Pauker, S. G. (1981) 'Whether or not to administer amphotericin to an 312' 71-72.
immunosuppressed patient with hematologic malignancy and undiagnosed fever: Scho.enbaum, S. c., McNeil, B. J. and Kavet, J. (1976) 'The swine-influenza decision',
Medical Decision Making, 1: 569-587. New England Journal ofMedicine,295: 759-765.
Hjelmgren, J., Berggren, F. and Andersson, F. (2001) 'Health economic guidelines - similarities, Schulman, K. A., Kinosian, B., Jacobson, T. A., Glick H., Willian M. K., Koffer H.
differences and some implications', Value in Health, 4: 225-250. and Eisenberg J. M. (1990) 'Reducing high blood cholesterol level with drugs:
Hoffmann, c., Grafvon der Schulenburg, J.-M. and on behalf of the EUROMET group cost-effectiveness of pharmacologic management', Journal of the AmericcU/ Medical
(2000) 'The influence of economic evaluation studies on decision making: a European Association,264: 3025-3033.
survey', Health Policy, 52: 179-192. Sculpher, M. J. and Buxton, M. J. (1993) 'The episode-free day as a composite measure of
Hunink, M., Glaziou, P., Siegel, J., Weeks, J., PUskin, J., Elstein, A., et al. (2001) Decision effectiveness', PharmacoEconomics, 4: 345-352 .
making in health and medicine. Integrating evidence and values. Cambridge, Cambridge Sculpher, M., Claxton, K. and Akehurst, R.(2005) 'It's just evaluation for decision making:
University Press. recent developments in, and challenges for, cost-effectiveness research' in P. C. Smith,
Johanesson, P. O. (1995) Evaluating health risks. Cambridge University Press, Cambridge. 1. Ginnelly and M. Sculpher (eds) Health policy and economics. opportunities and
Johannesson, M. and Weinstein, S.(1993) 'On the decision rules of cost-effectiveness analysis', challenges. Maidenhead, Open University Press.
Journal ofHealth Economics,12: 459-467 . Sheldon, T. A.(1996) 'Problems of using modelling in the economic evaluation of health
Loomes, G. and McKenzie, L. (1989) 'The use of QALYs in health care decision making', care', Health Economics, 5: 1-11.
Social Science and Medicine,28: 299-308. Sloan, F. E. (1995) Valuing health care: costs, benefits and effectiveness
McNeil, B. J., Hessel, S. J., Branch, W. T. and Bjork, 1. (1976) 'Measures of Clinical and other medical technologies. Cambridge, Cambridge University Press.
Efficiency. III. The Value of the lung scan in the evaluation ofyoung patients with pleuritic Sox, H. c., Blatt, M. A, Higgins, M. C. and Marton, K. L (1988) Medical decision making.
chest pain', Journal ofNuclear Medicine, 17(3): 163-169. Stoneham, MA, Butterworths.
Meltzer, D.(1997) 'Accounting for future costs in medical cost-effectiveness analysis', Stinnett, A. A. and Paltiel, A. D. (1996) 'Mathematical programming for the efficient
Journal of Health Economics, 16: 33-64. allocation of health care resources') Journal of Health Economic� 15: 641-653.
Ministry of Health(1994) Ontario guidelines for economic analysis ofpharmaceutical products. Sugden, R. and Williams, A. H. (1979) The principles ofpractical cost-benefit analysis.
Ontario, Ministry of Health. Oxford, Oxford University Press.
National Institute for Clinical Excellence (NICE) (2004 a) Guide to technology appraisal Sutton, A. J. and Abrams, K. R.(2001) 'Bayesian methods in meta-�malysis and evidence
process. London, NICE. synthesis', Statistical methods in medical research,10: 277-303.
National Institute for Clinical Excellence (NICE) (2004 b) Guide to the method;; of technology Thompson, S. G. and Barber, J. A.(2000) 'How should cost data in pragmatic randomised
appraisal. London, NICE. trials be analysed?', British Medical Journal, 320: 1197-1200.
Ng, Y. K. (1983) Welfare economics: introduction and development of basic concepts. London, Torrance, G. W.(1986) 'Measurement of health state utilities for economic appraisal­
Macmillan. a review� Journal ofHealth Economics, 5: 1-30.
Patrick, D. L. and Erickson, P.(1993) Health status and health policy. Allocating resources to Torrance, G. W. and Feeny. D.(1989) 'Utilities and quality-adjusted life years', International
health care. New York, Oxford University Press. Journal of Technology Assessment in Health Care,5: 559-575.
Pauly, M. V. (1995) 'Valuing health benefits in monetary terms' in F. A. Sloan (ed.) Valuing Weinstein, M. C. and Fineberg, H. V. (1980) Clinical decision analysis. Philadelphia, PA,
health care: costs, benefits and efef ctiveness WB Saunders Company. ofpharmaceuticals and other medical tec
Cambridge, Cambridge University Press. Weinstein, M. C. and Manning, W. G. (1997) 'Theoretical issues in cost-effectiveness
PUskin, J. S., Shepard, D. S. and Weinstein, M. C. (1980) 'Utility functions for life years and analysis', Journal ofHealth Economics, 16: 121-128.
health status', Operations Research, 28(1): 206-224.
Pritchard, C. (1998) 'Trends in economic evaluation'. OHE Briefing 36. London, Office of
Health Economics.
Raiffa, H. (1968) Decision analysis: introductory lectures on choices under uncertainty.
Reading, MA, Addison-Wesley.
Raiffa, H. and Schlaifer, R.(1959) Probability and statistics for business decisions. New York,
McGraw-Hill.
 Chapter 2

Key aspects of decision modelling

for economic evaluation

This chapter considers the basic elements of decision modelling for economic
evaluation. It considers the key stages in developing a decision analytic model
and describes the cohort model, the main type of decision model used in the
field. The decision tree and Markov model are described in detail and exam­
ples. provided of their use in economic evaluation.

2.1 . The stages of developing a decision model

It is possible to identify a series of stages in developing a decision model for
economic evaluation. In part, this will involve some general choices concern­
ing the nature of the evaluation. This will include the measure of effect and
the time horizon) but also the perspective of the analysis; that is, whose costs
and effects we are interested in? Below is a list of the stages in the development
process which relate specifically to the decision modelling.
2 1 . 1 . Specifying the decision problem
This involves clearly identifying the question to be addressed in the analysis.
This requires a definition of the recipient population and subpopulations. This
will typically be the relevant patients, but may include nonpatients (e.g. in the
case of screening and prevention programmes). This requires specific details
about the characteristics of individuals, but should also include information
about the locations (e.g. the uK NHS) and setting (e.g. secondary care) in
which the options are being delivered. The specific options being evaluated
also need to be detailed. These will usually be programmes or interventions,
but could include sequences of treatments with particular starting and
stopping rules.
Part of the definition of the decision problem relates to which institution(s)
islare (assumed to be) making the relevant decision. In some cases· this will be
explicitly stated -for example, in the case of a submission to a reimbursement
agency- but it will often have to be implied by the characteristics of the evalu­
ation) such as the sources of data used.
16 I KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION THE STAGES OF DEVELOPING A DECISION MODEL I 17

2 . 1 .2. Defining the boundaries of the model these definitions. This i s particularly important when modelling a baseline
All models are simplifications of reality and it will never be possible for (e.g. disease progression without treatment) and in extrapolating beyond the
a model to include all the possible ramifications of the particular option being data from randomized trials.
considered. Choices need to be taken, therefore, abollt which of the possible There are no general rules about appropriate model structure in a given
consequences of the options under evaluation will be formally modelled. situation. However, some of the features of a disease/technology that are likely
For example, should the possible implications of antibiotic resistance be to influence choices about structure include:
assessed in all economic evaluations of interventions for infectious diseases? • Whether the disease is acute or chronic and, if the latter} the number of
Another example relates to whether or not to include changes in disease possible health-related events which could occur over time.
transmission resulting from screening programmes for HIV. It has been shown • Whether the risks of events change over time or can reasonably be assumed
that including reductions in the horizontal transmission of HIV in such to be constant.
models has a marked impact on the cost-effectiveness of screening (Sanders • Whether the effectiveness of the intervention(s) (relative to some usual
et al. 2005). care baseline) can be assumed constant over time or time-limited in
some way.
2 . 1 .3. Structuring a decision model • If and when treatment is stopped, the appropriate assumptions about
Given a stated decision problem and set of model boundaries, choices have to the future profile of those changes in health that were achieved during
be made about how to structure the possible consequences of the options treatment. For example, would there be some sort of ,rebound' effect or
being evaluated. In part, this will be based on the nature of the interventions would the gains} relative to a comparator group, be maintained over time
themselves. For example, for an economic evaluation of alternative diagnostic (Drummond et' al. 2005).
strategies for urinary tract infection in children, it was necessary to use a • Whether the probability of health-related events over time is dependent on
complex decision tree to reflect the prior probability (prevalence) and diagnos­ what has happened to 'a patient' in the past.
tic accuracy (sensitivity and specificity) of the various single and sequential
screening tests (Downs 1999). In part, model structure will reflect what is 2 . 1 .4. Identifying and synthesizing evidence
known about the natural history of a particular condition and the impact of
The process of populating a model involves bringing together all relevant
the options on that process; for example, the future risks faced by patients
surviving a myocardial infarction and the impact of options for secondary evidence, given a selected structure, and synthesizing it appropriately in terms
of input parameters in the model. Consistent with the general principles of
prevention on those risks. evidence-based medicine (Sackett et al. 1996), there needs to be a systematic
As a general approach to structuring a decision model, there is value in approach to identifying relevant evidence. Evidence synthesis is a key area of
having some sort of underlying biological or clinical process driving the clinical evaluation in its own right (Sutton et al. 2000) which is of importance
model. Examples of the former include the use of CD4 counts or viral load in
HIV models (Sanders et al. 2005). The latter approach is more common and outside the requirements of economic evaluation. However, the requirements
of decision analytic models for economic evaluation have placed some
examples include the use of the Kurtzke Expanded Disability Status Scale in
multiple sclerosis (Chilcott et al. 2003), the Mini Mental State Examination in important demands on the methods of evidence synthesis. These include:
Alzheimer's disease (Neumann et al. 1999) and clinical events, such as myocar­ • The need to estimate the effectiveness of interventions despite the absence
dial infarction and revascularization in coronary heart disease (Palmer et al. of head-to-head randomized trials. This involves the use of indirect and
2005). The cost-effectiveness of the relevant interventions can then be assessed mixed treatment comparisons to create a network of evidence between
by attaching health-related quality-of-life weights and costs to states or path­ trials.
ways defined in this way. The advantage of using these biologically- or clini­ • The need to obtain probabilities of clinical events for models over a
cally-defined states is that they should be well-understood. In particular, there standardized period of follow-up despite the fact that clinical reports
should be good evidence about the natural history of a disease in terms of present these over varying follow-up times.
18 1 KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION THE STAGES OF DEVELOPING A DECISION MODEL 119

• The need for estimates of treatment effectiveness in terms of a common Box 2,.1 . Key concept in understanding uncertainty
endpoint although trials report various measures.
and.heterogeneity in de'cision models for
• The need to assess heterogeneity in measures between different . cost-effectiveness analysis
types of patients. Ideally this would be undertaken using individual . .
.
.

patient data, but metaregression can be used with summary data in some Variability: Individual patients will inevitably differ from one another
situations. in terms, for example, of the clinical events that they experience and the
These issues in evidence synthesis are being tackled by statisticians, often associated health-related quality of life. This variability cannot be reduced
within a Bayesian framework (Sutton and Abrams 200 1 ; Ades 2003; through the collection of additional data.
Spiegelhalter et al. 2004), and these are increasingly being used in decision Parameter uncertainty: The precision with which an input parameter
models for economic evaluation (Ades et al. 2006). An important area of is estimated (e.g. the probability of an event, a mean cost or a mean utility).
methodological research in the field relates to incorporating evidence synthesis The imprecision is a result of the fact that input parameters are estimated
and decision modelling into the same analytic framework - 'comprehensive for populations on the basis of limited available information. Hence uncer­
decision modelling' (Parmigiani 2002; Cooper et al. 2004). This has the advan­ tainty can, in principle, be reduced through the acquisition of additional
tage of facilitating a fuller expression of the uncertainty in the evidence base in evidence.
the economic evaluation. Decision uncertainty: The joint implications of parameter uncertainty in
a model result in a distribution of possible cost-effectiveness relating to the
2 . 1 .5. Dealing with uncertainty and heterogeneity options under comparison. There is a strong normative argument
Uncertainty and heterogeneity exist in all economic evaluations. This for basing decisions, given available evidence, on the expectation of this
is an area of economic evaluation methodology that has developed rapidly distribution. But the distribution can be used to indicate the probability
in recent years (Briggs 2001), and its implications for decision modelling that the correct decision has been taken.
represent an important element of this hook. Chapters 4 and 5 provide more Heterogeneity: Heterogeneity relates to the extent to which it is possible
detail about appropriate methods to handle uncertainty and heterogeneity. to explain a proportion of the interpatient variability in a particular meas­
Box 2 . 1 summarizes the key concepts, and these are further developed in urement on the basis of one or more patient characteristics. For example, a
Chapter 4. particular clinical event may be more likely in men and in individuals aged
over 60 years. It will then be possible to estimate input parameters
2 . 1 .6. Assessing the value of additional research (and cost-effectiveness and decision uncertainty) conditional on a
patient's characteristics (subgroup estimates) although uncertainty in
The purpose of evaluative research, such as randomized control trials) is to
those parameters will remain.
reduce uncertainty in decision making by measuring one or more parameters
(which may be specific to particular subgroups) with greater precision. This is
generally true in clinical research, but also in assessing cost-effectiveness.
Given limited resources) it is just as appropriate to use decision analytic of parameters. Formal value-of-information methods are considered fully in
models to assess the value for money of additional research projects as to Chapters 6 and 7. These methods have the strength of reflecting the joint
assess alternative approaches to patient management. In quantifying the uncertainty in all parameters. They also assess the extent to which reduction
decision uncertainty associated with a particular comparison) decision models in uncertainty through additional research would result in a change in deci­
can provide a framework within which it is possible to begin an assessment of sion about the lise of a technology and) if there is a change) its value in terms
the cost-effectiveness of additional research. This can be undertaken infor­ of improved health andlor reduced costs.
mally using simple sensitivity analysis by assessing the extent to which a Each of these stages is crucial to the development of a decision model that is
moders conclusions are sensitive to the uncertainty in one (or a small number) fit for the purpose of informing real policy decisions.
20 I KEY ASPECTS OF DECISION MODELLING fOR ECONOMIC EVALUATION SOME INTROD UCTORY CONCEPTS I N DECISION ANALYSIS I 21

2 . 2 . Some i ntroductory concepts in decision analysis model will indicate a number of mutually exclusive 'prognoses' for a given
Decision analysis is based on some key 'building blocks' which are common to patient and option (more generally, these are alternative 'states of the world'
all models. These are covered more fully in introductory texts (Weinstein and that could possibly occur with a given option). Depending on the type
Fineberg 1980; Hunink et al. 200 1 ; Drummond et al. 2005), and are only of model, these prognoses may be characterized) for example, as alternative
pathways or sequences of states. For a given option, the likelihood of each
summarized here.
possible prognosis can be quantified in terms of a probability, and their impli­
2.2. 1 . Probabilities cations in terms of cost and/or some measure of outcome. The calculation of
an expected value is shown in Box 2.2 using the example of costs. It is derived
In decision analysis, a probability is taken as a number indicating the likeli­
by adding together the cost of each of the possible prognoses weighted by the
hood of an event taking place in the future. As such, decision analysis shares probability of it occurring. This is analogous to a sample mean calculated on
the same perspective as Bayesian statistics (O'Hagan and Luce 2003). This the basis of patient-level data.
concept of probability can be generalized to represent a strength of belief
which, for a given individual, is based on their previous knowledge and expe­ 2 . 2 . 3 . Payoffs
rience. This more 'subjective) conceptualization of probabilities is consistent
with the philosophy of decision analysis, which recognizes that decisions As described in the previous section, each possible 'prognosis' or 'state of the
cannot be avoided just because data are unavailable to inform them, and world' can be given some sort of cost or outcome. These can be termed
'expert judgement' will frequently be necessary. 'payoffs', and expected values of these measures are calculated. The origins of
Specific probability concepts frequently used in decision analysis are:
• Joint probability. The probability of two events occurring concomitantly.
In terms of notation, the joint probability of events A and B occurring is Box 2.2. An illustration of the concept of expected
peA and B).
�alues using costs
• Conditional probability. The probability of an event A given that an event B ,
is known to have occurred. The notation is P(AIB).
I
Prognosis 1 Cost 25
Independence. Events A and B are independent if the probability of event A,
=

p eA), is the same as the probability of P(AIB). When the events are inde­ Probability = 0.4
pendent peA and B) = peA) x PCB).
Joint and conditional probabilities are related in the following equation:
I
• Prognosis 2 Cost 50
=

peA and B) = P(AIB) x PCB). Sometimes information is available on the Probability = 0.2
joint probability, and the above expression can be manipulated to 'condi­
tion out' the probabilities.
Prognosis 3 Cost 100
=

2.2.2. Expected values Probability 0.1

=

Central to the decision analytic approach to identifying a 'preferred' option

from those being compared under conditions of uncertainty is the concept of

I
expected value. If the options under comparison relate to alternative treatments Prognosis 4 Cost 75
=

for a given patient (or an apparently homogeneous group of patients), then Probability 0.3

J
=

the structure of the decision model will reflect the variability between patients
in the events that may occur with each of the treatments. The probabilities Expected cost (25 0.4) + (50 0.2) + (100 0.1 ) + (75 0.3) 52.50
= x x x x =

will show the likelihood of those events for a given patient. On this basis, the
22 I KEY ASPECTS OF DECISION MODELLING FOR ECONOM!C EVALUATION
COHORT MODELS I 23
decision analysis are closely tied to those of expected utility theory (Raiffa should explicitly consider the individual patient and allow for variability
1968), so the standard payoff would have been a 'utility' as defined by von
between patients. As described previously, the focus of economic evaluation is
Neumann-Morgenstern (von Neumann and Morgenstern 1944). In practice, on expected costs and effects, and uncertainty in those expected values. This
this would equate to a utility based on the standard gamble method of preference has resulted in most decision models focusing on the average patient experi­
elicitation (Torrance 1986). As used for economic evaluation in health care, ence these are referred to as cohort models. In certain circumstances, a more
the payoffs in decision models have been more broadly defined. Costs would appropriate way of estimating expected values may be to move away from the
typically be one form of payoff but, on the effects side, a range of outcomes cohort model, to models focused on characterizing variability between
may be defined depending on the type of study (see Chapter I). Increasingly, patients. These 'micro simulation' models are discussed in Chapter 3, but the
quality-adjusted life-years would be one of the payoffs in a decision model for focus of the remainder of this chapter is on cohort models.
cost-effectiveness analysis, which may or may not be based on utilities elicited The two most common forms of cohort model used in decision analysis for
using the standard gamble. economic evaluation are the decision tree and the Markov model. These are
The principle of identifying a preferred option on the basis of a decision considered below.
analytic model is on the basis of expected values. "Vhen payoffs are defined in
terms of ' von Neumann-Morgenstern utilities', this would equate with a 2.3 . 1 . The decision tree
preferred option having the highest expected utility; this is consistent with The decision tree is probably the simplest form of decision model. Box 2.3
expected utility theory as a normative framework for decision making under provides a brief revision of the key concepts using a simple example from the
uncertainty. Although a wider set of payoffs are used in decision models for management of migraine (Evans et al. 1997); the decision tree has been
economic evaluation, the focus on expected values as a basis for decision described in more detail elsewhere (Hunink et al. 200 1 ; Drummond et al.
making remains. This follows the normative theory presented by Arrow and 2005). The key features of a decision tree approach are:
Lind ( 1 970) arguing that public resource allocation decisions should exhibit
risk neutrality. For example, in cost-effectiveness analysis, the common incre­ • A square decision node -typically at the start of a tree -indicates a decision
mental cost-effectiveness ratio would be based on the differences between point between alternative options.
options in terms of their expected costs and expected effects. However, the + A circular chance node shows a point where two or more alternative events
uncertainty around expected values is also important for establishing the for a patient are possible; these are shown as branches coming out of the
value and design of future research, and this should also be quantified as part node. For an individual patient, which event they experience is uncertain.
of a decision analytic model. The methods for quantifying and presenting • Pathways are mutually exclusive sequences of events and are the routes
uncertainty in models are described in Chapters 4 and 5, respectively; and the through the tree.
uses of information on uncertainty for research prioritization are considered + Probabilities show the likelihood of a particular event occurring at a
in Chapters 6 and 7. chance node (or the proportion of a cohort of apparently homogeneous
patients expected to experience the event). Moving left to right, the first
probabilities in the tree show the probability of an event. Subsequent prob­
2.3. Cohort models
abilities are conditional; that is, the probability of an event given that an
The overall purpose of a model structure is to characterize the consequences earlier event has or has not occurred. Multiplying probabilities along path­
of alternative options in a way that is appropriate for the stated decision prob­ ways estimates the pathway probability which is a joint probability (as
lem and the boundaries of the model. The structure should also be consistent discussed previously).
with the key features of the economic evaluation, such as the perspectivE, time Expected costs and outcomes (utilities in Box 2.3) are based on the princi­
horizon and measure of outcome. There are several mathematical approaches ples in Box 2.2. Expected values are based on the summation of the pathway
to decision modelling from which the analyst can choose. One important values weighted by the pathway probabilities.
consideration is whether the model should seek to characterize the experience A somewhat more complicated decision tree model comes from a
of the 'average' patient from a population sharing the same characteristics, or cost-effectiveness analysis of alternative pharmaceutical therapies for
24 I KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUAT ION COHORT MODELS I 25

Box 2.3. Example of a decision tree based on Evans

(1997) .
•

et al.
.

Pathway
No recurrence (0 594)
A
Relief (0.558) ),
T Recurrence (OA06)
B
Sumatriptan
Endures attack (0.92)
c
No reHef l0.442) ), Relief (0.998)
o
T ER (0.08lA
Migraine T Hospita!isalion (0.002) E
attack

No recurrence (0.703)
F
Relief (0.379) ), Recurrence (0.297)
Caffeinel
G
ergotamine
Endures attack (0.92)
Relief 12:..998)
H
No reHel (0.621) )..
IER (0.08)
'I Hospitalisalion (0.002) J
Pathway
Sumatriptan Probability Cost Expected cost Utility Expected utility
A 0.331 16.10 5.34 1.00 0.33
B 0.227 32.20 7.29 0.90 0.20
C OA07 16.10 6.55 -0.30 -0.12
0 0.035 79.26 2.77 0.10 0.0035
E 0.0001 1 172.00 0.11 -0.30 -0.00003
Total 1 .0000 22.06 0.41

Caffeine/erogotamine
F 0.266 1 .32 0.35 1.00 0.27
G 0.113 2.64 0.30 0.90 0.10
H 0.571 1.32 0.75 -0.30 -0.17
I 0.050 64A5 3.22 0.10 0.0050
J 0.0001 1 1 57.00 0.11 -0.30 -0.00003
Total 1 .0000 4.73 0.20

gastro-oesophageal reflux disease (GORD) (Goeree et al. 1999). This model is

described in some detail here, both to ensure an understanding of the decision
tree, and to set up the case study used in Chapter 5 where the same model is
used to demonstrate the appropriate analysis of a probabilistic model.
As shown in Fig. 2 . 1 , six treatment options are considered in the form
of strategies, as they define sequences of treatments rather than individual
therapies:
• A: Intermittent proton-pump inhibitor (PPI). Patients would be given PPI and,
if this heals the GORD, they would taken off therapy. If they experience
26 I KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION
COHORT MODELS I 27

a recurrence of the condition, they would be returned to the PPI regimen. The structure of the decision tree used in the study is shown in Fig. 2.1. For
If patients fail to heal with their initial dose of PPI, the dose of the each strategy, the initial pathway shows whether their GORD initially heals
therapy would be doubled (DD PPI) and, once healed, patients would be and, if so, it indicates the maintenance therapy a patient will move to. If they
maintained on standard dose PPl. If the GORD recurs, they would be given do not heal, they move to step up therapy as defined for each of the five strate­
DD PPl. gies. The figure shows that, for each pathway on the tree, there is a probability
of GORD recurrence during two periods: 0-6 months and 6-12 months.
• B: Maintenance PPJ. Patients would initially be treated with PPI and, once
Should this happen, step-up therapy is used as defined above. It should be
healed, they would be maintained on PPI. If they fail to heal with their noted that the tree contains decision nodes to the right of chance nodes.
initial dose, or if the GORD recurs subsequent to healing, they would be However, this indicates a treatment decision defined by the strategies rather
given DD PPl. than a point in the tree where alternative courses of action are being compared.
• C: Maintenance H2 receptor antagonists (H 2RA). Patients are initially To populate the model, the authors undertook a meta-analysis of random­
treated with H,RA and, if they heal, they are maintained on that drug. If ized trials to estimate, for each drug, the proportion of patients healed at
their GORD subsequently recurs, they are placed on double-dose H2RA different time points. They also used available trial data to calculate the
(DD H2RA). If patients fail to heal on the initial dose, they are given PPI proportions of patients who recur with GORD over the two time periods.
and, if they then heal, are maintained with H,RA. If the GORD subse­ Routine evidence sources and clinical opinion were used to estimate the cost
quently recurs, they are given PPI to heal. If patients fail to heal initially of therapies and of recurrence.
with PPI, they are moved to DD PPI and, if they then heal, are maintained The decision tree was evaluated over a time horizon of 1 2 months. Costs
with PPI. If the GORD subsequently recurs on maintenance PPI, they are were considered from the perspective of the health system and outcomes were
given DD PPI to heal. expressed in terms of the expected number of weeks (out of 52) during which
• D: Step-down maintenance prokinetic agent (PA). Patients would be given a patient was free of GORD. Table 2.1 shows the base-case results of the analy­
PA for initial healing and, if this is successful, maintained on low dose (LD) sis. For each strategy over 1 year, it shows the expected costs and time with
PA; if their GORD subsequently recurs, they would be put on PA to heal (and without) GORD symptoms. The table shows the options that are domi­
again. If patients fail their initial healing dose of PA, they would be moved nated or subject to extended dominance (Johannesson and Weinstein 1993),
to PPI to heal and, if successful, maintained on LD PA. If their GORD and the incremental cost per week of GORD symptoms avoided are shown for
subsequently recurs, they would be treated with PPI for healing. If patients the remaining options. Figure 2 . 2 shows the base-case cost-effectiveness
fail their initial healing dose of PPI, they would be moved to DD PPI and, if results on the cost-effectiveness plane (Black 1990; Johannesson and Weinstein
successful, maintained on PPI. If the GORD subsequently recurs, they 1993). It shows that Option D is dominated as it is more costly and less effec­
would receive DD PPI to heal. tive than Options C, A and E . It also shows that Option F is subject to extended
• E: Step-down maintenance H2RA. Patients would initially receive PPI to dominance. That is, it lies to the left of the efficiency frontier defined by non­
heal and, if this is successful, they would be maintained 011 H,RA. If they dominated options. This means that it would be possible to give a proportion
subsequently recur, they would be given PPI to heal. Patients who initially of patients Option E and a proportion Option B and the combined costs and
fail on PPI would be given DD PPI and, if this heals the GORD, they would effects of this mixed option would dominate Option F (see the discussion of
be maintained on PPI. If their GORD subsequently recurs, healing would cost-effectiveness decision rules in Chapter iJ.
be attempted with DD PPl.
• P: Step-down maintenance PPJ. Patients would initially be treated with 2.3.2. Markov models
PPI and, if this heals the GORD, would move to LD PPI. In the case of a The advantages of Markov models
subsequent recurrence, patients would be given PPI to heal. Patients who
fail on their initial dose of PPI would be given DD PPI and, if successful, Although various aspects of the GORD case study, such as the choice of
maintained on PPl. If their GORD recurs, they would be given DD PPI outcome measure, can be criticized, the study provides a good example of a
to heal. cost-effectiveness model based around a decision tree structure. Some of the
28 ! KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION COHORT M O D E L S I 29

"iii +500
"
-0
w

<5
c
w
�"
B-Main PPJ
..:- g "
c (5.59,436)
�S >-
so 1>'
O w
", t: -0 -0 W �u
w u F-Step Down Main PPI
� �
'" ..,
*c _
m
c
c
m
c
"

�
(4.87,298) •
Q W • $ E E E ()
g; oo
"
"
en
N
0
0
o 0
0 -0
D-Step Down Main.PA
(-2.19,148) .. �w
� A-lnterm.PPI E-Step Down Main. H2AA
0 (2.63,21) (4.24,91)
� C-Main.H2RA �
tl
� lolW -6
(0,0) +6
;
Inoremental Benefils (Gord Weeks Averted)
�
-e
w
�
o '"
cr;
E ", W • +-
e _ t:: '" en
'" w '" W co co
c Q >
'" "'
- _ .. N

<i
"
'"
i"
w
0
�W O
Q
S E - +-
-500
>- � ti "
0
en Fig. 2.2 Base-case results of the gastro-oesophageal reflux disease case study shown
-0
" .s 8 iA- N "-
"
'"
on the cost-effectiveness plane (taken from Goeree et ai, ( 1 9 99)), The line joining
t:
'"
�
Strategies C, A, E and B is the 'efficiency frontier',
'"
u
w �
c

13 ;-" i
�
'"
, '" S potential limitations of the decision tree are also evident in the case study. The
'" cr;
-0 ... "' N c;;- oo " CD
N
(II ... � m
x
" "t",!:!
:: (II ;
Co :;j; "
'"
en " !'1
"
<0 first is that the elapse of time is not explicit in decision trees. GORD relapse
IV "'- "'- "'-
'@ 1j - Q >o
�E � -
:;; oo "- N
oo
0
"'
"'-
"
en
between 0 and 6 months and between 6 and 1 2 months has to be separately
13'" .ll 3 <.? 5
0 "
"- <0 " ;::! o<i built iuto the model as no element of the structure explicitly relates to failure
'"
.c
"-
rate over time.
0
�
'" t; = A second limitation of decision trees made evident in the GORD example is
\'
'" 8 ·8 the speed with which the tree format can become unwieldy. In the GORD case
g �W �.. .....
�
m ��t "-
en
oo
"-
oo
"
'"
m
0
"'
0
"'
"'
study, only three consequences of interventions are directly modelled: initial
IoU ..... Co -tA-
'"
'"
co '" "- oo m
healing, relapse between 0 and 6 months and relapse between 6 and
.c
-
12 months. GORD is a chronic condition and it can be argued that for this
E
£ � analysis) a lifetime time horizon may have been more appropriate than one of
N
;!f
"'
r
'" '"
0:
� 12 months. If a longer time horizon had been adopted, several further features
'3 u u '"
c
�
c
m
c
m
c
u
c
m
of the model structure would have been necessary_ The first is the need to
c
'"
� � E E E « reflect the continuing risk of GORD recurrence (and hence the need for step-up
is £' 0: w
0: 'ro >- �
'ro �
'"
�
'"
u
�
E ::;
E 'iij
E
rn
"
rn therapy) over time. The second is the requirement to allow for the competing
m
OJ
c
ro
E c c
m
c
� c
"
"
� risk of death as the cohort ages. The third is the consideration of other clinical
c .� �
0 c 0 � "
-
... E E E 6. 6.
0 "
w
" developments, such as the possible occurrence of oesophageal cancer in
.... 6. 'iij .�
�
'"
'ro
Bi z Bi � .w
� patients experiencing recurrent GORD over a period of years. This pattern of
� � z l'!
� �
f!1 til (j 4' u.; w i:i we •
�
� recurring-remitting disease over a period of many years and of competing
30 I KEY ASPECTS OF DECISION MODELliNG FOR ECONOMIC EVALUATION COHORT MODELS I 31
clinical risks is characteristic of many chronic diseases such as diabetes}
ischaemic heart disease and some forms of cancer. In such situations, the need State A
<cd4 500
<
to reflect a large number of possible consequences over time would result in 200
ceUs/mm3
the decision tree becoming very 'bushy' and, therefore, difficult to program
and to present. As such, a Markov framework was used to further develop the
GORD model described above (Goeree et al. 2002).
The Markov model is a commonly used approach in decision analysis to
handle the added complexity of modelling options with a multiplicity of
possible consequences. Such models have been used in the evaluation of State 8
screening programmes (Sanders et al. 2005), diagnostic technologies (Kuntz cd4 200
<

et al. 1999) and therapeutic interventions (Sculpher et al. 1996). The added ceUs/mm3
flexibility of the Markov model relates to the fact that it is structured around
mutually exclusive disease states, representing the possible consequences of
the options under evaluation. Instead of possible consequences over time
being modelled as a large number of possible pathways as in a decision tree,
a more complex prognosis is reflected as a set of possible transitions between
the disease states over a series of discrete time periods (cycles). Costs and
effects are typically incorporated into these models as a mean value per state
per cycle, and expected values are calculated by adding the costs and outcomes Fig. 2.3 The structure of the Markov mode! used in the case study (Chancellor
et al. 1997).
across the states and weighting according to the time the patient is expected to
be in each state.
A case study in HIV allowed for the fact that prognosis for these patients is now understood in
The details of the Markov model can be illustrated using a case study. This is a terms of viral load as well as CD4 count (Sanders et al. 2005).
cost-effectiveness analysis of zidovudine monotherapy compared with zidovu­ The transition probabilities governing the direction and speed of transitions
dine plus lamivudine (combination) therapy in patients with HIV infection between disease states in the model are shown in Table 2.2 where a cycle is
(Chancellor et al. 1997). This example has been used for didactic purposes taken as 1 year. For monotherapy, these 'baseline' (i.e. control group) proba­
before (Drummond et al. 2005), but is further developed here, and in Chapter 4 bilities are taken from a longitudinal cohort study where data were collected
for purposes of probabilistic analysis. prior to any use of combination therapy. The zeros indicate that backwards
The structure of the Markov model is shown in Fig. 2.3. This model charac­ transitions are assumed not to be feasible. The transition probabilities for
terizes a patient's prognosis in terms of four states. Two of the&e are based on combination therapy were based on an adjustment to the baseline values
CD4 count: 200-500 cells/mm3 (the least severe disease state - State A) and according to the treatment effect of combination therapy relative to monother­
less than 200 cells/mm3 (State B). The third state is AIDS (State C) and the apy. This treatment effect took the form of a relative risk (0.509) which was
final state is death (State D). The arrows on the Markov diagram indicate the deriv��;-;'t;;-analysis ;ft�ials: Although the treatment effect in the
transitions patients can make in the model. The key structural assumption in trials wassC;�ethi-�grather di({er�nt) it was assumed that the relative fisk
this early HIV model (now clinically doubtful, at least in developed countries) worEedto-rediice thefransrtlon-pro�.biljti�§jr;;;;- o_n�;tat�:t.'L��y';';;:;;;state�
is that patients can only remain in the same state or progress; it is not feasible The·caICliTa:;;;n;finese revisecncombination) transition probabilities is
for them to move back to a less severe state. More recent models have allowed shown in Table 2.2. Any probability relating to the movement to a worse state
patients to move back from an AIDS state to non-AIDS states and, through is multiplied by 0.509, and the probability of remaining in a state is corre­
therapy, to experience an increase in CD4 count. These models have also spondingly increased. The separation of baseline probabilities from a relative
32 I KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION COHORT MODELS I 33

Table 2.2 Transition probabilities and costs for the HIV M arkov model used in the £2278 (zidovudine) and £2086 (lamivudine). Outcomes were assessed
case study (Chancellor et al. 1 997) in terms of changes in mean survival duration so no health-related quality­
State at start of cycle State at end of cycle of-life weights (utilities) were included.
1 , Annual transition probabilities
(a) Monotherapy
cohort simulation
State A State B State C State D As explained previously, the calculation of expected costs and outcomes for all
State A 0.721 0.202 0.067 0.010 cohort models involves summing the costs and outcomes of all possible conse­
State B 0.000 0.581 0.407 0.012
quences weighted by the probability of the consequence. In the case of Markov
State C 0.000 0.000 0.750 0.250
State D 0.000 0.000 0.000 0.000 models, this involves calculating how long patients would spend in a given
(b) Combination therapy
disease state. This can be achieved using matrix algebra) but is more commonly
State A State B State C State D
undertaken using a process known as cohort simulation. This is illustrated, for
State A 0.858 0 . 1 03 0.034 0.005 the first two cycles for monotherapy, in Fig. 2.4. Exactly the same process
(1 - sum) (0.202 x RR) (0.067 x RR) (0.010 x RR) would be used for combination therapy based on the adjusted transition
State B 0.000 0.787 0.207 0.006 probabilities in Table 2 . 2 . This example uses a cohort size of 1 000, but
(1 sum)
- (0.407 x RR) (0.012 x RR) this number is arbitrary and, for a cohort model, the same answer will emerge
State C 0.000 0.000 0.873 0 . 1 27 for any size of starting cohort. Cohort simulation simply involves mUltiplying
(1 sum)
- (0.25 x RR) the proportion of the cohort ending in one state by the relevant transition
State D 0.000 0.000 0.000 1 .000 probability to derive the proportion starting in another state. In a spreadsheet,
2. Annual costs this is achieved by setting up the formulae for one cycle and then copying
Direct medica! £1 701 £1 774 £6948 down for subsequent cycles.
Community £1 055 £1 278 £2059 Calculating expected costs for a cohort model simply involves, for each
Total £2756 £3052 £9007
cycle, adding the costs of each state weighted by the proportion in the state,
RR, relative risk of disease progression. Estimated as 0.509 in a meta-analysis.
and then adding across cycles. This is shown in Table 2.3 for monotherapy
The drug costs were £2278 (zidovudine) and £2086 (Iamivudine).
based on the costs shown in Table 2.2 and annual drug costs for zidovudine of
£2278. The process of discounting to a present value is very straightforward in
treatment effect is a common feature of many decision models used for cost­ a cohort simulation and this is also shown in Table 2.3.
effectiveness. One advantage of this approach relates to the important task of Table 2.4 shows the calculation of expected survival duration (life
estimating cost-effectiveness for a particular location and population expectancy) for the monotherapy group. At its simplest, this involves adding
subgroup. Often it is assumed that baseline event probabilities should be as the proportion of the living cohort for each cycle, and adding across the cycles.
specific as possible to the location(s) and subgroup(s) of interest, but that the Of course, the units in which survival duration is estimated will depend on the
relative treatment effect is assumed fixed. cycle length. In the case study, all transitions are assumed to take place at the
It can be seen that all the transition probabilities are fixed with respect to start of the cycle, so the simple approach to calculating life expectancy
time. That is, the baseline annual probability of progressing from, for example, assumes those dying during a cycle do not survive for any proportion of the
State A to State B is 0.202, and this is the case 1 year after start of therapy and cycle. Strictly, an unbiased estimate of life expectancy would assume
it is also the case, for those remaining in State A, after 1 0 years. When these that deaths occur halfway through a cycle. It is possible, therefore, to employ
time invariant probabilities are used, this is sometimes referred to as a Markov a half-cycle correction as shown in the final column of Table 2.3. The half­
Chain. cyde correction is shown here just for life expectancy calculation. In principle,
Table 2.2 also shows the annual costs associated with the different states. however, it also applies to the calculation of expected costs. In the context of a
These are assumed identical for both treatment options being compared ­ cost-effectiveness analysis where the focus is on the incremental costs and
excluding the costs of the drugs being evaluated. The drug costs were outcomes of alternative options, it is unlikely that the half-cycle correction
34 I KEY ASPECTS OF DECISION M O D E LLING FOR ECONOMIC EVALUATION COHORT MODELS I 35

Table 2.3 Calculation of expected costs for monotherapy in the case study Markov
model based on the costs shown in Table 2 , 2 , annual drug costs of £2278 and an
annual'discount rate of 6%

cycle Proportion of cohort in each state (osts (f)

1.000=10 (year) A B ( 0 Undiscounted Discounted
Cl 0 0 '"
'" 1 000
�
0

({) 721 202 67 10 5463* 51 53t

2 520 263 181 36 6060 5393
3 376 258 277 89 6394 5368
4 271 226 338 1 65 6381 5055
""
W 5 1 95 1 86 364 255 6077 4541
,,'Y
6 141 147 361 350 5574 3929
() 0 "-
'" '"
�

7 102 114 341 444 4963 3301

* 8 73 87 309 531 431 6 2708t
?is -2
il: 9 53 65 272 610 3682 2 1 79
i"
'"
.c 10 38 49 234 679 3092 1727
0
0
c
11 28 36 1 98 739 2564 1 350
E 26 1 65 789
12 20 2 1 02 1 045
0
�

19 1 36 830
0
c 13 14 1 708 801
CO '"
'" '.;0
(\j '" 14 10 14 111 865 1 377 609
* "3
E
?is .�
15 7 10 90 893 1 1 03 460
t
0 16 5 8 72 915 878 346
.c
0
u 17 4 5 57 933 695 258
'"
,,�
IV' 0 18 3 4 45 948 548 1 92
�

"'Y" . � '"
,, -0 19 2 3 36 959 431 142
� .z
u �
o :Ii
2C 2 28 968 337 1 05
<{
<D � ::l
� "
.. 63 745 44 663
:g '"
�� �+ [202 x (3052 ... 2278)J
0
({) "" f:\" *{[721 x (2756-t.. + {67 x (9007 +� + {1 a x OJ} / 1000
'"
" '" £ -g r5463 / 111 + 0.06)IJ
" " �
"
�
o �

� § 1iJ
<D .;::: -g will make a lot of difference to results unless the cycle length is long as a
.0
E
� E
::> � > proportion of the model's time horizon.
0 (\j .2
= �0
c:
Of course) the cohort simulation and calculation of expected costs and
Q)
�
� �"
.;,
expected survival duration, shown in Fig, 2.4 and Tables 2.3 and 2.4 for the
() ,- .c
... �
mono therapy baseline, would also need to be undertaken for combination
therapy, The process would be exactly the same but would involve different
transition probabilities and costs.
36 I KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION S U M MARY I 37

Table 2.4 Calculation of life expectancy over 20 cycles for monotherapy for the case T 3 yrs after
=

initial treatment

. . . .�c:)
study Markov model. Calculation with and without a half-cycle correction is shown �ocal .. �__

recurrence
Cycle Proportion of cohort in each state Life years
B Die at start Die in middle
Remission
(year) A C D

0 1 000
721 202 67 10 0.990 0.995 Regional T 8 yrs after Transition probabilities equal
2 520 263 181 36 0.964 0.977 recurrence =

initial treatment regardless of timings or

0.9 1 1 0.937
directions of earlier transitions
3 376 258 . 277 89

4 271 226 338 1 65 0.835' 0.873

Note: T time Since initial operation (i.e. since start of model)
=

Fig. 2.5 Example of a simple Markov model to illustrate the Markov assumption.
5 1 95 1 86 364 255 0.745 0.790

6 141 1 47 361 350 0.650 0.697

7 1 02 1 14 341 444 0.556 0 603t

the Markov model will have '112- meJILOLi.-r.e.gardinJi where the patient has
8 73 87 309 531 0.469 0.513
co'!'e_from.otlheli.millROf that transition. This is illustrated in Fig. 2.5 using­
0.390 0.429
9 53 65 272 61 0 a simple Markov model to characterize prognosis following treatment for
10 38 49 234 679 0.321 0.355 cancer. It shows that patients can have two forms of cancer recurrence: local
11 28 36 1 98 739 0.261 0.291 and regional. Once treated for their recurrence) patients then move into a
12 20 26 1 65 789 0.2 1 1 0.236 remission state but, once in that state, all patients are considered homoge­
13 14 19 1 36 830 0 . 1 70 0.1 90 neous regardless of where they have come from (what type of recurrence) or
14 10 14 111 865 0.135 0.1 52 the timing of the recurrence after initial treatment for the cancer. The implica­
90 893 0 . 1 07 0.121
tion of this is that it is not possible to have different probabilities of future
15 7 10
transitions (e.g. the probability of death in Fig. 2.5) according to the nature or
16 5 8 72 915 0.085 0.096
timing of the recurrence.
57 933 0.067 0.076
17 4 5 In general, the Markov assumption means that it may not be straightfor­
18 3 4 45 948 0.052 0.059 ward to build 'history' into this type of model. That is, to model a process
19 2 3 36 959 0.041 0.047 where future events depend on past events. This may be possible by adding
20 2 28 968 0.032 0.036 additional states to the model and incorporating time dependency into transi­
7.996 8.475 tion probabilities. Both of these extensions to the Markov model are consid­
ered in Chapter 3.
"'(271 + 226 + 338) / 1 000
t{102 + 1 1 4 + 341 + (0.5 x (444 - 350)l}/1000
2 .4. Summary
This chapter has described some of the key concepts for the use of decision
The Markov assumption analysis for economic evaluation. The majority of cost-effectiveness studies
Although the Markov version of a cohort model provides greater flexibility are based on cohort modelling. This provides a flexible framework which can
than a decision tree, it also has some important restrictions in the context of be programmed relatively easily (often in spreadsheets) and evaluated rapidly.
structuring complex prognoses. The restriction relates to what is known as the However, the cohort framework can be restricting. In the case of Markov
Markov assum..p!i.9_n, or 'memoryless' feature of Markov models. This assump-
-
model." for example, the memoryless feature of these models can cause diffi­
culties in modelling a complicated prognosis.
-----

tion means that once a notional patient has moved from one state to another,
38 I KEY ASPECTS OF DECISION MODELLING FOR ECONOMIC EVALUATION EXERCISE: REPLICATING A MARKOV MODEL OF HIV/AIDS I 39

2 . 5 . Exercise: replicating a Markov model of HIV/AIDS 5. Discounting. The original analysis was based on discounting the costs at
6 per cent but not discounting the estimates of life years.
2.5.1. Overview
The aim of this exercise is to get you to replicate a previously published
Markov model by Chancellor and colleagues ( 1997). The model itself is some­ 2 .5.2 . Step-by-step guide to constructing the model
what out of date, in that it compares combination therapy (lamivudine and Open the file 'Exercise 2.5 - Template' and start at the <Parameters> worksheet.
AZT) with monotherapy (AZT alone). Nevertheless, the model is straightfor­ You will see that the cells to be completed are coloured yellow.
ward enough to serve as a simple example that it is possible to replicate in a Using the figures from Table 2.5, complete columns C and D on the
short period of time. <Parameters> worksheet for the transition probabilities (rows 9-17). Column
The basic structure of the model was given in Fig. 2.3, which shows that the C should contain the number of events of interest and column D should
disease process is structured as chronic, such that patients can move to succes­ contain the complement, such that columns C plus D should equal the appro­
sively more serious disease states, but cannot recover. priate row total from Table 2.5 above. Now calculate the respective transition
The cycle length of the model is 1 year and is evaluated over 20 years (after probabilities in column B using the counts in columns C and D (the reason for
which more than 95 per cent of patients are expected to have died). structuring the spreadsheet like this will become apparent in the exercise in
Use the data given below to populate the model and calculate the incremental Chapter 4).
cost-effectiveness ratio for AZT monotherapy. You should be able to replicate Enter the other information for the state costs) drug costs, treatment effect
a figure for the ICER of £6276. and discounting given above in the appropriate place in the template.
1. Transition probabilities. These were calculated from the counts of individuals Having entered the input parameters of the model, the task now is to
that were observed to move between the four health states each construct the Markov models for the treatment alternatives: combination and
year in a longitudinal data set from the Chelsea and Westminster Hospital monotherapy. Note that the parameters you have just entered are in named
in London. These counts are given in Table 2.5 and you should be able cells, such that they can be referred to by the shortened name in column A of
to verify that these counts give the transition probabilities presented in the <Parameters> worksheet.
Table 2.2 part 1 (a). If you move to the Markov worksheet, you will see the structure of the
2. State costs. These are given in the original article separately by state for Markov model laid out for you for each treatment alternative. Start with the
(direct medical' and (community care' costs and these were reproduced in monotherapy model.
Table 2.2 part 2. 1 . Generating the Markov trace
3. Costs of drugs. The yearly cost of AZT monotheray is given as £2278 and
The initial concern is with generating the Markov trace, that is, showing the
lamivudine is given as £2086.50. proportions of patients that are in any one state at any one time. This is to be
4. Treatment effect. A meta-analysis of four trials is reported in the paper entered in columns C to F representing the four model states, and G provides
giving a pooled relative risk of disease progression as 0.509. Note from a check (as the sum across C to F must always equal the size of the original
Table 2.2 part 1 (b) how this treatment effect is applied to the monother­ cohort - which is set to 1 in cell C7 such that the Markov trace represents
apy transitions in the model. proportions of patients).
The first step in building a Markov model is to define the transition matrix.
Table 2.5 Counts of transition between the four model states per year This is a matrix that shows the probability of transition from states
A B C 0 Total represented in the rows to states represented in the columns. To save you
A 350 116 17 1 734 time, a copy of the transition matrix is reproduced in Table 2.6 for the
B 0 731 512 15 1 258
monotherapy arm.
C 0 0 1312 437 i. Start by making sure that you understand the transition matrix. In
0
particular, make sure you can replicate it trom the information given
in the diagram of the model in Fig. 2.3.
40 I KEY ASPECTS OF DECISION MODE LUNG FOR ECONOMIC EVALUATION EXERCISE: REPLICATING A MARKOV MODEL O F H IV/AIDS I 41

Table 2.6 Transition matrix for the monotherapy arm in row 29, sum the columns to give the expected life years both discounted and
Transition matrix A B C D
undis<;:ounted.
A tpAZA tpA2B tpAZC tpA2D 3 . Est; mating costs
B 0 tpBZB tpB2C tpBZD In column K, calculate the costs for each time period by applying the state
C 0 0 tpCZC tpCZD costs, not forgetting that AZT is given for the whole time period. In column L,
D 0 0 0 apply the standard rate of discount for costs. Again, in Row 29, sum the
columns to give costs both undiscounted and discounted.
4. Adapting the model for combination therapy
You now need to repeat the steps above, but this time for combination therapy.
ii. Use the transition matrix to populate the Markov model. This will
involve representing the transitions between the different states repre­ i. Start by copying the whole monotherapy model to the corresponding
sented in columns C to F. combination therapy model. This will minimize repetition as tasks
Hints: Start with just one row at a time (e.g. the first- row 8). Once you now relate to adjusting the model for combination treatment. (Note:
have this row correct you can simply copy it down to the other rows. remember to anchor the year (column A) in the discounting formula
Although it is tempting to look across the rows of the transition matrix - before copying).
consider looking down the columns as this indicates, for a given state, ii. Firstly, the treatment effect needs to be added in. In the original article,
where the people entering that state come from. Indeed, the transition the relative risk parameter was applied to all transitions. The corre­
matrix describes the proportion ofpatients in the other states in the previ­ sponding transition matrix for combination therapy is, therefore,
ous cycle arriving in the given state for this cycle. For example, looking given in Table 2.7.
down for 'State B' we find that tpA2B patients in 'State A' last period iii. Use this transition matrix to adjust the trace in rows 8 and 9 only (as
arrive in State B this period and that tpB2B patients in 'State B' last the base case assumption in the originally reported model was that
period stay in 'State B' this period. The formula for cell D8 should there­ treatment effect was limited to 2 years).
fore read: =C7'tpA2B+D7'tpB2B Now add in the cost of lamivudine for these 2 years only in column V (as the
iii. When you get to the dead state, do not be tempted to make this the drug is assumed to be given for only 2 years)
remainder of the other cells in the row - the use of remainders in this
way will mean that any errors could go unnoticed. Instead, it is good
practice to complete all states as you think they should be then sum 1 Recall that the standard discount rate is given by11(1 + r)twhere ris the discount rate and
trepresents time (in years).
across the states to make sure the total sums to one. Do this check in
column G and make sure it does!
iv. When you think you have the first row correct, copy this row down to Table 2.7 Transition matrix for the combination therapy arm
the 19 rows below. If your check in column G is stiil looking good then B C D
Transition matrix A
most likely you have done it correctly.
A HpAZB*RR- tpA2B*RR tpAZCRR tpAZD*RR
By far the most tricky bit of building the Markov model in Excel is
now complete. Now that you have the Markov trace you can calculate tpA2C*RR·
the cost and effects. tpA2D*RR
B 0 l -tpB2C*RR- tpBZC*RR tpB2D*RR
2 . Estimating life years tpBZD*RR
In column H calculate the proportion of the cohort that is in one of the 'alive' C 0 0 t -tpC2D*RR tpCZD*RR
states for each of the years in the model. In column I apply the standard
discount formula! (although we have set this to zero in the base case). Finally, D 0 0 0 1
42 I KEY ASPECTS OF DECISION MODELliNG FOR ECONOMIC EVALUATION REFERENCES I 43

Hunink, M., Glaziou, P., Siegel, J., Weeks, J., Pliskin, J., Elstein, A., et at. ( 2001 ) Decision
5. Cost-effectiveness estimates
making in health and medicine. Integrating evidence and values. Cambridge, Cambridge
The final task is simply to copy the corresponding results from the sums at University Press.
the bottom of the Markov model sheet (row 29) to the <Analysis> worksheet Johannesson, M. and Weinstein, S. (1993 ) 'On the decision rules of cost-effectiveness
(row 5) and to calculate the appropriate increments and ICER. analysis: Journal of Health Economics,1 2: 459-467 .
Congratulations, you have now replicated the Markov model. Compare Kuntz, K. M., Fleishmann, K. E., Hunink, M. G. M. and Douglas, P. S.(1999 ) 'Cost­
effectiveness of diagnostic strategies for patients with chest pain', Annals of Internal
your result for the ICER to that given in the solution (£6276). Is any debugging
Medicine,130: 709-718.
required? If it is, then you may want to compare your Markov trace and stage Neumann, P. J., I-Iermann, R. C. and Kuntz, K. M. (1999 ) 'Cost-effectiveness of donepezil
costs for monotherapy against those reported in Table 2.3. in the treatment of mild or moderate Alzheimer's disease', Neurology,5 2: 1138-45 .
O'Hagan, A. and Luce, B.( 2003 ) A primer on Bayesian statistics in health economics and
References outcomes research. Bethesda, MD, Medtap International.
Ades, A. E. ( 2003 ) A chain of evidence with mixed comparisons: models for multi-parameter Palmer, S., Sculpher, M., Philips, Z., Robinson, M., Ginnelly, L., Bakhai, A., et al.( 2005)
evidence synthesis and consistency of evidence, Statistics in Medicine, 22: 2995-3016 . 'Management of non-ST-elevation acute coronary syndromes: how cost-effective are
Ades, A. E., Sculpher, M. J., Sutton, A., Abrams, K., Cooper, N., Welton, N . , et al.( 2006 ) glycoprotein IIb/IIIa antagonists in the UK National Health Service?: International
'Bayesian methods for evidence synthesis in cost-effectiveness analysis', Journal of Cardiology,100: 229- 240.
PharmacoEconomics 24: 1-19. Parmigiani, G. ( 2002) Modeling in medical decision making: a Bayesian approach.
Arrow, K. J. and Lind, R. C.(1970 ) 'Risk and uncertainty: uncertainty and the evaluation of Chichester, Wiley.
public investment decisions', American Economic Review,60: 364-78 . Raiffa, H. (1968 ) Decision analysis: introductory lectures on choices under uncertainty.
Black, W. C. (1990 ) 'The CE plane: a graphic representation of cost-effectiveness', Medical Reading, MA, Addison-Wesley.
Decision Making, 10: 212-214. Sackett, D. 1., Rosenberg, W. M. c., Gray, J. A. M., 1--faynes, R. B. and Richardson, W. S.
Briggs, A. ( 2001 ) 'Handling uncertainty in economic evaluation and presenting the results' (1996 ) 'Evidence-based medicine: what it is and what it isn't', British Medical Journal,
in M. Drummond and A.I- McGuire (eds) &onomic evaluation in health care: merging 312: 71-7 2.
theory with practice. Oxford, Oxford University Press; pp.17 2- 214. Sanders, G. D., Bayoumi, A. M., Sundaram, v., Bilir, S. P., Neukermans, C. P. and et al.
Chancellor, J. v., Hill, A. M., Sabin, C. A., Simpson, K. N. and Youle, M. (1997 ) 'Modelling ( 2005) 'Cost-effectiveness of screening for HIV in the era of highly active antiretroviral
the cost effectiveness oflamivudinelzidovudine combination therapy in HIV infection', therapy', New England Journal of Medicine, 352: 570-585.
PharmacoEconomics,1 2: 1-13. Sculpher, M., Michaels, J., McKenna, M. and Minor, J. (1996 ) 'A cost-utility analysis of
Chilcott, J., McCabe, c., Tappenden, P., O'Hagan, A., Cooper, N. J., Abrams, K., et aL laser-assisted angioplasty for peripheral arterial occlusions', International Journal of
( 2003) 'Modelling the cost effectiveness of interferon beta and glatiramer acete in the Technology Assessment in Health Care1 2: 10 4--1 25.
management of multiple sclerosis', British Medical Journal, 3 26: 522. Spiegelhaltcr, D. J., Abrams, K. R. and Myles, J. P.( 2004 ). Bayesian approaches to clinical
Cooper, N. J., Sutton, A. J., Abrams, K. R., Turner, D. and Wailoo, A.( 2004 ) 'Comprehensive trials and health-care evaluation. Chichester, Wiley.
decision analytical modelling in economic evaluation: A Bayesian approach: Health Sutton, A. J. and Abrams, K. R. ( 2001 ) 'Bayesian methods in meta-analysis and evidence
Economics,13: 203- 226 . synthesis', Statistical Methods in Medical Research, 10: 277-303.
Downs, S.(1999 ) 'Technical report: urinary tract infections in febrile infants and young Sutton, A. J., Abrams, K. R., Jones, D. R., Sheldon, T. A. and Song, 1. A.( 2000) Methods for
children', Pediatrics, 103: 54 . meta-analysis in medical research. Chichester, Wiley.
Drummond, M. E, Sculpher, M. J., Torrance, G. W. , O'Brien, B. and Stoddart, G. L. ( 2005 ). 1orrance, G. W.(1986 ) 'Measurement of health state utilities for economic appraisal ­
Methods for the economic evaluation of health care programmes. Oxford, Oxford a review� Journal of Health Economics, S: 1-30.
University Press. von Neumann, l and Morgenstern, O.(1944 ) Theory of games and economic behavior.
Evans, K. W., Boan, J. A., Evans, J. L. and Shuaib,A. (1997 ) 'Economic evaluation of oral New York, Princeton UniverSity Press.
sumatriptan compared with oral caffeine/ergotamine for migraine', Weinstein, M. C. and Fineberg, H. V. (1980 ) Clinical decision analysis. Philadelphia, PA,
PharmacoEconomics,1 2: 565-577 . WB Saunders Company.
Goeree, R., O'Brien, B. J Hunt, R., Blackhouse, G., Willan, A. and Watson, J. (1999 )
.•

'Economic evaluation oflong term management strategies for erosive oesphagitis',

PharmacoEconomics, 16: 679-697 .
Goeree, R., O'Brien, B. J., Blackhouse, G., Marshall, J.. Briggs, A. H. and Lad, R. (200 2)
'Cost-effectiveness and cost-utility oflong-term management strategies for heartburn',
Value in Health, S: 31 2-3 28.
 Chapter 3

Further developments in
decision analytic models
for economic evaluation

As a framework for economic evaluation, decision analysis needs adequately to

reflect the key features of the natural history of a disease and the impact of alter­
native programmes and interventions. Models are, however, simplifications of
reality and will always approximate these phenomena. In short, to be 'fit for
purpose' for decision making, decision models need to strike a balance between
realism and flexibility in terms of computation and data requirements. Chapter 2
described the two basic types of decision model that predominate in economic
evaluation: the decision tree and the Markov model. These models can be used
for a wide range of decision problems. In some circumstances) however, their
structural assumptions are clearly at odds with the theory and/or evidence associ­
ated with a particular disease or technology. In such cases, it may be appropriate
to develop the models further in one or more key respects.
This chapter considers some of the further developments that might be
considered for decision models to increase the extent to which they fit with
what is known about a disease or the relevant interventions. The following
section describes ways in which greater 'complexity' can be added to cohort
models. The next main section describes the use of patient-level simulation
models in economic evaluation, and the subsequent section considers the use
of dynamic models in the context of infectious diseases.

3 . 1 . Extensions to cohort models

Cohort models are the mainstay of decision analysis for economic evaluation.
This is because they can balance an adequate reflection of available knowledge
and evidence regarding a disease and interventions with parsimony in terms
of computation and data. For certain types of decision problem) it is impor­
tant to go beyond the simple decision tree or Markov model. Here, three
extensions to the simple cohort model are described: the combination of the
decision tree and the cohort model; adding time-dependency; and loosening
the memoryless restriction of the Markov model.
46 I FURTHER DEVELOPMENTS IN DECISION ANALYTIC MODELS
EXTENSIONS TO COHORT MODELS I 47
3 . 1 . 1 . Combining decision trees and Markov models
Although they are described as distinct forms of model in Chapter 2, it would
be a mistake to see Markov models and decision trees as mutually exclusive;
indeed, Markov models are really just a form of recursive decision tree. For
some evaluations, they can be used jointly. One form of this is when transitions
between Markov states are characterized in terms of a tree - these have been
called Markov cycle trees (Hollenberg 1984). A model can also be made up of
multiple elements involving One or more tree and Markov model. An example
of this is a cost-effectiveness analysis of glycoprotein lIb/IlIa (GPA) drugs in
non-ST-elevation acute coronary syndrome (NSTACS) (Palmer et al. 2005).
As shown in Fig. 3 . 1 , a decision tree was used to characterize the possible
short-term events following an acute episode of NSTACS under one of four
strategies (three based on the use of GPAs and one 'usual care' strategy without
these drugs). The short-term period was 6 months, which represented the
longest period of follow-up in most of the GPA trials.
The ultimate purpose of the decision tree was to estimate, for each of the four
strategies, the proportion of the patient cohort in each of three states at
6 months: dead, alive having not experienced a myocardial infarction but remain­
ing with ischaemic heart disease (IHD), and alive having experienced a myocar­
dial infarction. The objective of the Markov model was to estimate the expected B
m
�
quality-adjusted survival duration and lifetime costs for tlle cohort. Hence, the
differences between therapies only relate to the first 6 months; the Markov model a
1
,
a
estimates long-term survival duration and costs conditional on the patient's t3
,
e-
m o
,
i,
status at 6 months, but does not differentiate between initial treatments. The
J
•
�
Markov model allows for the fact that, over the longer term, patients will !
g
,
,
�
t
w

continue to be at risk of myocardial infarction. The reason for there being two
myocardial infarction states in the Markov model is that evidence suggested that
�
a
i
"
a
"
�
0
B
patients' risk of death and their costs are higher following myocardial infarction � ro

� �
than in subsequent periods (this is discussed further below). �, ,
� ,
,

1 •

! �
m
3 . 1 . 2 . Building time-dependency into Markov transition u
i, Z

probabilities �, !,
The HIV case study used to illustrate Markov models in Chapter 2 assumed
that all transition probabilities were fIxed with respect to time. For example,
1
m
0
'"

the annual probability of dying when in the 'AIDS state' was the same regard­
�
less of how long the patient had been in that state. As for all modelling
assumptions, this is a simplifIcation of reality. The critical issue is whether j
<

the simplification is markedly at odds with reality and may generate poten­
tially misleading results in terms of the cost-effectiveness of the options
48 I FURTHER DEVELOPMENTS IN DECISION ANALYTIC MODELS EXTENSIONS TO COHORT MODELS I 49

being compared. For those models where the assumption of constant Table 3.1 Revised transition probability matrix for the HIV case study where transitions
transition probabilities is considered too strong, there are ways of letting some to a dead state vary with the age of the cohort. See Table 2 . 2 for initial transitions
transition probabilities vary with time. State A State B State C State 0 State E

Types of time-dependency State A 1 0.202 -

- 0.202 0.067 0.0 1 0 P(age)
0.067 - 0.01 0 -
To understand how this is possible, it is important to distinguish between two P(age)
types of time-dependency in transition probabilities in Markov models; these State B 0.000 1 - 0.407 - 0.407 0.012 P(age)
are described below. 0.012 -
Probabilities that vary according to time in model. The simplest form of time­ P(age)
dependency in Markov models is where transition probabilities vary according State C 0.000 0.000 1 - 0.250 - 0.250 P(age)
to how long the cohort has been modelled. In other words, the probability P(age)

of one or more transitions changes as the cohort ages. Returning to the HIV State D . .
0.000
. . . ...
0.000
. . . . .. . . . . . . . .
. . . . . . ...... ..... . . . . . . . ........ . . . . . . . . . . . . . . . .
0.000
. . .... . . .... . . ........ ......
1 .0 0.000
case study in Chapter 2, transitions to State D (i.e. the probability of dying) State E 0.000 0.000 0.000 0.000 1 .0
were assumed fIxed with respect to time (Fig. 2.3, Table 2.2). However, it may P(age1) P(age2) P(age3) P(age4) P(ageS) P(age6) P(age7) P(age8) P(age9) P(agel0)
be considered more appropriate to assume that, as the cohort ages, the annual 0.001 0.001 0.002 0.003 0.003 0.005 0.005 0.006 0.006 0.007
probability of death increases. This would probably make more sense if the
probability of death was divided into two: HIV/AIDS-specific mortality States A. B and C are as defined in Fig. 2.3, State D is Hlv/AIDS-specific mortality and State E is
other·case mortality. The numbers in the table below the main transition probability matrix are the
and other-case mortality. With such a division, it might be reasonable to age-dependent probabilities of other-cause mortality over a 1 O·year period.
have constant transition probabilities for the first, but to increase transition
probabilities as the cohort ages for the second.
The starting age of the cohort should be explicitly stated, as this would
typically be an important element in characterizing the individuals to whom time in a state increases. As a result of the Markov assumption, this form of
the intervention applies, especially when its impact on mortality is being time-dependency is less straightforward to implement. In the HIV case study,
modelled. Furthermore, for cost-effectiveness modelling, the cohort should be it may be the case that the probability of dying while in the AIDS state
homogeneous, as the decision problem is to establish the most cost-effective increases as the time the patient has had AIDS increases. Ideally, it would be
option for specifically defIned groups of individuals. If the starting age of the possible to increase the transition probability from State C to State D in Fig. 2.3
cohort is known, then so is the age of the cohort at any cycle of the model, to reflect this. However, the Markov assumption is such that, once a patient
whatever the state. In order to implement age-dependent probabilities, therefore, has entered State C, the model cannot establish at which point they entered
it is simply necessary to have a unique transition probability for each cycle. so it cannot 'keep track' of how long they have been in the state. Therefore,
This is shown in Table 3.1, which revises the transition probability matrix for as currently structured, it would not be possible to implement this form of
the HIV case study shown in Table 2.2. Instead of one 'dead' state, there are time-dependency between States C and D.
now two: HIV/AIDS-related (State D) and other-cause (State E). Transitions However, if it was the case that the probability of moving between State A
to State E are assumed to be age-dependent as defIned in the table below the and one or more of the other states in the model was considered to vary
main transition probability matrix. These age-dependent probabilities might depending on how long the patient had been in State A, this could be imple­
be taken from national life tables. As the cycle number increases, so does the mented quite easily. This is because all patients start the model in State A at
probability of moving to State E. As the transition probabilities need to sum to a given time point and, once they have left that state, they are unable to return.
1 horizontally, the probability of remaining in a given state for the next cycle It is, therefore, always possible to keep track of how long any patient has been
has to allow for the age-dependent probability. in State A. In this case, of course, making a transition probability dependent
Probabilities which vary according to time in state. The second type of time­ on time in state is exactly the same as making it dependent on the age of the
dependency relates to situations where a transition probability changes as cohort because they are in that state from the start of the model.
50 I FURTHER DEVELOPMENTS IN DECISION ANALYTIC MODELS
EXTENSIONS TO COHORT MODELS I 51

Using survival analysis to i mplement time-dependency Box 3.1 :A summary oHhe d istinction between rates
The example of time-dependency illustrated in Table 3.1 simply uses a 'look-up' and probabilities
. , .
table to define the time-dependent probabilities. This would be appropriate
when data are readily presented in such a format in the literature or in routine Rates
statistics (e.g. routine life tables). It is often the case, however, that the rela­ The instantaneous potential for the occurrence of an event, expressed per
tionship between a transition probability and time can be explicitly estimated number of patients at risk.
from patient-level data. Such data would typically come from some sort of Rates can be added and subtracted.
longitudinal study (e.g. a trial or cohort study) which has recorded the time to
one or more events for each patient. The appropriate way of analysing such Probabilities
<time to event' data is survival analysis. A number ranging between 0 and 1 .
The use of survival models in the medical field is widespread and well Represents the likelihood o fa n event happening over a specific period of
understood. Their key feature is that they can handle (noninformative) time.
censoring that frequently occurs in follow-up studies - that is, observation It is possible to convert an instantaneous rate to a probability over a
ceases before some patients have experienced the event of interest. The most particular time period) if the rate is assumed to be constant over that period:
popular form of survival model is the semiparametric Cox proportional hazards
model (Cox and Oakes 1984). As this form of model does not specify how the p � 1- exp {-rtl
risk of an event changes over time (i.e. the functional form of the hazard func­
where p is the probability, r is the rate and t is the time period of interest.
tion), its use to inform time-dependency in Markov models is limited.
Similarly, it is possible to convert a probability over a period of time to
Alternatively) parametric survivor functions can be employed and can
a (constant) instantaneous rate:
be fitted in many popular statistical packages (Collett 1994). It is important
to emphasize, however, that standard survival analyses are based on r � - [In (l-P)J/t
hazard rates, but Markov models employ transition probabilities.
For example, assume 100 patients are followed up for 5 years after which
As summarized in Box 3 . 1 ) there are important differences between rates
20 have had a particular event. Hence the 5-year probability of the event is
and probabilities that have to be recognized in decision analysis, although
0.2. Assuming a fixed rate with respect to time) what is the instantaneous
they can readily be translated one to the other (Miller and Homan 1994).
event rate?
An example of the importance of this distinction is when flexibility is required
in terms of the cycle length in a Markov model. Moving, say, from a I -year to Rate � - [In (I-O.2)J/5 � 0.04463
a 6-month cycle length involves more than merely dividing the transition
probabilities by 2. Rather, instantaneous rates need to be derived and, from It is then possible to calculate a I-year probability of the event:
that, 6-month probabilities calculated. Understanding the differences between I-year probability � 1 - exp (-0.04463 x I ) � 0.043648
rates and probabilities is particularly important in the context of survival
analysis. Note that this I-year event probability is not the same as the 5-year event
p;;:'
. :
____�"--." "." . � . -)

Fundamental relationships in ival angiJ!§.il In order to understand how probability divided by 5.

to use survival analysis to derive transition probabilities) some important
concepts in survival analysis need to be introduced. The first of these is
a probability density function (pdf) for survival data, j(t), with an associated
cumulative density function:
52 I FURTHER DEVELOPMENTS IN DECISION ANALYTIC MODELS EXTENSIONS TO COHORT MODELS I 53

which gives the cumulative probability of failure up to time t. The complement Estimating discrete time transition probabilities from instantaneous hazard rates.
of this function is the survivor function, that is, the probability of surviving The re.lationship in eqn 3.3 is central to the process used to derive transition
for a period of time greater than t probabilities for Markov models. Consider a very simple Markov model with
just two states: alive and dead. Therefore, only one transition probability needs
() ( )
S t =P T> t =I-F t . () to be estimated: the transition to death. Define the length of the Markov cycle
Note that it is, therefore, straightforward to relate j( t) and S( t) as u and let the instantaneous hazard of death at time t be represented by h(t),
as above. The challenge, therefore, is to estimate the appropriate discrete
()
[ t =
( ) d (I - S( t ))
dF t
()
S' t . (3.1 )
transition probability between time-points t - u and t, call this tp(t") where tu
dt dt indicates that t is now measured as integer multiples of the cycle length of the
It is now possible to define the hazard function (where the hazard is a rate), model, u.
which is the instantaneous chance of failure at time t, conditional on having The baseline transition probability of the event of interest can simply be
survived to time t, or, algebraically: defined as one minus the ratio of the survivor function at the end of the interval
to the survivor function at the beginning of the interval:
() (
. P t + 8t 2: T > t ! T > t
h t = hm .
) (J ( )/ (
tp t = I - S t S t - u )
Ot-')O 8t
(
The standard rule for conditional probabilities is P AI B ) = peA and B)I PCB)
which can be rewritten in terms of the cumulative hazard (from eqn 3.3) as:
(see Chapter 2), so this can be re-written as: ( ) { ( )}/exp { -R( t - u)}
tp t,, = 1 - exp -R t

() (
. P t + 8t 2: T > t 1 ) = l - exp { R ( t - u) - R ( t )} .
(3.4)
h t = hm
()
.-

&...,0 8t S t

=
( ).
[ t
(3.2)
To see why it is important to move from the arguably more simple specifica­
tion of the transition probability as the ratio of two points on the survival
S[t) curve to the specification in eqn 3.4, consider the application of a treatment
We now define the cumulative hazard function: effect. It is common to include treatment effects as multiplicative, perhaps as
relative risks, or in terms of treatment effects estimated directly from a
survival analysis, in terms of a hazard ratio. Such a treatment effect should not
be applied directly to the baseline transition probability; rather, the treatment
effect (call this �) should instead be applied to the hazard rate. Therefore, if
noting that the cumulative hazard up to time tis not the same as the probability eqn 3.4 represents the baseline transition probability in the absence of treat­
of failure up to time t which is given by F(t). Using the result from eqn 3 . 1 ment, the treatment transition would be given by:
and the standard rules o f calculus, we can express the cumulative hazard a s a
function of the survivor function: (
tPT t") = l - exp { 1'[ R ( t - u) - R ( t )J}
()
'
S' t
f () = 1 - exp { R ( t - u ) - R ( t Jr
R t =-
St
o
() ", 1 - . exp {R ( t - u) - R ( t )} .
l'

= - In {S( t )}
The last line above indicates that it would be incorrect to multiply the treat­
ment effect by the baseline probability. To illustrate how these methods are
or) alternatively, as the survivor function in terms of the cumulative hazard:
used with actual data to populate specific decision models, two examples are
() { ( )}
S t = exp -R t . (3.3) described below.
54 I FURTHER DEVELOPMENTS IN DECISION ANALYTIC MODELS EXTENSIONS TO (OHORT MODELS I 55

Example 1 : acute myocardial infarction in patients with NSTACS. As part of Table 3.2 STATA output for a Weibull regression to derive time-dependent
the process of populating the Markov model shown in Fig. 3.1 in the GPA proba �i!ities of acute myocardial infarction in patients with non-ST-elevation
example above (Palmer et al. 2005), patient-level time-to-event (survival) data acute coronary syndrome
were available from an observational study where the event was acute myocar­ Weibul! regression - log relative-hazard form
dial infarction. These data were from 916 patients with (non-ST elevation No. subjects 916 No. obs 916
acute coronary syndrome NSTACS ). The maximum follow-up time was No. failures 52
Time at risk* 1 058651 LR chi' (0) -0.00
5 years and 5 1 events were reported during that time. The aim was to estimate
Log likelihood -262.00507 Prob chi2 >
the risk of acute myocardial infarction in a model with yearly transition prob­
Coef. SE z P >Izl (95% Confidence interval)
abilities and to extrapolate to a "lO-year timeframe.
-8.028897 0.6855026 -1 1 .71 0.000 -9.372457 .. 6.685336
A parametric Weibull model was employed to model these data (Collett -cons
1994; Carroll 2003). The Weibull pdf is given by: /In_p -0.3040257 0.1 268008 -2.40 0.01 6 -0.5525507 -0.0555008

f(t) ; l1.ytY-] exp{-l1.tY} P 0.7378419 0.0935589 0.5754801 0.9460 1 1 3

l/p 1 .355304 0.1 7 1 8536 1 .05707 1 .73768
where the 11. parameter gives the scale of the distribution and the (ancillary)
.;.: Note that the unit of time for this regression model is measured in days.
yparameter defines the shape. From the fundamental relationships of survival
analysis described above, we can easily divide the pdf into component parts:
a and 1 , the hazard rate falls over time. Also, the Weibull model nests the
h(t); l1.ytY-] exponential as a special case when r I . =

S(t);exp{-iltr} Table 3.2 shows the results of a Weibull survival analysis using the acute
H(t);l1.tY• myocardial infarction data based on the software package STATA v7.0.
It should be noted that, in the STATA output, the constant _cons is the In(l1.)
Fig. 3.2 shows different time-dependent hazard functions that can be value and p y. Note that the 'p' parameter in the table (equivalent to the
=

modelled using the Weibull function. The shape of these alternative functions shape parameter y in the exposition above) is less than I (and statistically
depends on the value of the yparameter. For example, when yis between significantly so) indicating evidence for a decreasing hazard over time. Hence)
in the Markov model, it is possible to have transition probabilities to
y> 2 the myocardial infarction state which decrease as cycles elapse. These yearly
transitions estimated from these data are given in Table 3.3. These are gener­
ated from the following equation, which is derived from the general form in
eqn 3.4 above, substituting the expression for the cumulative hazard from the
Weibull as given above:

tP (tJ � I - {
exp A (t - U Y - M' } .

Example 2: Cause-specific death from early breast cancer following surgery.

�77�-==--=::::==== r= 1
A patient-level data set was obtained from the Medical Oncology Unit at the
Churchill Hospital, Oxford, containing personal and clinical details of breast
0< 1'<1 cancer patients treated at this centre from 1986 onwards. Patients were
�----
--
Time selected who had operable breast cancer, and who had at least 5 years of
Fig. 3.2 Illustration of the different time-dependent hazard rates that can be follow-up information subsequent to initial surgery (maximum follow-up was
generated with the Weibull model depending on the value of 1. the shape 1 5 years). The aim was to predict death from early breast cancer for women
parameter (Collett 1 994). with different prognostic characteristics (Campbell et al. 2003).
56 I FURTHER DEVELOPMENTS IN DECISION ANALYTIC MODELS EXTENSIONS TO COHORT MODELS I 57

Table 3.3 Estimated yearly transition probabilities 3 . 1 .3. Relaxing the Markov assumption
to myocardial infarction, based on the Weibu!l
Chapter 2 introduced the concept of the Markov assumption, which relates to
model, in patients with non-5T-elevation acute
coronary syndrome the fact that, in determining future transitions in such a model, there is no
memory of what has happened to 'patients' in the past. In many situations,
Year Transition probabilities (%) this feature of Markov models will not be a limiting factor. However, this may
2.47 not be the case in modelling some diseases and technologies. Given the
2 1 .68 computational ease of using cohort models, it would be desirable to be able to
3 1 .46 relax the Markov assumption while retaining the cohort feature. Figure 2.5
4 1 .34 illustrated the Markov assumption using the example of a model evaluating
5 1 .25 cancer therapy. The key feature of this model was that once a patient has been
treated for either local or regional recurrence of their cancer and entered the
6 1.19
'remission' state, the model is unable to take into account the details of those
7 1 .14
past events in determining future transitions.
8 1.1
One way to 'build memory' into the model is to add additional states. This is
9 1 .06 illustrated in Fig. 3.3 using the example introduced in Fig. 2.5. Compared with
10 1 .03 the model structure shown in Fig. 2.5, which had one state representing remis­
sion following treatment for a cancer recurrence, the revised model now has
six. The use of multiple 'remission' states allows the model to reflect that
An initial exploration of the model used the Weibull function, but found
future transitions (here, the risk of death), costs and health-related quality of
that the hypothesis of a constant hazard could not be rejected. Therefore, a
life may differ according to two factors: when a patient was treated for a cancer
straightforward exponential model (constant transition probabilities) was
recurrence (in terms of time since initial treatment or start of the model); and
employed. The pdf for an exponential distribution is given by:
whether their recurrence was local or regional to the initial primary tumour.
J(t) = ;(exp{-;(t} The first of these is a form of time-dependency with the future transitions,
and this can be split into component parts: costs and health-related quality of life dependent on how long a patient has
been in remission. As discussed in Section 3.1 .2, this form of time-dependency
h(t)=;(
S(t) = exp {-;(t}
H(t) = ;(t.
Remission Remission Remission
The Oxford data were employed to estimate the baseline transition Local recurrence (Y1 LR) (Y2 LR) (Y>2 LR)
probabilities of breast cancer death in the absence of adjuvant therapy. Thus
the baseline transitions were estimated as:
tpo (t.)= l - exp{ ;((t -u)-;(t}
{
= 1- exp -;(u }
where In(;() is the linear predictor of covariates. This is consistent with the
Regional Remission Remission Remission
process of moving between the instantaneous (constant) rate and the transition recurrence (Y1 RR) (Y2 RR) (Y>2 RR)
probability, as described above and in Box 3.1. The exponential model repre­
sents the standard assumption in Markov models: the hazard rate and transition Fig. 3.3 Relaxing the Markov assumption by adding additional states to the model
probability are fixed with respect to time. shown in Fig. 2.5.
58 I FURTHER DEVELOPMENTS IN DECISION A NALYTIC MODELS
PATIENT-LEVEL SIMULATION MODELS I 59 ·
proportions of a cohort, the memoryless feature of the Markov does not apply.
cannot be implemented in the standard Markov model in Fig. 2.5 because it
depends on time in a state that is not the initial state in the model. The addi­ Specifically, with patient-level simulation, the accumulating history of the patient
tion of extra states to the model makes this form of time-dependency possible can be used to determine transitions, costs and health-related quality of life.
because 'memory' is added. The time-dependency is characterized in terms of This can be seen in the context of the cancer model illustrated in Fig. 2.5.
two tunnel states where patients remain for only one cycle (here, 1 year), A given individual patient can experience, say, a local recurrence 3 years after
during which they may have different transition probabilities, costs and initial surgery. They can then go into the remission state, during which time
health-related quality of life to the other. If they survive the tunnel states, they are at risk of further recurrence and death. If the understanding of the
patients go into a third remission state, at which point future transitions, costs disease suggests that once in the remission state, a patient's risks of further
and health-related quality of life may be different from those in the tunnel recurrence and death are a function of whether they had a local or regional
states and are constant with respect to time. recurrence, and the time since that recurrence, additional states need to be
I-Ience the tunnel state is a way of implementing time-dependency by added to the standard Markov modeL
adding memory to a Markov model. Such models have been referred to as For patient-level simulation, the fact that patients are tracked individually
semi-Markov processes (Billingham et al. 1999). If extensive time-dependency means that it is possible for the model to reflect a patient's history (in terms,
is required, however, a large number of tunnel states may be necessary. In such for example, of details of and times since past events). There are several alterna­
situations, programming the model in a spreadsheet can be very difficult. One tive forms of this sort of model (Davies 1985; Barton et al. 2004). For example,
way around this is to use a mathematical programming language such as R, some models are similar to the standard Markov model in using states,
which facilitates multidimensional transition probability matrices (Ihaka discrete time periods and transition probabilities. Discrete-event simulation,
however, is structured around how long the individual patient remains in a
and Gentleman 1996). This sort of approach has been used in applied cost­
effectiveness studies, although it is uncommon. For example, in a recent cost­ state rather than the probability of moving to another state. More detailed
effectiveness analysis of new treatments for epilepsy in adults, a semi-Markov overviews of the use of patient-level simulation models in economic evaluation
model was used, reflecting the fact that available evidence indicated the prob­ are available (Karnon 2003; Barton et al. 2004).
ability of treatment failure reduced with time (Hawkins et al. 2005a; Hawkins There are, however) disadvantages to the lise of patient-level simulation. The
et al. 2005b). To implement this using standard tunnel states would have been first is that they can be more demanding of data. That is, if various aspects of
too complicated to program in a spreadsheet, not least because the model patient history are used to determine future prognosis, then the model will
related to sequences of treatments rather than individual therapies. Therefore, require input parameters, which are conditional on those patient characteristics.
the model was implemented in R using a transition probability matrix with For example, for the cancer model discussed above, it would be necessary to
three dimensions: one for the state where the patient starts the cycle, one for estimate the risk of death as a function of time in remission and dependent on
the state where they finish the cycle and one for number of cycles in the start­ the type of recurrence the patient had experienced. There is another perspective
ing state. Multidimensional matrices can be used to reflect other types of on this apparent drawback. Sometimes a large and detailed dataset is available
'patient history' into cohort models such as previous events. in a given area, and patient-level simulation can be used to reflect the richness
of the data. Models developed in such a way can then be used to address a
range of specific decision problems in the disease area as they emerge. This is a
3 . 2 . Patient-level simulation models
different premise for the development of decision models than is usually the case
3 . 2 . 1 . The features of patient-level simulation when a decision analysis is undertaken to inform a specific decision problem at
Although cohort models predominate in cost-effectiveness modelling, some a particular point in time.
studies focus on modelling the progression of individual patients through the An example of a patient-level simulation used to model a specific dataset
model rather than a cohort. These patient-level simulation models (sometimes and hence to become a 'generic' model available for a range of decision problems
termed microsimulations or individual sampling models) can be attractive when is the UKPDS Outcomes Model (Clarke et al. 2004; http://www.dtu.ox.ac.uk/
the assumptions of cohort models (in particular, Markov chains) prove too limit­ index.html?maindoc=/outcomesmodel/; accessed June 2006). A schematic of
ing. As individuals are moved through the model one at a time, rather than as the model's structure is presented in Fig. 3.4. This is a patient-level simulation
60 I FURTHER DEVELOPMENTS IN DECISION ANALYTIC MODELS PATIENT·LEVEL S I M U LATION MODELS I 61

'-' Start: define the folloWing pade'itt cliaraCte�,istk.s: ':.'.' : :'» >" :','::'�:':C': ": complexity to be captured in their structure that may not be possible with
�
Age at diagnosis, ethnicity, sex, BM( 'Hl:JA1c, total;HDLcholestero ' (Lipids), blood �ressur?; cohort models.
smoking status, atrial fibrillation at diagnosis, peripheral vascular disease (PVD) at dlagnosls
History of diabetes-related events:
A third limitation, which is related to the second, is the flexibility of
Ischaemic heart disease (tHD), congestive heart failure (CHF), blindness, these models to assess uncertainty. Chapter 2 outlined the key concepts in
amputation renal failure, m�ocardial infarction (MI), stroke
uncertainty analysis in decision modelling (see Box 2.1), and these are further
t developed in Chapter 4. Patient-level simulations are primarily focused on
f�,
Update:pat nt' nskfacto'rs using risk
factor equations: �[ Commence model cy�le modelling variability in events (and hence costs and outcomes) between patients.
Their rationale is that, by more fully reflecting the complexity of between-patient
HbA1cq Eq. l1 variability, a more appropriate estimate of expected cost-effectiveness will
Blood pressure Eq. 12
Total: HDL cholesterol Eq. 13
result. There is less flexibility in these models, however, in assessing parameter
Smoking Eq. 14 'Randomly ol"der:�d',riin:el;(!nt:,'�qii,a ti� : ps uncertainty. A given set of simulations (e.g. 10 000 individual patients) will be
Ischaemic heart disease' (rHD) Eq. 1
based on a set of fIxed input parameters. The conventional way in which the
t Myocardial infarctio'n 'CMI) Eq.2 implications of uncertainty in those parameters is assessed in patient simula­
Congestive heart failure (CHF) Eq. 3
Update history of diabetes-related events
Stroke Eq. 4
tion models is simple sensitivity analysis that is, to change a small number of
Amputation Eg. 5 parameter values and consider the change in results. For larger models, even
Blindness Eq. 6
-
Eq. 7
this can represent a computational burden, as the full set of simulations would

J
Renal failure
. Calculate life ,
years & QALYs Diabetes-related mortality Eq. 8 typically have to be run for all changes in the input parameters.
(conditional on CHF, amputation, Eq. 9
The major burden, however, comes when the joint uncertainty in all param­
& renalfailure, MI or stroke

"
t Yes No having occurred) eters (and its implications for decision uncertainty) is assessed using proba­

r t Other mortality Eq. lO bilistic sensitivity analysis (PSA) (Briggs et al. 2003; Claxton et al. 2005). These
, , Dead?
J methods are described fully in Chapters 4 and 5. In the context of patient-level
Fig. 3.4 A schematic of a patient-level simulation used as a 'generic' diabetes model simulation, a full assessment of parameter and decision uncertainty using PSA
using data from the UKPDS Outcomes Model (Clarke et al. 2004). will require two levels of simulation: one level based on fixed parameters to
estimate a single expected value; and the second to sample from a distribution
of possible input values. With, say, 1 0 000 simulations for the two levels of
simulations, this would result in 100 million individual simulations. This is
based on 3642 individual patients from the United Kingdom Prospective only likely to be feasible for a small proportion of patient-level simulation
Diabetes Study (UKPDS Study Group 1998) which is implemented as a series models implemented in a fast programming language. As discussed further in
of interrelated parametric equations. The value of patient· level simulation in Chapter 4, even if the focus of a model is to estimate expected costs and
this case is that it provides a flexible way of reflecting that patients are at effects, when a model exhibits a nonlinear relationship between inputs and
competing risk of seven diabetes-related complications over time (e.g. renal fail­ outputs (as it true, for example, with a Markov model), simulating across the
ure, myocardial infarction, blindness and stroke), and that experiencing one of distributions of the parameters using probabilistic methods is necessary to get
these events may change the patient's risk of experiencing one of the others in appropriate expectations.
the future. When a model is parameterized using individual patient data from a A shortcut to avoid this time-consuming computation has been developed
single study, it is important that all relevant evidence has been used. As discussed using a statistical model called a Gaussian process (Stevenson et aZ. 2004),
in Chapter 1, this can be regarded as a requirement of decision making. although this is yet to be widely used. Methods are also being developed to use
A second potential drawback with patient-level simulation is the computa­ analysis of variance (ANOVA) methods to make PSA more practical in
tional burden. To estimate expected costs and expected effects in such models patient-level simulation methods (O'Hagan et aZ. 2005). The simulation asso­
requires a large cohort of patients to be simulated. Compared with the cohort ciated with undertaking value of information analysis (see Chapters 6 and 7)
simulation, this can be time-consuming, but such models can allow a level of represents an even greater magnitude of computational burden.
62 I FURTHER DEVELOPMENTS IN DECISION ANALYTIC MODELS
PATIENT-LEVEL S I M U LATION MODELS I 63

3.2.2. Examples of a patient-level simulations

and comparisons with cohort models
Patient-level simulation models are quite widely used to evaluate changes in
health service organization. For example, a discrete-event simulation was used 1;:
c
.� e
to assess the cost-effectiveness of reorganizing surgical and anaesthesia serv­
<'iE 1;5� �
oS
ices (Stahl et al. 2004). The model sought to balance changes in safety (in 0- '"
� .8
terms of avoidable medical errors) with gains in efficiency (in terms of
z0 o�
C �
�
throughput) as a result of increasing patient volume. The authors argued for �
c
1;5
J!l E
0

the advantages of discrete-event simulation for their research in terms of the �

0
" 1;5J!l ��
need to reflect the fact that the probability of events, the duration of processes .9� .9 � t 0
c

• :r: J!l1;5�
� 0
and the rate at which patients enter the model are stochastic processes. In �
.c 0
addition, this sort of model allows patients to interact or compete for finite •�
0
� �
resources which, in turn, can stimulate further effects in the model. a:
Patient-level simulation models are also heing used for the economic evalu­
�� ��
ation of medical interventions. An example is a model looking at the cost­
@£
effectiveness of population screening for Helicobacter pylori infection to �
prevent gastric cancer and peptic ulcer disease (Roderick et al. 2003). -5
.215.
A schematic of the model is shown in Fig. 3.5. Unlike the usual Markov-type
a.�
model, the 'states' in the model reflect both health-related events (e.g. gastric �
"
cancer) and health service delivery (e.g. give triple therapy). The authors USe
the model to simulate a series of cohorts of individual patients: a prevalence
cohort and then a series of incidence cohorts. The authors do not describe
why the use of patient-level simulation has major advantages over cohort
modelling in their research. The potential limitations of the methods in
handling uncertainty are clearly demonstrated, in that only one or two param­
eters were varied at a time. Furthermore, although the authors emphasize the
major parameter uncertainties in the model, the use of formal value of infor­
mation methods would be difficult to achieve, and certainly was not attempted ,
,
in the model. ,
, . .
. . '
There is a small literature comparing cohort models and patient-level simu­ \ �-
lations. One study compared these two types of model in their predictions of
20-year events in patients with type 2 diabetes (Brown et al. 2000). The
authors compared 12 starting cohorts (or patients) in terms of baseline char­
acteristics such as age, blood sugar and blood pressure. Both models generated
clinically realistic results. The comparison indicated that the cohort model
predicted higher rates of mortality and myocardial infarction, but fewer
strokes. The extent to which this comparison was able show which type of
model is superior was, however, limited by the fact that, in addition to differ­
ent model structures, they also used different evidence and assumptions.
64 I FURTHER DEVELOPMENTS IN DECISION ANALYTIC MODELS STATIC VERSUS DYNAMIC MODELS IN I N FECTIOUS DISEASE / 65

In another study, a Markov model and discrete-event simulation were are constructed is more difficult. From a pragmatic viewpoint, the analyst
compared in the context of a cost-effectiveness analysis of chemotherapy plus needs to get a clear idea of the sort of structural features that the model should
tamoxifen versus tamoxifen alone in node-positive breast cancer in post­ exhibit, and assess how far a standard cohort-type model will take them in
menopausal women under the age of 65 years (Karnon 2003). Unlike the reflecting that structure. If such a model falls short of the requirements, there
comparison of models in type 2 diabetes, the evidence used in the two models needs to be a consideration of whether building on the cohort concept will
was the same. As a result of the differing characteristics of the Markov and bridge the gap, for example, using additional states, time-dependent probabilities
discrete-event models, however, some of the assumptions differed. For example, or a semi-Markov process. If the enhanced cohort model still falls short of
)
the Markov model had a cycle length of one month, but the discrete-event what is required for an 'adequate structure , then patient-level simulation may
simulation had a minimum time in a state of one month; in addition, survival be necessary if the computational burden involved with reflecting parameter
times with metastatic breast cancer assumed a constant mortality rate but the and decision uncertainty can be dealt with.
discrete-event simulation used time-specific rates available in the literature.
Many of the structural assumptions in the two models were, however, the
3.3. Static versus dynamic models in infectious disease
same. Probabilistic sensitivity analysis was undertaken on both types of
model, but the differences in computational burden were clear as the simula­ An important area where different types of model can potentially have a
tions for the Markov model took 1 hour and the discrete-event simulation major impact on results is in the economic evaluation of interventions for
took 3 days. The results of the two types of model were remarkably similar in infectious diseases, in particular evaluations of vaccination programmes.
terms of both the direction of the differences and the absolute values. The However, the key distinction in this area is not between cohort models and
incremental cost per quality-adjusted life-year (QALY) gained for the Markov patient-level simulations, but between dynamic and static models (Edmunds
model was £3365 compared with £3483 with the discrete-event simulation. et al. 1999; Brisson and Edmunds 2003). Most models used in the economic
In terms of uncertainty, the cost-effectiveness acceptability curves (see evaluation of vaccination programmes are static, in that they assume that the
Chapter 5) generated by the two model types were very similar. Despite the rate of infection among a susceptible population is fixed. Dynamic models
similarity in overall results, however, "[he author emphasized that there were more accurately reflect the reality of infectious diseases, in that the rate of
important differences in the models but that, in terms of the final results, these infection is a function of the number of infected individuals in the community_
tended to cancel each other out. A key difference was in the estimated survival Therefore, dynamic models explicitly allow for the effects of herd immunity in
times once a patient had developed metastatic cancer. a way that cannot be achieved with a static model. Furthermore, as described
In selecting between a cohort model and a patient-level simulation, it is in Chapter 2, static models will typically focus on a cohort that ages as it
important to be aware of their strengths and weaknesses. The methods litera­ progresses through the model. Dynamic models, on the other hand, are run
ture comparing the two types of model is small, and further research of this over many years on the basis of multiple cohorts.
type is required. A key consideration in the selection is that all models are Brisson and Edmunds used the example of an economic evaluation of a
approximations of reality, so how far should a model be pushed in reflecting simplified infant varicella vaccination programme to highlight the major
the features of a disease or the interventions? Patient-level simulations may go differences in results that can be generated by these two types of model
further in this respect, but this comes at a cost in terms of data requirements (Brisson and Edmunds 2003). The number of infected cases was predicted
and computational burden. 'A'hen alternative types of model have been devel­ over time with the two types of model (see Fig. 3.6). Fully reflecting the
oped for a given decision problem, it is quite straightforward to see whether dynamic nature of the disease with respect to herd immunity results in three
the patient-level models are worth their extra 'cost' in terms of whether the distinct phases. These are marked as 2, 3 and 4 in Fig. 3.6a, with period I being
results and conclusions differ. In the case of the Kamon comparison described the prevaccination period. The first phase is referred to as the 'honeymoon
above, the simpler Markov model would have sufficed, as the results were very period' shortly after the vaccination programme begins, when the number of
similar between the two model types (Kamon 2 0 0 3 ) . However, assessing susceptibles falls sufficiently to prevent endemic transmission and the number
potential differences in results from different model types before the models of infected cases is very low. The second is the post-honeymoon epidemic,
66 I FURTHER DEVELOPMENTS IN DECISION ANALYTIC MODELS EXERCISE: CONSTRUCTING A MARKOV MODEL OF TOTAL HIP REPLACE M E N T I 67

(4) when the number of susceptibles increases (through births) above a threshold
(a) ,
14,000 r that results in an increase in infections to epidemic level. The third period is
the postvaccination endemic equilibrium where a long�term equilibrium is
<i; 12,000 reached with lower infection levels than prior to the vaccination programme.
� 10,000
C
fZl Q·18Yls
Q 19·44yrs
The static model cannot capture these three distinct periods (Fig. 3.Sb) and, as
a result, underestimates the number of infections avoided (in the example in
o
II 45yrs+
�
c �Q) the paper, by 1 0 million over 80 years for England and Wales). The difference
.gs � 8,000 between the two models is due to the static model's inability to reflect herd
.g g immunity) the effect of which depends on the extent of continuing infection
i=> g 6,000
0 in the community. Therefore, for decision problems relating to infectious
§ ....�
<t il diseases where herd immunity is important) dynamic models are essentiaL It is
a. 4,000
possible to develop both cohort and patient-level simulation versions of
8
i 2,000 dynamic models.

-5 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 3.4. Summary
Year after start of vaccination For a given decision problem, there is a need for the analyst to decide how to
reflect the essential features of a disease) and the impact of interventions. This
( b) 14,000 has to be done in a way that recognizes that all models are approximations of
reality. It is not possible to approach this structuring process in a formulaic
<i; 12,000
iii way. Rather, judgements need to be taken in such a way as to find an appropriate
;r 10,000 balance between two key attributes. The first is to provide clinical 'face validity'
�
cQ) §Q)a. 8,000
in the sense that assumptions used in the model are plausible given what is
known about the disease and the interventions. The second is that model
:gc O
g needs to be fit for purpose in terms of its outputs, That is, it needs to provide
� g� 6,000 estimates of expected costs and effects of all relevant options over an appro­
=> 0
§ ....� priate time horizon. It also needs to be sufficiently tractable to facilitate
« ill 4,000
a.
detailed sensitivity analysis including probabilistic methods.
�
>0 2,000
8
3 . 5 . Exercise: constructing a Markov model
-5 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 of total hip replacement
Year after start of vaccination
3.5. 1 . Overview
Fig. 3.6 Comparison of the number cases of varice!la infection over time, following
the introduction of an infant vaccination programme. (a) Shows the predicted results The purpose of this exercise is to build another Markov model in Excel. The
with a dynamic model which is able explicitly to reflect herd immunity; and (b) model will illustrate a number of facets of model building in the Markov
shows the predictions with a standard static model (Brisson and Edmunds 2003), framework that complement the HIV/AIDS model example from Chapter 2.
Of particular interest is the lise of time-dependent transition functions to
represent probabilities. The exercise will get you to construct two types of
transition probabilities: time-varying transitions from a life table) and time­
varying transitions estimated from a parametric survival model.
68 I FURTHER DEVELOPMENTS IN DECISION ANALYTIC MODELS EXERCISE: CONSTRUCTING A MARKOV MODEL O F TOTAL H I P REPLAC E M E N T I 69
The step-by-step guide below will take you through a number of stages of arms of the model and for the life tables used for mortality rates. O f course,
developing the model, which is based on a simplified version of a published there are many ways to structure models - much being down to personal
cost-effectiveness model (Briggs et al. 2004). These steps are: preference. However, we hope that you will find that our chosen way of
1. Preparing parameters and naming cells. presenting the model does not deviate too much from your own preferences!
2. Parametric time.:dependent transitions from a survival analysis.
3. Life table transitions for background mortality. 3.5.2. Step-by-step guide

4. Building a Markov model for the standard prosthesis. All cells for you to complete in the exercise are coloured yellow. However,
make sure that you also pay attention to the content of the surrounding cells,
5. Adapting the model for a new prosthesis.
as although these may not require changes, they provide important information,
6. Estimating cost-effectiveness (deterministically). much of which you will have to use.
Before beginning the exercise take a moment to familiarize yourself with the
Excel template for the exercise, 'Exercise 3.5 - template.xlf. The first worksheet 1 . Preparing parameters and naming cells
in the workbook <Model Figure> contains the overall structure of the model The aim of this section of the exercise is to prepare the <Parameters> work­
you are about to build. It will be worth spending a little time familiarizing sheet to facilitate the use of parameter names throughout the rest of the
yourself with this figure; it is also presented in Fig. 3.7. The figure contains a modelling exercise. On opening the worksheet you will see that much of the
state transition diagram of the four-state Markov model you will be building work has been done for you, including entry of patient characteristics (linked
and contains a key to the different parameters of the model, which are colour­ to the <Analysis> worksheet), the inclusion of standard discounting rates, and
coded into three categories: transition probabilities, state costs and state utilities. the cost of the two prostheses that we are seeking to evaluate. This section will
The model itself is constructed across a number of sheets. There are separate guide you through the completion of the parameter information for the
sheets for the parameters of the model and the final analysis, for the survival (constant) transition probabilities, the state costs, the state utilities and the
analysis used to estimate prosthesis failure rates, for each of the Markov model naming of parameters.
i. We will start with the three transition probabilities that will be assumed
to be constant in the model: the operative mortality rates and the
re-revision rate parameters in cells B 1 6:B18. We will assume that the
operative mortality is 2 per cent for both primary and revision proce­
dures and that the re-revision rate is 4 per cent per annum. Enter these
values now.
Note that we will be returning to the time-dependent risk of prosthesis failure
(revision risk) below, so for the moment leave B22:B27 blank.
Successful ii. Now we deal with the costs in the model. Note that the costs of the
Death
Primary prostheses (standard and new) have already been entered for you. That
leaves the costs for the states of the model. However, the state cost for
the primary procedure is assumed to be the same irrespective of the
implant - we therefore leave this out (as it is received by both arms and
Successful
nets out of the analysis). For the purpose of this model we also assume
Revision THR
Revision that there is no ongoing cost to the health service if people have
a successful implant in place. Therefore the only cost to consider at
this point is the cost of the revision procedure itself. This has been
Fig. 3.7 Diagram showing the structure of the model (Briggs et al. 2004, Fig. 1). estimated as £5294 and this should be entered into cell B33.
70 I FURTHER DEVELOPMENTS IN DECISION ANALYTIC MODELS
EXERCISE: CONSTRUCTING A MARKOV MODEL OF TOTAL HIP R E P LAC E M E N T I 71
iii. The quality-of-life utilities of a year spent in each of the model health Take a moment to understand this survival model. If you are having trouble
states has been estimated to be 0.S5, 0.75 and 0.30 for the successful
with the appropriate interpretation then do re-read the appropriate section on
primary, successful revision and revision states respectively. Enter these pages 50-56 we are about to implement this model and you may struggle if
�

values into cells B41:B43. you are not clear on the interpretation from the start.
Notice that the <Parameters> worksheet bas been structured with columns i. To start with, generate a link from the <Parameters> worksheet (cells
for 'name', 'value', and 'description' for each parameter. We will exploit the fact B22:B24 and B26:B27) to the corresponding results of the survival
that we can automate the naming of parameters by giving cells the name of
analysis on the <Hazard function> worksheet (cells C6: C I O ) .
the label in an adjacent cell. Naming is very powerful - it is much easier to
Remember that these values are on the log hazard scale.
remember parameter names than cell references and this means that mistakes
ii. We now want to calculate the lambda value of Weibull distribution
are less likely to happen.
(this, together with the gamma parameter, enahles us to specify the
iv. Highlight the 'name' and 'value' of the patient characteristics (i.e. high­ baseline hazard). In estimating a standard survival analysis, it is the log
light cells AS:B9). Now choose Insert > Name > Create from the menu of the lambda parameter that is a linear sum of the coefficients multi­
bar. Excel will recognize that you may be trying to create names from
plied by the explanatory variables. Therefore, to get the value of log
adjacent labels in the left column, so answer 'OK' to the pop-up box.
lambda, multiply the coefficients of age and sex (cells B23:B24) hy the
v. Check that the naming has been successful by selecting, in turn, cells age and sex characteristics (cells BS:B9) and add together, not forget­
BS and B9. If you have successfully named these cells, then, when a celi ting to also add the constant term (cell B22). Enter this formula into
is selected, you will see the name appear in the 'name box' at the top cell B25.
left of the screen immediately above the workbook.
iii. Remember that the parameters in cells B25:B27 are on the log scale.
vi. Now continue to name all of the parameter cells on this worksheet Exponentiate each of these cells to give the value of the lambda and
down to row 43. (Note that you can even name the cells of the revision gamma parameters of the Weibull distribution and the hazard ratio for
rate regression parameters that we have not yet entered.) the new prosthesis compared with the standard.
An alternative way of naming cells is to use Insert> Name > Define from the Note that what we have estimated so far are the constant parts of the paramet­
menu bar, which allows individual cell/area naming. We will use this method ric Weibull hazard function. The task now is to make this time-dependent
in the next section for naming a column, however it is much quicker to input a we cannot do this on the <Parameters> worksheet - instead this must be done
large number ofparameter names using the automatic method. on the model worksheet itself. Open the <Standard> worksheet that contains
the outline of the Markov model for the standard prosthesis.
2. Parametric time-dependent transitions from a survival analysis
In preparation for huilding the Markov model we are going to specify the
If you open the <Hazard function> worksheet you will see the output of a time-dependent transition probability for prosthesis failure. Note the (annual)
regression analysis on prosthesis failure. A parametric Weibull model was cycle is listed in column A for 60 years - this is our time variable. Recall from
fitted to patient level survival time data to estimate this function. The regression earlier that the cumulative hazard rate for the Weibull is given by:
model shows that prosthesis survival time is significantly related to age, sex
and type of prosthesis (new versus standard). In addition, the significance of the H(t)= ;tt1
gamma parameter indicates that there is an important time-dependency to and that the time-dependent transition probability (for a yearly cycle length
the risk of failure, which increases over time. and time measured in years) is given by:
Note that the estimation of this regression model was undertaken on the log
hazard scale. We therefore have to exponentiate the results to get the actual tp(t) = l-exp{H(t - 1)- If(t)}
{ }
hazard rate. The exponents of the individual coefficients are interpreted
as hazard ratios (column E). For example, the new prosthesis has a hazard =l-eXP A(t - l)' -AtY
= l-eXP {A [(t-lY -t1 ]}.
ratio of 0.26, indicating that it is a much lower hazard compared with the
standard prosthesis.
72 I FURTHER DEVELOPMENTS !N DEC!SION ANALYT!C MODELS EXERC!SE: CONSTRUCTING A MARKOV MODEL OF TOTAL H I P REPLACEME NT I 73
iv. Use this formula to calculate the time-dependent transition probability Hints: you need to use two VLOOKUP( . ) functions within the IF(. .. )
. .

for the standard prosthesis (referred to in the spreadsheet as 'revision statement dependent on the value of the sex of the subject (cell C9 on the
risk') using the cycle number for the time variable and the lambda and <Parameters> worksheet). You have to add starting age to the cycle
gamma parameters already defined (you can now refer to these by number to get current age which is used as the index in the
name). Remember that Excel recognizes only one type of parenthesis, VLOOKUP(. . . ) function.
whereas the formula above uses different types for clarity. iii. In preparation for using this newly entered information name the
v. Finally, in preparation for using these time-dependent transitions vector E7:E66 'mr' (for mortality rate).
select this vector (cells C7:C66) and use the Insert > Name > Define You have now implemented all three key types of transition probability:
pull-down menu to label the vector 'standardRR: constant, time-dependent (function) and time-dependent (tabular).
You have just implemented a relatively sophisticated use of survival analysis
to model a time-dependent transition in a Markov model. Note that it only
4. Building a Markov model for the standard prosthesis
works because we will be using this time-dependent transition from the
initial state of the model, that is, we know exactly how long subjects have Having specified the time-dependent transition parameters on the
spent in the initial model state. Take a moment to make sure you understand <Standard> worksheet we can now proceed to construct the Markov model
what you have just implemented and the limitations on its use in a Markov proper. Inilially, we are concerned only with generating the Markov trace, that
framework. is, shOWing the numbers of patients that are in any one state at any one time.
This is the concern of columns G to L with H to K representing the four main
3. Life table transitions for background mortality model states, G represents the initial procedure, and K provides a check (as the
We can still employ time-dependent transitions in other model states providing sum across G to K must always equal the size of the original cohort).
The first step in building the Markov model is to define the transition
the time-dependency relates to time in the model rather than time in the state
matrix. This proceeds in exactly the same way as the ElV/AIDS example from
itself. This is the case for a background mortality rate, for example, which
Chapter 2.
depends on the age of a subject rather than which state of the model they are
in. In this section we will illustrate the use of a time-dependent transition to i. Start by defining the transition matrix in terms of the appropriate
model background mortality in the model from a life table. variable names, just as done for the HIV/AIDS model, using the infor­
Open the <Life tables> worksheet and familiarize yourself with the contents. mation given in the state transition diagram of Fig. 3.7.
Rows 3-9 contain figures on age-sex specific mortality rates taken from a ii. Use the transition matrix to populate the Markov model. This
standard life table, published as deaths per thousand per year. These are will involve representing the transitions between the different states
converted to the corresponding annual probabilities in rows 14-20. Notice the represented in columns G to K. (You might want to review the hints
addition of the 'Index' column - the reason for this will become apparent. given in Chapter 2 as to how to go about this). Remember not to use a
i. As a first step, name the table containing the transition probabilities remainder for the final (dead) state and then to check that all states
'Lifetable' taking care to include the index, but not the header row of sum to one in column L to make sure all people in your cohort are
labels (i.e. cells C15:E20). accounted for.
You now need to familiarize yourself with the VLOOKUP( . . . ) function in iii. When you think you have the first row correct, copy this row down to
Excel - it would be a good idea to look up the function in Excel's help files. the 59 rows below. If your check in column L is still looking good,
Also open the <Standard> worksheet and note that the time-dependent back­ then you have most likely done it correctly.
ground mortality is to be entered in column E (labelled 'Death Risk'). Now that we have the Markov trace, we can calculate the cost and effects for
ii. Use the VLOOKUP( . . . ) command nested within in an IF( . . . ) function each cycle of the model.
in order to choose a value to be entered in the 'Death Risk' column iv. In column M, calculate the cost of each cycle of the model (this is
based on the age and sex of the patient at that point in time. just the number of people in each state multiplied by the state cost).
74 I FURTHER DEVELOPMENTS IN DECISION ANALYTIC MODELS
REFERENCES I 75
6. Estimating cost-effectiveness (deterministically)
Don't forget to include the cost discount rate and the cost of the origi­
nal prosthesis in row 6. The final section of this exercise is very straightforward. We simply want to
v. In column N, calculate the life years. By doing this without quality bring all the results together onto the <Analysis> worksheet for easy viewing.
adjustment or discounting, this column can be used to calculate life 1. On the <Analysis> worksheet, link the cells for costs and QALYs for
expectancy (which is often useful, although not in this example). the two different models using the named results cells.
vi. In column 0, calculate quality-adjusted life years by cycle. Again, don't ii. Calculate the incremental cost, incremental effect and the ICER.
forget to discount. That's it! Your Markov model is now complete. I-lave a play with the patient
vii. Finally, in row 68, sum the columns and divide by 1000 to get the per characteristics to see how these influence the results. Make sure you are clear
patient predictions of cost, life expectancy and QALYs for this arm of in your own mind why patient characteristics influence the results. We will be
the model. Use the automatic naming feature to generate the names returning to this issue of heterogeneity in Chapters 4 and 5.
given in row 67.
References
5. Adapting the model for a new prosthesis Barton, P. , Bryan, S. and Robinson, S. (2004 ) 'Modelling in the economic evaluation of
health care: selecting the appropriate approach', Journal ofHealth Services Research and
Having constructed the standard prosthesis Markov model from scratch, it Policy,9: 1I0- lIB_
would be tedious to have to repeat much of the same to create the new pros­ Billingham, 1. J., Abrams, K. R. and Jones, D. R. (1999) 'Methods for the analysis of quality­
thesis arm model. Instead we will adapt the standard model. of-life and survival data in health technology assessment� Health Technology Assessment,
1. In the
<Standard> worksheet, select the topmost left-hand corner cell 3 (entire issue).

(the one without a row or column label) - this highlights the whole Briggs, A. H., Ades, A. E. and Price, M. J.(2003 ) 'Probabilistic sensitivity analysis for decision
trees with multiple branches: Use of the Dirichlet distribution in a Bayesian framework',
worksheet. Copy the worksheet and highlight the same area on the
Medical Decision Making, 23: 34 1-350.
<NPI> worksheet. Now paste and you should have an exact duplicate Briggs A, Sculpher M, Dawson J, Fitzpatrick R, Murray D, Malchau B. (2004) 'Are new
model ready to adapt. cemented prostheses cost-effective? A comparison of the Spectron and the Charnley',
ii. Firstly) we must introduce the treatment effect of the new prosthesis. Applied Health Economics & Health Policy,3 .' 78-89.

This is easily achieved as the treatment effect reduces the baseline Brisson, M. and Edmunds, W. J. (2003 ) 'Economic evaluation of vaccination programs: the
impact of herd-immunity', Medical Decision Making, 23: 76-82.
(standard) hazard ratio by the factor RRnpl (a value of 26 per cent).
Brown, J. B., Palmer, A. J., Bisgaard, P., Chan, W'J Pedula, K. and Russell, A.(2000 ) 'The Mt.
Therefore apply this treatment effect parameter RRnp I to the expres­
Hood challenge: cross-testing two diabetes simulation models', Diabetes Research and
sion in column C taking care to reduce the hazard rate by this factor) Clinical Practice, 50 (SuppL 3 ): S5 7- S64.
not the probability (see page 53). Campbell, H. E., Gray, A. M . , Briggs, A . H . and Harris, A. (2003 ) 'Cost-effectiveness of using
iii. Now rename the revision risk vector from 'standardRR) to 'np lRR). different prognostic information and decision criteria to select women with early breast
cancer for adjuvant systemic therapy. Health Economists' Study Group Conference,
iv. Now update cells H7 and 17 to refer to np lRR rather than standard RR Canterbury, July.
and copy this adjustment down to the 59 rows below. Carroll, K. J. (2003 ) 'On the use and utility of the Weibull model in the analysis of survival
v. Update cell M6 to refer to the cost of the new prosthesis rather than data', Controlled Clinical Trials,24: 682- 701.
the standard prosthesis. Clarke, P. M., Gray, A. M.) Briggs, A., Farmer, A. J., Fenn, P., Stevens, R. J., et aI. (2004 ) 'A
model to estimate the lifetime health outcomes of patients with type2 diabetes: the
vi. Finally, update the labels in row 67 to an NP 1 prefIx and use the auto­ )
United Kingdom Prospective Diabetes Study (UKPDS) Outcomes Model , Diabetologia,
matic naming feature to rename the row 68 cells below. 4 7: 1 74 7-1 759.
You should now have successfully adapted the Markov model to generate Claxton, K . , Sculpher, M., McCabe, c., Briggs,A.,Akehurst, R . , Buxton) M., et al.(2005 )
results for the cost, life expectancy and QALYs associated with using the new 'Probabilistic sensitivity analysis for NICE technology assessment: not an optional
extra', Health Economics, 14: 339-347.
prosthesis.
76 I FURTHER DEVELOPMENTS IN DECISION ANALYTIC MODELS

Collett, 0.(1994 ) Modelling survival data in medical research, London, Chapman and Hall/CRe.
Chapter 4
Cox, 0. R. and Oakes, D.(1984 ) Analysis of survival data, London, Chapman & HalL
Davies, R.( 1985 ) 'An assessment of models of a health system', Journal ofthe Operational
Research Society, 36: 679- 687 . Making decision models
Edmunds, W. J., Medley, G. F. and Nokes, D. J,(1999 ) 'Evaluating the cost-effectiveness of
vaccination programmes: a dynamic perspective', Statistics in Medicine,18: 32 63- 3282 . probabilistic
Hawkins, N" Epstein, D" Drummond, M" Wilby, J., Kainth, A" Chadwick, D., et al.(2005a)
'Assessing the cost-effectiveness of new pharmaceuticals in epilepsy in adults: the results
of a probabilistic decision model', Medical Decision Making 25: 4 9 3-510.
Hawkins, N., Sculpher, M. J. and Epstein, D. (2005b) 'Cost-effectiveness analysis of treat­
ments for chronic disease - using R to incorporate time-dependency of treatment In this chapter, we describe how models can be made probabilistic in order to
response', Medical Decision Making25: 511-519. capture parameter uncertainty. In particular, we review in detail how analysts
Hollenberg, p, (1984 ) 'Markov cycle trees: a new representation for complex Markov should choose distributions for parameters, arguing that the choice of distri­
processes'. Abstract from the Sixth Annual Meeting of the Society for Medical Decision
bution is far from arbitrary, and that there are in fact only a small number of
Making, Medical Decision Making, 4 .
candidate distributions for each type of parameter and that the method of
Ihaka, R. and Gentleman, R. (1996) 'R: A language for data analysis and graphics', Journal of
Computational and Graphical Statistics, s: 299 -314 . estimation will usually determine the appropriate distributional choice.
Karnon, J.(2003) 'Alternative decision modelling techniques for evaluation of health care Before this review, however, consideration is given to the rationale for making
technologies: Markov processes verses discrete event simulation', Health Economics, decision models probabilistic, in particular the role of uncertainty in the
12: 8 37-848. decision process. There then follows a discussion of the different types of
Miller, D. K, and Homan, S. M. (1994 ) 'Determining transition probabilities: confusion and uncertainty, which focuses on the important distinction between variability,
suggestions', Medical Decision Making, 14: 52-58 .
heterogeneity and uncertainty.
O'Hagan, A" Stevenson, M. and Madan, J. S.( 200 5) 'Monte Carlo probabilistic sensitivity
analysis for patient-level simulation models'. Sheffield Statistics Research Report
5 61 /05 . Sheffield, University of Sheffield. 4. 1 . The role of probabilistic models
Palmer, S" Sculpher, M., Philips, Z., Robinson, M., Ginnelly, L., Bakhai, A., et at.(2005 ) The purpose of probabilistic modelling is to reflect the uncertainty in the
'Management of non-ST-elevation acute coronary syndromes: how cost-effective are
input parameters of the decision model and describe what this means for
glycoprotein lIb/IIIa antagonists in the UK National Health Service?', International
Journal ofCardiology,100:229-240. uncertainty over the outputs of interest: measures of cost, effect and cost­
Roderick, P., Davies, R., Raftery, J., Crabbe, D., Pearce, R., Patel, P., et al.(2003) 'Cost­ effectiveness (whether incremental cost-effectiveness ratios or net-benefit
effectiveness of population screening for Helicobacter pylori in preventing gastric cancer measures). However, there is a legitimate question as to what the role of
and peptic ulcer disease, using simulation', Journal ofMedical Screening, 10; 148-156. uncertainty is in decision making at the societal level. In a seminal article,
Stahl, J. E., Rattner, D., Wiklund, R., Lester, J" Beinfeld, M. and Gazelle, G. S. (2004 ) Arrow and Lind ( 1970) argued that governments should have a risk-bearing
'Reorganising the system of care surrounding laparoscpoic surgery: a cost-effectiveness
role when it comes to public investment decisions. The Arrow-Lind theorem
analysis using discrete-event simulation', Medical Decision Making, 24: 4 61 -471.
Stevenson, M. D., Oakley, J. and Chilcott, J. B. (2004 ) 'Gaussian process modelling in
would suggest therefore that decision makers might only be concerned with
conjunction with individual patient simulation modelling: a case study describing the expected value decision making, and we might therefore question why uncer­
calculation of cost-effectiveness ratios for the treatment of established osteoporosis', tainty in cost-effectiveness modelling should concern us at all? We propose
Medical Decision Making, 24: 89-100. three main reasons why it is important to consider uncertainty, even if the
UKPDS Study Group (1998 ) 'Intensive blood glucose control with sulphonylureas or concern of the decision maker is expected values. Firstly) most models involve
insulin compared with conventional treatment and risk of complications in patients
combining input parameters in ways that are not only additive, but also multi­
with type 2 diabetes', Lancet, 352: 8 37-853.
plicative and as power functions, resulting in models that are nonlinear in
those input parameters. Secondly, uncertainty over the results of an analysis
implies the possibility of incorrect decision making which imposes a cost in
terms of benefits forgone, such that there may be value of obtaining more
78 I MAKING DECISION MODELS PROBABILISTIC
THE ROLE OF PROBABILISTIC MODELS I 79
information (thereby reducing uncertainty) even in a world where our only
4.1 .2. Value of information
interest is in expected values. Finally, policy changes are rarely costless exer­
cises and decision reversal may be problematic, such that there may exist value It is often wrongly assumed that a fundamental interest in making decisions
associated with delaying a decision that may be impossible, or problematic, to on the basis of expected cost-effectiveness means that uncertainty is not
reverse. Each of these reasons is explored in more detail below) and we relevant to the decision-making process. For example, in a provocatively
conclude the section with a discussion of potential biases if the estimation of titled article, Claxton (1999) proclaimed the 'irrelevance of inference', where
expected value were the only interest in decision models. he argued against the arbitrary use of 5 per cent error rates in standard
medical tests of hypothesis. Unfortunately, some readers have interpreted
the irrelevance of inference argument to mean that decisions should be taken
4 . 1 . 1 . Uncertainty and nonlinear models
only on the basis of the balance of probabilities, that is, that a greater than
Once we move beyond simple decision trees, which are linear in the input 50 per cent probability is sufficient justification for recommending a treat­
parameters, to Markov models and other more sophisticated models, the ment option.ln fact, Claxton argues for a much more rational approach to
model structure essentially becomes nonlinear. This is due to the fact that the handling uncertainty that avoids arbitrary error rates. Uncertainty is argued to
outputs of the model can be a multiplicative function of the input parameters. be costly in that there is always a risk that any decision made is the wrong one.
For example, even a very simple Markov model will involve multiplying the V\'here incorrect decisions are made, then society will suffer a loss as a conse­
underlying transition probability parameters together to generate the Markov quence. Therefore, in the decision theoretic approach (Pratt et al. 1995), value
trace) which is an inherently nonlinear construct. is ascribed to the reduction of uncertainty (or the creation of additional
It is common, in statistics) to be reminded that, for a nonlinear transforma­ information) such that a decision may include the option to acquire more
tion, g(.), the following equality does not hold: information.
Note that the decision theoretic approach ackuowledges that information
gathering entails a cost; therefore the decision to acquire more information
That is, the expectation of the transformation does not equal the transfor­ involves balancing the costs of acquiring more information with its value,
mation of the expectation (Rice 1995). The same is true of decision models as such that a decision to collect more information is not simply an exercise in
statistical models. We can consider OUf model a nonlinear transformation delaying a decision. Value-of-information methods are described in detail in
function (albeit a complex one). Our fundamental interest is in the expected Chapter 7. For the purposes of this chapter, we simply note that the value-of­
value of the output parameters (costs, effects and cost-effectiveness), but we information approach has at its heart a well-specified probabilistic model that
will not obtain this expectation by evaluating the model at the expected values captures parameter uncertainty in such a way that uncertainty in the decision
of the input parameters. Instead, it will be necessary to specify input distribu­ can be adequately reflected and presented to the decision maker.
tions for input parameters of the model and propagate this uncertainty
through the model to obtain a distribution over the output parameters. 4 . 1 .3. Option values and policy decisions
It is then the expectation over the output parameters that represent the point Value-of-information methods implicitly assume that current decisions
estimate for the decision model. should be made on the basis of expected cost-effectiveness, but that where
For this reason, even if the decision maker is convinced that their only interest there is value in reducing uncertainty and collecting additional information)
is in the expected value of the model, it is still necessary to consider uncer­ future decisions might change the expected value such that the current
tainty in the input parameters of a nonlinear model rather than simply employ decision needs to be overturned. This notion of expected value decision
the point estimates. Nevertheless, in all but the most nonlinear models, the differ­ making may not adequately reflect that policy changes are not costless, and
ence between the expectation over the output of a probabilistic model and that more importantly, may be difficult or impossible to reverse. Palmer and Smith
model evaluated at the mean values of the input parameters, is likely to be (2000) argue that the options approach to investment appraisal can
modest, suggesting the bias in the latter approach is usually not a major concern. offer insight into the handling of uncertainty and decision making in health
80 I MAKING DECISION M O D E LS PROBABILISTIC VARIABILI TY, HETEROGEN EITY AND UNCERTAINTY I 81

technology assessment. They argue that most investment decisions exhibit the more difficult to manipulate analyses directly in favour of a treatment because
following characteristics: of the clear and direct link to the evidence base.
+ Future states of the world are uncertain.
• Investing resources is essentially irreversible. 4.2. Variability, heterogeneity and uncertainty
• There is some discretion Over investment timing. Having argued that it is important for analysts to incorporate uncertainty esti­
mates for the parameters into their models) it is worth considering precisely
Under these conditions, they argue that it would be optimal to adjust cost­
the form of the uncertainty that is to be captured. Unfortunately, there exists
effectiveness estimates in line with an options approach to valuing the 'option'
much confusion surrounding concepts related to uncertainty in economic
to make a decision) assuming that as time progresses additional information
evaluation and it is often the case that the literature does not use terms consis­
will become available that may reduce some of the uncertainties inherent
tently. In this section, we distinguish between variability (the differences that
in the decision. The authors acknowledge the essentially passive nature of
occur between patients by chance) and heterogeneity (differences that occur
the emergence of additional information in the options literature and suggest
between patients that can be explained) from decision uncertainty - the
that it would be optimal to integrate the options approach into the value-of­
fundamental quantity that we wish to capture from our decision models. We
information approach discussed previously.
begin by considering the concepts of variability, heterogeneity and uncer­
4 . 1 .4. Potential for bias in decision models tainty. We then introduce an analogy with a simple regression model to
explain each concept and argue that the concepts can equally be applied to a
In the three subsections above) we argue the case for why it is important for
decision model.
analysts to represent uncertainty in their decision models even if decision
makers are only interested in expected values. An implicit assumption behind
4.2. 1 . Variability
expected value decision making is that unbiased estimates of cost-effectiveness
are available. Yet in some settings) such as in pharmaceutical company submis­ Wben we consider patient outcomes) there will always be variation between
sions to reimbursement agencies) there may be clear incentives that may different patients. For example, suppose a simple case series follows a group of
encourage bias in decision models. For example) there are two bodies in the severe asthmatics with the aim of estimating the proportion of patients that
UK that consider cost-effectiveness evidence: the Scottish Medicines experience an exacerbation of their asthma in a 1 2-week period compared
Consortium (SMC) in Scotland and the National Institute for Health and with those that did not. Suppose that of 20 patients followed, four experience
Clinical Excellence (NICE) in England and Wales. The SMC considers cost­ an exacerbation within the 12-week period such that the estimated proportion
effectiveness at the launch of products and invites submissions only from the is 0.2 or 20 per cent. If we consider the group to be homogeneous then we would
manufacturers of products. NICE considers products later in the lifecycle of consider that each patient has a 20 per cent chance of having an exacerbation
technologies, and commissions an independent appraisal of the technology over the follow-up period. However, each individual patient will either have an
from an academic group in addition to considering evidence from manufac­ exacerbation or not, such that there will be variability between patients even if
turers. It is clear that there is a keen incentive for manufacturers to 'make a we know that the true probability of an exacerbation is 20 per cent. This vari­
case' for their product, which could lead to potential bias in favour of their ability between subjects has been referred to as first order uncertainty in some of
products. the medical decision-making literature (Stinnett and Paltie1 1997), however it
If the only requirement to claim cost-effectiveness was a model that showed may be best to avoid such terminology as it is not employed in other disciplines.
a point estimate of cost-effectiveness that falls into an acceptable range, then it
is possible that combinations of parameter values could be chosen to generate 4 . 2 . 2 . Heterogeneity
the required result with no consideration of the underlying uncertainty. While While variability is defined above as the random chance that patients with
the requirement to explicitly include a probabilistic assessment of uncertainty the same underlying parameters will experience a different outcome, hetero­
(as is now required by NICE (2004)) does not guarantee that models produced geneity relates to differences between patients that can, in part, be explained.
will be unbiased, direct consideration of uncertainty may make it slightly For example, as we can see from any standard life table, age and sex affect
82 I MAKING DECISION MODELS PROBABILISTIC VARIABILITY, H E T E ROGENEITY AND UNCERTAINTY I 83

mortality rates - women have lower mortality than men (of the same age) observing four events from 20, we had instead observed 400 events out
and, beyond the age of 30, mortality increases approximately exponentially of 2000. As the proportion is still 20 per cent, the variability between
with age. Note that if we condition on age and sex, there will still be variability patients remains unchanged, however, the uncertainty in our estimate of the
between individuals in terms of whether or not they will die over a specified proportion is much reduced at se(ji) � 0.009 and associated 95% confidence
period of, for example, 20 years. The distinction between heterogeneity and interval (0.18-0.22).
variability is important - we often seek to understand and model heterogene­ It is not just parameter uncertainty that is important, however) we must
ity, as it is quite possible that policy decisions will vary between individuals also consider the importance of model (or structural) uncertainty. Model
with different characteristics. Heterogeneity is not a source of uncertainty as it (or structural) uncertainty relates not to the parameters themselves, but to
relates to differences that can be explained. For example, mortality rates may the assumptions imposed by the modelling framework. Any estimate of
vary by age and sex and, given age and sex, there may be uncertainty in the uncertainty based on propagating parameter uncertainty through the model
mortality rate. For a given individual, however, their age and sex will be will be conditional on the structural assumptions of the model and it is
known with certainty. important to recognize that different assumptions could impact the estimated
uncertainty.
4.2.3. U ncertainty

In terms of the previous section, it is uncertainty that we are seeking to 4.2.4. An analogy with regression models
capture in our decision models, rather than variability or heterogeneity. It is In the three previous subsections we distinguish between the concepts of vari­
worth differentiating two forms of uncertainty: parameter uncertainty and ability, heterogeneity and (parameter/model) uncertainty. Part of the reason
model (or structural) uncertainty. The first of these is internal to the model for confusion with these concepts in the literature is a lack of consistency in
and the second is effectively external to the mode1. terminology. In this section, therefore} we offer an analogy with a standard
To the extent that parameters of a given model are estimated, they will be regression model that should be applicable across a variety of disciplines.
subject to uncertainty as to their true value; this is known as parameter uncer­ Consider the standard regression model:
tainty. This type of uncertainty has sometimes been termed second order
p
uncertainty to distinguish it from first-order uncertainty (or variability) as y � a + IAXj + e
discussed above. An example would be the previously discussed estimate of i""l
the proportion of exacerbations. The estimated proportion was 20 per cent
based on the observation of four events out of 20. We are concerned with the which relates a dependent variable Y to p independent variables Xi' j � 1 . . p.
.

certainty of this estimate and we could employ standard statistical methods to The intercept a and coefficients f3 are the parameters to the model and can be
represent uncertainty in our estimate. The standard approach is to recognize estimated by ordinary least squares. The error term is given by e.
that the data informing the parameter estimate follow a binomial distribution All of the concepts described above in relation to decision models can be
and that the standard error of the proportion can be obtained from the bino­ given an analogous explanation in terms of the standard regression equation
mial distribution: given above. The estimated value of Y represents the output parameter of the
model, while the a and f3 values are the input parameters. The f3 coefficients
se (ji) � �p( l - p)ln represent heterogeneity in that different values of the covariates X (e.g. patient
characteristics) will give different fitted values. Note that additional parame­
where p is the estimated proportion and n is the sample size. In the example ters are required to model the heterogeneity. Parameter uncertainty is given by
above, where ji � 0.2 and n � 20, se(ji) � 0.09 and the 95% confidence interval the estimates of standard error for the ex and f3 parameters of the model.
(0.02-0.38) is obtained by taking 1.96 standard errors either side of the point Variability (or unexplained heterogeneity) is encapsulated by the error term of
estimate. the modeL Finally, note that the fitted values of the model and the uncertainty
To understand the distinction between variability between patients of the estimated parameters are conditional on the model itself. The model
and uncertainty in the estimate of the proportion, consider that instead of above assumes a ,simple additive relationship between the independent and
84 ! MAKING DECISION MODELS PROBABILISTIC CHOOSIN G DISTRIBU TIONS FOR PARAM ETERS 85

dependent variables. We might also consider a multiplicative model, for exam­ 4.3. 1 . The normal distribution and the central
ple, by assuming that covariates relate to the log of the dependent variable. limit theorem
This would lead to a different set of estimated fitted values; the difference
As was introduced in Chapter 1 and further discussed earlier in this chapter,
between the two estimates relates to model uncertainty (Draper 1995). Note
the fundamental interest for the cost-effectiveness analyst is with expectations
that the focus of the remaining sections of the chapter is the characterization
(mean values). In capturing parameter uncertainty in the estimation of the
of parameter uncertainty. We return to representing model uncertainty in
expected value of a parameter, we need to represent the sampling distribution
Chapter 5.
of the mean. The central limit theorem is an important theorem in respect of
!he sampling distribution of the mean. The theorem essentially states that the
4.3. Choosing d istributions for parameters sampling distribution of the mean will be normally distributed irrespective of
the underlying distribution of the data with sufficient sample size. This has
In this section we consider how to choose and fit distributions for
profound implications for our choice of distribution for any of our parame­
parameters of decision models ) under the assumption of a homogeneous
ters in that the normal distribution is a candidate distribution for representing
sample of patients informing parameter estimation. The use of regres­
!he uncertainty in any parameter of the model.
sion modelling to handle parameters that are a function of covariates is
In deciding whether to use the normal distribution, the issue becomes one
handled subsequently, including the use of survival analysis to estimate
of whether the level of data informing the estimation of the parameter is of
probabilities.
sufficient sample size to justify the normal assumption. Recall the previously
One criticism often levelled at probabilistic decision models is that the
introduced example of observing four exacerbation results out of 20, where it
choice of distribution to reflect uncertainty in a given parameter is essentially
was argued that the data informing the estimation of the proportion was
arbitrary, which adds an additional layer of uncertainty that must itself be
binomially distributed. In Fig. 4. 1 , the discrete binomial distribution for possi­
subjected to sensitivity analysis. In this section we argue that this is not the
ble values of the proportion is shown as the grey bars. Overlaid is the normal
case. Rather, the type of parameter and its method of estimation will usually
distribution based on the estimated standard error of 0.09 and above the
reveal a small number of (often similar) candidate distributions that should
normal distribution is an I-bar representing the previously calculated 95%
be used to represent uncertainty. These distributions will often reflect the
confidence interval (0.02-0.38).
standard distributional assumptions employed to estimate confidence inter­
vals, as described in almost any introductory medical statistics text (Altman
1991; Armitage and Berry 1994). Indeed, we would argue that by following
0.25 95% CI
standard approaches to distributional assumptions whenever possible, the
quality and credibility of the analysis will be enhanced. We begin by describing
the use of !he normal distribution, as this is effectively a candidate distribution G' 0.20
events=4
c
for any parameter through the central limit theorem, and !hen go on to consider Q) n=20
g. 0.15
individual parameter types. &
Note that although a formal Bayesian perspective could (and some would Q)
:5 0.10
"'
argue should) be adopted when fitting distributions to parameters, it is not
the purpose of this chapter to review formal Bayesian methods. We therefore £
0.05
adopt a rather informal approach to fitting distributions to parameters based
on the evidence available which will, in general, lead to very similar distribu­ 0.00 l-<4+-+-+--+--+-�-+-_f-+-I-+-4f-+-+-+-t--+--
tions to a formal Bayesian analysis with uninformative prior distributions. ()
;:>ry ()":- \)ry.; \)'? \)'?- ()� ()fp ()� C)'P (;)9J "-
Excellent introductions to the formalities of the Bayesian approach can be Proportion of patients
found in !he book by Gelman and colleagues (1995) and a specific primer on Fig. 4.1 Binomial distribution for estimated proportion based on four events out of
cost-effectiveness analysis by O'Hagan and Luce (2003). 20 with normal distribution overlaid.
86 I MAKING DECISION MODELS PROBABILISTIC CHOOSING DISTRIBUTIONS FOR PARAM ETERS I 87
It is clear from Fig. 4.1 that the use of the normal distribution in this context 0.25
is not appropriate as there would be a non-negligible probability of sampling
an impossible value - in this case a probability below zero. Note that it would 0.20 events=4
n=20
not be appropriate to simply discard impossible values in this context. If we
were to use the normal distribution shown in Fig. 4. 1, but discarded any value
fg. 0.15
that was drawn that was iess than zero, we would effectively be drawing from a Jg
w
truncated normal distribution. However, the problem is that the truncated :5 0.10
'"
normal distribution would no longer have the mean and variance that was £
originally chosen. We return to how the uncertainty in the estimation of the 0.05
proportion in this example should be handled below.
0.00 f-f-f.--i-J-..++-i-f-I-++""I'-I-..,._
. _____
4.3.2. Distributions for probabil ity parameters :;>'Y " ,,':- ,,'Y ,,'? ,,� ,,':> ,,'1> ,,') ,,'P ,,'I>
Proportion of patients
Probability parameters have an important constraint - all probabilities can
Fig. 4.2 Binomial distribution for estimated proportion based on four events out of
onlytake values between the range of zero and one. Furthermore, probabilities
20 with a beta(4. 1 6) distribution overlaid.
of mutually exclusive events must sum to one. When selecting distributions
for probability parameters, it is therefore important that the probability
parameters continue to obey these rules, even once the random variation is
beta distribution turns out to be extremely straightforward. If the data are
introduced. Exactly how the distribution of probability parameters is deter­
represented by a number of events of interest r, observed from a given sample
mined depends on the method of estimation. Below we deal with probabilities
size n, the proportion of events to the total sample gives the point estimate of
estimated from a binomial proportion and from the multinomial equivalent
the probability. Uncertainty in this probability can be represented by a beta
in a univariate context.
( a,[3) distribution, simply by setting a ; r and [3 ; n-r. Figure 4.2 shows the
Beta distribution for binomial data result of fitting a beta(4, 16) distribution to the data of the previous example.
It is clear that the beta distribution fits the da!., very well and exhibits the desired
The earlier example illustrated in Fig. 4.1 related to a proportion of 0.2 based
properties of not allowing probabilities outside of the logical constraints.
on four events out of 20. This proportion is the natural estimate of the proba­
bility of an event. However, we have already seen that a normal approximation Dirichlet distribution for multinomial data
is not appropriate for these data - so how should the uncertainty in the esti­ Instead of having binomial data, it is often common for the data to be divided
mated probability be represented? The solution to the problem comes from a into a number of categories; that is ) the data are multinomial. Recall the exam­
standard result in Bayesian statistics - the beta distribution enjoys a special ple of the HIV/AIDS model developed in Chapter 2 and with the transition
relationship with binomial data, such that if a prior is specified as a beta distri­ matrix reported in Table 2.2. The original article reported that
bution, then that distribution can be updated when binomial data are observed the transitions were estimated from a cohort of patients from the Chelsea
to give a beta distributed posterior distribution (Gelman et al. 1995). This and Westminster Hospital in the UK (Chancellor et al. 1997). The counts from
relationship between the beta and binomial distribution is termed conjugacy, which the probabilities were estimated were reported in Table 2.5 (as part
with the beta distribution described as conjugate to binomial data. For our of Exercise 2.5), where each row of the table represents the data from which
purposes, the technicalities of the Bayesian solution are not required (although transition probabilities can be estimated. It should be clear, for example, that
the interested reader is directed to the technical appendix). Rather, it is simply the probability of transition from AIDS (State C) to death (State D) is esti­
that the beta distribution is a natural choice for representing uncertainty in a mated by 1 3 12/1749 ; 0.75 and that the uncertainty in this estimate can be
probability parameter where the data informing that parameter are binomial. represented by a beta( l3 12, 437) distribution - but what about the transitions
The beta distribution is constrained on the interval 0-1 and is characterized from States A and B? The data informing these transitions are naturally
by two parameters, <X and �. When data are binomially distributed, fitting the multinomial - with four and three categories respectively. The Dirichlet
88 I MAKING DECISION MODELS PROBABILISTIC CHOOSING DISTRIBUTIONS FOR PARAMETERS 1 89

distribution, which is the multivariate generalization of the beta distribution and rearrange to give the unknown parameters as a function of the known
with parameters equal to the number of categories in the multinomial distri­ sampk mean'- and variance:

I/)
bution, can be used to represent, in a probabilistic fashion, the transition
Ji(
probabilities for transiting among polytomous categories (or model states). (a+ (3)� I
Mathematical details for the fitting of Dirichlet distribution are given in the
technical appendix, but note that the fitting of the Dirichlet distribution is just
a� Ji(a+[3).
as easy as that of the beta distribution, with parameters having the interpreta­ For example, if instead of knowing that four events out of 20 were observed,
tion of 'effective sample sizes' (Gelman et al. 1995). Thus the uncertainty in only the proportion of 0.2 and the binomial estimate of standard error of 0.09
the transition probabilities from State A of the model are simply represented were reported, it would still be possible to fit the beta distribution parameters.
by a Dirichlet( ! 2 5 1 , 350, 1 16, 1 7 ) distribution and from State B by a From the equation above, we calculate a� 0.22 . 0.8 / 0.092 0.2 � 3.75 and we
-

Dirichlet(731, 512, IS). The ease of fitting is the primary reason for employing can then calculate:
the Dirichlet distribution, as a series of conditional beta distributions will
(I )
generate exactly the same results (see appendix). Further details on the use of [3�a. -j1
the Dirichlet distribution, including a discussion of the use of priors to j1
counter potential zero observations in the transition matrix can be found in � 3.75 ·0.8/ 0.2
the paper by Briggs and colleagues (2003). ; 15

Fitting beta distributions by method of moments and, therefore, fit a beta(3.75,l5). Note that the slight difference in the fitted
In the two previous subsections, we emphasize just how easy it is to fit beta distribution comes about due to the rounding errors introduced by using
(Dirichlet) distributions to binomial (multinomial) data when counts of the figures from a published source.
event of interest plus its complement are available. However, it is not always 4.3.3. Distributions for relative risk parameters
the case that such counts are available or even appropriate. For example, when
Relative risk parameters are one of the most common parameter types used to
fitting beta distributions to secondary data or meta-analysis results, it may
incorporate the effect of treatment into models. In order to understand the
only be the mean/proportion and standard error/variance that are reported. If
appropriate distribution to choose and how to fit it, it is worth looking at the
this is the case then it is still possible to fit the beta distribution using an
background of how a relative risk is calculated and how the confidence interval
approach known as method of moments. For e-beta(a,j3) the moments of
is constructed. A generic two-by-two table is shown in Table 4.1 with n repre­
the beta distribution are given by:
senting the total observations and a, b, c and d representing the cell counts.
a The relative risk is defined:
E[el� ­
a+[3 RR _ _a
- a+c b+d
/
b _

a[3
var[8l�
(a+[3)2 (a +[3+I) a b+d
a+c '-b- '
If the sample moments Ji and s2 are known then we simply equate the
sample moments to the distribution moments: Table 4.1 Standard two-by-two table for estimating relative risk

� a
_
Treatment group Control group Total

j1 a+[3 Event present a b a + b

Event absent c d c +d
S2 � a[3
(a+[3)2 (a+[3+I) Total a + c b+d n
90 I MAKING � ECIS!ON MODELS PROBABILISTIC CHOOSING DISTRIBUTIONS FOR PARAMETERS I 91

As the relative risk is made up of ratios, it is natural to take the log to give: 4.3.4. Distributions for costs
In (RR) = In(a) - In(a + c) + In( b + d) - In(b). Just as our choice of distribution for probability data was based upon the
The standard error of this expression is given by: range of the data, so it should be noted that cost data are constrained to be
non-negative and are made up of counts of resource use weighted by unit
[ 1 1 1
se In ( RR ) = - - - + - - -
1 1 costs. Count data are often represented by the Poisson distribution (which is
J a a+c b b+d discrete) in standard statistical methods. The gamma distribution also enjoys
a special relationship with Poisson data in Bayesian statistics (the gamma is
which can be used to calculate the confidence interval on the log scale in the conjugate to the Poisson, which means that posterior parameter distributions
usual manner. To obtain the confidence interval on the relative risk scale, the for Poisson data are often characterized by gamma distributions). This
log scale confidence limits are simply exponentiated. suggests that a gamma distribution, which is constrained on the interval 0 to
Knowing that confidence limits for the relative risk parameters are positive infinity, might be used to represent uncertainty in cost parameters.
calculated on the log scale suggests that the appropriate distributional assump­ Another alternative, which is often employed in regression analyses, is the
tion is lognormal. To fit the distribution to reported data is simply a case lognormal distribution. Both the lognormal and the gamma distributions can
of deconstructing the calculated confidence internal. Continuing with be highly skewed to reflect the skew often found in cost data. Here we illustrate
the HIV/AIDS model example - the authors of the original paper employ a fitting a gamma distribution to represent uncertainty in a skewed cost param­
relative risk estimate of 0.5 1 with a quoted 95% confidence interval of eter. The use of the lognormal distribution for costs is illustrated later in this
0.365-0.710 (Chancellor et al. 1997), which they apply to the baseline transi­ chapter as part of a discussion of regression methods.
tion probabilities in the Markov model. Taking the natural logs of the point To fit a gamma distribution to cost data we can again make use of the
and interval estimates generates the following log scale estimates: -0.675 method of moments approach. The gamma distribution is parameterized as
(-1.008, - 0.342). Dividing the range through by 2 x 1.96 recovers the estimate gamma( a,f3) in Excel and the expectation and variance of the distribution can
of log scale standard error: be expressed as functions of these parameters as given below:

e�gamma( IX,I3)
[ RR)]
se In (
-D.342 - -1.008
2 x 1.96
0.173.
E[II] = af3
varl II] = ap2
Now we simply take a random draw from a N(-0.675, 0.170) distribution
and exponentiate the result. A note of caution is that some software packages (e.g. TreeAge DATA)
The approach described above is based on replicating the standard approach parameterize the gamma distribution with the reciprocal of the beta parame­
of reporting relative risks in the medical literature. Note, however, that ter (i.e. /3'= 1/13) and so care must be taken about which particular form
the mean of a lognormal distribution calculated in this way will not return the is used.
original relative risk estimate. For example, the quoted relative risk in the The approach is again to take the observed sample mean and variance and
equation above is 0.509, but the mean of a lognormal distribution with mean set these equal to the corresponding expressions for mean and variance of the
of -0.675 and standard error of 0.170 (both on the natural log scale) will give distribution:
an expected value on the relative risk scale of 0.517. This simply reflects
11 = af3,
the fact that the standard reporting of relative risk is for the modal value on
the relative risk scale rather than the mean. As is clear from this example, the It is then simply a case of rearranging the expressions and solving the two
difference between the mode and the mean on the relative risk scale is rather equations for the two unknowns simultaneously:
small. We return to the issue of estimating costs on the log scale later in this
chapter, where the difference between the mean and mode on the original
scale is of much greater importance.
(X = P:,
5
92 I MAKING DECISION MODELS PROBA8)LlSTIC DRAWING VALU E S FROM T H E CHOSEN DISTRIBUTION I 93

I
Again taking as an example a parameter from the HIVIAIDS model, consider The central point of the triangular distribution is the mode, not the mean.
the direct medical costs associated with the AIDS state of the model, which Therefore, if the modal value is not the central point between the minimum
is reported in the original article as £6948 (Chancellor et al. 1997). and maximum values, the distribution is non-symmetric and the mean will
f
Unfortunately, although this estimate seems to have been taken from a
patient-level cost data set, no standard error was reported. For the purposes of f not equal the mode. The distribution itself has three points of discontinuity at
each of the minimum, mode and maximum, which is unlikely to represent
this example, suppose that the standard error is the same value as the mean.
I our beliefs about the uncertainty in a parameter (is there really absolutely zero

II
We can estimate the parameters of the gamma distribution from the equations chance of being below the minimum?) Finally, minima and maxima are poor
above as a = 6948/6948 = 1 and f3 = 6948'16948 = 6948, hence we fit a statistics in the sense that the range of variation measured tends to increase
gamma( 1,6948). with sample size as there is more of a chance of observing an extreme value. It

Distributions for utilities

I is generally considered desirable for the variance of a parameter distribution
4.3.5.
! to diminish when we have greater information on that parameter.

I
Utility parameters are clearly important in health economic evaluation, but Consider how we might fit a triangular distribution to the simple proportion
represent slightly unusual parameters in terms of their range. The theoretical parameter where four events are observed out of 20. We certainly might make
constraints on utility parameters in terms of their construction are infinity at the modal value 0.2, but how would we set the minimum and maximum
r values? Setting them equal to the logical constraints of 0 and 1 for the parame­
the lower end (representing the worse possible health state) and 1 at the upper !
end (representing perfect health).
! ter is not advisable - by doing so, the mean of the distribution would become

I!
A pragmatic approach, often employed when health state utilities are far 1.2/3 = 0.4 which is not what we would want to use. We could try setting the
from zero, is to use a beta distribution. However, this is not appropriate for minimum and maximum as 0 and 1 then solve for the mode, giving a mean

I
states close to death where values less than one are possible. of 0.2 - but it should be clear that there is no such result for this example.
A simple transformation of D = 1 - U, such that D is a utility decrement or a A method of moments type approach, based on a symmetric distribution
disutility provides the solution. This utility decrement is now constrained on I would result in a minimum value of less than zero. The point of this example

!
the interval 0 to positive infinity and the previous methods of fitting a lognormal is to emphasize that while the triangular distribution is apparently simple, and
or gamma distribution can be applied. might therefore be considered an appealing choice of distribution, the lack of
a link to the statistical nature of the estimation process hampers rather than

I
4.3.6. Is there a role for triangular distributions?
helps the choice of parameters of the distribution.

It is common to see probabilistic analyses presented with parameter distribu­

tions represented by the triangular distribution. Triangular distributions 4.4. Drawing values from the chosen d istribution
are typically represented by three parameters: a minimum, a maximum and
a mode. The distribution itself is simply a triangle with unit area with the apex I Having considered which distributions may be appropriate for representing
uncertainty in different types of parameter distribution, we now consider the
at the mode and the probability going down to zero at the minimum and mechanics of how we can draw random values from these distributions.
maximum values. The mean and variance of the triangular distribution are Most software packages include some form of random number generator
given by: (RNG) that will give a pseudo random number on the interval 0-1. As all
values within the interval are equally likely, this RNG gives a uniform distribu­
mean =!:.( min+ mode + max ) tion. This uniform distribution forms the building block of random sampling.
3
var = J...
18
(min2 + mode2 + rnax2- min· mode -· min· max-mode · max). 4 . 4. 1 . The uniform distribution and the cumulative
distribution function
The apparent popularity of the triangular distribution may come from We need a way of mapping a random draw from a uniform distribution to a
the fact that it is simple to fit, requiring only three points to be specified. random draw from a distribution that we specify. To do this, we need a little
However, as a representation of uncertainty, it often leaves a lot to be desired. background on distribution functions.
94 I MAKING DECISION M O D E LS PROBABILISTIC DRAWING VALU ES FROM T H E CHOSEN DISTRIBUTION 95

The distribution that we wish to draw from is a probability density function cdf function returns the integrated probability denSity to that point. Instead,
(pdf) which defines (for a continuous function) the probability that a variable we specify the integrated probability and get the value. I n Excel, the
falls within a particular interval by means of the area under the curve. NORMDIST(x,O, l , l ) gives the integrated probability p from the standard
The characteristic feature of a pdf is that the total area under the curve normaL The inverse cdfis NORMINV(p,O,I) � x, which can be made proba­
integrates to 1. bilistic by replacing p with the RANDO function.
The cumulative distribution function (cdf) defines the area under the pdf to
a given point. Therefore the cdf is constrained on the interval 0-1. It is the cdf 4.4.2. Correlating parameters
that can provide the appropriate mapping from a uniform distribution to the A common criticism of probabilistic sensitivity analysis is that parameters are
corresponding pdf. assumed to be independent. However, it is important to realize that while this
To see how this mapping occurs, consider Fig. 4.3. The top right-hand panel is commonly what happens in applied analyses, it is possible to correlate
shows the cdf for the standard normal distribution (i.e. N(O,!) . Note that the parameters if the covariance structure between them is known. Rather, the
vertical axis runs from zero to one. problem is more that analysts usually have no data on the covariance structure
Now consider a random draw from the uniform distribution (top left panel and so choose not to model covariance (in either direction).
of the figure) from this 0-1 intervaL By reading across from the vertical axis to One example where we clearly do know the covariance relationship
the cdf curve and down to the horizontal axis, we map a random draw from of parameters is in a regression framework, where we have access to the
the uniform distribution to a random draw from the N(O, I ) distribution. variance-covariance matrix. In this situation, we can employ a technique
Repeating this process a large number of times will generate an empirical known as Cholesky decomposition to provide correlated draws from a multivariate
picture of the N(O, I) pdf shown in the bottom right panel of the figure. normal distribution, Obtaining correlations between oilier distributional forms
Of note is that in using the mapping process above we are effectively using is not straightforward; instead the approach taken (in common with many
the inverse of the cdf function. Usually, specifying a value of the variable in a other statistical procedures) is to search for a scale on which multivariate
normality is reasonable.

�.- , l-� ��
Cholesky decomposition for multivariate normal distributions
The starting point for the Cholesky decomposition method is the
1· · ·--·----·-· · --., .-. - . ..-:;/ variance-covariance matrix, such as would be obtained from a standard
regression, call this matrix V. The Cholesky decomposition of matrix V is a

JJo :::
lower triangular matrix (a matrixwhere all cells above the leading diagonal are
zero), call this matrix T, such that T multiplied by its transpose gives the
+- _ __ = _- -- _-'- _ - --_ --
covariance matrix, V. In this sense, we can think ofT as being like the square
-3 -2 -1 0 2 3 root of the covariance matrix.
Once matrix T has been calculated, it is straightforward to use it to generate
a vector of correlated variables (call this vector x). We start by generating
a vector (z) of independent standard normal variates and apply the formula:
x ::::: y + Tz, where y is the vector of parameter mean values.

It can sometimes be difficult to understand conceptually how Cholesky

decomposition works. Indeed, the best way to understand Cholesky decompo­
-3 -2 sition is to see it working in practice by doing a small example by hand and
Fig. 4.3 Using a uniform random number generator and the inverse cumulative working through the algebra. In this simple example we assume just two
distribution function to generate a random draw from the corresponding probability parameters to be correlated. The starting point is to write down the general
density function. form for a Cholesky decomposition matrix, T and to multiply T by its
96 I T
>­ REGRESSION MODELS FOR HANDLING HETEROGENElTY I 97
II
MAKING DECISION MODELS PROBABILISTIC

transpose to give a 2 x 2 matrix. This matrix can then be set equal to the correlation coefficient. This can be achieved by setting up a matrix of correla­
variance-covariance matrix:
I tions t9 be achieved and then using this correlation matrix to draw random

(a o)(a b) = (a2 )( )
I
parameters from the multivariate standard normal distribution using the
ab var (xJ cov ( Xl'X2 ) Cholesky decomposition method as described above. Having generated these
b c ° c ab b 2 +c 2 = pse(x, )se(x,) var(x2 ) .
random draws which have the required correlation structure, they can be back
transformed to a uniform distribution using the standard normal distribu­

I!
For a known variance-covariance matrix, it is straightforward to solve for tion. This vector of draws from the uniform distribution has rank order
a, c
the unknown b and components of the Cholesky decomposition matrix in correlation between the parameters given by the original correlation matrix.

)l
terms of the known variance and covariance: The individual elements can then be combined with the desired inverse

(a ) = ( cumulative density functions to generate rank order correlations between

�
J
( se ( x, )
0 0 0
I non-normal distributions.
b c l cov(xl'x2 )/a �var(x2) -b2 = p· se(x2) JI-p2 .se(x2) .
To generate correlated random variables we go back to the original Cholesky (I 4.5. Regression models for handling heterogeneity
Previously, our concern was with choosing distributions when fitting parame­
= +
equation of x Y Tz: ters to a single group of homogeneous patients. However, it is rare that

! patients are truly homogeneous and it is common to use regression methods

to explore heterogeneity between patients. In the sections below a number of

I
regression models are introduced. All share the same basic structure, being
Multiplying this expression out gives based around a Hnear predictor, but the form and scale of the regressions
differ according to the nature of the data.

! 4 . 5 . 1 . Logistic regression to estimate probabilities

I
from binomial data
and then substituting in the definitions of a, b and c we have defined previ­ In a paper looking at the cost-effectiveness of ACE-inhibitors for the treat­
ously gives: al.
ment of stable coronary disease, Briggs et (2006) employed a regression
model to calculate the probability of death conditional on having had a
primary clinical endpoint from the clinical trial (a combined endpoint of
myocardial infarction, cardiac arrest and cardiovascular death). The data to
which the model was fitted were binomial - from a total of 1091 primary clin­
from which it is apparent how the procedure is working. For example, the first ical endpoints, 406 (37 per cent) were fatal. The fitted model was a standard
random variable will clearly have the mean and standard error required. The logistic regression model of the form:
second random variable will also have a mean and standard error given
)
by the associated parameter s mean and standard error. The correlation is
introduced through the shared component of variance 2, in proportion to the
In
(-) = a+"
It 1
I - It
p
.:.. f3)X)
)=1
overall correlation.
where It represents the probability of death, the term on the left is the log-odds
Generating rank order correlation of death) and where the covariates are assumed to have an additive effect on
As mentioned above, while Cholesky decomposition can be employed to the log-odds scale. The results are shown in Table 4.2, which suggests that age,
correlate multivariate normal parameters, the correlation of other types of cholesterol level and a history of a previous myocardial infarction all increase
distribution are less straightforward. A practical solution is to correlate the the odds of an event being fatal, given that the primary clinical endpoint has
ranks of random draws from distributions rather than focus on the (Pearson) occurred.
98 I MAKING DECISION MODELS PROBABILISTIC REGRESSION MODELS FOR HANDLING HETEROGEN EITY I 99

Table 4.2 Logistic regression for the probability of having a fatal event given that a Although the probability of a fatal event is the parameter of interest in the
primary clinical endpoint has occurred above model, it could be considered as an endogenous parameter, being fully
determined by other, exogenous, parameters - the coefficients from the
Covariate Coefficient Standard error Odds ratio 95% Confidence
interval regression results, If we want to represent uncertainty around the probability
parameter there are two broad approaches that could be used. Firstly, the
Age 0,040 0,007 1 .G40 1 .026-1.054
Cholesky decomposition method could be employed as described previously,
Cho!esterol 0 , 1 87 0,057 1 .206 1 ,079-1347
under the assumption that the coefficients in Table 4.2 follow a multivariate
Previous myocardial 0,467 0,1 50 1 596 1 . 1 88-2 , 1 42
normal distribution,
i nfarction
The covariance-correlation matrix for the logistic regression model above is
Intercept -4373 0598
shown in Table 4.3. The leading diagonal shows the variances and above
the leading diagonal, the covariances between the coefficients are shown, As the
covariance matrix is symmetric, the cells below the leading diagonal show the
correlation coefficients. It is clear from these coefficients that there is consider­
It should be clear from the form of the logistic regression model above that
aole correlation betvveen the estimated intercept term and the other parameters,
exponentiating both sides will generate a multiplicative model of the odds,
and moderate correlation between the age and cholesterol coefficient parameters,
such that the exponentiated coefficients from Table 4.2 are odds ratios, These
An alternative, but equivalent, method for estimating the uncertainty in the
are shown in the fourth column of the table together with their corresponding
probability is to note that the general (matrix) formula for the variance of a
95% confidence intervals, In order to generate the probability of a fatal event
it is necessary to go one step further and rearrange the equation above to give
( )
linear predictor (LP) is given by var LPo X;rVXo' where Xo is the column
=

vector of covariates for a given individual patient (and X6' a fOW vector being

{ �fljXj} [ O ""f]
the following expression for the estimated probability:
its transpose) and V representing the variance-covariance matrix for the
coefficient parameters. For our example:

'
exp a +

� IAXj }
0000" ,000075 ,000047
r ,000075 ,003201 -,000402 -.022327 6
l + exp a + var(LPo) = ( 65 6 1 I) ,000047 -,000402 ,022592 -.017710 1
l )""1
-,003421 ·-,022327 -,017710 .357748 1
So, for example, to estimate the probability of a fatal event for a 65-year-old,
= 0,0059
with a cholesterol level of 6 mmolll and a history of previous myocardial
infarction, we first estimate the linear predictor (LP) as
Table 4.3 Covariance (correlation) matrix for the logistic regreSSion model of
LP ; -4.373+65 x 0,040+6x 0.187 + 0.467 Tatle 4.2. Leading diagonal shows the variances and above the l eadi ng diagonal
; -0.184 the covariance between parameters. Below the leading diagonal, the correlation
between parameters is shown
and then substitute this estimate into the equation above to give the proba­
Age Cholesterol Previous myocardial Intercept
bility as:
infarction
{
exp -o,184 } Age 0,000047 0,000075 0,000047 -0,003421
l+exp -o.l84{ } Cholesterol 0. 1 9 0,003201 -0,000402 -0,022327
; 0.454 Previous myocard'ial 0.05 -0,05 0,022592 -0,01 77 1
i nfarction
that is, an elevated risk compared with the overall mean of 37 per cent chance
Intercept -0.83 -0,66 -0,20 0.357748
of a fatal event.
100 ! MAKING DECISION MODELS PR08ABIUSTIC
REGRESS!ON MODELS FOR HANDLING HETEROGENEITY I 1 01

(a) Using Cholesky decomposition 4.5.2. Survival analysis to estimate probabilities from
time-to-event data
50 In Chapter 3, the use of standard survival analysis methods for the estimation
of transition probabilities in Markov models is discussed. In particular, the use
g 40
of the Weibull model for the hazard of failure is highlighted. Recall that the
'"
�
Weibull model is a two parameter proportional hazards model, with shape
I parameter y, and a scale parameter A. Also recall from Fig. 3.2 how different
values of y lead to very different shapes for the underlying hazard function.
The formulae for the Weibull distribution and the corresponding hazard and
survivor functions are given below:

Probability of fatal event /(t) = AytY-1 exp{-At!}

(b) Assuming normal linear predictor h(t)=Ayty-1
50

45
S(t)=exp{-AtY}.
40
Our concern in this section is to illustrate how heterogeneity between
5
,., 3
o patients can be modelled in the context of a parametric survival analysis. This
lE $
0
=>
,,.,5
is usually achieved by assuming that covariates act (in proportionate fashion)
�
"- 20 on the scale parameter, but that the shape parameter remains the same for all
15 patients'! In algebraic terms we specify:
10
p

71 76 $! 86 "1 96 !O!
InA=a+ lAXi)
j=l
Probability of fatal event
Fig. 4.4 Simulating two equivalent approaches to generating a probabilistic probability such that the value of the lambda is constructed from the exponential of the
parameter from a logistic regression model. linear predictor from a regression.
As an example, consider the Weibull regression of late prosthesis failure
from the total hip replacement model reported in Table 4.4 (Briggs et al.
2004). The second column gives the estimated coefficients from the Weibull
which can be used as the variance in a normal distribution with mean equal to model on the log scale with the log scale standard error in column three.
the point estimate of the linear predictor. Random draws from this distribu­ Exponentiating the estimated coefficients give the hazard ratios in the case
tion can be used to estimate the uncertainty in the probability once the inverse of the covariates, the baseline hazard in the case of the intercept and the value
logistic transformation is applied. of the shape parameter. The intercept (or log of the baseline hazard) relates to
The equivalence of the two approaches can be seen in Fig. 4.4 which
shows the histogram of a Monte Carlo simulation for the probability of a
fatal event using each approach. Although there are some slight differences
1 It is possible to parameterise the shape parameter in terms of covariates. However, this
apparent in the histograms, this results from random chance. The mean makes the representation much more complex (for example, by parameterizing the shape
values (standard deviations) of the Cholesky and normal linear predictor parameter of the Weibull model it would no longer be a proportional hazards model). In
approaches from the simulations were 0.448 (0.019) and 0.446 (0.0 1 9 ) , practice, such models are rarely reported and we are unaware of any application in the
respectively. health economic evaluation field that have yet used such models.
1 02 I MAKING DECISION M O D E LS PROBABILISTIC
REGRESSION MODELS FOR HANDLING HETEROGENEITY 1 1 03

Table 4.4 Weibull survival model for the hazard of late prosthesis failure
Table 4.5 Covariance (correlation) matrix for Weibull model of Table 4 . 5
Covariate Coefficient Standard error exp(coefficient) 95% Confidence
Spectron Years over 40 Male Intercept Gamma
interval
Spectron 0 . 1 46369 -0.0001 1 1 0.000184 -0.000642 0.000259
Spectron prosthesis -1 .34 0.383 0.26 0 . 1 2-0.55
Years over 40 -0.06 0.000027 0.000033 -7.83E-04 2.80E-08
Years over age 40 -0.Q4 0.005 0.96 0.95·-0.97
Male 0.00 0.06 0.0 1 1 895 -0.007247 0.000051
Male 0.77 0 . 1 09 2.16 1 . 74-2.67
Intercept -0.01 -0.72 -0.32 0.043 2 1 9 -0.005691
Intercept -5.49 0.208 0.0041 0.0027-0.0062
Gamma 0.01 0.00 0.01 0.58
Gamma parameter 0.37 0.047 1 .45 1 .32- 1 .60

uncertainty in a Weibull model to estimate a transition probability, we would

be neglecting the correlation between the two parameters. Such a case is
a 40-year-old woman given the standard (Charnley) prosthesis.
illustrated in the upper part of Fig. 4.5 where the left-hand panel shows the
Exponentiating the linear predictor gives the value of A for the Weibull model.
independent random sampling of the A and yparameters (assuming normal­
For example, the baseline hazard is the A value for a 40-year-old woman with a
ity on the log scale), and the right-hand panel shows the associated transition
Charnley prosthesis at the point of the primary replacement (i.e. time point
probability estimates over time including estimated 95% confidence intervals.
zero). The value of A for a 60-year-old man who has been given the Spectron
prosthesis is calculated as:
0.6 0.03

{
A = exp -5.49 + (60-40 ) x -O.04 + 0.77 - 1 .34 } O.S ..§' 0.025
""", 95% Confidence Interval
- Trans�ion probability

= 0.001 1. 0.4 � I 0.02

The gamma parameter of the Wcibull distribution is also presented in 0.3

$ 01
e
�
� 0.015

Table 4.4 and is also estimated on the log scale. The exponentiated coefficient 0.2
.'l :�§ 0.01
is significantly greater than one, but also less than two, suggesting that the eo
0.1 0.005
hazard function increases over time, but at a decreasing rate.
���-_��-' o
Once the shape and scale parameters are estimated, the calculation of the -6.5 -6 -5.5 -5 -4.5 -4 2 3 4 5 6 7 8 9 10
transition probability (as a function of the patient characteristics) proceeds Log Lambda Years
just as described in Chapter 3. The question becomes how to incorporate 0.6 0.03
,'.',.j 95% Confidence Interval
uncertainty into the estimation of the transition probability. The os .i!' 0.Q25 -Transilion protJabiiity

covariance-correlation matrix for the Weibull model of Table 4.4 is shown in 15

, 0.4 ct! jg 0.Q2
Table 4.5 with variances of the estimated coefficients on the leading diagonal,
covariance above the leading diagonal and correlation coefficients below.
0.3
�
(!i
e
�
g 0.015
go ..
In the previous section, in the context of logistic regression, it was argued 0.2 ...lffi 0.01
eo

t:
that there were two approaches to incorporating uncertainty in the estimated 0.005
probability - either to use the Cholesky decomposition of the covariance
.
� ���� -4 '
matrix. Or to estimate the variance of the linear predictor from the covariance -6.5 -6 -5.5 -5 -4.5 -4 2 3 4 5 6 7 8
Log Lambda Years
matrix directly. Note, however, that the latter approach is not appropriate for
survival models with more than a single parameter. If we were to estimate Fig. 4.5 The impact of correlating parameters in a Weibu!l survival regression on the

A and its associated uncertainty from the linear predictor directly and then uncertainty of estimated transition probabilities The top panel shows the conse­
quences of failing to account for correlation between the parameters. Appropriate
employ that parameter along with the estimated value of ytogether with its
correlation is shown in the bottom panel.
1 04 I MAKING DECISION MODELS PROBABILISTIC REGRESSION MODELS FOR HANDLING HETEROGENE ITY 1 1 05

Table 4.6 Ordinary least squares regression of age on cost of primary care over four
month period of the UKPDS
Covariate Coefficient Standard error 95% Confidence interval
Age 0.83 0.24 0.36 1 .30
.....................................
Constant 22.40 1 5.08 -7.1 7-51 .97

500 1006 1500 2000

Residuals

Contrast this with the lower part of Fig. 4.5, which shows the equivalent (a) Distribution of residuals (b) Residual versus fitted plot
results but where Cholesky decomposition is employed to ensure that the A. Fig. 4.6 Regression diagnostics for the ordinary least squares on 4�month primary
and y parameters are appropriately correlated on the log scale (assuming joint care cost.
normality). Notice that by including the appropriate correlation, the uncer­
tainty in the estimated transition probabilities is substantially reduced. This model suggests that each year of age increases cost by 0.7 per cent of
cost, rather by a fixed amount. The regression diagnostics for this model look
4 . 5 . 3 . Regression models to estimate cost/utility data
much better behaved on the log scale (see Fig. 4.7). However, in any proba­
As with considering univariate parameters for cost or utility decrements! bilistic model it will be necessary to estimate costs on the original scale, but
disutilities (see earlier), the normal distribution is a candidate when using retransformation back from the log scale is not straightforward due to the
regression methods to adjust for heterogeneity (through the central limit nonlinear nature of the transformation. For example, the estimate of expected
theorem). However, both costs and utility decrements can exhibit substantial cost on the original scale is not obtained by exponentiating the linear predictor
skewness which may cause concern for the validity of the normal approxima­ from eqn 4.1 above:
tion. Consider the ordinary least squares regression model reported in
Table 4.6, showing the relationship between age and primary care costs over a
4-month period in patients participating in the UK Prospective Diabetes Rather, it turns out that a smearing correction must be applied (Duan 1983;
Study (1998).
Manning 1998), which in the case of the log transformation of costs corre­
The results suggest that each year of age increases the 4-month cost of
sponds to a factor equal to the mean of the exponentiated log scale residuals:
primary care by 83 pence. If these results were to be used in a probabilistic
model, then it would be possible to estimate cost from the above regression
using either Cholesky decomposition or the variance of linear predictor
E[Co J � exp{a + !3 x ageo } x -n1 I" exp{eJ
i=!
method described above. However, examination of standard regression diag­
nostic plots presented in Fig. 4.6 suggests reasons to be cautious of the normal This added factor complicates the adaptation of log transformed regression
assumption. The residual distribution is clearly highly skewed and the residual models as the smearing factor becomes an additional parameter to the model.
versus fitted plot suggests possible heteroskedacticity. An alternative approach to modelling costs is to use a generalized linear model
A common approach to handling skewness is to consider fitting a regression
model to the natural log of cost, in the case of the UK Prospective Diabetes Table 4.7 Ordinary least squares regression of age on the natura! log of primary care
Study (UKPDS) data, the model would be: cost over a 4-month period in the UKPDS
In(Ci) � a+!3x agei +ei ' (4.1) Covariate Coefficient Standard error 95% Confidence interval
Age 0.007 0.002 0.003-0.0 1 1
This model is presented in Table 4.7, with the corresponding regression ........................................................................................
Constant 3.315 0 . 1 37 3.047-3.583
diagnostics (on the log scale) in Fig. 4.7.
1 06 I MAKING DECISION MODELS PROBABIUSTIC
SUM MARY I 1 07

4 . 5 . 4 . A note o n the use of GLMs for proba bilistic

analysis
However, care needs to be taken when using a GLM with a nonlinear link
function as part of a probabilistic model. Examination of the covariance
matrix reveals the covariance between the age coefficient and the constant
term in the regression model is -0.000577. For the example of a 60-year-old,
the linear predictor is 4.294 = 3.634 + 60 X 0.0 1 1 and its variance is estimated
as 0.033 = 0.01 372 + (60 X 0.002) 2 + 2 x 60 0.002 x -0.000577. Therefore, it
would be natural in a probabilistic model to assume a lognormal distribution
(a) Distribution of residuals (b) Residual versus fitted plot for mean cost by drawing from a normal distribution with mean 4.294 and
Fig. 4.7 Regression diagnostics for the ordinary least squares on the natural log of variance 0.033 before exponentiating the result. However, it is important to
4�month primary care cost.
recognize that the mean of a lognormal distribution is given by the exponent
of the log mean plus half the log variance. Therefore, if the expectation of the
probabilistic distribution is to correspond to the mean of the cost from a log
(GLM) where a distribution for the underlying data is assumed together with link GLM, it is necessary to adjust the log mean by taking away half the log
a scale for the linear predictor. GLMs have gained in popularity for modelling variance. For the example of the 60-year-old, by sampling from the normal
costs as they model the expectation of the cost directly. For the UKPDS example, distribution with mean 4.294 - 0.033 / 2 and variance 0.033 the resulting
a GLM can be written as: lognormal cost distribution will have the desired mean of £73.26. (In this
example the correction factor is unimportant as the mean of the lognormal
g(E[C; ])=a+/3 x age; (4.2) distribution with mean 4.294 and variance 0.033 on the log scale is only £74.48,
however, for much higher costs or more skewed/uncertain parameter estimates
where the function g(.) is known as the link function. This link function spec­
the difference could be much more substantiaL)
ifies the scale of measurement on which the covariates are assumed to act in
linear fashion. In the case of the log link, it should be clear from eqn 4.2 that
the expected cost is obtained by simple exponentiation of the linear predictor. 4.6. Summary
In Table 4.8, a GLM is presented for the UKPDS example assuming a gamma The purpose of this chapter has been to argue that the choice of probability
distribution for the cost data and a log link. The coefficients are therefore distribution to represent uncertainty in decision analytic models is far from
estimated on the log scale and are similar to those reported in Table 4.7. arbitrary. As the focus of the analysis is on estimating expected values of
In contrast to the results of the log transformed cost data, however, the back parameters) the normal distribution is always a candidate distribution due to
transformation to the raw cost scale is straightforward. For example, the the central limit theorem. However, this theorem is based on asymptotics and
expected primary care cost for a 60-year-old would be estimated as £73.26, decision models are often constructed with parameters that are informed by
which is calculated by exp{3.634 + 60 x O.O l l } . very little data. In such situations, the choice of distribution should be
informed by the logical constraints on the parameter and the form of the
data/estimation process. Although other distributions, such as the triangular
Table 4.8 Generalized linear mode! regression of age on primary care cost over a
distribution, have proved a popular choice in the past, we argue against the
4-month period in the UKPDS use of these distributions on the basis that they bear little relationship to
the sampling distribution of the mean of sampled data. As an aide memoir,
Covariate Coefficient Standard error 95% Confidence interval
Table 4.9, provides a summary of the types of parameters commonly encoun­
Age 0.0 1 1 0.002 O.ooS-0 . Q 1 7 tered in health economic models, logical constraints on those parameters and
Constant 3.634 0.137 3 .259-4. 0 1 0 candidate distributions for the parameters based on the data/estimation process.
I MAKING DECISION MODELS PROBABILISTIC
EXERCISE: MAKING THE H IV/AIDS MODEL PROBABILISTIC I 1 09

x In addition, we have attempted to illustrate the appropriate handling of

]5 '" '" '" heterogeneity through regression methods and the importance of including
...
c 0. > :;- ----:- > ----:- >
'" 0. 2 2 - 2 - z parameter correlations into any probabilistic model. We have argued that
E > - > - "
:i' :;-
m -
z :;: ;;; :;:
E :;:
� �
:;- � '"'" « '" <{ '" z the distinction between variability, uncertainty and heterogeneity is impor�
8 z -'0 0 0
� z z
:;: 0 :;: 0
:i' tant, and although variability is not as important for decision making per se
:;: b :;: b '"
e � e ll:: ll::
<{
� <{ �
0
.ll
w
co
O x
Z w
x
w "' ;;) "' ;;) Z as it is in the environmental literature (Thompson and Graham 1996;
Thompson 2002), the separation of variability and heterogeneity from
uncertainty underlies better decision making, as we discuss in the remaining
chapters.

4,7 . Exercise: making the HIV/AIDS model probabilistic

.,:' 0 0 0 0
e o A A A A 4.7. 1 . Overview
0
A2 2 2 2
'"
'.p c
.. - o . o E. 0
E .'" a:' S E
AS A A The aim of this exercise is to demonstrate how the deterministic HIV/AIDS
'15;
"
'" ... 0 S' � :?:' c:c:t. -j "'- -' 'b model from Exercise 2.5 can be made probabilistic by fitting distributions to
'l:
A
>
" -. t:i' 2:,
10:
"'- E '"
.�
E -' E -' E
'6 " ",- ", "'- '" 'b parameters. Emphasis is on choosing the correct distribution for different
.. €� m "m " m
"" IO E .5
m
,� lii -,o E E - E
m � types of parameters and in particular on the use of the Dirichlet distribution
E 0 m
... � �
" a E 0
fj
e - 0
E e
'" '" §0
E e
=> '6 m e e
'" '" for generating a probabilistic transition matrix, where constant transitions are
c
.'C 01 '"
", .9 ", .9 Z
.s a
1;;
.9
- 0
co Q � appropriate,
�
e The step-by-step guide below will take you through a number of stages of
0
'+,
=>
.0
making the model probabilistic:
·c
t: e
'5
�
e '"'" 1. Using a lognormal distribution for relative risk.
-'" 0
'f' E
2, Using a gamma distribution for cost,
'" e
�
0
'0
'5 e
0
"f:
0.
e
�=> 3, Using a beta distribution for the transition from AIDS to death.
0
8. '�
0 0.
VI
0
l3 'S v c

'0 ...
'f'
0 � -
o '" � '" '"
m
4, Using a Dirichlet distribution for the transition matrix,
0. 0. 0 �
0
'" ""
..
e '0 _
m 10
E => E

�I
0
� 1ii 0. � m
1! e A template for the exercise, (Exercise 4.7 - template.xIs', is available, based on
E
- >
E " 2: "�
'"
'" '0
'0 '"
0.
'" ... '0 '0 � the solution to Exercise 2.5 adapted to allow additional information to be
c 10 .� � e
:v '" E Q) .fJ
.:..: C
'
0
" E 0
0 .g entered.
t;
E
III
'" c
... ..
om '" .- '" >
f" �
'0
�
=>
=> ,
.0 1
E '" 0 'C

I
i" 12 m => t:
'"0. '!; '; E
o 0
0 _ E �'" e '0 4.7.2, Step-by-step guide
5
E .10: 0 E Q) 0 ""E �
e
c
0
E
..
... '"
c
iii
=> E
� i= iii
0
� 0
u <{ On opening the exercise template spreadsheet, take a moment to examine the
E changes from the spreadsheet solution to Exercise 2.5. In particular, you
8

I
will see that the parameter list is now labelled 'live' and that cell C3 allows
'0
2c switching between a probabilistic model and generating a deterministic point
'" ....
E .S "-
ffi o
estimate (from columns C and D respectively). Make sure the switch for the
E::> f! �

I
v>
10:
�
E AI '" probabilistic model is set to zero, so that the deterministic column D appears
c � ;;> '" 1;;
� "
.f; ·c
v _
'" � E in the live cells (column B). The general task in this exercise is to complete the
::; -W � :B VI
"i - -0 .t: f"
E i3 .0
'" 0 S
AI
m probabilistic column C so that when the switch in C3 is set to I the model
- .-
.p
� .. '" e
"
VI .!2 A
'
Vi £ 'S
';::; . ;2
0.

{! � g � 8 "'
"' -
''' u _
0 ",
::l Q � I becomes fully probabilistic.
110 I MAKING DECISION MODELS PROBABILISTIC
EXERCISE: MAKING THE TOTAL H I P REPLACEMENT MODEL PROBABILISTIC 1111
1 . Using a lognormal distribution for relative risk is simply the usual two parameter distribution with the second (beta)
In the original Chancellor et al. ( 1 997) article, the quoted relative risk was . pararheter set to 1 . Therefore in cells EI3 to E15, enter a random draw
0.509 with 95% confidence interval 0.365-0.710. Use this information to from a gamma distribution with the alpha parameter set to the value in
calculate the appropriate log mean and log standard error in cells F35 and column F and the beta parameter equal to 1 (do not be alarmed if Excel
G35. In cell C35, take a random draw from the normal distribution (with the returns an error - see below).
mean and standard error you just calculated) and exponentiate to give the ii. Now there is a problem in Excel: it solves the gamma distribution by
10gnormaUy distributed relative risk. iteration and when the value of alpha is very high (over about 300) it
2. Using a gamma distribution for cost will often fail to find a solution (what you see is a '!NUM' error
message) due to the iterative procedure Excel uses to solve the gamma
In rows 2 1 through 26 the direct medical and community care costs associated
function. However, for high values of the alpha parameter in the single
with each state of the model are listed. Unfortunately, little information was given
gamma function, the resulting distribution is not at all skewed and
in the original article about the variance of costs (despite the fact that mean costs
tends toward the normal distribution. Therein lays the solution to the
were derived from a patient level dataset). A simple assumption has therefore
problem. For high values of alpha, we substitute a random draw from
been made that the standard error of the costs is equal to the mean. Hence, the
the normal distribution with the equivalent mean and variance of the
entry in column E is simply the same as that in column D. Using the information
gamma distribution. Recall that the variance of the gamma distribu­
in columns D and E and the formula given in the section 'Distributions for costs',
tion (from the presentation) is alpha*betaA2. Therefore, as beta � I for
use the method of moments approach to solve for the parameters of the gamma
the single parameter gamma, the variance is simply alpha. Either do
distribution and enter these values in columns F and G. Now generate a random
this by hand in cells E13 and E14, or else set up an I F statement to
draw from the gamma distribution using these parameters and enter in column C.
choose the correct distribution automatically.
3. Using a beta distribution for the transition from AIDS to death iiI.. Once, you have valid numbers in cells E13 to E15, the corresponding
As mentioned in the presentation, a beta distribution can be used for a probability in column C is calculated as the corresponding draw in
dichotomous decision probability and is straightforward to fit using the column E divided by the sum of E13:EI5. This ensures that all proba­
counts already given in columns F and G of the number of events (deaths) bilities are in the range 0-1 and that the three probabilities sum to 1 .
observed and the complement (total - deaths). Therefore, in cell CI7 generate Enter the appropriate formula i n cells C 1 3 to CIS. Congratulations!
a random draw from the beta distribution using the values in cells F I 7 and You have just constructed your first Dirichlet distribution.
G17 as parameters. Note that the value in cell CI6 must, by definition, be I minus iv. Now repeat the three steps above, but this time for the four probabili­
the value of CI6 (in other words do not sample from two independent beta ties associated with State A in rows 9-12.
distributions as probabilities must sum to one). You now have a fully probabilistic model. If you switch cell C3 to equal 1
and then turn to the <Analysis> worksheet you should find that a single draw
4. Using a Dirichlet distribution for polytomous transitions from the model is made each time you press <F9>.
This chapter previously detailed how the Dirichlet distribution is the multino­
mial equivalent of the beta distribution that can be used when more than two
4.8. Exercise: making the total hip replacement
transitions are possible. Unfortunately, Excel does not include the Dirichlet
directly, therefore it must be constructed. This is relatively straightforward, model probabilistic
but does require an extra step. 4.8. 1 . Overview
i. Start with the three-way transition from State B to remaining in State B, The aim of this exercise is to demonstrate how the deterministic total hip
or moving to State C (AIDS), or moving to State D (death), as detailed in replacement (THR) model from Exercise 3.5 can be made probabilistic by
rows 13-15. The method is to normalize a draw from three independent fitting distributions to parameters. Emphasis is on choosing the correct distri­
single parameter gamma distributions, where the single parameter is bution for different types of parameters and on correlating parameters using
the number of events. Note the single parameter gamma distribution the Cholesky decomposition of a variance-covariance matrix.
112 I MAKING DECISION MODELS PROBABILISTIC EXERCISE: MAKING T H E TOTAL H I P REPLACEMENT MODEL PROBABILISTIC 1113

The step-by-step guide below will take you through a number of stages of iii. Now enter 0 in cell D3 and verify that the results on the <Analysis> work­
making the model probabilistic: . sheet and match the results from Exercise 3-5. When cell D3 is switched
I. Setting up the <Parameters> worksheet for probabilistic analysis. to the value I, column B should be populated with empty values.
2. Fitting beta distributions to (constant) probability parameters. Now that the restructuring of the <Parameters> worksheet is complete, the rest
of the exercise is concerned with fitting probabilistic values to the parameters in
3. Fitting gamma distributions to cost parameters.
column C such that when cell D3 is set to I, the model becomes fully probabilistic.
4. Fitting beta distributions to utility parameters.
2. Fitting beta distributions to (constant) probability parameters
5. Correlating parameters by Cholesky decomposition and application to the
regression model for prosthesis survival. The following information has been provided to you concerning operative
mortality rates following primary surgery:
A template for the exercise, 'Exercise 4.8 template.xls, is available, based on
the solution to Exercise 3.5. The hospital records of a sample of100 patients receiving a primary THR were exam­
ined retrospectively. Of these patients, two patients died either during or immediately
4.8.2. Step-by-step guide following the procedure. The operative mortality for the procedure is therefore esti­
mated to be2 per cent.
1 . Setting u p the <Parameters> worksheet for probabilistic analysis
On opening the file you will find that the parameters screen has changed and i. Use this information to specify the parameters (alpha and beta in
the point estimates of the parameters that you used in the previous exercise columns G and H) of a beta distribution for the probability of opera­
have been moved to column D (labelled 'deterministic'). This means that the tive mortality during primary surgery.
model no longer works, as all the 'live' parameter cells are currently blank. ii. Use the BETAINV( . . . ) function in cell CI6 to generate a random draw
Take a moment to familiarize yourself with the new layout of this worksheet. from this distribution.
In particular, note the addition of cell D3 that, in conjunction with the new iii. As no information is available on operative mortality following a revision
layout, will enable us to 'switch' between a deterministic and probabilistic procedure, use the same distribution again for this parameter (row 17).
version of the model. The following information has been provided to you concerning revision
i. Firstly, note that not all parameters have been labelled as potential procedures among patients already having had a revision operation.
candidates for the probabilistic analysis: patient characteristics (age The hospital records of a sample of100 patients having experienced a revision procedure
and sex), discount rates and prosthesis costs have all been coloured to replace a failed primary THR were reviewed at one year. During this time, four
black indicating they are to be left deterministic. Before starting on the patients had undergone a further revision procedure.
exercise, ensure that you are totally happy with why these parameters
iv. Use this information to fit a constant transition probability for the
will not form part of the probabilistic analysis.
annual re-revision rate parameter (row 18).
ii. The first step is to use an IF( . . . ) statement to determine what appears
You should now have three probabilistic parameters in C16:C18 that should
in column B dependent on the value in cell D3. The idea is to be able
automatically generate a new value from the specified distribution when you
to switch between deterministic values (column D) and probabilistic
hit the <F9> key.
values (column C) easily. Note that cell B25 should be excluded from
v. Use the formula for the moments of a beta distribution from the
this (in addition to the deterministic parameters referred to above).
section on 'Distributions for probability parameters' in order to specify
You will recall that the lambda parameter from the regression is
the mean and standard error (columns D and E) for these distributions.
endogenous to the model - therefore, it will by definition be determin­
istic or probabilistic depending on the nature of the other parameters vi. How do these compare with the mean values and standard errors that
that make it up. Note that just as B25 is exponentiated, so must B26 be would have been calculated from these data using standard methods?
(to get a lognormal distribution). To do this, you can exponentiate the Hint: in order to answer this question you need to recall that the binomial
whole of the IF( . . . ) statement. �
standard error is estimatedfrom theformula: se = p(l- p) I n .
114 I MAKING DECISION MODELS PROBABILISTIC
T E C H N ICAL APPENDIX I 115

3. Fitting gamma distributions to cost parameters

5. Application of Cholesky to the regression model
for prosthesis survival
As the costs of prostheses are handled deterministically, the cost of the
We now implement the Cholesky decomposition matrix method in the
primary procedure is ignored (as it is assumed to be the same in each arm).
model. Look at the <Hazard junction> worksheet, which gives the results of a
In addition, there is assumed to be no cost of monitoring successful proce­
regression analysis for the hazard ratio. You have used these results before
dures, so there is only really one probabilistic cost to specify for the model ­
when constructing the deterministic model, but now we have added the
the cost of the revision procedure.
covariance matrix for the parameters (together with the correlation matrix,
You are given the following information regarding the cost of revising a
which is a little easier to interpret). It should be clear that, while some param­
failed primary THR:
eters do not have a strong relationship) there are strong relationships within
A number of units involved in THR were reviewed and the mean cost of a revision this set of parameters - particularly between the regression constant and the
procedure was found to be £52 94 with a standard error of £1 487 .
other parameters.
i. Add the value for the standard error into the spreadsheet (cell E33). i. Note that the covariance matrix you have been given is a 5 x 5 matrix.
ii. Use the method of moments approach described in the presentation to You have been given a lower triangular Cholesky decomposition of
calculate the parameters (cells G33 and H33) of a gamma distribution the covariance matrix in cells C26:G 3 1 . Take a few moments to
that has the corresponding mean and standard error. understand how this was obtained (you may want to refer back to the
iii. Use the GAMMAINV( ... ) function in cell C33 to generate a random section on correlating parameters).
draw from this distribution. ii. In cells C36:C40, use the NORMINV( . . . ) function to generate five
An alternative approach would have been to fit a normal distribution for random draws from the standard normal distribution.
this parameter - do you think this would be valid in this case? iii. In cells E36:E40, enter the solution of the Tz matrix calculation, that
is the Cholesky decomposition matrix multiplied by the vector of
4. Fitting beta distributions to utility parameters
standard normal variates.
As was indicated in Section 4.3.2, there are two methods to fitting the beta
iv. Now add the estimated mean values from the regression to Tz and
distribution. Here the second approach is taken using method of moments to
enter into cells G36:G40.
fit beta distributions to the utility data.
You are given the following information on utilities for patients experienc­ v. You have now created a vector of five multivariate normal parameters
ing different states of the model: that are correlated according to the estimated covariance matrix. The
The following information has been provided to you concerning operative final task is to go to the <Parameters> worksheet and link the proba­
mortality rates following primary surgery: bilistic cells C22:C24 and C26:C7 to the relevant cell of the vector you
have just created.
A study was instigated to explore the utility weights subjects placed on different
outcomes of THR related to the states of the Markov model - the following results
vi. Although not essential to the working of the parameters worksheet,
were calculated in terms of mean (standard error) by state: for completeness you should add in the standard errors and distribu­
Successful primary 0.85 (0.03)
- tion for the parameters you have just entered.
Successful revision 0.75 (0,04)
-
vii. Remember that we have been working on the log hazard scale for the
Revision 0.3 (0,03)
Weibull regression. We need to make sure that the model parameters
-

i. Enter the standard errors into the worksheet (E41:E43). are exponentiated. This should already have been done for the
ii. Obtain the beta parameters corresponding to these mean values and endogenous lambda parameter in cell B25 (check that this is the
standard errors by method of moments. case). The gamma parameter and the parameter relating to the rela­
iii. Use the BETAINV( . . . ) function in cells C41:C43 to generate random tive risk of the new prosthesis should also be exponentiated. Note this
draws from these distributions can either be done by exponentiating the IF( . . . ) function in column B
116 I MAKING DECISION MODELS PROBABILISTIC
TECHN ICAL APPENDIX 1 117
or by separately exponentiating the deterministic and probabilistic of the beta distribution representing prior beliefs and to use these quantities as
values in columns C and D. the updated parameters of the (posterior) beta distribution. Where the prior
viii. Congratulations! The model should now be probabilistic. Verify this distribution is formulated to be minimally informative, as with the beta(1,l)
by switching cell D3 to the value 1 and examining the <Analysis> (which gives a uniform distribution over the 0-1 interval), the data will domi­
worksheet. nate that prior. Note that in the text, the technicalities of the Bayes approach
were omitted such that the beta distribution was fitted directly to the data.
4.9. Technical appendix
4 . 9. 1 . The polytomous case
Probabilities for dichotomous branches in decision trees are often estimated
Where data occur naturally across k > 2 categories, the data are said to follow
from data where r events of interest are observed from n observations. Such data
the multinomial distribution, given by:
are characterized by the two parameter binomial distribution, Bin(n,n), where
n is the probability of an event and n is the number of observations. If the r --- multin (n;n!,n2) ··· , Jrk)
number of events r follows the binomial distribution then:
p(r ) J
n
l r1 r2 · .. rk
) n;' n;, . . . n�

for

In a probabilistic model, the challenge is to specify a (continuous) distribu­

tion for the probability n of the event. In a Bayesian framework, this is
most naturally achieved by specifying a prior distribution for n from the beta It is now natural to formulate the decision problem with multiple branches
distribution beta( a,jJ) , as this continuous distribution is conjugate to the and consider the distrihution of the probabilities nl, n 2' . . . , nk for each of the
binomial distribution and is bounded on the 0-1 interval (as probabilities individual branches. This is achieved using the Dirichlet distribution, which is
must be). The prior belief of a probability of an event can be specified by the the multivariate generalization of the beta distribution and is given by:
beta distribution:
n - Dirichlet ( al 'az "" ,ak )
p(n ) r(al + · .. +ak) na,-I a,-i
p (n ) - n
r (al ) r (ak ) l . . . k
_

• • •

where r(u) represents the single parameter gamma function.

where
Bayes' theorem states that the posterior probability of a parameter is
proportional to the prior probability of the parameter times the data likeli­
hood. Applying Bayes' theorem to the example of binomial data with a beta
prior gives:
The application of Bayes' theorem gives the following derivation for the
-
p(n ) oc n' (I-nr' na-I ( I- n )� 1 posterior distribution of the probabilities:
f3+n-r-1
oc no:+r-1 ( 1 - n )

which identifies the posterior distribution as beta( a + r, jJ + n - r) . In other

words, in order to generate a continuous distribution for the probability of an which identifies the posterior as Dirichlet (al + rp ... , a+ rk)' Specifying a
event that is bounded on the 0-1 interval, it is necessary only to add the minimally informative prior, such as Dirichlet (1, . . . , 1 ) means that the data
observed number of events and non-events, respectively, to the two parameters dominate the prior.
118 I MAKING DECISION MODELS PROBABILISTIC REFERENCES I 119

Note that for k = 2 categories, the Dirichlet is exactly equivalent to the beta References
distribution. Altman, D. C. ( 1991) Practical statistics for medical research, London, Chapman and HalL
Armitage, P. arid Berry, G, ( 1994) Statistical methods in medical research, 3rd edn. Oxford,
4.9.2. Sampling from the Dirichlet distribution
Blackwell Scientific Publications.
In this chapter, we have emphasized the convenience of the Dirichlet Arrow, K. J. and Lind, R. C. (1970) 'Uncertainty and the evaluation of public investment
distribution in modelling uncertainty when parameters are estimated from decisions', American Economic Review, 60: 364-378.
multinomial data. However, sampling from the Dirichlet distribution is Briggs, A, H., Ades,A. E. and Price, M. J. (2003) 'Probabilistic sensitivity analysis for decision
less straightforward as it is not commonly represented in standard software trees with multiple branches: use of the Dirichlet distribution in a Bayesian framework:
Medical Decision Making, 23: 341-350.
packages. One exception to this is the WinBUGS software (Spiegelhalter
Briggs, A., Sculpher, M" Dawson, J" Fitzpatrick, R., Murray, D. and Malchau, H. (2004) 'The
et al. 1998), designed for Bayesian analysis and freely available from
use of probabilistic decision models in technology assessment: the case of total hip
http://www.mrc-bsu.cam.ac.uk/bugs/welcome.shtmI. This software was origi­ replacement', Applied Health Economics and Health Policy, 3: 78-89.
nally designed to allow Bayesian analysis for nonconjugate prior distributions; Briggs, A. H., Mihaylova, B., Sculpher, M., Hall, A., Wolstenholme, J., Simoons, M" Ferrari,
however, as the Dirichlet and multinomial distributions are directly available R" Remme, W. J., Bertrand, M. and Fox, K. (2006) 'The cost�effectiveness of perindopril
within the package, sampling from the appropriate Dirichlet distribution is in reducing cardiovascular events in patients with stable coronary artery disease using
data from the EUROPA Study� Heart92(SuppI 2), A36.
especially easy in WinBUGS.
Chancellor, J. V" Hill, A. M" Sabin, C. A., Simpson, K. N. and Youle, M, (1997) 'Modelling
For other software packages that do not include a direct representation of
the cost effectiveness of lamivudinelzidovudine combination therapy in HIV infection',
the Dirichlet distribution there are two approaches to generating a Dirichlet PharmacoEconomics, 12: 54-66.
distribution (Gelman et aI. 1995). Claxton, K ( 1999) 'The irrelevance of inference: a decision-making approach to the
stochastic evaluation of health care technologies', Journal ofHealth Economics, 18:
1 . Normalized sum of independent gamma variables
341-364.
Draw variables Xl' X2' . . . Xk from independent gamma distributions with Draper, D. (1995) 'Assessment and propagation of model uncertainty', Journal of the Royal
shape parameters aI' ex 2' . . . a k and a common scale parameter: Statistical Society, Series B, 57: 45-97.
Duan, N, ( 1983) 'Smearing estimate: a nonparametric retransformation method', Journal
Xj - Gamma (aj,/3). of the American Statistical Association, 78: 605-610.
Gelman, A., Carlin, J. B., Stern, H, S. and Rubin, D. B. (1995) Bayesian data analysis.
Then the required probabilities are simply: London, Chapman and Hall.
Manning, W. G. (1998) 'The logged dependent variable, heteroscedasticity, and the
retransformation problem', Journal ofHealth Economics 17: 283-295.

Ikj=l a·J
Na.tional Institute for Clinical Excellence (NICE) (2004) Guide to the methods of technology
assessment, London, NICE,

Equivalently, the xi can be drawn from a chi-squared distribution with 2ai

O'Hagan, A. and Luce, B. (2003) A primer on Bayesian statistics in health economics and
outcomes research. Bethesda, MD, MedTap International.
degrees of freedom. Palmer, S. and Smith, P. C. (2000) 'Incorporating option values into the economic evaluation
of health care technologies', Journal ofHealth Economics, 19: 755-766,
2. Series of conditional beta distributions
Pratt, 1. W., Raiffa, H. and Schlaifer, R. (1995) Introduction to statistical decision theory,
First ttl is drawn from a beta (a" I)" a/ Next, for each ttj in turn, Cambridge, MA. MIT Press.
j = 2 . . . k - l, draw 'Pi from a beta (aj,ILj+! a,), and then set 1/:j = Rice, J. A. (1995) Mathematical statistics and data analysis, 2nd edn, Belmont. Duxbury Press.

(! - Ii�; 1/:;) 'Pj . Finally, set 1/:k = 1 - Ii;;l 1/:;. Spiegelhalter, D. J., Thomas, A. and Best, N, (1998) WinBUGS [1.3}. MRC Biostatistics Unit,
University of Cambridge.
It is this method that corresponds to the decomposition of a multi-branch
Stinnett, A. A. and Paltiel, A. D. (1997) 'Estimating CE ratios under second�order uncertainty:
node into a series of conditional dichotomous nodes. However, it should be the mean ratio versus the ratio of means', Medical Decision Making, 17: 483-489.
clear from the derivations above that if the data are multinomial, this approach Thompson, K. M, (2002) 'Variability and uncertainty meet risk management and risk
is far less convenient computationally. communication', Risk Analysis, 22: 647-654.
120 I MAKING DECISION MODELS PROBABILISTIC

Thompson, K. M. and Graham, J. D. (1996) 'Going beyond the single number: using Chapter 5
probabilistic assessment to improve risk management', Human and Ecological Risk
Assessment, 2: 1008-1034.
UK Prospective Diabetes Study (UKPDS) Group', ( 1998) 'Intensive blood-glucose control Analysing and presenting
with sulphonylureas or insulin compared with conventional treatment and risk of
complications in patients with type 2 diabetes (UKPDS 33). Lancet, 352: 837-853. simulation output from
probabilistic models

The aim of this chapter is to explore exactly how probabilistic models, with all
parameters being represented by their respective distributions, should be
handled in terms o f producing, analysing and then presenting the results of
the probabilistic analysis. The chapter starts by looking at how to analyse and
present large numbers o f Monte Carlo simulation results o f the output
parameters of interest for the straightforward comparison of two alternative
treatments. We then turn to the issue of calculating the sensitivity of the overall
results to individual parameter values using standard an alysis o f covariance
(ANCOVA) methods. These methods are illustrated using the AIDS/HIV
model that was developed in the previous chapters based on the work
by Chancellor et al. ( 1997). The cost-effectiveness acceptability curve (CEAC)
is introduced and the advantages over the standard presentation o f interval
estimates of uncertainty are outlined.
The next section deals with the use of multiple cost-effectiveness acceptabil­
ity curves to represent uncertainty and distinguishes the use o f multiple
curves in two very different situations. In the first, multiple acceptability
curves are used to reflect the same treatment choice between alternative
treatments in the presence of heterogeneity. In the second, multiple curves are
employed to represent mutually exclusive treatment options. The final
two sections present worked case studies o f probabilistic modelling which
highlight a number of features of probabilistic modelling, but serve to illus­
trate the two multiple curve situations in particular. The first example relates
to estimating heterogeneity in cost-effectiveness analysis of the use of an ACE
inhibitor in stable coronary heart disease for patients at different levels o f
cardiovascular risk (Briggs e t al. 2006); and the second relates to the use o f
different management strategies for erosive oesophagitis (Goeree e t al. 1999;
Briggs et al. 2002a).
1 22 ! ANALYSING A N D PRESENTING S I M U LATION OUTPUT FROM PROBABILISTIC MODELS ANALYSING S I M U LATION RESULTS: TWO TREATMENT ALTERNATIVES ! 123

5 . 1 . Analysing simulation results: two treatment If we consider the ideal circumstance of knowing the position of an interven­
alternatives tion o� the ,cost-effectiveness plane with no uncertainty, then a number of
eventualities can arise. A new treatment is said to <dominate' control, being less
In this section we review the standard approaches to decision making in cost­ costly and more effective if it is located in the SE quadrant and vice versa ­
effectiveness using the graphical cost-effectiveness plane, going on to consider control dominates treatment if it is located in the NW quadrant. In these two
how we present information under conditions of uncertainty, including an illus­ circumstances it is clearly appropriate to implement the least costly and most
tration of how to write macros in Excel to record simulation results. The use of effective (or dominant) treatment and no recourse to cost-effectiveness ratios is
interval estimates for cost-effectiveness is illustrated and we argue for the use of required. However) far more common is for one treatment to be more effective
CEACs as a more general solutian to the problem of presenting uncertainty in than the other, but also more costly. In such circumstances, a decision must be
cast-effectiveness analysis. We complete the section by introducing the net­ made as to whether the additional health benefits of the more effective treat­
benefit statistic and showing how it can be used to generate acceptability curves. ment are worth the additional cost. If the ICER of the new therapy (Ll.C/Ll.E) ­
the slope of a straight line from the origin that passes through the (Ll.E, Ll.C)
5.1 . 1 . The cost-effectiveness plane
coordinate - is less than the acceptable 'threshold ratio' of the decision maker
The cost-effectiveness plane (Black 1990) is presented in Fig. 5.1 and shows the (representing the willingness-to-pay for a unit of health gain) then the treat··
difference (treatment minus control) in effectiveness (Ll.E) per patient against ment should be adopted.
the difference in cost (Ll.C) per patient. Plotting the effectiveness difference on Of course, the above discussion assumes that we know with certainty the
the horizontal axis has the advantage that the slope of the line joining any cost, effect and cost-effectiveness of an intervention, such that the appropriate
point on the plane to the origin is equal to the incremental cost-effectiveness decision is clear. In practice, the outputs of our probabilistic models give us
ratio (lCER) ; Ll.CILl.E, the statistic of interest in cost-effectiveness studies. The the distribution over incremental cost, incremental effect and the joint cost­
plane can be defined as four separate quadrants that are labelled using the effect distribution. For example, Fig. 5.2 shows 1000 Monte Carlo simulations
points of the compass to minimize confusion that can arise from alternative
numbering schemes.
£12,000
NW Quad rant ,NE Quadrant
•

(New treatment dominated)

£10,000
. ' • (T;a,de-off)
��->..
/
ICER",,£6,200
NW Quadrant NE Quadrant
(New treatment dominated) (Trade-off) £8,000

/
Threshold ratio: A
0
<l
1i
£6,000

c
� £4,000
$
;;;
"0
175 £2,000
0
0
. '"
-0.4 0.1 0.6 1.1 1.6
-£2,000
SW Quadrant SE Quadrant
(Trade-oft) (New treatment dominated)
-£4,000
Effect difference, �E
SW Quadrant SE Quadrant Fig. 5.2 Estimated jOint cost-effectiveness density for the HIV/AIDS
model plotted on
(Trade-off) (New treatment dominates) the cost-effectiveness plane. Ellipses assume multivariate normality for the joint
Effect difference, dE density with the same covariance structure as the simulations and cover 5%, 50%
Fig. 5.1 Cost-effectiveness plane for new treatment compared with existing treatment and 95% of the estimated joint density on the cost-effectiveness plane.
1 24 I ANALYS I N G AND PRESENTING SIMULATION OUTPUT FROM PROBABILISTIC MODELS
ANALYS I N G S I M U LATION RESULTS: TWO TREATMENT ALTERNATIVES 1 125

of the output of the AIDS/HIV model together with the elliptical joint density
estimates assuming multivariate normality with the same variance-covariance
as the simulations.
The presentation of the simulated joint density in Fig. 5.2 requires that the counter variable. The basic loop starts with a 'Do' command and ends with
probabilistic output of our probabilistic models is recorded. As this is a a 'Loop' command. We then have a choice about how to incorporate the
straight forward repetitive process, it can be easily achieved in Excel using qualifying statement 'while' followed by a logical expression. Suppose that
simple programs known as 'macros'. Box 5 . 1 outlines a simple approach to we want to undertake 1000 simulations. We can either specify 'Do While
macro writing in Excel using the 'record' feature of the softw'are. Index < 1000' in the first line of the loop, or we can specify 'Loop While
Index <= 1 000' in the last line of the loop. Remember to augment the
counter variable in the penultimate line of the loop with a line of code:
, .
Index = Index + 1.
Box 5.1 : Example code for visual basic macro .
Suppose the purpose of the macro is to copy a probabilistic variable and
to record simulation results .
\ record it to a results worksheet. If we record the copy and paste routine, the
The idea of writing macros can sound rather intimidating, particularly as Visual Basic editor will default to using absolute cell referencing which speci­
macros are effectively written in the Visual Basic programming language. fies an alphanumeric code relating to the column (letter) and row (number)
The macros themselves are stored in the 'Visual Basic Editor' which is sepa­ of the cell. If this is embedded within a loop then results will continually be
rate from the main workbook (although all programming is saved as part pasted into the same cell. What we wish to do, however, is to record the results
of the workbook). Fortunately, to program macros, you need to know very into different cells for each cycle of the loop. We therefore augment the code
little visual basic, particularly if you follow two simple principles. Firstly, with a relative cell reference that specifies an offset ofY rows and X columns
keep calculations in the spreadsheet rather than in the macros - this keeps from a starting absolute cell reference. The code for this is shown below using
your working transparent to others. Secondly, use the record feature - by the counter variable, Index, to offset an extra row for each cycle of the loop.
recording macros you can generate the code you need to perform different Range ( ' G6 ' ) . Select
tasks without having to know how to write the code. ActiveCell . Offset ( Index , O ) . Range ( ' Al ' ) . Select
Macro commands are accessed from the Tools > Macro menu. Selecting The first line sets the absolute reference and the second line specifies a
'Record new macro' from this menu starts the record feature. A pop-up box relative reference using the absolute reference as a placeholder.
asks you to complete a name for the macro (no spaces allowed) after which The code for this simple illustrative example is shown below as it would
a 'stop recording' button appears on your worksheet. From this point on, appear in the visual basic editor in Excel. Lines in bold indicate lines that
all actions that you perform will be converted to the equivalent visual basic were added (see above). The remaining lines can be recorded by carrying '
code until you press the stop button. This is useful when you know what out the actions in Excel with the record feature on.
you want the macro to do, but not the code for how to do it, for example to
Sub Simulation ( )
record the costs and outcomes predicted by a probabilistic model each time
Dim Index as Integer
the parameter values are simulated. These simulation results can then be
stored for further analysis. In order to undertake the repetitive task of Index " 0

recording these results, we set up a simple loop. Do While Index < 1000
To complement the record feature we need to know just two basic sets SheetS ( ' Results ' ) . Select
of commands. These are the 'Looping' commands and the ability to use Range ( ' D 1 0 : E 1 0 ' ) . Select
relative cell referencing rather than absolute cell referencing.
Selection. Copy
In setting up a loop, we start by defining an index variable (Dim Index)
Range ( ' D1 8 : E 1 8 ' ) . Select
J
and setting this index variable to zero (Index = 0) - this is our
126 1 ANALYS ING AND PRESENTING S I M U LATION OUTPUT FROM PROBABILISTIC MODELS ANALYSING S I M U LATION RESULTS: TWO TREATMENT ALTERNATIVES 1 1 27

Box 5.1: Example code for visual basic macro to record simulation ICER=£10,700 NE Quadrant
results (continued)
(Trade-off)

£1 0,000
. '
.-/'
ActiveCell . Offset ( Index, O ) . Range ( ' Al ' ) . Select ICEA=£6,200

Selection . PasteSpecial Paste : = xlValue s , £8,000

Operation: = xlNon e , SkipBlanks : = False ,
£6,000
Transpose : = False
Index = Index + 1 £4,000 ICER=£2,500

Loop
End Sub i £2,000
.
--------. --- - � '.-

-0.4 0.1 0.6 1.1 1 .6

-£2,000
SW Quadrant SE Quadrant
5.1 .2. Interval estimates for cost·effectiveness ratios (Trade-off) (New treatment dominates)
-£4,000
With patient-level information on the costs and effects of treatment interven­ Effect difference, L\.E
tions it is natural to consider representing uncertainty in the leER using Fig. 5,3 Monte Carlo simuiation results on the cost�effectiveness plane for the
interval estimates. Note that) as decision models are inherently Bayesian, it H!V/AIDS model showing i nterval estimates for cost, effect and incremental
would be better to describe such intervals as 'credible intervals' rather then use cost�effectiveness ratio.
the frequentist 'confidence interval'. We choose to talk about general 'interval'
estimation, reflecting that the output parameters under discussion are from a
decision model rather than statistical model. qualitatively different from negative leERs in the SE quadrant ( favouring
Interval estimates can be obtained for outcomes o f interest using the simu­ the new treatment) yet will be grouped together in any naYve rank-ordering
lation results by using the percentile method. This simply involves taking the exercise. Similarly, positive ratios o f the same magnitude in the SW and NE
al2 and ( 1 - (12) percentiles o f the simulation vector as the ( 1 - a) lOO% quadrants have precisely the opposite interpretation from the point o f view of
uncertainty interval for outcome of interest. Percentile intervals for the the intervention under evaluation. This is because the decision rule in the SW
HIV/AIDS example are shown in Fig. 5.3, with the horizontal I-bar representing quadrant is the opposite o f that in the NE. For example, an leER of 500 may
the 95 per cent uncertainty interval on life-years gained, the vertical I-bar be considered as supporting a new treatment in the NE quadrant if society has
representing the 95 per cent uncertainty interval on incremental cost, and the set a threshold ratio of 1000. In the SW quadrant, however, this value of the
'wedge' representing the 95 per cent uncertainty interval on the leER. leER would be considered as support o f the existing treatment rather than the
new treatment. Again, any naYve ranking exercise could easily conflate leERs
5 . 1 .3. Beyond interval estimation: acceptability curves
with the same magnitude but with different implications for decision making.
Although it was possible to calculate the 95 per cent uncertainty interval for A solution to this problem can be found by returning to the original decision
the leER in the case of the HIV/AIDS model as illustrated in Fig. 5.3, the over­ rule introduced in Fig. 5.1. I f the estimated leER lies below the threshold
a!l level o f uncertainty will often be such that problems are caused for interval ratio It, reflecting the maximum that decision makers are willing to invest to
estimation for the leER. This occurs when simulation results cross the vertical achieve a unit of effectiveness, then it should be implemented. Therefore,
axis suggesting that there is non-negligible probability that the new treatment in terms o f the Monte Carlo simulations on the cost-effectiveness plane in
is less effective than the existing treatment. The problem is that ratios o f the Fig. 5.3, we could summarize uncertainty by considering how many of the
same sign) but from different quadrants, are not strictly comparable. Negative simulations fall below and to the right of a line with slope equal to A, lending
leERs in the NW quadrant of the plane ( favouring the existing treatment) are support to the cost-effectiveness o f the intervention. Of course, the appropriate
1 28 I ANALYSING AND PRESENTING S I M U LATION OUTPUT FROM PROBABILISTIC MODELS ANALYSING S I M U LATION RESULTS: TWO TREATMENT ALTERNATIVES 1 1 29

1 .00 associated with ICERs (Claxton and Posnett 1996; Stinnett and Mullahy 1998;
0.90 Tambour et al. 1998; Claxton 1999). In particular, Stinnett and Mullahy (1998)

1 �::�
<l> 0.80 offer a comprehensive account of the net-benefit framework and make a
convincing case for employing the net-benefit statistic to handle uncertainty
in stochastic cost-effectiveness analysis. The algebraic formulation of the deci­
sion rule for cost-effectiveness analysis that a new treatment should be imple­
� mented only if its ICER lies below the threshold ratio, 1lc/IlE < A" can be
i 0.30
0040
rearranged in two equivalent ways to give two alternative inequalities on either
Q. 0.20 the cost scale (net monetary benefit, NMB) (Claxton and Posnett 1996;
0.10 £2,523 Tambour et al. 1998; Claxton 1999) or on the effect scale (net health benefit,
£1 0,740
0.00 NHB) (Stinnett and Mullahy 1998):
£. £5,000 £10,000 £15,000 £20,000
Value of threshold ratio (A)
NMB : A, · IlE - Il C > O
Fig. 5.4 Cost-effectiveness acceptability curve for the HIV/AIDS model showing the IlC
NHB : IlE- > O'
recovery of uncertainty interval estimates. T
These decision rules are entirely equivalent to the standard rule in terms of
the ICER but have the advantage that when applying the net-benefit statistics
value of A. is itself unknown. However} it can be varied in order to show how to Monte Carlo simulation data, they will unambiguously sort out the 'accept­
the evidence in favour of cost-effectiveness of the intervention varies with ,t ability' of an individual simulation trial on the cost-effectiveness plane. By
The resulting curve for the HIV / AIDS example is shown in Fig, 5.4 and using a net-benefit formulation, we can avoid the problem apparent with
has been termed a cost-effectiveness acceptability curve (van Hout et al. 1994) ICERs of conflating simulations of the same sign but in opposite quadrants of
as it directly summarizes the evidence in support of the intervention being the cost-effectiveness plane. As such, it turns out to be much simpler to calculate
cost-effective (i.e. acceptable) for all potential values of the decision rule. acceptability curves from Monte Carlo simulations using net benefits than
This 'acceptability curve' presents much more information on uncertainty using the joint distribution of costs and effects.
than do confidence intervals. The curve cuts the horizontal axis at the probabil­ The net-benefit framework also overcomes a particular problem associated
ity that the intervention under evaluation is cost-saving, as a value of zero for A, with mean cost-effectiveness ratios. Introductory textbooks emphasize the
implies that only the cost is important in the cost-effectiveness calculation. importance of taking an incremental approach (Weinstein and Fineberg 1980;
The curve is tending towards the probability that treatment is effective, as an Drummond et al. 2005) rather than comparing mean cost-effectiveness ratios.
infinite value for A, implies that effect only is important in the cost-effectiveness A recent contribution to the literature highlighted the fundamental problem
calculation, As well as summarizing, for every value of A., the evidence in favour of taking patient-level mean ratios: the mean of ratios is not equal to the ratio
of the intervention being cost-effective, acceptability curves can also be employed of the mean values (Stinnett and Paltiel I997). The consequence is that:

CI - -
Co * CI -Co
to obtain the uncertainty interval on cost-effectiveness. The limits are obtained
-
EI Eo EI - Eo
by looking across from the vertical axis to the curve at the appropriate points --­

for the desired confidence level and reading off the associated cost-effectiveness
value from the horizontal axis. For the HlV/AIDS example, we can recover the demonstrating that the incremental ratio cannot be constructed from the
previously calculated interval of £2500/LYG to £10 700/LYG. difference between the mean cost-effectiveness ratios in each afm of a trial or
model.
5 . 1 .4. The net-benefit framework By contrast, the difference in the average net-benefit of the experimental
Relatively recently, a number of researchers have employed a simple rearrange­ treatment and the average net-benefit of standard care treatment will give
ment of the cost-effectiveness decision rule in order to overcome the problems the overall incremental net-benefit statistic introduced above_ This is
130 I ANALYSING A N D PRESENTING SIMULATION OUTPUT FROM PROBABILISTIC MODELS
T
1 U S I N G ANCOVA TO ESTI MATE T H E I M PORTANCE OF I N DIVIDUAL PARAMETERS I 1 31

1
t

t
straightforward to see algebraically through simple manipulation of the net­ Incremental costs Parameters Incremental life years gained
benefit expressions

1I
NMB, -NMBo = (A E, -CJ-(A.Eo -Co )
= A(E, - Eo )-(C, -Co)
= A !1E-!1C
= ANMB,
for the experimental (subscript 1 ) and control (subscript 0) interventions.
Therefore the usefulness of the average net-benefit is not directly in terms of 1
the average figures themselves, but in the simple linear relationship between I
I
average and incremental net-benefit. The full power of this simple relation­
ship will become apparent subsequently when we turn to analysing multiple
mutually exclusive treatment options. ,

\
100% 80%> 60% 40% 20% 0% 0% 20% 40% 60% 80% 1 00%
Proportion of sum of squares Proportion of sum of squares
5 . 2 . Using ANCOVA to estimate the importance Fig. 5.5 ANCOVA analysis of proportion of sum of squares for incremental cost
of individual parameters I (left-hand side) and incremental life years gained (right-hand side) explained by

I!
uncertainty in the model input parameters.
The presentation of simulation results on the cost-effectiveness plane and the
use of CEACs gives a useful impression of the overall uncertainty in a model.
However, it may still be important to understand the relative effect of the
different parameters in terms of contribution to this overall uncertainty. in the figure. The figure clearly shows that the medical costs of states A and C,
In order to explore this issue) we need to record not only the 'output' param­ and the community care cost of state A in the model are most important for
eters (the costs and effects of the different interventions), but also the input explaining the uncertainty of incremental cost. For incremental life years, only

I
parameters that go into providing each simulation of the outputs. When we the treatment effect (expressed in the model as a relative risk) is important.
have both the inputs and outputs of the model recorded, there are three Note that all of the transition probabilities from a given state of the model are
grouped together, reflecting the fact that these are all estimated from a single

I
approaches that we might use to explore individual parameter uncertainty. We
might summarize the correlations between the inputs and outputs. However) Dirichlet distribution. For example, the 'transitions from P( include all four
this is not a recommended approach as correlations can be high even if transition parameters from A to each of the other states in the model (including

\
the overall variance induced is low. A better approach is to use analysis of remaining in A) as these are all obtained from the same Dirichlet distribution.
covariance methods (ANCOVA) which can summarize the proportion of the ANCOVA analysis easily allows grouping of related variables in this way. That
variance in the output parameter 'explained' by variation in the input parame­ the transition probabilities do not seem to impact the uncertainty in either
ter. However, in its simple form) this assumes a linear relationship between incremental costs or effects reflects the relative precision with which these
inputs and outputs, such that for nonlinear models this approach is only an parameters of the model are estimated.
approximation. Although the ANCOVA approach is only an approximation to the individual
Figure 5.5 shows an ANCOVA analysis for the HlV/AIDS model applied parameter contribution for nonlinear models) its ease of implementation has
to both incremental cost only (left-hand side of figure) and incremental much to recommend it. Having recorded the input parameters and the corre­
life-years gained only (right-hand side). Although the ANCOVA could also sponding output parameters of the model, it is a very simple step to run an
be applied to net-benefit conditional on a specific value of the threshold ratio, ANCOVA for a given output parameter using the input parameters as explana­
the results would reflect a middle ground between the two extremes shown tory variables. This can be done in any standard software package including
1 32 I ANALYSING AND PRESENTING S I M U LATION OUTPUT FROM PROBABILISTIC MODELS REPRESENTING U N C E RTAINTY WITH M U LTIPLE CEACS I 133

many spreadsheet packages. Furthermore, the R2 statistic provides a summary the lack of memory in a Markov model, which fundamentally assumes that all
of the extent of approximation in nonlinear models (as for a linear model, the patients in ·a'given state are homogeneous, and due to the fact that most
uncertainty in the inputs should perfectly explain the uncertainty in the modelling exercises are based on secondary analyses with parameter estimates
outputs resulting in an R' of 100%). based on aggregated statistics across patient samples. In Chapter 4, much
It is important to recognize that an ANCOVA analysis only summarizes the attention was given to the potential use of regression analysis to understand
individual parameter contribution to the variance of the output of interest heterogeneity in parameter estimates. However, such analyses are dependent
(incremental costs, incremental effects or net-benefit) when our real concern on the analyst having access to patient-level data.
is with decision uncertainty. In the next chapter, a more sophisticated 'Where patient characteristics influence the parameters of a model, then it
approach based upon value-of-information methods will be introduced. This is clear that the resulting cost-effectiveness can differ. Where the aim of
allows the individual parameter contribution to decision uncertainty to be cost-effectiveness analysis is to estimate the lifetime cost per quality-adjusted
assessed. Nevertheless, the straightforward nature of the ANCOVA approach life-year (QALY) of an intervention, it should be clear that life expectancy will
is likely still to be useful for a swift understanding of the main parameter influence potential QALY gains and that this in turn is influenced by (among
contributions to uncertainty in the model. other things) the age and sex of the subject. Furthermore, health-related quality
of life (HRQoL) is also dependent on age and possibly on sex, as is evident
5 . 3 . Representing uncertainty with multiple CEACs from the published HRQoL norms for the UK (Kind et al. 1999). Therefore,
at the most fundamental level, we might expect heterogeneity in all cost-per­
In the AIDS/HIV example presented above, the presentation of the CEAC is
QALY figures even before considering heterogeneity in the parameters of the
perhaps the most common form found in the literature, with a single curve
disease process, treatment effect and costs.
representing the incremental analysis of a new treatment alternative for an
As it is possible to implement different treatment decisions for patients with
assumed homogeneous group of patients. However, it is ofteu the case that
different characteristics, all cost-effectiveness models should at least consider
patient characteristics can affect the potential outcomes of the model, such
the potential for their results to vary across different subgroups and, in principle,
that treatment choices may be different for patients with different characteris­
each subgroup of patients should be represented by a different CEAC in order
tics. Furthermore, it is rarely true that there is only one treatment alternative
to facilitate different policy decisions. This general approach will be illustrated
that is relevant for a single patient group. In this section, we provide an
later in this chapter by showing how a series of statistical equations that model
overview of the two distinct situations that may lead to the presentation of
heterogeneity can be combined to estimate cost-effectiveness that varies by
multiple acceptability curves in the same figure, representing rather different
patient characteristics for a stable coronary disease population treated with an
situations. These two situations reflect a common distinction in economic
ACE inhibitor. Although the example relates to cardiovascular disease, the
evaluation that governs how the incremental analysis is performed (Karlsson
implications of heterogeneity are much more general and are likely to impact
and Johannesson 1996; Briggs 2000). Where the same intervention can be
almost all potential evaluations. Indeed, true homogeneity of patient popula­
provided to different patients, the decision to implement that intervention
tions is rare and consideration should always be given as to whether difterent
can be made independently based on the characteristics of the patient.
characteristics could result in different treatment decisions for different cate­
This contrasts with the situation where different (though possibly related)
gories of patient.
interventions are possible treatment options for the same group of patients,
such that a choice of one intervention excludes the others. 5 . 3 . 2 . Multiple curves for mUltiple treatment options

As argued above, the standard presentation of the CEAC reflects the standard
5.3. 1 . Multiple curves for patient subgroups concern of cost�effectiveness analysis with the incremental comparison of an
(modelling heterogeneity) experimental treatment against a comparator treatment. Similarly, in clinical
In the past, many economic evaluations, including many cost-effectiveness evaluation, randomized control trials commonly have just two arms. However,
modelling exercises, have assumed that patients eligible for treatment are it is rarely the case that decision makers face such a restricted set of options.
essentially homogeneous. This approach has, most likely, been encouraged by Furthermore, in decision modelling (in direct contrast to clinical trial research)
134 I ANALYSING AND PRESENTING S I M ULATION OUTPUT FROM PROBABILISTIC MODELS A MODEL OF THE COST-EFFECTIVENESS OF ACE-INHIBITION 135

the cost of including additional options in an evaluation is small. As was The model was designed to assess the cost-effectiveness of perindopril
argued in Chapter I, economic evaluation should include all relevant treatment 8 mg once daily from the perspective of the UK National Health Service.
comparisons if they are to reliably inform decision making. Outcomes frdrn treatment are assessed in terms of QALYs. The time horizon
The consequence is that, in a fully specified economic model, there are of the analysis was 50 years and costs and future QALYs were discounted at an
likely to be more than two treatment alternatives being compared. When this annual rate of 3.5 per cent (NICE 2004). The majority of the data used in
is the case, multiple CEACs can be presented. These curves are conceptually the analysis are taken from the EUROPA trial. An important objective of
the same as the use of acceptability curves to summarize uncertainty on the the analysis was to assess how cost-effectiveness varies according to patients'
cost-effectiveness plane in a two-treatment decision problem except that there baseline risks of the primary EUROPA endpoints (nonfatal myocardial
is now a curve relating to each treatment option.l The characterizing feature infarction, cardiac arrest or cardiovascular death - hereafter referred to as
of the presentation of multiple CEACs to represent multiple and mutually 'primary events').
exclusive treatment options is that the curves sum to a probability of one
vertically. 5.4. 1 . Model structure
Later in this chapter we will return to the example of gastro-oesophageal A Markov model was chosen as the preferred structure for the model and a
reflux disease (GORD) management introduced in Chapter 2 to illustrate the state transition diagram for the Markov model of the EUROPA study is shown
use of multiple CEACs in practice. Particular attention is given to the role of in Fig. 5.6. The general principle was to stay as close to the trial data as possible,
the mean net-benefit statistic as a tool for calculating the curves. therefore patients enter the model into the 'trial entry' state. Over the course of
the model (employing a yearly cycle for 50 years until the vast majority of
5.4. A model of the cost·effectiveness of ACE·inhibition patients have died), patients are predicted to suffer a 'first event', which is
in stable coronary heart disease (case study) represented by the rectangular box. This corresponds to the primary combined
In this section, we describe a model for the treatment of stable coronary artery endpoint of the trial of cardiovascular mortality together with nonfatal
disease with an ACE-inhibitor. Of particular note is the use of regression myocardial infarction or cardiac arrest. Note that the use of the rectangular
methods (both parametric survival analysis and ordinary least squares) to box is to emphasize that this is a predicted event and not a state of the model.
represent heterogeneity following the methods outlined in Chapters 3 and 4. Patients suffering events will either experience a fatal event, in which case they
The model outlined has been described in detail elsewhere (technical report will move to the 'cardiovascular death' state, with the remainder deemed to
available from the authors) and the results have been published separately have survived the event and so move to the 'nonfatal event history' state. In the
(Briggs et al. 2006). The model was constructed from data obtained from the
EUropean trial on Reduction Of cardiac events with Perindopril in patients
with stable coronary Artery disease (EUROPA) study. The trial randomized 12 Lifetable First event
Trial entry
218 patients with stable coronary heart disease to the ACE inhibitor perindopril Eqn 1

8 mg once daily or to matching placebo. Over a mean follow-up of' 4.2 years, the
trial showed that the use of perindopril resulted in a 20 per cent relative risk Ufetable NFE Non-fatal Fatal
Non CVD Mil CA /CVD CVD
reduction in the primary endpoint of cardiovascular death, myocardial infarc­ history ( 1 ) Eqn 2 Eqn2
tion or cardiac arrest (from a mean risk of 9.9 per cent in the placebo arm to SUb. event
8.0 per cent in the perindopril arm) (The EUROPA Investigators, 2003). Eqn 3

Ufetable Sub. event

Eqn 4

1 In the two alternative cases, the standard presentation of CEACs involves plotting the
probability that the experimental observation under evaluation is cost-effective. Note, fig. 5.6 State transition diagram for the Markov model of the EUROPA study.
however, that the probability that the control intervention is cost-effective could be MI,myocardial infarction; CA, cardiac arrest; CVD, cardiovascular death; NFE,
plotted, but would simply amount to the perfect complement of the first curve. nonfatal event; non-CVD, noncardiovascular death.
136 1 ANALYSING A N D PRESENTING S I M U LATION OUTPUT FROM PROBABILISTIC MODELS A MODEL OF THE COST-EFFEGIVENESS OF ACE-INHIBITION 1 1 37

first year after a nonfatal event) patients are assumed to have an elevated risk was at much higher risk of a subsequent event. The risk of a further primary
of a subsequent event. However) if patients do not experience a subsequent event ,o ne or more years after the initial event is based on the first risk
event, they move, in the next cycle (year) to a second 'nonfatal event history' equation, updated to reflect the fact that all patients would have experienced
state. From this state, they can again experience a subsequent event, but at a a nonfatal event.
lower rate than the immediate year after the first event. From any of the states Table 5.1 presents the results of fitting each of these risk models, and shows
where patients are alive) they are deemed at a competing risk of a noncardio­ which of a set of baseline risk factors were predictive in the equations (choice
vascular death. of predictive factors was based on a mix of statistical significance and clinical
With the addition of mean costs and HRQoL scores for each state as judgement). The first equation shows the hazard ratios associated with the
described above, the model was able to estimate expected (mean) costs and risk of a first cardiac event (the trial primary endpoint of cardiovascular
QALYs over a 50-year period for the perindopril and standard management death, myocardial infarction or cardiac arrest). Of note is the 20 per cent risk
options. The model assumes that perindopril is only potentially effective while reduction associated with being randomized to perindopril, as reported in the
it is being taken by patients. In the base case analysis it was assumed that main trial report (The EUROPA investigators 2003).
patients would take the drug only for 5 years, after which they would not incur Other characteristics found to be protective were younger age) being female,
the costs of treatment. Once treatment stops) event rates are assumed to be the previous revascularization and cholesterol lowering therapy. Among charac­
same for the two arms of the model. teristics found to increase risk were being a smoker) having had a previous
myocardial infarction and symptomatic angina. The second equation shows a
5.4.2. Risk equations underlying the model logistic regression estimating the odds of the first cardiac event being fatal. It
The analysis is based on a series of risk equations estimated using EUROPA can be seen that only three characteristics were important enough to enter this
data. These estimate the relationship between the primary event and patients' equation: being older, having had a previous myocardial infarction and
characteristics, including to which arm of the trial they were randomized. Of increased levels of total cholesterol were all found to increase the odds of the
the 1 2 218 patients in the original EUROPA study, 292 did not have a complete event being fatal. Importantly, the use of perindopril was found not to influ­
set of data on covariates. The model development is therefore based on the ence the risk of the event being fatal. The third equation considered the risk of
1 1 926 remaining patients for whom full information was available. With the a subsequent event in the year after an initial event. Just one characteristic was
addition of data from UK life tables on mortality rates for noncardiovascular found to be important in explaining this risk: the presence of angina symp­
reasons, the risk equations facilitate simulation of the fatal and nonfatal events toms (levels 2, 3 or 4 on the Canadian Cardiovascular Society's angina scale)
that a cohort of patients is expected to experience with and without perindopril. or previous history of heart failure elevated the risk of a subsequent event. The
The equations are based on a mean follow-up of 4.2 years in EUROPA, but the ancillary parameter of the Weibull model was less than one, indicating a
statistical relationships they represent are assumed to apply over the 50-year sharply falling hazard of subsequent events over time. The first equation is
time horizon of the analysis. The mean costs and QALYs associated with the used to estimate the risk of subsequent primary events one or more years after
use of perindopril, relative to standard management, are estimated by attaching an initial nonfatal event (with the nonfatal event covariate having been
costs and HROoL values to the events patients experience over time. updated) as the trial itself had very little data in on the long-term risk of
Three risk equations were estimated from the EUROPA data. The first is a events subsequent to the first primary event. The assumption is that after the
standard parametric time-to-event survival analysis relating to the patient's first year, patients will have stabilized and the risks of subsequent events will
risk of a first primary event following randomization. It is based on 1069 be similar to the risk of the first trial event. As this first equation also includes
primary events observed in the 1 1 926 patients (592 in the placebo group and a treatment effect of perindopril, the model effectively assumes that continued
477 in the perindopril group). The second equation is a logistic regression treatment will reduce the risk of subsequent, as well as initial, events.
estimating the probability that a given first primary event would be fatal. This
5.4.3. Quality-adjusted life-years
is based on the 1069 primary events of which 400 (38 per cent) were fatal. The
third equation estimates the risk of a further primary event in the year following The mean difference in QALYs between perindopril versus placebo is the area
an initial nonfatal event, a period during which the data suggested a patient between the two quality-adjusted survival curves; three elements are involved
 w
00
]>
z
l>
Table 5.1 Estimated risk equations for the risk of a first primary event (equation 1 ), the odds of that event being fatal (equation 2) and the risk '<
of a further primary event in the first year after a first nonfatal event (equation 3) '"
Z
G\
]>
Explanatory Equation 1: Risk of first primary Equation 2� Odds that first event Equation3 : Risk of subsequent Z
"
baseline characteristics event (1069 events)* is fatal (400 events) event in first year following initial �

Hazard lower Upper Odds Lower Upper nonfatal event �

m
�
ratio 95% limit 951)/0 limit ratio 95% limit 95% limit Hazard Lower Upper
Z
m

ratio 95% limit 95% limit ::!

z
Use of perindopri! 0.81 0.71 0.91
G\
�

1 .04 1 .03 1 .05

<;:
Age in years c

Years greater than age 65 1 .06 1 .04 1 .08 �

6
Male 1 . 54 1 .2 8 1 .87 z
o
c
Smoker 1 .49 1 .2 7 1 .74 -;
�
c
Previous myocardial infarction 1 .44 1 .26 1 .66 1 .60 1.19 2.14 -;
�

o
�
Previous revascuJarization 0.88 0.77 0.99
<;:
Existing vascular diseaset 1 .69 1 .44 1 .98 �

o
�

]>
Diabetes mellitus 1 .49 1 .28 1 .74 ro

ro
Family history of coronary 1 .2 1 1 .05 1 .38 r
�
artery disease -;
n
Symptomatic anginat 1 .32 1 .1 6 1 .5 1 1 .85 1 .29 2.64 :;:
or history of heart failure o
;;;
r
*Primary trial endpoint of cardiovascular mortality, myocardial infarction or cardiac arrest; tany of stroke, transient ischaemic attack or peripheral vascular disease �

Table 5.1 (continued) Estimated risk equations for the risk of a first primary event (equation 1 ), the odds of that event being fatal (equation 2)
and the risk of a further primary event i n the first year after a first nonfatal event (equation 3)

Explanatory Equation 1 : Risk of first primary Equation 2: Odds that first event Equation 3: Risk of subsequ�"nt "
,
baseline characteristics event (1069 events)* is fatal (400 events) event in first year follow.ing initial
Hazard lower Upper Odds Lower Upper nonfatal event
ratio 95% limit 95% limit ratio 95% limit 95% limit Hazard Lower Upper
ratio 95% limit 95%, limit
Systolic blood pressure 1 .00 1 .00 1 .01

Units creatinine clearance 1 .01 1 .00 1 .02 ]>

below 80mJ/min :;:
0
8MI > 30 (obese) 1 .4 1 1 .22 1 .63 ;;;
r

Total cholesterol 1.13 1 .07 1 .20 1 .2 1 1 .08 1 .3 5 'i:

-;
I
Using nitrates at baseline 1 .42 1 .2 5 1.63 m

n
0
Using calcium channel 1 .20 1 .06 1 .36 �

blockers at baseline
Using lipid lowering
�:z;
0.86 0.75 0.97 -;
therapy at baseline "
Z
m

Constant term -12.27 -12.97 -1 1 .57 -4.37 -5.54 -3.20 -6.46 ··7.25 -5.67
m
�
�

(on the Jog scale)

'i:
Ancillary parameter 0.70 0.59 0.82
�
Z
*Primary trial endpoint of cardiovascular mortCll"lty, myocardial "Infarctbn or cardiac arrest; tany of stroke, tranS"lent ischaemic attack or per"lpheral vascular disease I
in
"
0
z

w
'"
1 40 I ANALYSING A N D PRESENTING S I M U LATION OUTPUT FROM PROBABILISTIC MODELS A MODEL OF THE COST·EFFECTIVENESS OF ACE-INHIBITION 1 1 41

in this calculation. The first is the risk of cardiovascular mortality - this is The third element of costs related to days of inpatient hospitalization for
based On the risk equations described above. The second is the mortality rate any reason which were recorded at follow-up in EUROPA, together with ICD-9
for noncardiovascular causes. This is based on official life tables for England codes on reasons for admission. In order to translate these data into costs, one
and Wales and for Scotland (http://www.statistics.gov.uk/methods_quality/ of the clinical team, blinded to treatment allocation, mapped all relevant ICD-9
publications.asp), with deaths for cardiovascular reasons removed. Data are codes to UK hospital specialties. The cost per day for each specialty was taken
combined to give the rate of noncardiac death, by age and sex, for Great from the UK Trust Financial Returns (NHS Executive 2004) which generated a
Britain. It is assumed that perindopril does not affect the mortality rate from cost for each hospitalization episode in EUROPA.
non cardiovascular causes. Indeed, in the EUROPA trial, death from noncar­ The implications of the cost of concomitant medications and inpatient
diovascular causes was not Significantly different (2.8% versus 2.6% for hospitalizations for the cost-effectiveness of perindopril were assessed using a
placebo and perindopril, respectively). linear regression analysis. Its purpose was to estimate the mean costs associated
The third element in estimating QALYs is the HRQoL experienced by with the events considered in the risk equations defined above - for example,
patients over time. Given that no HRQoL data were collected in EUROPA, the to estimate the mean cost incurred in the year a patient experiences a first
following approach was taken. Mean age- and sex-specific HRQoL scores for primary event. The covariates for the cost regression were selected using the
the UK population were identified based on the EQ-5D instrument, a generic Same process as for the risk equations.
instrument which provides an index which runs between 0 (equivalent to The regression analysis on costs is reported in Table 5.2 and indicates a
death) and 1 (equivalent to good health), where negative values are permitted, 'background' annual cost per surviving patient that depends on age, presence
based on the preferences of a sample of the UK public (Kind et aZ. 1999). To of any existing disease, presence of angina symptoms, creatinine clearance and
represent the mean decrement in HRQoL associated with coronary heart
disease, relative to the population mean, the mean baseline EQ-5D score
Table 5.2 Results of the cost regression showing costs for the different mode! states
measured in all patients in a trial comparing bypass surgery with coronary
and the impact of covariates
stenting was used (Serruys et af. 2001). Patients in this trial had a mean base­
line age of 61 years and, on average, their HRQoL score was 14 per cent below Explanatory variable Cost (£) Standard Lower Upper
the same aged group in the UK population. This decrement was employed to errOr 95% limit 95% limit
represent the HRQoL of all living patients in the analysis regardless of which Nonfatal primary endpoint 9776 1 24 9533 10 019
cardiac events they had experienced. History of nonfatal event 818 91 640 997
Fatal primary endpoint 301 9 1 53 2719 3318
5.4.4. Costs
Non·CVD death 1 0 284 183 9924 1 0 643
Three elements of costs were included in the analysis, all of which are expressed
Age 11 2 7 14
in 2004 UK pounds. The first was the acquisition cost of perindopril. This was
Existing vascular disease 326 47 234 418
based on the use of 8 mg of perindopril once daily over a period of 5 years.
Diabetes mellitus 215 43 131 298
Perindopril is costed at £10.95 per 30-tablet pack, or 37p per day, which repre­
sents its UK price from 1st January 2005 (Joint Formulary Committee 2005). Symptomatic angina 229 34 1 63 295
The second cost element related to concomitant medications, the use of Units creatinine clearance below 7 2 3 10
SOml/min
which was recorded at each EUROPA follow-up visit, based on 13 cardiovas­
cular categories. The British National Formulary (no.48) (Joint Formulary Using nitrates at baseline 230 29 1 73 288
Committee 2004) was used to ascertain the typical individual drug prepara­ Using calcium channel blockers at 1 52 30 93 21 1
tions in each category and their recommended daily doses. The Department of baseline

Health's Prescription Cost Analysis 2003 database (Department of Health Using lipid lowering therapy at 95 28 40 1 50
baseline
2004) was used to estimate a mean daily cost for each of the concomitant drug
Constant -· 1 7 1 06
categories.
1 42 I ANALYSING A N D PRESENTING S I M U LATION OUTPUT FROM PROBABILISTIC MODELS A MODEL OF THE COST-EFFECTIVENESS OF ACE-INHIBITION I 1 43

the use of nitrates, calcium channel blockers or lipid lowering agents at Heterogeneity of cost-effectiveness within EU ROPA
baseline. In addition to these background costs, the regression model predicts The cost-effectiveness model, structured as described above from a series of
the additional costs associated with the modelled events in the trial. In the interlinking'covariate adjusted risk equations and life tables, is able to predict
year in which a nonfatal primary event occurs, £9776 is added to the cost-effectiveness as a function of the covariate pattern. Therefore, for each
background cost. In subsequent years, the addition to the background cost patient in EUROPA the cost-effectiveness model was used to generate a
is £8 18. In the year that a fatal cardiovascular event occurs, the additional prediction of the cost-effectiveness for that set of patient characteristics based
cost is estimated as £3019, which contrasts with an additional cost of on a comparison of treating with perindopril versus not treating that type of
£ 1 0 284 in the year of a noncardiovascular death. This difference can be patient. A histogram of the distribution of predicted cost-effectiveness results
explained by the fact that cardiovascular death is often relatively quick (in terms of incremental cost per QALY gained from perindopril) for each of
compared with other types of death. The advantage of this regression approach the individuals in EUROPA is presented in Fig. 5.7. It is important to recognize
rather than just costing the events of interest is that the full cost to the health that this distribution represents the estimated heterogeneity in the EUROPA
service is captured, including the 'background' costs of extending the life of study with regard to cost-effectiveness - it does not relate to uncertainty as the
patients. data points in Fig. 5.7 relate only to point estimates. This heterogeneity in
cost-effectiveness arises from the heterogeneity of baseline risk of primary
5.4.5. Distributions for model parameters events combined with a constant relative risk reduction associated with
Following the methods outlined in Chapter 4, the distributional assumptions treatment, resulting in differing absolute risk reductions for patients with
are chosen to reflect the form of the data and the way in which the parameters different characteristics. The median (IQ range) cost-effectiveness across the
were estimated. Standard statistical assumptions relating the estimation of the heterogeneous population of EUROPA was estimated as £9500 per QALY
regression models were used for the risk equations. For the survival analysis (£6500-14 400) per QALY. Overall, 89 per cent of the EUROPA population
models the assumption was multivariate normality of the log hazards scale. were estimated to have a point estimate of incremental cost per QALY below
For the logistic regression model, the assumption was multivariate normality £20 000 and 97 per cent below £30 000.
on the log odds scale. For the cost equation, the assumption was multivariate
normality on the raw cost scale. In all cases, Cholesky decomposition of the
variance-covariance matrices was used to capture correlation between coeffi­
£9,500 median cost per QALY
cients in the regression models. For the population norms for EQ-5D, the 1400
assumption is normality within the age/sex-defined strata. For the assumption 1200
of 14 per cent reduction in utility, the assumption is a gamma distribution
89% patients fall below £20,000 per QALY
with variance equal to the mean reduction. No uncertainty is assigned for the 1000

risk of non cardiac death as these estimates are based on national death regis­ is
c 800
•
0
ters where the numbers are very large. IT

3: 600 below £30,000 per QALY

Representing parameter u ncertainty and

5.4.6. 400
heterogeneity in the results of the EU ROPA model
200
One of the main features of the modelling work described above is the direct
o
modelling of heterogeneity between patients, in addition to the handling of
� G� G� �� G& G� �� G� G� O& G� G� G� 0� � v G& O� G� G� O� G�
uncertainty. These two aspects are presented separately below. Firstly, the cf)' q,'<J' rf)'ey"CV'q.,,,0,¢'O'r.!j"'<iJ'c,.<V 'c,.,�f:)'q.,r:;Cl.)''v0fo'<f)0;'cJ'CV'<:S'0·cJ
t:;)'rf:;
; '' '<f;?�'$ 'q.roO°
results across different characteristics of patients in EUROPA are presented. leER
Secondly, the importance of uncertainty for selected types of patient (based on Fig. 5.7 Distribution of predicted cost-effectiveness results for individuals in the
their ranking of estimated cost-effectiveness) is illustrated. EUROPA study.
j
'Il
H
144 I ANALYS!NG A N D PRESENTING S I M U LATION OUTPUT FROM PROBABILISTIC MODELS ANALYSING S I M U LATION R E S U LTS: MULTIPLE TREATMENT OPTIONS 1 1 45

Probabilistic assessment of parameter u ncertainty Table 5.3 Illustrative cost-effectiveness results for five covariate patient profiles

As highlighted above, the analysis of heterogeneity was based only on point Percentile of ranked cost-effectiveness results
estimates of the cost-effectiveness of perindopril. As the model is based on a 2.5th 25th Median 75th 97.5th
linked set of statistical equations, the probabilistic analysis is reasonably Age in years 51 35 59 52 55
straightforward, being based on standard statistical assumptions concerning Male
the distribution of parameters, as described previously. However, the need to
Smoker 0 0 o o o
separate out heterogeneity from uncertainty complicates the presentation.
Previous myocardial infarction o o
The approach taken is to illustrate uncertainty for five individuals, chosen on
Previous revascularization 0
the basis of their location within a rank ordering of the cost-effectiveness
results (the 2.5th, 25th, 50th (median), 75th and 97.5th percentiles). The Existing vascular disease 0 o o o

5-year risk of a primary trial endpoint and the covariate patterns for each of Diabetes mellitus 0 o o o

these individuals are set out in Table 5.3 along with the point estimates of Family history of coronary artery 0
incremental cost, QALYs gained and estimated cost per QALY ratio for these disease
five individuals. These individuals are then used to illustrate uncertainty in the Symptomatic angina/history of 0 0 0 0
remainder of the section. heart failure

Propagating the statistical parameter uncertainty through the model for Systolic blood pressure 1 35 1 25 1 25 125 122
each of the individual covariate patterns presented in Table 5.3 generates an Creatinine clearance mllmin 74 125 1 09 1 08 76
estimated joint distribution in the uncertainty of additional cost and QALY 8MI > 30 (obese) 0 0 0 0
gain estimates. These are presented on the cost-effectiveness plane in Fig. 5.8 Total cholesterol 4.2 6.3 6.6 3.7 4.48
(for just the median, 2.5th and 97.5th percentiles from Table 5.3 for clarity). Using nitrates at baseline 0 0
Each joint density plot is summarized by an acceptability curve (van Hout
Using calcium channel blockers at o o 0 0
et al. 1994) in Fig. 5.9. baseiine
By effectively presenting uncertainty and heterogeneity on the same plot, Using lipid lowering therapy at o 0
these figures emphasize the importance of understanding the difference baseline
.•..•..............•..•...... • . . . . . . • . . . . . . . . . . . . . . . . . . . . . ...............................
between the two concepts. In particular, note how the separate acceptability Predicted 5-year risk of primary 23 10 9 6 4
curve for each type of patient within Fig. 5.9 encourages the decision maker to endpoint (%)
make an independent treatment decision across different categories of patient, Incremental cost (f) 404 352 484 447 500
....... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

based on both the expected cost-effectiveness for that set of characteristics and QAlYs gained 0.1 08 0.054 0.051 0.031 0.0 1 6
the associated uncertainty. ......................................

Incremental cost-effectiveness 3700 6500 9500 1 4 400 32 1 00

ratio (£)
5 . 5 . Analysing simulation results: multiple
treatment options
This section employs the model for assessing the cost-effectiveness of six 5.5. 1 . A model for assessing the cost-effectiveness
management strategies for the treatment of GORD that was introduced in of GORD treatment
Chapter 2 in order to illustrate the issues that arise when evaluating multiple Gastro-oesophageal reflux disease is a common condition that results
treatment options that are mutually exclusive for a given set of patients. Full from regurgitation of acid from the stomach into the oesophagus. The
details of the model have been presented in detail previously (Goeree et al. most frequent symptom of GORD is heartburn and the majority of patients
1999; Briggs et al. 2002a), and so the focus of this section is on a discussion of with GORD require pharmacotherapy to reduce acid secretion. Currently,
the appropriate probabilistic assessment of the model. the choice of first-line antisecretory therapy is between the H2-receptor
146 I ANALYSING A N D PRESENTING S I M U LATION OUTPUT FROM PROBABILISTIC MODELS ANALYSING S I M ULATION RESULT S : MULTIPLE TREATMENT OPTIONS ! 147

£700 antagonists (H 2RAs), such as ranitidine and cimetidine, and proton pump
inhibitors (PPIs) such as omeprazole. Although they have higher acquisition
£600 costs, PPIs have been found to be more efficacious than H2RAs in terms of
both the rate and speed of healing.
£500 ,
w
, ':
The objective of the original study was to compare, over a I -year period, the
;;
0
u
£400
expected costs and outcomes of alternative drug treatment strategies for the
" ' .
c
0
.'
:�"".:'i; .v.";:�·:: I,cr :,: '. .
•
,
" management of patients with erosive oesophagitis confirmed by endoscopy,
� . but without complications such as Barrett's oesophagus or stricture. Outcomes
:':: :.: . : ,
£300
' '
.... .
..
:'.

I
"C : �
'" are quantified in terms of GORD recurrence and weeks per year without
, ": ",'
�
£200 GORD as indicated by data from clinical trials on healing and recurrence of
oesophagitis.
£100 � 97.5th Percentile of risk I
Six strategies involving different combinations of first-line agents and
I
• 50th Percentile (median) risk
. 2.5th Percentiie of risk
change of therapy conditional on failure to heal or recurrence of GORD were
-0_05 0 0.05 0.1 0.15 0.2 0.25 ! modelled:

I
QALYs gained
• Strategy A: Intermittent PPJ. Acute treatment with a PPI for 8 weeks and
Fig. 5.8 Cost-effectiveness plane showing the scatter plot of 1000 Monte Carlo trials then no further treatment with prescription medication until recurrence.
of the probabilistic model for the !OW, moderate and high-risk individuals from
EUROPA
• Strategy B: Maintenance PPJ. Acute treatment with a PPI for 8 weeks then

I
continuous maintenance treatment with a PPI (same dose).
• Strategy C: Maintenance H2RA. Acute treatment with an H2RA for 8 weeks
and then continuous maintenance treatment with an H2RA (same dose).
0.9 • Strategy D: Step-down maintenance prokinetic agent. Acute treatment
I with a prokinetic agent (PA) for 12 weeks and then continuous mainte­
I

I
0.8

�R
nance treatment with a lower dose of PA.

�:: ! • Strategy E: Step-down maintenance H,RA. Acute treatment with a PPI for
8 weeks and then continuous maintenance treatment with a H2RA.
_ 0.5
• Strategy F: Step-down maintenance PPJ. Acute treatment with a PPI for 8
g OA
�
:0
0.3 -0- 2.Sth Percentile
weeks and then continuous maintenance treatment with a lower dose PPJ.
Treatment options A-F represent clinical strategies rather than single drug
25th Percentile
-<.>-
a.. 0.2 - 50th Percentile (Median)
-.- 75th Percentile
treatments for the management of erosive oesophagitis where the physician is
0.1 -.- 97.5th Percentile assumed to increase the dose of a drug or switch to another drug if the patient
fails to respond to the tlrst-line treatment. The structure of the decision tree that
£30,000 £40,000 £50,000
was developed was shown in Fig. 2.1. The model is recursive in two 6-month
Decision threshold
periods; hence, probabilities of recurrence in the period to 12 months are condi­
Fig. 5.9 Cost-effectiveness acceptability curves for the five covariate patterns repre­
senting percentile points from the cost-effectiveness distribution. tional upon recurrence or nonrecurrence in the period from 0 to 6 months.
Treatment outcomes
For GORD, the most commonly used formulation of outcome for economic
evaluation has been either oesophagitis-free or symptom-free time in a
period of follow-up. The advantage of such a measure is that it combines two
148 I ANALYSING A N D PRESENTING S I M U LATION OUTPUT FROM PROBABIUSTIC MODELS ANALYS I N G S I M U LATION R E S U LT S : MUlTIPLE TREATM ENT OPTIONS 1 149

important aspects of efficacy: (i) the speed with which oesophagitis is healed; incremental cost-effectiveness ratios are calculated by comparing each option
and (ii) the likelihood of oesophagitis recurring. In this analysis, the primary with the next more costly and more effective intervention. This process
outcome measure is GORD-free time during the 1 2-month period of the produces an_:_efficiency frontier' of increasingly more costly and more effective
model, defined as the time where the oesophagitis is healed. A meta-analysis strategies. The results of this analysis for the GORD model are presented on
of healing and recurrence studies published to November 1997 was under­ the cost-effectiveness plane in Fig. 2.2, which also shows the efficiency frontier
taken to estimate healing and recurrence probabilities together with associated between nondominated options.
GORD-free time. Full details of this analysis are given in the original study The figure clearly shows that step-down maintenance PA (strategy D) is
(Goeree etal. 1999). dominated by maintenance H2RA, (strategy C), intermittent PPJ (strategy A)
and step-down maintenance H2 RA (strategy E). The efficiency frontier is
Resou rce use and unit costs
given by the lines joining strategies C (the origin) A, E and B. Strategy F is
Generic prices wereysed for drugs where a generic equivalent is available, internal to this frontier indicating that it also can be ruled out through the
employing the 'best available price' from the Ontario Drug Benefit (ODB) principle of extended dominance (I.e. a linear combination of strategies E and
programme together with a 10 per cent pharmacy mark-up charge. A dispensing B would strongly dominate F). The slope of the frontier at any point reflects
fee of Can$4 . 1 1 was used (i.e. ODB programme fee of Can$6. 1 1 less a incremental cost-effectiveness - the additional cost at which additional effects
Can$2.00 patient co-payment). Cost estimates for physician fees were taken
can be purchased.
from the physician fee schedule for Ontario, and procedure costs, such as
endoscopy, were estimated from a hospital participating in the Ontario Case Probabilistic assessment of the GORD model
Costing Project in Southwestern Ontario. Following the principles outlined in Chapter 4, distributions were specified
To estimate the costs associated with the management of patients with for all the relevant parameters of the model - the details of which can be
symptoms of GORD recurrence, information on clinical practice patterns and found in the original article (Briggs et al. 2002b). The probabilistic analysis
resource utilization was obtained by convening an expert physician panel and was undertaken by randomly sampling from each of the parameter distributions
using a modified Delphi technique. Estimated resource utilization was then and calculating the expected costs and expected weeks free from GORD for that
combined with unit cost information to give the mean cost associated with combination of parameter values. This process formed a single replication of the
each recurrence under each management strategy. model results and a total of 1 0 000 replications were performed in order to
examine the distribution of the resulting cost and outcomes for each strategy.
5.5.2. Results from the GORD model The results of these 1 0 000 replications from the model are presented on the
Before demonstrating the probabilistic assessment of multiple treatment cost-effectiveness plane in Fig. 5.10 together with the baseline estimate of the
options, the standard approach to generating a cost-effectiveness frontier of efficient frontier.
mutually exclusive options is reviewed. For each of the individual replications, an efficient frontier could be calculated
together with the incremental cost-effectiveness ratios for treatments on the
Deterministic cost-effectiveness acceptability frontier. In particular, Fig. 5.10 suggests that it may not be possible to rule out
The decision tree model outlined in Fig. 2.1 was evaluated to estimate the strategy F, the strategy based on step-down maintenance PPI, as it potentially
expected costs and the expected weeks without GORD in the 1 2-month forms part of the frontier in many replications. Note, however, that it is
period of the model. The conventional approach to examining the cost­ impossible to gain a dear view from Fig. 5.10 as to how often strategy F forms
effectiveness of the alternative strategies involves first determining whether part of the frontier. This is because there can be substantial covariance
any strategies are strictly dominated by other strategies having both lower between the simulations plotted in the figure (it turns out that strategy F
costs and greater therapeutic effects, and secondly determining whether any forms part of the frontier in 27 pel' cent of simulations although it is not clear
strategies were dominated through the principles of extended dominance, that how this result should be interpreted). The potential for covariance between
is, whether linear combinations of other strategies can produce greater benefit simulation points limits the usefulness of presenting simulation results
at lower cost (Cantor 1994). Then, among nondominated treatment options, for multiple strategies on the cost-effectiveness plane. Only for cases such as
1 50 I ANALYSING A N D PRESENTING S I M U LATION OUTPUT FROM PROBABILISTIC MODELS ANALYSING S I M U LATION R E S U LTS: MULTIPLE TREATMENT OPTIONS 1 1 51
I
$1 ,200

$1,100
A: Intermittent PPI
B: Maintenance PPI
I
1
any other treatment alternative. The proportion of times a strategy has the
highest net'benefit among the 1 0 000 replications of the model gives the
strength of evidence in favour of that strategy being cost-effective, This ability

I
C: Maintenance H2RA of the net-benefit framework to handle multiple mutually exclusive treatment
0: Step�down maintenance PA
$1 ,000 E: Step-down maintenance H2RA options is a very strong advantage of the approach and provides a much more

§
F: Step-down maintenance PPI straightforward solution to finding the cost-effective option from among
multiple treatment alternatives than the conventional approach outlined
1;; $900

j
above, We simply formulate the mean net-benefit for each option and choose
I the option with the greatest mean net-benefit. Only this option will have a
$800
! positive incremental net-benefit relative to any other option, Note that there is
i
no need to consider dominance (strict or extended) and there is no need to
$700 I specify the appropriate comparator, Therefore it is very easy to implement in a
form that can automatically find the optimal option in a large simulation
$600 experiment.
38.00 39.00 40.00 41.00 42.00 43.00 44.00 45.00 46.00 47.00 48.00
For example, ten simulations from the GORD model of six management
Weeks free of GERD
Fig. 5.10 Results of 1 0 000 Monte Carlo simulation evaluations of the gastro.. options are shown in Table 5.4. The corresponding net-benefit for each option
oesophageal reflux disease model presented on the cost-effectiveness plane. is shown on the left-hand side of Table 5,5, assuming a willingness to pay
of $200 per week free from GORD symptoms, The indicator for whether
a strategy is optimal is shown on the right-hand side of the table and corre­
strategy D where there is clear separation ofthe (marginal) probability densities sponds to an indicator of whether the option has the highest mean net-benefit
is it possible to draw a firm conclusion, Note, for example, that it would be among all the options,
impossible to infer the slope ofany points along the simulated cost-effectiveness I n reality, of course, the threshold ratio for a week free from GORD
frontiers as it is not clear which simulated points should be joined together, symptoms is not known. However, by plotting out the proportion of times
For this reason we would suggest that the cost-effectiveness plane is not the intervention has the greatest net-benefit, for all possible values of -t, much
the appropriate vehicle for presenting simulation output for more than two can be learned concerning the implications of the estimated uncertainty
treatment options. for the treatment decision. This generates a series of CEACs for the multiple
Of course, it is possible to analyse the simulation results appropriately as option case, To calculate these CEACs, we have only to average across the
this covariance information is recorded. Conditional on knowing the thresh­ trials to fmd the proportion of times that each option is optimal for a given
old ratio - in this example the willingness to pay for a week free from GORD ceiling ratio, We then repeat the process with a new threshold ratio and plot
symptoms - it is possible to identify the efficient frontier, calculate the incre­ the results.
mental cost-effectiveness ratios and choose one strategy from the six available Figure 5 , 1 1 shows the result of just such an exercise for the probabilistic
for each of the 10 000 replications. However, the application of the traditional evaluation of the GORD modeL Note the use of the log scale for the threshold
construction of a cost-effectiveness frontier is rather involved and it turns out ratio to help present the curves more clearly and also note that the summation
that a more straightforward approach exists, which is simple to implement for across the curves at any point gives a total of one: this must be the case as the
large numbers of simulation results, options are mutually exclusive.
This illustration of the usual case for two treatment alternatives is easily As expected, strategy D does not feature in Fig, 5, l ! , indicating that it is
generalized to the multiple option case: the management option of choice never a contender for cost-effectiveness, Strategy F does feature, although it
from the six strategies under evaluation in the GORD example will be the never achieves more than 1 3 per cent of simulations, suggesting it is cost­
option with the greatest mean net-benefit. This must be the case, as only that effective, even at the most favourable threshold ratio (about $260 per day free
treatment will have a positive incremental net-benefit when compared with from GORD symptoms).
 "'
"

"
z
"
'<
�
z
G>
"
z
o
Table 5.4 Example of ten simulations of the six different management options �
�
m
�
F
Z
A B C D E m

SFWks Cost ($) SFWks Cost ($) SFWks Cost ($) SFWks Cost ($) SFWks Cost ($) SFWks Cost ($) ='
Z
43.96 714 47.00 1116 41 .78 649 38.95 825 46.05 738 46.49 959 G>
'"
44.34 642 47. 1 8 1 085 40.92 671 39.63 799 45.50 753 46.60 936 ;;:
C
r
44.37 695 47.22 1 1 00 42.33 634 39.54 822 46.29 727 46.29 982 "<
6
44. 1 7 667 47. 1 0 1 097 40.82 679 39.02 816 45.47 762 46.43 949 z
o
c
44.01 676 47. 1 2 1 09 1 42.17 622 39.03 799 45.94 726 46.40 948 �
c
�
44.1 4 709 47.42 1 079 41 . 1 0 693 39.62 813 45.69 773 46.56 961 �
�

o
46.72 950 �
44.21 692 47.26 1 1 09 41.69 656 39.42 807 46.06 740
;;:
44.41 719 46.99 1 1 30 41.59 681 39.09 840 46.06 766 46.51 978 �

o
�
44.27 693 47. 1 8 1 088 40.69 717 40.72 804 45.36 796 46.68 950
'i:
ro
44.59 651 47.46 1 083 40.66 717 39.39 809 45.35 792 46.66 953 r
Vi
SFWks, symptom free weeks ='
n
;;:
o
o
m
r
�

&.

'" ...· 00 00 00 00 00 00 00 00 00 00 » 2
3 tE Probability cost-effective N � - � � � � � N 0
!!> .g r;}
-0 • m m m � N m � N � !::I. C"
00 � w � � � m � m 00
��
l
3 ;­
� � o
" o
"
o
c.,
0
:r:..
0
c.n
0
en
0
�
0
Co
o
'"
i
..
. � U1
- » 00 00 00 00 00 00 00 00 00 co m � a u'
o � .j::,. W N W .f::,. W W W W N ....
�
'" � & $ t g � � t ;:g � . � :: �
-
�
0- -0
",
;::; C­ il: _
'"
0-
'"
� 2: o '-oJ -...; -...j -....J -....J � -....J -....J -....J -...J ! n. 9- co co
::;.;; � .j::,. � Ql,
»
.j::,. m m � 00 .j::,. 00 � -...J
�r a:
�
c '< � N W 00 N 00 w � 0
z
�
»
� 8
� 00 W m � N W 00
c 0 "
l §:' -h
w �
- "
m
'" '" C '"
'<
� '-J '-J m -...J � '-oJ m � � m '< �
-
=e - � � :.s 8 g � g � � § g' 9- Z
'" Q G>
o -
- =e
o 0 00 � '-oJ � � -...J �
s' :a. '"
� '"
0 m rul §l ;;:
co 0- '" c
& 9- -l
00
N
00
N
00
.j::,. �
00
w
00
�
00
w
co
�
00
w
00
�
- <.0
- 0
C _. OJ"
.....
'-oJ
'-J
�
'-J
�
m
'-oJ
N
m
�
m
N
w w
N
�
OO
'-J
N
"' '<
"' ru �
'" � <D
� 5.. 6
�

� (j) g} ... z
Q..
-
Q. 00 00 00 00 00 00 00 00 00 00 5' �
0. w W W W W W W N W W 0. m
� 00 N � � w W -...J 00 � �
;;;
ru JJ � m � � 0 N 00 � W 0 f,'J' C
"i
'" m
3 g. �
):o �
'" �
�

" '"
- -
o 0 0 0 0 0 0 0 0 0 =e ;;:
3 c
'" o
o "' !!!. � "i
§l 0 0 0 o 0 0 0 01 5'
�
.�
<.0

� n
�
=e �
n;­ o 0 0 0 0 0 0 0 0 0 �
� m
o
<.0
=e
"1 �
OJ o 0 0 0 0 0 0 0 0 01 0 � ;;:
Z
i': m
ru
ro §l �
Vi'
o 0 o o o m
Jll o
�
- '<
='
o
o Vi'
z
o
o o o 0 0 0 o 0 ' '''
�

"'
w
 T
,
f

I
1 54 1 ANALYSING A N D PRESENTING SIMU LATION OUTPUT FROM PROBABILISTIC MODELS
EXERCISE: ANALYSING S I M U LATION RESULTS 1 1 55

!
While this sort of presentation of the choice between mutually exclusive It is important to note that in moving towards the use of acceptability
treatments in the face of many options is a natural extension of the use of curves .to represent uncertainty in treatment decisions, we are encouraging
analysts and d.ecision makers to think beyond conventional error rates in stan­
CEACs in the two treatment cases) the issue arises of how exactly decision 1
makers should use this information to choose between the remaining strate­ dard statistical a)lalysis to guide decision making. It is clear from the curves in
gies that form part of the frontier. One approach would be to say that. for any Fig. 5.1 1 that no options can be distinguished at the 95 per cent level of uncer­
given value of the shadow price, the optimal decision would be to choose the tainty within a reasonable range of willingness to pay for a week free from
strategy that is most likely to be cost-effective. But of course, this decision rule GORD. Of course, the arbitrary nature of the conventional approach to deci­
gives the exact same treatment recommendations as the baseline estimates in sion making under uncertainty emphasizes the inadequacies of such a simple
Fig. 2.2, where uncertainty was 'not considered. Furthermore, Fenwick and decision rule based on statistical significance, indeed, Claxton ( 1 999) has
colleagues (20 0 1 ) have pointed out that such an approach is not consistent argued that significance testing of this sort is irrelevant. Instead he suggests
with maximizing net-benefit, if the distribution of net-benefit is skewed. that decision making should be fundamentally concerned with expected
Instead, they propose presenting the (cost-effectiveness acceptability frontier', values. That is not to say that the decisions should be made on the basis of the
baseline point estimates as presented in Fig. 2.2 without reference to uncertainty
which is plotted for the GORD example in Fig. 5 . 1 2 . The frontier shows
the options that would be chosen under the rule of maximizing net-benefit, in obtaining those estimates. Rather, that the expected returns to obtaining
and although it closely accords with an approach of choosing the option further information should be assessed in order to determine whether it
that has the maximum probability of being cost-effective, there are some is worth commissioning more research to obtain improved estimates of the
discrepancies. decision parameters and it is to this issue that we turn in the next chapter.
For example, if the willingness to pay to avoid GORD symptoms was $270,
then due to the skew in the distribution of net-benefit, strategy B is preferred 5 . 6. Summary
to strategy E in terms of expected net-benefit, despite the fact that strategy E In summary, probabilistic modelling of deterministic models is a practical
has a higher probability of being cost-effective. solution to the problems of conventional simple sensitivity analysis. Presenting
uncertainty as CEACs encourages analysts and users to think carefully about
the state of evidence relating to the parameters of the model. The use of
acceptability curves to present information on the probability of multiple
treatment options being (ost�effective is a natural extension of the (two alterna­

'\\
0.9
tives' case usually presented in the literature. We will argue in the next chapter
0.8 that an appropriate approach to decision making under uncertainty requires an
"
E
> 0.7 C
understanding of the value of collecting additional information to inform deci­
t5 A sion making and that CEACs are not sufficient tools on their own. Nevertheless,
:m 0.6
"
1i;
it should become clear in the next chapter that expected value of information
0 0.5
" methods will have to be predicated on a well specified probabilistic model.
g 0.4
:0
'"
'"
Ie 0.3
n.
5 . 7 . Exercise: Analysing simulation results
0.2 from the total hip replacement model
0.1
5.7. 1 . Overview
0
100 1 00 0
The purpose of this exercise is to show how Excel can be used to run simula­
10
F tions drawing from the distributions chosen for the previously constructed
Threshold Ratio (A)
probabilistic model of total hip replacement (Exercise 4.8) using macros to
Fig. 5.12 Cost-effectiveness acceptability frontier for the gastro-oesophageal reflux
disease example.
record the results.

•
1 56 J ANALYSI N G A N D PRESENTING S I M U LATION OUTPUT FROM PROBABILISTIC MODELS EXERCISE: ANALYSING S I M U LATION RESULTS J 1 57

The step-by-step guide covers four main areas: vii. Now use the on-screen button to stop the macro (this should have
1. Structuring the simulation and writing macros. opened in a separate window when you started recording - if it did
nodor any reason, you need to stop the macro using the Tools >
2. Analysing results: the cost-effectiveness plane.
Macro > Stop Recording menu).
3. Another macro for calculating CEACs.
viii. Open the Visual Basic Editor (from the Tools> Macro menu) and you
4. Multiple curves: different patient subgroups. will find that the basics of your macro have been recorded for you. As
The template for this exercise is ' Exercise 5.7 - template.xls'. demonstrated in Box 5 . 1 , you need to add a variable to count the
number of trials ( 1 000), a Do Loop and an Active Cell Offset.
5.7.2. Step-by-step guide ' When complete, set your macro running from the Tools > Macro > Run
1 . Structuring the simulation and writing macros
Macro menu and your macro should repeatedly copy row 4 into the 1000 rows
below - this is the result of the probabilistic sensitivity analysis that can now
In the template file you will find a new worksheet <Simulation>. This sheet
be analysed.
provides the structure for much of this exercise. The aim is to record repeated
Note) you may want to examine the code in the solution file to see how to
results from the probabilistic analysis constructed in Exercise 4.8 and to
turn off the screen updating so that the macro runs in the background without
analyse the results. Results of each round of simulations (trials) are to be )
showing the actions and causing the annoying (screen flicker effect.
placed in rows) while columns record results for each trial.
i. Note that columns C to 0 are labelled with parameter names. The aim
here is to record the input parameter values along with the overall cost­ 2. Analysing the results: the cost-effectiveness plane
effectiveness results (in columns Q to T). The first task is to link the cell Now that we have the results of 1000 probabilistic trials of the model, the
under the parameter label with the relevant probabilistic parameter results can be analysed.
(first making sure that the 'probabilistic switch' is set to 1). Although 1. Begin by calculating the incremental cost and effect results (columns
you could do this by pointing to the live cells of the <Parameters> U and V of the <Simulation> worksheet) for each trial from the cost
sheet, the most straightforward method is to use the parameter names. and effect in each arm of the model.
ii. Now do the same with the cost and effect output parameters. ii. Now plot these incremental results on the cost-effectiveness plane.
By this stage, the row 4 for columns C to T should contain the corresponding Either plot them directly, or use the <CE plane> worksheet, where
probabilistic parameter. These cells should update to new parameters when some of the formatting has been done for you.
the <F9> key is pressed providing the 'switch cell' on the <Parameters> Notice the slightly strange shape of the 'cloud' of points. This is due to the
worksheet (D3) is set to 1. The next step is to construct a macro that will copy nonlinearity relating input and output parameters inherent in the model
row 4 and paste it into the rows below 1000 times from row 6 to row 1005. You structure. As a consequence) the point estimates from the deterministic model
may find it useful to refer to Box 5.1 relating to macro writing. (evaluated at the mean values of the distributions) are not the same as the
iii. Start by setting a new macro to record from the Tools > Macro menu. expectations across the output parameters.
iv. First enter 1 in the 'switch' cell D3 on the <Parameters> worksheet to iii. In cells Q1007:VI007 calculate the mean values across the 1000 trials
make sure the model is in probabilistic mode. of the model.
Hint: do this even ifit is already set at 1, as the purpose is to record the action. iv. Now link the probabilistic results table on the <Analysis> sheet to
v. Now switch to the <Simulation> worksheet and select cells C4:T4 and these results and compare with the deterministic results (ensuring that
copy and paste these cells to row 6. Remember to use the paste special the model is in deterministic mode).
command in order to paste only the values and not the formulae. You should notice that the probabilistic results are slightly different from the
vi. Finally, set the switch cell back to 0 so that the model is in determinis­ deterministic results. In fact, as argued in Chapter 4) it is the expectation
tic mode. across the probabilistic results that should be used as point estimates.
1 58 I ANALYS ING AND PRESENTING S I M ULATION OUTPUT FROM PROBABILISTIC MODELS
EXERCISE: ANALYS ING SIMULATION R E S U LTS 1 1 59

3. Another macro for calculating CEACs

Notice that columns AB and AC are direct complements of one another. As
Once probabilistic results are obtained as in the form of the first step above, in the . simple two treatment alternatives case, if one treatment is not cost­
the analysis of the results can proceed much like the analysis of bootstrap effective, then the other must be. Also notice that columns AC and AD give the
replications of real data in a statistical analysis (Briggs and Fenn 1998). This same results for the new prosthesis. This is because the difference between
includes the potential for calculating confidence intervals for the ICER (using, average net benefits is the incremental net-benefit.
for example, simple percentile methods) or CEACs. We focus on the calculation As the results in AB4:AD4 have been linked to the threshold ratio cell Zl,
of CEACs as the appropriate way of presenting uncertainty in cost-effectiveness a simple macro can be used to change the value in Zl and to record the
acceptability as a method that avoids the problems that can arise when corresponding results.
uncertainty covers all four quadrants of the cost-effectiveness plane.
vi. Start by selecting a new macro to record from the Tools > Macro menu.
The calculation of CEACs will involve another macro (not essential, but
vii. First copy the value of the ceiling ratio from cell AF6 to ZI.
useful in this context). 2 However, some preparation of the worksheet is
required first. In particular, the presentation mentioned that two alternative viii. Now copy the results from AB4:AD4 to AG6:AI6 remembering to
approaches to the use of the net-benefit statistic could be used to generate specify paste special: values.
acceptability curves. This exercise will calculate both in order to demonstrate ix. Now stop the macro using the on-screen button and open the Visual
that the methods are equivalent. Basic Editor.
i. On worksheet <Simulation> cell ZI contains the 'threshold ratio' (i.e. x. The basics of your macro have been recorded for you - in the same
the maximum acceptable willingness to pay for health gain). Use this way as before you need to add a variable to count the number of ceil­
value to calculate the mean net monetary benefit in columns X and Y ing ratio iterations (58), a Do Loop and an Active Cell Offset.
for the standard and new prostheses, respectively (that is, using the xi. Once the macro has worked its magic, you can plot the resulting
costs and effects for the standard prosthesis from columns Q and R, CEAC using the <CEA Curve> worksheet. Plot both the standard and
and similarly for the new prosthesis). new prosthesis curves.
ii. In column Z, calculate in the incremental net monetary benefit of the
new prosthesis using the incremental costs and effects from columns 4. Multiple curves: different patient subgroups
U andV. Up to this point we have been working with a model where the patient charac­
iii. In columns AB and AC use an IF( . . . ) function to generate an indicator teristics have been set for a female aged 60 years. In this part of the exercise, we
variable to show a value of 1 if the prosthesis (standard and new, respec­ want you to repeatedly obtain results for men and women aged 40, 60 and
tively) is the most cost-effective for that trial, 0 if it is not. Remember, 80 years and record the results and present them as multiple acceptability
the most cost-effective prosthesis has the greatest average net-benefit. curves.
iv. In column AD use an IF( . . . ) function to generate an indicator variable i. Your worksheet should already have the results for women aged 60.
to show whether the new prosthesis is cost-effective compared with Copy and paste both the (probabilistic) point estimates and the CEAC
the standard prosthesis. This occurs when the incremental net-benefit results (for the new prosthesis only) into the table template in the
of the new treatment is positive. appropriate point in the <Sub-group results> worksheet.
v. Finally, calculate the mean of columns AB, AC and AD in row 4. This ii. Now change the patient characteristics on the <Analysis> worksheet
represents the proportions of simulation trials in which the associated and re-run both your simulation and CEAC macros.
prosthesis is cost-effective given the value of the ceiling ratio in cell Zl. Hint: you could set up a button to do this automatically and even link the
two macros together with a single macro. Also make sure that you
2 For example, it is possible to create CEACs in Excel by using the 'data table' command. provided a sheet reference to your CEAC macro - otherwise running it
The disadvantage ofthis method, however, is that it includes large amounts of'active' cells from another sheet can have devastating results!
to the spreadsheet, all of which recalculate whenever an action is undertaken. In large iii. Finally, plot the resulting CEACs for the different patient characteristics
models this can greatly increase the time it takes to run simulations,
on the <Sub-group CEACs> worksheet.
1 60 I ANALYSING AND PRESENTING S I M U LATION OUTPUT FROM PROBABILISTIC MODELS REFERENCES I 1 61

You should find that the curves can be quite different for different patient iii. On the <Simulation> worksheet, link the input and output parameters
characteristics. For example, the new prosthesis seems to be more cost�effec­ in F4 'and U4:V4 to the relevant cells (note that the naming has already
tive generally for men (due to the fact that on average men have higher failure beert done for you).
rates and so the absolute benefits of reducing failure risk are greater) and that iv. The simulation macro has already been edited to take into account the
the new prosthesis is not so cost-effective for elderly patients (where death is wider range of results to be recorded (C4:V4) so you can now run the
an important competing risk for prosthesis failure). 'MCsimulation' macro.
Now that the adaptations relating to the third option are complete the
updated results can be produced.
5.8. Exercise: introducing a third prosthesis
v. Calculate the mean values across the simulations in cells Q 1 007:VlO07
into the total hip replacement model
of the <Simulation> worksheet and use these values to update the point
5.8. 1 . Overview estimates for the probabilistic analysis on the <Analysis> worksheet.
The purpose of this exercise is to introduce a further prosthesis into the THR vi. Plot the three options on the cost-effectiveness plane using the <CE
model in order to illustrate the approach to calculating multiple CEACs for plane> worksheet. (Note that you should do this by plotting three
multiple treatment options. point clouds from columns Q to V, rather than plotting increments).
The step-by-step guide covers two main tasks: Note that one of the problems of looking at the three clouds of points on
1. Adding in a third prosthesis option. the cost-effectiveness plane is that we lose the perception of how the points are
correlated between the three options.
2. Multiple curves for mutually exclusive treatment options.
The template for this exercise is 'Exercise 5.8 - template.xls'. 2. Multiple curves for mutually exclusive treatment options

5.8.2. Step-by-step guide The results from the probabilistic analysis of mutually exclusive options can
be presented as multiple acceptability curves. However) in contrast to the inde­
1 . Adding in a third prosthesis option
pendent subgroup analysis of the first part of this exercise, these multiple
Although in the vast majority of trials just two treatment alternatives curves must sum to one. We must therefore repeat the analysis based on
are compared, in modelling it is much more common to look at multiple net-benefit, but this time for three options, which is where the use of mean
treatment options. If you open Exercise 5.8 - template. xis, you will find that net-benefit becomes useful.
an additional worksheet <NP2> has been added that contains the model
i. Calculate the average net monetary benefit for each option in columns
for a second new prosthesis. Take a moment to examine the <Hazard
X to Z.
function> worksheet, which contains additional information concerning
this prosthesis in terms of its effectiveness and correlation with other ii. Now create indicator variables in columns AB to AD to show whether
parameters. each of the three prostheses are cost-effective for that trial (only the
The aim in this section is to update the model so that the third pros­ prosthesis with the greatest net-benefit is cost-effective).
thesis option is fully integrated. This involves updating the parameter, analy­ iii. Calculate the proportions of trials in which each of the prostheses are
sis and simulation sheets before re-running the probabilistic sensitivity cost-effective in AB4:AD4.
analysis. iv. Now run the 'CEACurve' macro, which will generate the data for the
i. Starting with the <Hazard function> worksheet, update the last row of acceptability curves in columns AF to AI.
the Cholesky decomposition matrix (row 32). Also update the last row v. Finally, plot the three acceptability curves using the worksheet <CEA
of the random variable generating table (row 41). curve>.
ii. Move to the <Parameters> worksheet and add in the information As was noted previously, the interpretation of multiple acceptability curves
concerning the effectiveness of the new prosthesis in row 28. can be tricky - especially when the model is nonlinear. Fenwick et al. (2001)

I
J
 T
1 62 I ANALYSING A N D PRESENTING SIMULATION OUTPUT FROM PROBABILISTIC MODELS

I REFEREN C E S I 1 63

f
Karlsson, G. and Johannesson, M. ( 1996) 'The decision rules of cost-effectiveness analysis',
have argued for the use of a cost-effectiveness acceptability frontier in these
PharmacoI{�9nomics, 9: 1 13-120.
situations. An example of this applied to the model you have just constructed
'1
Kind, P., Hardm<l:!1 G. and Macran, S. ( 1999) UK population norms for EQ-5D, University of
'
can be downloaded from the book website. York, Centre"for �ealth Economics, Discussion Paper 172.

I
National Institute f6r Clinical Excellence (NICE) (2004) Guide to the methods of technology
asseSS/Hent. London, NICE.
References
NHS Executive (2004) Trustfinancial returns. Leeds, NHS Executive.
Black, W. C. ( 1990) 'The CE plane: a graphic representation of cost-effectiveness', Medical Serruys, P. W., Unger, F., Sousa, J. E., Jatine, A., Bonnier, H. J. R. M., Schonberger, J. P. A. M.
Decision Making, 10: 212-214. and et al. (2001) 'Comparison of coronary artery bypass surgery and stenting for
Briggs, A. H. (2000) <Handling uncertainty in cost-effectiveness models: I the treatment of multivessel disease', New England Journal of Medicine, 344:
PharmacoEconomics, 17: 479-500. i 1 ! 1 7-1124.
I
Briggs, A. and Fenn, P. ( 1998) 'Confidence intervals or surfaces? Uncertainty on the Stinnett, A. A. and Mullahy, J. ( l 998) 'Net health benefits: a new framework for the analysis
cost-effectiveness plane', Health Economics, 7: 723-740. ! of uncertainty in cost-effectiveness analysis', Medical Decision Making, 18 (Suppl):

I
Briggs, A. H., Goeree, R., Blackhouse, G. and O'Brien, B. J. (2002a) 'Probabilistic analysis of S68-S80.
cost-effectiveness models: choosing between treatment strategies for gastroesophageal Stinnett, A. A. and Paltiel, A. D. (1997) 'Estimating CE ratios under second-order
reflux disease', Medical Decision Making, 22: 290-308. uncertainty: the mean ratio versus the ratio of means', Medical Decision Making,
Briggs, A. H., Goeree, R., Blackhouse, G. and O'Brien, B. J. (2002b) 'Probabilistic analysis of 17; 483-489.
cost-effectiveness models: choosing between treatment strategies for gastroesophageal Tambour, M., Zethraeus, N. and Johannesson, M. (1998) 'A note on confidence intervals in
reflux disease', Medical Decision Making, 22: 290-308. cost-effectiveness analysis: International/ournal of Technology Assessment in Health
Briggs,A H., Mihaylova, B., Sculpher, M., Hall, A, Wolstenholme, J., Simoons, M., Ferrari, R., Care, 14: 467--471.
Remme, W. J., Bertrand, M. and Fox, K. (2006) 'The cost-effectiveness of perindopril in The EUROPA investigators (2003) 'Efficacy of perindopril in reduction of cardiovascular
reducing cardiovascular events in patients with stable coronary artery disease using events among patients with stable coronary artery disease: randomised, double-blind,
data from the EUROPA Study', Heart 92(SuppI 2);A36. placebo-controlled, multicentre trial (the EUROPA study)" Lancet, 362: 782-788.
Cantor, S. B. ( 1994) 'Cost-effectiveness analysis, extended dominance and ethics: a van Hout, B., A!, M. J., Gordon, G. S. and Rutten, F. F. (1994) 'Costs, effects and CJE-ratios
quantitative assessment', Medical Decision Melking, 14: 259-265. alongside a clinical trial', Health Economics, 3: 309-319.
Chancellor, J. v., Hill, A M., Sabin, C. A., Simpson, K. N. and Youle, M. ( 1997) 'Modelling Weinstein, M. C. and Fineberg, H. V. ( 1980) Clinical decision analysis. Philadelphia, PA,
the cost effectiveness oflamivudine/zidovudine combination therapy in HIV infection', WB Saunders Company.
PharmacoEconomics, 12: 54-66.
Claxton, K. (1999) 'The irrelevance of inference: a decision-making approach to the stochas­
tic evaluation of health care technologies', Journal ofI-Iealth Economics, 18: 341-364.
Claxton, K. and Posnett, J. (1996) '.An economic approach to clinical trial design and
research priority-setting: Health Economics, 5: 5 13-524.
Department of Health (2004) Prescription cost analysis: England 2003, London, Department
of Health.
Drummond, M. F., Sculpher, M. J., O'Brien, B., Stoddart, G. L. and Torrance, G. W. (2005)
Methods for the economic evaluation of health care programmes, 3rd edn. Oxford,
Oxford University Press.
Fenwick, E., Claxton, K. and Sculpher, M. (2001) 'Representing uncertainty: the role of
cost-effectiveness acceptability curves', Health Economics, 10: 779-787.
Goeree, R., O'Brien, B" Hunt, R., Blackhouse, G., Willan, A and Watson, J. ( 1999)
'Economic evaluation of long-term management strategies for erosive oesophagitis',
PharmacoEconomics, 16: 679-697.
Joint FormularyCommittee (2004) British national formulary 48. London, British Medical
Association and Royal Pharmaceutical Society of Great Britain.
Joint Formulary Committee (2005) British national formulary 49. London, British Medical
Association and Royal Pharmaceutical Society of Great Britain.
 Chapter 6

Decision making, uncertainty

and the value of information

In this chapter we discuss how the results of probabilistic decision modelling

should be interpreted and how decisions should be made in the light of the
type of analysis described in the previous chapters. We identify two decisions
which must be made: (i) should a technology be adopted on the basis of existing
evidence, and (ii) whether further evidence is required to support this decision
in the future. The first decision is dealt with in the following section. Measures
of value of information, which can inform the second decision are covered
subsequently. The role of this type of analysis in identifying research priorities,
including the results of two recent UK pilot studies are then presented,
followed by exercises to take you through this type of analysis.

6. 1 . Decision makil1Jg with uncertainty

In this section we discuss how the results of probabilistic decision modelling
should be interpreted and how decisions should be made in the light of the
type of analysis described in the previous chapters. In order to do so we must
first identify the decisions faced by health care systems and then consider what
type of decision rules will be consistent with both the objectives and the
constraint of the health care system.

6. 1 . 1 . What are the decisions?

There are two conceptually distinct but simultaneous decisions that must be
made within any health care system. Firstly, should a technology be adopted or
reimbursed given the existing evidence and the current uncertainty surround­
ing outcomes and resource use? In particular, which of the range of possible
strategies, some of which may use the technology in different ways (e.g. at
different points in the sequence of patient management and have different
starting and discontinuation criteria), should be adopted? This question must
also be addressed for the range of possible patient groups based on character­
istics that may be related to expected costs and outcome (e.g. baseline and
competing risks, age, as well as those which may be related to relative effects of
the interventions).

I
J
1 66 I DECISION MAKING, U N C E RTAINTY AND THE VALUE OF I N FORMATION
DECISION MAKING WITH U N C E RTAINTY I 1 67

Secondly, is additional evidence required to support these adoption or reim­ research priorities for the UK NHS and informing the National Institute
bursement decisions? As well as deciding whether more evidence is required in for Clinical Excellence (NICE) in making research recommendations (Claxton
general, it is also important to know what type of evidence (for particular et al. 2004; Claxton et al. 2005a).
uncertain parameters) would be most valuable; what design of studies would The application of these methods requires three core tasks to be completed:
be most useful in generating the type of evidence required, and how much of (i) the construction of a decision analytic model to represent the decision
this evidence is required. There is also a need to address what type and level of problem; (ii) a probabilistic analysis of this model to characterize the current
evidence may be required across the range of relevant patient groups. So, as decision uncertainty; and (iii) establishing the value of additional information.
well as addressing whether more evidence is required there is also a range of The first two tasks have been dealt with in previous chapters. In this chapter
decisions about the appropriate design of any subsequent research. we address how the results of probabilistic modelling should be interpreted by
By clearly distinguishing these two separate but related decisions in health decision makers and how to address the question of whether more evidence is
technology assessment, it is possible to identify analytic methods that can required.
inform the adoption decision, the decision to acquire more evidence and As we saw at the end of Chapter 5, the uncertainty surrounding the decision
inform research design in a way which is consistent with the objective and problem can be characterized by 'propagating' these distributions through the
constraints on health care provision. This is not possible using more traditional model using Monte Carlo simulation methods. The output of these simulations
evaluation approaches based on hypothesis testing and levels of statistical provides the joint distribution of expected costs and outcomes for each strategy
significance which combine these separate decisions and consequently fail to being compared. The uncertainty surrounding the cost-effectiveness of a tech­
provide an adequate guide to either. nology, for a range of thresholds for cost-effectiveness, can be represented as a
Requi rements for decision making cost-effectiveness acceptability curve (CEAC). Figure 6.1 illustrates an example
of a CEAC for a simple model of zanamivir for the treatment of influenza.
These policy decisions indicate a number of requirements that an analytic
framework must meet. These include a clear specification of the objectives
and constraints of health care provision; a means of structuring the decision
1 .0 r----.___..
problem; a characterization of the uncertainty surrounding the decision; and
0.9 Zanamivir
a means of interpreting the results of the analysis so that decisions about __

-...... Standard Care

adoption and additional evidence can be made in a way which is consistent 0.8
with both the objectives and constraints on health care provision.
� 0.7
�
1i5 0.6
6. 1 .2. A framework for ana lysis
i;
Bayesian decision theory and value of information analysis provides an 8 0.5
Z.
analytic framework which can address whether a technology should be
'3 0.4
'"
adopted based on current evidence and whether more evidence is required to .0

support this decision in the future (Claxton et al. 2002). These methods have £ 0.3
firm foundations in statistical decision theory (Raiffa and Schlaifer 1959; Pratt 0.2
et al. 1995) and have been successfully used in other areas of research, such as
0.1
engineering and environmental risk analysis (Howard 1966; Thompson and leER " £51 .682
Evans 1997). More recently these methods have been extended to setting
0 0 0 0
priorities in the evaluation of health care technologies (Claxton and Posnett s g
o
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0,
c5 c5 c5 c5 c5 c5 c5 c5 c5
;;; '" ,.., '" 0
1996; Claxton 1999a; Claxton et al. 2002). In addition, they have been usefully '" '" '" '" � '" gj '"
co " co 0
;;;
applied to a number of different health technologies (Yokota and Thompson Threshold for cost-effectiveness
2004; Ginnelly et al. 2005), and have been piloted as a means of setting Fig. 6.1 Cost-effectiveness acceptability curve.
 T
1
1 68 I DECISION MAKING, UNCERTAINTY AND T H E VALUE OF I N FORMATION i DECISION MAKING WITH UNCERTAINTY I 1 69

This is a probabilistic reanalysis of the model contained in the independent

Technology Assessment Report, which was submitted as part of the NICE
appraisal of zanamivir in 2000 (BurIes et al. 2000; NICE 2000). In this example,
the probability that the intervention is cost�effective increases as the willingness
to pay for additional health (quality-adjusted life-years) or the threshold for
I If the objective underlying health technology assessment is to make
decisions that are consistent with maximizing health gains from available
resources, then decisions should be based on expected cost-effectiveness (net­
benefit) given the existing information. This is irrespective of whether any
differences are regarded as statistically significant or fall outside a Bayesian
cost-effectiveness increases. The incremental cost-effectiveness ratio, based on range of equivalence. The traditional rules of inference impose unnecessary
expected incremental costs and incremental QALYs, is £5 1 682. Thus, if the costs on individual patients and on the population of current and future
threshold for cost-effectiveness was just greater than £51 682, then the inter­ patients and are irrelevant to the choice between alternative technologies
vention could be regarded as cost�effective, even though the probability that it (Claxton 1999b).
is cost-effective is less than 0.5 (0.472). This is because the distributiou of the I
This is because one of the (mutually exclusive) alternatives must be chosen
additional net benefits is positively skewed, with a mean greater than its
median value. The decision uncertainty can be represented by including a iI and this decision cannot be deferred. For example, by rejecting the new tech­
nology we are selecting current practice and vice versa. The opportunity cost
cost-effectiveness frontier (Fenwick etal. 200 1 ) . As discussed in Chapter 5, the I of failing to make the correct decision based on expectation is symmetrical
frontier indicates the probability that the alternative with the highest net­ I and the historical accident that dictates which of the alternatives is regarded as

I
benefit will be cost-effective. The decision uncertainty or the error probability (current practice' is irrelevant. No one would ever suggest that a new treatment
is then 1 minus the value of the frontier. should be adopted when its mean net-benefit is less than current practice but,
The question is how should this be interpreted by decision makers and how in terms of opportunity cost, this is precisely what is implied by the arbitrary
should the decision be made under uncertainty? For example, if society was rules of inference.
willing to pay £60 000 per QALY gained then the probability that zanamivir is Of course, the measure of net-benefit is based on a particular objective
cost-effective is only 0.52 and the error probability is therefore 0.48. In these (maximize health outcomes) or social welfare function which may be judged
circumstances, what decision should be made about zanamivir? inappropriate, and consequently, this measure may be regarded as incomplete.
However, if there are equity issues that need to be incorporated, they can be
6. 1 . 3. The i rrelevance of inference made explicit with appropriate adjustments to the measure of outcome.
If we base decisions on the traditional rules of inference (either frequentist Similarly, if there is particular concern for safety and rare but catastrophic
or Bayesian) then we would conclude that the apparent cost-effectiveness events, then these undesirable outcomes should be given appropriate weight
of zanamivir is not statistically significant or falls within a Bayesian range in the calculation of expected net benefits. If decision makers wish to adopt
of equivalence; we cannot reject the null hypothesis and the result is indeter­ some type of voting rule, then the decision may focus on the median rather
minate. In practice this would lead to the rejection of zan amivir, despite than the mean net benefits, and the treatment which benefits the greatest
the fact that it is the alternative with the highest probability of being number could be adopted. Attitudes to risk can be incorporated in the measures
cost-effective and, more importantly, has the highest expected net-benefit of outcome, and if we wish to incorporate the fact that some individuals'
on the basis of the information currently available. If we fail to adopt an inter­ preferences violate the axioms of expected utility theory, then prospect theory
vention simply because the differences in net-benefit are not regarded or some notion of regret can be used in the measure of net benefits. It is not
as statistically significant or the error probability is 'too high' then we will necessary to take a position on the appropriate social welfare function, different
impose opportunity costs on patients who could benefit from it. These definitions of need, the normative content of expected utility theory, or the
opportunity costs are the forgone expected net benefits resulting from selecting best way to incorporate important and legitimate equity concerns to accept
a technology with a lower expected net-benefit. For an individual patient in the irrelevance of inference. It is worth noting, however, that these consider­
the zanamivir example, these forgone expected net benefits are £3.07 or ations will all have implications for the measures of outcome and the design
0.000052 QALYs at a cost-effectiveness threshold of £60 000. For the UK of studies ) not just their interpretation. Whatever view is taken, confidence
population of current and future patients they could be valued at £2 748 276 intervals) P-values and levels of significance still remain entirely irrelevant to
or 45.8 QALYs the decision.

�...
170 I DECISION MAKING, U N C E RTAINTY AND THE VALUE OF I N FORMATION EXPECTED VALUE OF PERFECT I N FORMATION 1 171

6.2. Expected value of perfect information

'"
Although decisions should be based on expected cost-effectiveness given the
existing information, this does not mean that adoption decisions can simply "
be based on little, or poor quality, evidence, as long as the decision to conduct
further research to support adoption (or rejection) is made simultaneously. (Q �
'" w
Decisions based on existing information will be uncertain, and there will oS
always be a chance that the 'wrong' decision will be made. That is, although we - �c:
"'
w
make the correct decision now based on our current estimate of expected net­ o _
w

!I
benefit, there is a chance that another alternative would have had higher net­ c:

benefit once our current uncertainties are resolved. If our decision based on , :$
'"
c:

i
current information turns out to be 'wrong', there will be costs in terms of '" E
I
"
�
health benefit and resources forgone. Therefore, the expected cost of uncer­ '" oS
I
tainty is determined jointly by the probability that a decision based on existing
information will be wrong and the consequences of a wrong decision. With 1
estimates of the probability of error and the opportunity costs of error we can '"
at![
0 0 0 0 0 I
0 0 0 0 0 is o o 0
calculate the expected cost of uncertainty or the expected opportunity loss 0 0 0 0 0 o is o
o
0<

surrounding the decisions. The expected costs of uncertainty can be interpreted '"
6 6
'"
6 6 6
'" g 6
'" g
'" '"
0 0
0< i;:I 0< '" [;j '" 0<
as the expected value of perfect information (EVP1), as perfect information can
SSOl lil!UnlJoddO OJ SSOl lil!unlJoddO C
eliminate the possibility of making the wrong decision. if the objective of the
health care system is to maximize gains in health outcome subject to a budget
constraint, then this is also the maximum that the health care system should
be willing to pay for additional evidence to inform this decision in the future,
and it places an upper bound on the value of conducting further research.
If there are other objectives of health care provision, such as equity, which can
be identified and valued, then these can be incorporated into the analysis and
the societal value of information derived.
This general idea is illustrated in Fig. 6.2 for the choice between two alterna­
tives. Three things determine the value of information. Firstly, how cost-effective
(or ineffective) the technology appears given current or prior information
(O'Hagan and Luce 2003), that is, the position of the expected (prior) incre­
mental net-benefit that is valued in terms of health in Fig.6.2. Secondly, the
uncertainty surrounding cost-effectiveness (or cost-ineffectiveness), that is, c:
o
the distribution of the prior mean incremental net-benefit. Thirdly, the slope ''8
c:
of the loss function which values the consequences of decision error based .i!
o
g6 g 8
�

on current information. In Fig. 6.2a the expected incremental net-benefit is is S

6 6 6
positive ) so the new technology appears to be cost-effective given existing o '" 0 "i
information. However, this decision is uncertain, and there is a probability of
� � w � '"

error represented by the tail area where incremental net-benefit is less than
SSOl lil!UnlJoddo <C (,) i
zero. If this is the way our current uncertainties resolve then it would have
1 72 I DECISION MAKING, U N C ERTAINTY A N D T H E VALUE OF I N FORMATION EXPECTED VALUE OF PERFECT I N FORMATION 1 1 73

been better to reject the technology and there will be opportunity losses in the these analytic solutions) which have been used in the earlier literature on
form of net-benefit forgone. These opportunity losses are represented by the value-of-information analysis (Raiffa and 5chlaifer 1959; Claxton and Posnett
loss function, which has a slope equal to the cost-effectiveness threshold and 1996). Each .of the three issues discussed in the previous section are repre­
simply converts the incremental net-benefit measured in health terms on the sented in the expression for EVPI:
X-axis into monetary terms on the Y-axis.
For example, in Fig. 6.2a it is possible that our decision based on expected
incremental net benefits will be wrong, as there is a probability that the incre­ where:
mental net-benefit will be less than zero. If it is, then the cost of this error or A. = cost-effectiveness threshold
the opportunity loss is the difference between the net-benefit we could have Do = 1 1)0 1/0"0
gained by making the right decision in these circumstances (reject the new 1)0 = prior mean incremental net health benefit
technology) and the net-benefit we get from the decision based on the expected 0"02 prior variance of the incremental net-benefit (1)0)
incremental net-benefit. This is represented by the height of the loss function L(Do) = unit normal loss integral for Do
and illustrates that there is a small chance that incremental net-benefit will be The unit normal loss integral is the integration of a loss function with slope
a lot less than zero (-5 QALY), in which case the opportunity loss will be very of 1 with a standard normal distribution (mean = 0 and 0"0 = I ). Therefore,
high (£250 000). However, there is a greater chance that the incremental net­ L(Do) represents probability of decision error when incremental net-benefit
benefit will be just less than zero (-1 QALY), in which case the opportunity has a mean of Do and 0"0 = 1. The probability of decision error is determined
loss will be lower (£50 000). The expected opportunity loss (EVPI) is essentially by: (i) the standardized distance (Do) of the prior mean incremental net benefits
taking a weighted mean of all these losses where the weights are the probability of the technology (1)0) from the point at which the decision based on current
of incurring each of these losses. information would switch (where 1)0 = 0), that is, how cost-effective the tech­
Figure 6.2b illustrates that where there is more uncertainty (greater variance nology appears to be; and (ii) the uncertainty surrounding 1)0 (Var(1)o) = all.
in incremental net-benefit) then the probability of error (tail area) will The slope of the loss function is the cost-effectiveness threshold (A.) and repre­
increase and expected opportunity loss and, therefore, EVPI will be higher. In sents the monetary value placed on opportunity losses when they are incurred.
Fig. 6.2c the technology appears to be more cost-effective (mean of the distri­ So, this expression for the EVPI reflects the determinants of the probability of
bution is further to the right), the probability of error falls and EVPI also falls; error and the consequences of error illustrated in Fig. 6.2.
the technology appears so cost-effective there is little uncertainty about its You will find this approach taken in earlier papers on value-of-information
adoption. In Fig. 6.2d the cost-effectiveness threshold is higher (the slope of analysis (Claxton and Posnett 1996, Claxton 1999a, Claxton et al. 2001). It
the loss function is greater), opportunity losses (in health terms) are valued will be appropriate if incremental net-benefit is normally distributed. This
more highly and this will tend to increase the EVPI. However, increasing the may be the case if all the evidence for costs and benefits comes from the same
cost-effectiveness threshold will also increase the expected incremental sample) in which case central limit theorem may assure normality. However)
net-benefit (the distribution will shift to the right) and the probability of error decision models combine evidence from a variety of sources and, as discussed
will fall, therefore, the EVPI may fall. Figure 6.2 also indicates that the shape of in Chapter 4, use a range of distributions assigned to parameters. The resulting
the distribution also matters. In this case) with positive expected incremental net-benefit from probabilistic analysis is a mixture of these distributions,
net-benefit, if it was negatively skewed but with the same variance, there through a model structure which is likely to be nonlinear and may contain
would be a higher chance of larger errors with greater opportunity loss and discontinuities generated by logical functions. In addition, there may be
higher EVPI. complex correlation structures generated between parameter estimates when
evidence is synthesized. In these circumstances there is no reason to believe
6.2 . 1 . EVPI using the normal distribution that net-benefit will be normally distributed. In fact, it is unlikely to have a
If incremental net-benefit can be regarded as normally distributed, then parametric distribution at all. Therefore, a nonparametric approach to estab­
simple analytic solutions can be used to establish the EVPI surrounding the lishing EVPI is required. This type of approach is outlined in the next section
decision. The discussion of Fig. 6.2 provides some intuition when considering and is now most commonly used in the evaluation of health technologies.
1 74 I DECISION MAKING, UNCERTAINTY A N D T H E VALUE OF I N FORMATION EXPECTED VALUE OF PERFECT IN FORMATION I 1 75

6. 2 . 1 . Nonparametric a pproach to EVPI Table 6.1 Calculating EVPI

We do not need to be restricted to analytic solutions that require assumptions Treatment Optimal choice Maximum net-benefit Opportunity loss
of normality because we can work out the EVPI directly from the simulated 11,. B
output from our model. \Vith current information, decisions must be made Iteration 1 9 12 B 12 0
before we know how the uncertainties will be resolved, that is, we must make a Iteration 2 12 A 12
10 2
decision now based on the expected net benefits of each of the alternatives.
Iteration 3 14 20 B 20 0
However, with perfect information, we could make our decisions once we
Iteration 4 11 10 A 11
know how the uncertainties in the model will resolve, so we could make differ­
Iteration 5 14 13 A
ent decisions for different resolut'ions of net-benefit. The EVPI is simply the . . . . " " . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
. . . . " . . . . . . . . . . . . . . . . . . . . . . . .

difference between the payoff (expected net-benefit) with perfect and current Expectation 12 13 1 3 .8 0.8
information (Pratt et al. 1995; Thompson and Graham 1996; Felli and Hazen
1998; Ades et al. 2004; Sculpher and Claxton 2005).
For example, ifthere are j alternative interventions, with unknown parameters
0, then given the existing evidence, the optimal decision is the intervention This approach is illustrated in Table 6.1 for two alternative treatments A and B.
that generates the maximum expected net-benefit: The table represents simulated output from five iterations generating a net�
benefit for each of the treatments. With current information, the best a decision
maxj EeNB(j, Ii), maker can do is to choose the alternative with the highest expected net-benefit
that is, choose j with the maximum net benefits over all the iterations from the (maxj Be NB(j, e)) which, in this case, is to choose treatment B with expected
simulation because each iteration represents a possible future realization of net-benefit of 13. With perfect information the decision maker could choose
the existing uncertainty (a possible value of Ii). With perfect information, the the alternative with the maximum net-benefit (columns 3 and 4) for each
decision maker would know how the uncertainties would resolve (which value resolution of uncertainty (maxj NB(j, e)), that is, choose B for iteration 1 ;
Ii will take) before making a decision and could select the intervention that A for iteration 2 and B for iteration 3 etc. However) we do not know in advance
maximizes the net-benefit given a particular value of e: which of these possibilities will turn out to be true, so the expected net-benefit
with perfect information is simply the expectation of the maximum net-benefit
maxj NB( j, e). (£13.80). The EVPl is then simply the difference between the expected net­
However, the true values of 0 are unknown (we don't know in advance benefit with perfect information and the expected net-benefit with current
which value e will take). Therefore, the expected value of a decision taken with information (£13.80 - £13.00 = £0.80). This is entirely equivalent to taking the
perfect information is found by averaging the maximum net-benefit over the expectation of the opportunity losses in column 5, where opportunity loss is
joint distribution of e: simply the difference between the net-benefit of optimal choice for that iteration
and the alternative that would be chosen based on current information. This
confirms the earlier discussion that EVPI is also the expected opportunity loss
In other words, first calculate the maximum net-benefit for each iteration or the expected costs of the uncertainty surrounding the decision. It should be
from the simulation (for a particular value of e), then take the mean over noted that in this example, although alternative B has the highest expected
these maximum net benefits (over the possible values of Ii). The expected net-benefit, it also has a lower probability of being cost-effective (P= 0.4) than
value of perfect information for an individual patient is simply the difference alternative A ( P = 0.6), demonstrating the importance of cost-effectiveness
between the expected value of the decision made with perfect information frontiers as discussed in Chapter 5.
about the uncertain parameters 8, and the decision made on the basis of exist­ This provides the EVPI surrounding the decision as a whole for each time
ing evidence: this decision is made (for an individual patient or individual patient episode).
However, once information is generated to inform the decision for an individual
patient or patient episode then it is also available to inform the management
176 I DECISION MAKING, UNCERTAINTY A N D THE VALUE OF I N FORMATION
�
.1
EXPECTED VALUE OF PERFECT I N FORMATION 1 1 77

of all other current and future patients (it has public good characteristics
and is nonrival). Therefore, it is important that EVPI is expressed for the total
population of patients who stand to benefit from additional information I The relationship between the EVPI and the cost-effectiveness threshold in
Fig. 6.3 has an intuitive interpretation. When the threshold for cost-effectiveness
is low, the technology is not expected to be cost-effective and additional infor­

I
over the expected lifetime of the technology. This requires some assessment of mation is unlikely to change that decision (EVPI is low). Thus, current
the effective lifetime of the technology, the period over which information evidence can be regarded as sufficient to support the decision to reject the
about the decision will be useful ( T) , and estimates of incidence over this technology. In these circumstances the EVPI increases with the threshold
period (I,). because the decision uncertainty (probability of error) increases and the

EVPI for the population = EVPI . I,:,, 2. . . ,T I, 1(l+r)'.

! consequences of decision error (opportunity loss) are valued more highly.
I
I
Conversely, when the threshold is higher than the incremental cost-effectiveness
ratio (ICER), the intervention is expected to be cost-effective and this decision
The EVPI associated with future patients is discounted at rate r to provide I

I
the total EVPI for the population of current and future patients (see Chapter 7 is less likely to be changed by further research as the threshold is increased.
for further discussion of this issue). The decision uncertainty falls because the technology appears to be increasingly
Figure 6.3 illustrates the population EVPI for the example used in Fig. 6.1, cost-effective as the threshold is increased (probability of error falls, tending
If this population EVPI exceeds the expected costs of additional research, then ,
to reduce the EVPI), but the consequences of error are valued more highly
it is potentially cost-effective to conduct further research. For example, if (tending to increase the EVPI). In this example the reduction in decision
additional research is expected to cost £lm then additional research is potentially uncertainty offsets the increased value of the consequences of error. However,
cost-effective if the threshold is greater than £31 000 per QALY. At lower values the EVPl will ultimately increase with very high values of the threshold
of the threshold, the new technology should be rejected based on current because the decision uncertainty falls at a declining rate, but the value of
evidence and further research is required to support this decision because the opportunity losses increases at a constant rate and will ultimately offset this.
returns from further research cannot offset the costs, In this particular example, the population EVPI reaches a maximum when the
threshold is equal to the expected incremental cost-effectiveness ratio of this
technology. In other words, in this case the EVPl reaches a maximum when we
are most uncertain about whether to adopt or reject the technology based on
£5,000,000 existing evidence.
£4,500,000 The EVPI curve for two alternatives is easy to interpret. However) most deci­
£4,000,000 sion problems involve more than two alternatives. The principles of calculating
£3,500,000 EVPI remain the same (see the section on EVPl for strategies later in this
a: chapter) but the EVPl curve can take a variety of shapes depending on
iii £3,000,000 whether the alternatives being considered are cost-effective at some value of
5 £2,500,000
'iii the threshold (in which case there will be a number of peaks or changes in the
§. £2,000,000
a.
slope of the EVPl curve at threshold values equal to the lCER of each of the
alternatives), or if some of the alternatives are dominated or extendedly domi­
£1 ,500,000
- nated (the peak or discontinuity will be in negative threshold space: we would
£1 ,000,000 - - -- -- -
-- - - - -- -- - -- - - - - - - - - -- -
only wish to adopt the alternative if we were willing to pay to reduce health
£500,000 outcome).
£O ��+-��--�---- It should be clear from this discussion of EVPl that the value of further
� § § § § § § § § § 8 research will depend on both the uncertainty surrounding estimates of cost

3 g Threshold
5 * for cost�effectiveness
g §- t]' gf gi � and effect but also on how cost-effective or cost-ineffective a technOlogy is
expected to be, given existing evidence and the size of the patient population
Fig. 6.3 Population expected value of perfect information curve. that could benefit from additional research, One implication is that it is
178 I QEClSION MAKING, U NCERTAINTY AN D T H E VALUE OF I N FORMATION
EXPECTED VALUE OF PERFECT I N FORMATION FOR PARAM ETERS 1 179

perfectly possible that the value of additional evidence about a new technol­ decision uncertainty and no value of information because there is very little
ogy, which is substantially cost-effective based on existing evidence, will be doubt t,hat this is not cost-effective, despite the uncertainty in costs and outcomes.
very low even ifthere is uncertainty surrounding the parameters. That is, there Of coursei·there is no such thing as perfect information but it does place an
may be uncertainty in cost and outcomes, but the decision uncertainty - and upper bound on the returns to research. The population EVPI can be used as a
therefore the EVPI - may still be low. In these circumstances, the technology first hurdle to identify research that is potentially cost-effective and rule out
should be adopted and no further research is required to support this research which will not be worthwhile. However, we should be cautious when
decision. comparing EVPI across technologies as in Fig. 6.4. This is because the cost of
This discussion of value-of-information analysis suggests that there will be research may differ and the marginal benefits of actllal research may also
different values of information for different technologies applied to different differ across these technologies.
patient populations and for the same technology in different circumstances, i The framework of analysis, which has been outlined above allows us to take
such as for different indications, patient populations and in different health 1· a rational approach to decision making with uncertainty. By separating the
care systems with different cost-effectiveness thresholds. This is illustrated in I two questions of whether to adopt a technology given existing information
Fig. 6.4 which shows the EVPI associated with five stylized examples of health
care technologies, which were appraised by NICE between 1999 and 2000. I from the question of whether more information is needed to inform this
choice in the future, we can get sensible answers to both. As EVPI is so easy to
establish directly from the simulation once a probabilistic model has been
It shows that we should demand different amounts of evidence for different
technologies and that there can be no fixed rules for regulatory or reimburse­ developed, there is no reason why decision EVPI should not be a routine part
ment decisions (Claxton 1999b). How much evidence is required to support of any evaluation.
a decision is essentially an empirical question. Figure 6.4 also illustrates the
difference between parameter and decision uncertainty. For example, the 6.3. Expected value of perfect information
prophylactic extraction of wisdom teeth (WT) has the most parameter uncer­ for parameters
tainty (there was very little and poor quality evidence). However, there is no
The EVPI surrounding the decision problem can indicate whether further
research is potentially worthwhile. However, it would be useful to have an
£35,000,000
indication of what type of additional evidence would be most valuable.
In Chapter 5 we introduced ANCOVA as a measure of the importance of
£30,000,000
--Ritalin a parameter. As was noted, however, these rely on linearity, are not directly
....,r,.- Zanamivir
related to decision uncertainty, and cannot provide a measure of value that

i
£25,000,000 -Y.- RillJzole
can be compared with the costs of further investigation. The value of reducing
� --Orlistat
£20,000,000 the uncertainty surrounding particular parameters in the decision model can
be established using a similar approach to the EVPI for the decision problem
1l. £1 5,000,000 as a whole. This type of analysis can be used to fOCllS research on the type of
&. evidence that will be most important by identifying those parameters for
£1 0,000,000
which more precise estimates would be most valuable. In some circumstances,
£5,000,000
tllis will indicate which endpoints should be included in further experimental
research. In other circumstances, it may focus research on getting more precise
0 0 0 0 0 0 0
estimates of particular parameters that may not necessarily require experi­
0 0 0 0 0 0 0
0 0 0 0 0 0 0 mental design and may be provided relatively quickly. The analysis of the
cS cO cS cO cS cO cS
O'i O'i
en
�
'" '"
'"
'" '" '" '" value of information associated with each of the model inputs is, in principle,
Threshold for cost-effectiveness conducted in a very similar way to the EVPI for the decision as a whole (Pratt
Fig. 6.4 Population expected value of perfect information for NICE case studies. et al. 1995; Thompson and Evans 1997; Ades et al. 2004).
 1
I EXPECTED VALUE OF PERFECT I N FORMATION FOR PARAMETERS j 181
180 I DECISION MAKING. U N C E RTAINTY AND T H E VALUE OF I N FORMATION

The expected value of perfect information for a parameter(s) (EVPPI) is It should be apparent that although this is conceptually very similar to the
simply the difference between the expected value with perfect and current initially discussed calculation for decision EVPI, it is also more computation­
information about the parameter(s). The expected value with perfect infor­ ally intensive...This is because we have an inner and outer loop of expectation:
mation is found by taking the maximum expected net-benefit given perfect first we must run-the simulation for parameters IfI but with a particular value
information only about the parameter(s) of interest (calculating expected of !p (an inner loop), then we must sample a new value of !p (an outer loop)
net benefits over all the other uncertain parameters in the model) and then and rerun the simulation. This must be repeated until we have sampled
calculating the mean of all the possible values of the parameter of interest. The sufficiently from the distribution (or joint distribution) of!p.
EVPI for the parameter is simply the difference between the expected This suggests that when choosing how many iterations to run on the inner
net-benefit with perfect information and the expected value with current and outer loops, as you will do in the exercise, there is no reason why we
information (the same as for decision EVPI). However, this does require should choose an equal number of iterations. Although there can be no clear
substantial additional computation for models where the relationship between rules, it seems reasonable that we will need more iterations when there are
the parameters and the expected cost and outcomes is not linear, for example more uncertain parameters in order to sample from the whole of the joint
in Markov models. distribution. For example, if <p is only one parameter and lJI contains, say,
More formally, suppose we were interested in the value of perfect information 20 parameters, then it makes more sense to run more iterations on the inner
about a parameter or a subset (!p) of all the uncertain parameters O. With than the outer loop. However, there are many circumstances where we will
perfect information we would know how !p would resolve (which value it will Want to calculate the EVPPI for groups of parameters together. In this case
take) and the expected net-benefit of a decision would be found by choosing there may be ten parameters in IJf and ten in !p. In this case an equal number of
the alternative with the maximum expected net-benefit when those expected simulations in the inner and outer loop would Seem more reasonable.
net benefits are averaged over the remaining uncertain parameters in the One implication of this discussion is that, just as with decision uncertainty,
model ( 1Jf). In other words, we take a value of !p and then calculate expected information about a parameter is only valuable if it is possible for it to change
net benefits over the remaining uncertainties ( 1Jf) and choose the alternative the decision (which alternative has the maximum expected net-benefit).
j that has the maximum expected net-benefit: Therefore, information is only valuable insofar as uncertain parameters do
not resolve at their expected value. This suggests that those parameters that
are more uncertain (can resolve at extreme values) and are closely related to
differences in net-benefit between the alternatives, are likely to have a greater
As before, however, the true values of !p are unknown (we do not know what value of information associated with them. However) parameters that may be
value !p will take), therefore, the expected value of a decision taken with very uncertain but have little effect on the differences in net-benefit (e.g.
perfect information is found by averaging the maximum expected net benefits changes the cost or the effects of each alternative to an equal extent or in way
over the distribution of !p: that offsets each other) may have a very low or even zero EVPPI. This, as well
as the issues raised in the next two paragraphs, sometimes makes it very difficult
to 'guess' which parameters or groups of parameters will have high EVPPI
The expected value with current information is the same as before because associated with them in anything but the simplest of models. Of course, COunter­
!p u 1Jf= 0): intuitive results should prompt careful scrutiny of the model and the analysis
to understand the results. However, we should take no comfort from what
appear to be intuitive results, because if it really were that easy then there
would be no point in building a model.
So the EVPPI for the parameter or group of parameters !p is simply
the difference between the expected value of a decision made with perfect 6.3. 1 . Groups of parameters
information and the expected value with current information: It is very important to recognise that the EVPPI for the individual parameters
do not sum to the decision EVPI. Equally, the EVPPI for a group of parameters

'lI.
 1
182 I DECISION MAKING, UNCERTAINTY AND T H E VALUE OF I N FORMATION
EXPECTED VALUE OF PERFECT I N FORMATION FOR PARAMETERS 1 1 83

is not the sum of the individual parameter EVPPIs. This is because the indi­
multilinear relationship between the parameters and net-benefit. If the model
vidual parameters when considered in isolation may not resolve in such a way as
is linear in '1'; and <p and 'I' are independent then:
to have a sufficient impact on differences in net-benefit to change the decision.
However, when a group of parameters is considered together, then the joint

I\
effect of each resolving at an extreme value may have sufficient impact on
differences in net-benefit to change the decision and generate a value of infor­ The expected net-benefit for a particular value of <p can be calculated based
mation. It is perfectly possible that the EVPPI for all the individual parameters on the mean values of the other parameters '1', and the inner loop of simula­
may be zero, but for the decision and for groups of parameters it may be tion is unnecessary. This will also be true when net-benefit is multilinear in 1fI
substantial. and where there are no correlations between the parameters in lfI and when
This discussion indicates that it is best to approach EVPPI analysis by <p and 'I' are also independent. Multilinearity allows the net-benefit function to
initially conducting EVPPI analysis 011 a small number of groups of parameters. I contain products of (independent) parameters. This 'short cut' may, therefore,
The groups should be chosen to match the type of research that would be [
,
be used for many standard decision tree models, with branching path proba­
conducted. For example, it would be reasonable to group all the parameters bilities and independent parameters (Felli and Hazen 1998; Ades et al. 2004).
that may be vulnerable to selection bias and would require randomized Reduction in opportunity loss
design. It would also make sense to group all those parameters associated with
In Table 6.1 we demonstrated that EVPI was also the expected opportunity
natural history or baseline risk, which may be informed by observational or
loss or the expected cost of uncertainty. It is natural to interpret EVPPI as the
epidemiological studies, and group all those associated with quality of life that reduction in opportunity loss or the reduction in the cost of uncertainty if we
could be informed by a survey. Then, if there is significant EVPPI associated had perfect information about the parameters of interest. In other words it is
with a group of parameters, it can then be broken down by conducting EVPPI the difference between the decision EVPI and the EVPI for the decision when
on smaller groups or individual parameters to identify where the value of
we have perfect information about the parameters of interest. This is true as
information lies. This is much more efficient than starting with EVPPI for all long as we calculate the EVPI for the decision over all the ways the parameter
parameters and then combining them in different ways. More importantly it is can resolve and not simply its expected value. It can easily be shown that the
much more relevant to informing decisions about research priorities. terms in these expressions cancel and this approach does lead (in the limit and
Another important reason to group parameters is if they are correlated. The without spurious correlation) to the EVPPI outlined above. It is consistent
EVPPI for the groups of correlated parameters will preserve the correlation with the well-known identity of maximizing expected value (net-benefit) and
structure. However, conducting EVPPI in the way described above on only
minimizing opportunity loss. Although this approach has been taken in the
one of the correlated parameters will break this correlation. The EVPPI could past) the approach set out above is recommended as it is easier to implement
either be under-or overestimated depending on the type correlation between and avoids unnecessary noise and spurious correlation in the simulation.
the parameters, the direction of the relationship of each parameter to differ­
ences in net-benefit, and the decision based on current information. Where 6.3.3. EVPI for multiple treatment options
there is correlation between parameters it is also possible for the EVPPI
The initially discussed requirements for decision making included the estimation
on one of the correlated parameters to be greater than the EVPI for the of the expected costs and health outcomes for each of the possible strategies
group of correlated parameters and in some circumstances greater than the
that could be adopted, including the range of possible strategies (not just
decision EVPI. those directly compared in current trial evidence), and for a range of clinical
policies (how a technology could be used). The inclusion of all possible alter­
6.3.2. EVPPI for linear models
natives or strategies within the analysis is necessary to inform the adoption
The approach to EVPPI outlined above provides a general solution for non­ decision but it is also crucial for the estimation of the value of information.
linear models. However, it is also more computationally intensive because Even when an alternative is never regarded as cost-effective over a range of
it requires an inner and outer loop of simulation. The computational possible threshold values, due to dominance or extended dominance, there
requirements can be somewhat simplified if the model has either a linear or may be a chance that it could be cost-effective. Indeed, an alternative which is
1 84 I DECISJON MAKING, UNCERTAINTY A N D T H E VALUE OF I N FORMATION EXPECTED VALUE OF PERFECT INFORMATION FOR PARAMETERS I 1 85

regarded as dominated or extendedly dominated may have a higher chance of The difference between the EVPI for the full decision problem (all strategies)
being cost-effective at a particular threshold value than an alternative which is and the EVPI when a strategy is excluded from the analysis can indicate the
not dominated (Claxton and Thompson 200 1 ) . In these circumstances, bias in EVPIthat is introduced by excluding the strategy. For example, excluding
excluding the alternative from the analysis may not change the decision about C from the analysis would lead to the EVPI being underestimated by £0.60.
which strategy should be regarded as cost-effective based on current informa­ However, this difference in EVPI should not be interpreted as the EVPI associ­
tion, but it will have a substantial impact on the EVPI. Indeed, the exclusion of ated with the strategy. This is for two reasons: (i) this difference conflates the
any strategy from the analysis will reduce the decision EVPI if there is a non­ value of the strategy itself (improvement in net-benefit) and its contribution
zero probability that it could be cost-effective, because there may be particular to decision uncertainty; and (ii) many model parameters are relevant to more
resolutions of the uncertain parameters where the excluded strategy would than one strategy. The only meaningful measure is the value of information
provide the maximum net-benefit.
The principles of calculating EVPI for multiple alternatives remains the
1,
I
associated with those parameters that are specific to the strategy, such as the
relative effectiveness of a technology that is specific to the strategy.

II
same as was illustrated in Table 6.1 for the choice between two alternatives It is possible to use value-of-information analysis to decide whether a strategy
A and B. Table 6.2 illustrates how to calculate the EVPI when alternatives should be included as relevant or can be safely excluded from the analysis. For
C and D are also available. example, in Table 6.2, if after 1 0 000 iterations, D was still never optimal (the
As before, the table represents simulated output from five iterations} gener­ probability it is cost-effective is effectively zero) then it could be safely
ating net benefits for each of the now four strategies. With current informa­ excluded from further analysis because its exclusion has no effect on the EVPI.
tion, the alternative with the highest expected net-benefit remains B with an In this sense, we can use value of information to define relevant comparators
expected net-benefit of 13. With perfect information, the decision maker can and exclude others, that is, any strategy is potentially a relevant comparator if
choose the alternative with the maximum net-benefit, which will now be alter­ it has some chance of being cost-effective and therefore contributes to the
native C for iterations J and 4. Therefore, the expected net-benefit with EVPI. However, to decide whether particular strategies that contribute to
perfect information (the expectation of the maximum net benefits) has EVPI should be compared in subsequent research requires measures of value­
increased from £ 1 3 .80 to £14.40 and the EVPI has increased from £0.80 to of-sample information; this is discussed in Chapter 8 (Claxton and Thompson
£1 .40. So, although alternative C is not cost-effective based on current infor­ 2001). lt should be noted that what is and what is not relevant will depend on
mation, there is a chance that it will be. Excluding C will not effect the deci­ the particular value of the cost-effectiveness threshold.
sion to adopt B but it may affect the decision to conduct further research.
Alternative D, on the other hand) is never the optimal choice in these five iter­ 6.3.4. An exa mple of EVPPI
ations (there is no chance that it will be cost-effective) and its inclusion or Figure 6.5 illustrates the EVPPls associated with the decision EVPI in Fig. 6.3
exclusion has no effect on the EVPI. at a threshold of £40 000 per QALY. In this example, the EVPPI associated
with reduction in symptom days on zanamivir is relatively high. A more
precise estimate of the reduction in symptom days will require experimental
Table 6.2 Calculating EVPI for multiple alternatives
design to avoid selection bias and suggests that further randomized trials may
A B C 0 Optimal Maximum Opportunity be worthwhile. However, other parameters with lower EVPPI such as the qual­
choice net-benefit loss ity of life associated with influenza symptoms (QALY/day); the probability of
Iteration 1 9 12 14 8 C 14 2 hospitalization with standard care (P(hospitalization)); and the probability
Iteration 2 12 10 8 7 A 12 2 that a patient presenting with influenza-like symptoms does in fact have
Iteration 3 14 20 15 12 B 20 0 influenza (P(disease)) may not require experimental research and may be
9
important if the costs of further investigation (resources and delay) are low.
Iteration 4 11 10 12 C 12 2
Other parameters such as the probability of complications requiring antibi­
Iteration 5 14 10 A 14
. . . . . . . . . . . ... . . . . . . . .13 11
. , . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . otics with standard care (P(complications)) have negligible EVPPI and are
Expectation 12 13 12 9 14.4 1 .4
relatively unimportant. The impact of zanamivir on complications requiring
 �.
I

I
186 I DECISION MAKING, U N C E RTAINTY AN D T H E VALUE OF I N FORMATION
EXPECTED VALUE OF PERFECT I N FORMATION FOR PARAM ETERS I 1 87

£4,500,000
£4,000,000
£3,500,ob6
!
-+-p(complications)
£4,000,000
;" -p(hospltallsation)

£3,500,000 £3,000,000 •-.lrReduction in symptom days

a:: £3,000,000 ->*-Log odds ratio (complications)

a.
>
£2,500,000 -"'-Log odds ratio (hospitalisation)

l �. /i�\
� £2,500,000
.Q
§ £2,000.000
� £2,000,000 �
& £1 ,500,000
£1 ,000,000
� :::::::::: '
.
ILl ,
!
,! £500,000
£500,000
"
£0 £0 �..l-.�. - ��
PX-�i!(--x""':'i�;;;-
" - -;;; 'JJ:�*" ,'*

£0 8 g 0 0
0 0 0g 8
0 g
0 8
0 80 8 8 8
0
0 0 0 0 0 0 0 � o o
� � Q � � ffi � � � 0
W
Threshold for cost·effectiveness
Fig. 6.6 Expected value of perfect information for parameters.

Fig. 6.5 Expected value of perfect information for parameters.

In general, those parameters that are more closely related to differences in

antibiotics (log odds ratio (complications» has negligible EVPPI but the costs will be more important at lower thresholds; those that are more closely
impact on hospitalization (log odds ratio (hospitalization» has positive related to differences in effect will be more important at higher threshold
EVPPI. This suggests that if a trial is conducted to provide a more precise esti­ values. This demonstrates that the value of research and the type of research
mate of the reduction in symptom days then also measuring the impact on required depends on the threshold for cost-effectiveness. In other words, not
hospitalizations could be worthwhile. It is clear in Fig. 6.5 that the EVPPIs will only is the value of further research essentially an economic question, but the
not necessarily sum to the overall decision EVPI for the reasons outlined type and design of future research is an economic question too.
above. The earlier discussion of decision EVPI suggested that what might be
Figure 6.6 illustrates the relationship between EVPPI for the model parame­ regarded as sufficient evidence is an empirical question, which will differ
ters and the cost-effectiveness threshold. The EVPPIs for the model parame­ across technologies and will depend on the cost-effectiveness threshold used
ters are related to the cost-effectiveness threshold in the same way as decision in decision making) as well as the amount and quality of evidence available.
EVPI (illustrated in Fig. 6.3). In this example, just as with decision EVPI, the This was illustrated using five stylised examples of health care technologies
EVPPIs reach a maximum when the cost-effectiveness threshold is equal to the which were considered by NICE in Fig. 6.4. Figure 6,7 illustrates the EVPPI for
leER for zanamivir) that is, when the prior mean incremental net-benefit is these stylised examples, where parameters have been grouped into effect, utili­
zero. At low cost-effectiveness thresholds « £25 000) the EVPPI is negligible ties and cost. The EVPI associated with effect, utilities and cost differs across
because zanamivir is not cost-effective at these values and standard care is these technologies. For example, it is additional evidence about the relation­
chosen based on current information. In these circumstances) even if individual ship between clinical measures of effect and quality of life which is most valu­
model parameters resolve at their extreme values, they are unlikely to change this able to inform the decisions to adopt Ritalin® for the treatment of ADHD and
decision. Clearly the absolute value of EVPPI changes with the threshold just orlistat for the treatment of obesity. However, it is additional evidence about
as decision EVPI changes. However, the relative importance of each parameter clinical effect which is most valuable when considering the use of zanamivir
(and the ranking) may also change with the cost-effectiveness threshold. for the treatment of influenza, and additional evidence about the expected
 -,-
188 I DECISION MAKING, UNCERTAiNTY AND T H E VALUE OF I N FORMATION IDENTIFYING RESEARCH PRIORITIES ! 189

£20,000,000 additional research. We have already shown how the EVPPI can start to indi­
cate which research designs may be worth considering; however, establishing
£1 8,000,000
t2 Effectiveness whether a particular type of research should be conducted requires an analysis
£1 6,000,000 � Utillies
1
, of the value of sample information. This is introduced in Chapter 7.
_"- £1 4,000,000 B Cost

g;
w
£12,000,000
6.4. Identifying research priorities
§ £10,000,000
� A number of methods for setting priorities in research and development of
"5 £8,000,000 health care technologies have been proposed and some have been used to
� £6,000,000 identify priority areas for research. These include measures of the burden of
disease or the technology (Gross et al. 1999; Michaud et al. 2001); measures of
£4,000,000
the expected 'payback' from research (Buxton and Hannay 1997; Townsend
et al. 1997; Davies et al. 2000); and estimates of the welfare losses due to varia­
£0 .f----c=-
c �-"" O rlistat tions in clinical practice (Phelps and Parente 1990). However, each of these
WT Ritalin Zanamivir
the stylized proposed methods has serious methodological problems. Firstly, all of these
Fig. 6.7 Expected value of perfect information for parameters for NiCE
examples. proposed approaches view research simply as a means of changing clinical
practice rather than considering research as providing additional information,
which will reduce the uncertainty about what is appropriate clinical practice.
costs of motor neurone disease when considering the use of riluzole. This Indeed, measures of 'payback' or welfare losses due to variations in clinical
demonstrates that not only will we want different amounts of evidence for practice require the analysis to identify 'appropriate utilization' or which tech­
different technologies (and different amounts of the same technology but in nology should be adopted a priori. These methods therefore implicitly assume
different circumstances), we will also want different types of evidence. Again, that there is no uncertainty surrounding the decision that the proposed
there are clear implications for the efficient regulation and reimbursement of research is supposed to inform.
health technologies (Claxton 1999b). Secondly, these approaches, particularly measures of burden, attempt to
In this section we have demonstrated that the value of information associated identify research priorities using aggregate measures across broad clinical
with individual parameters or groups of parameters can be established using areas. However, the information generated by evaluative research is only valu­
the same principles as decision EVPI as previously discussed. The EVPPI able if it informs specific clinical decisions for specific groups of patients.
is essentially a measure of the sensitivity of the decision problem to the uncer­ Aggregate measures fail to recognize that the value of research in a clinical area
tainty in particular parameters. However, it has a number of advantages over is simply made up of the value oi research about each of the constituent clinical
other measures of sensitivity that are available. Firstly, it does not require or decision problems faced within that area. Therefore, simply because aggregate
assume a linear relationship between the parameter and the model output measures such as burden of disease may suggest a clinical area is a 'high' priority,
(net-benefit). Indeed, by using the general approach outlined above, EVPPI it does not mean that specific evaluative research relating to anyone clinical
can be established for any nonlinear or discontinuous relationship with net­ decision problem will be valuable. Similarly, proposed research to inform a
benefit. Secondly, unlike other measures of sensitivity, the EVPPI is driven by particular decision in a 'low' priority disease area may be very valuable.
decision uncertainty and the impact of parameter uncertainty on decisions. In this sense, attempts to identify research priorities across broad clinical
In this sense it combines both the notion of the importance of a parameter areas using aggregate indicators will be mistaken. 'What is required is a measure
and the uncertainty surrounding its value, Finally, the approach generates of the societal value of particular research, which can inform specific clinical
estimates of the values of information which are consistent with the objectives decisions for defined groups of patients and value the additional information
of and the budget constraint on health care provision. Therefore, the value generated by research in a way, that is consistent with the objectives and the
of information can be directly compared with the costs of conducting resource constraints of health care provision.

,J
I
 I
190 1 DECISION MAKING, UNCERTAINTY A N D T H E VALUE OF INFOR MATION
I I D E NTIFYING RESEARCH PRIORITIES 1 191
I

I
As we have seen in earlier sections of this chapter, the EVPI places an upper explicitly, and of synthesizing evidence from a range of sources. One feature of
bound on the societal returns to additional evidence. These values are consis­ the gu.idance is the development of (reference case' requirements for analysis.
tent with both the objective of the health care system and the budget constraint The reference case in the NICE guidance includes decision analytic modelling,
(represented by the cost-effectiveness threshold). Therefore, the value of i synthesis of evidence from various sources, and the characterization of uncer­
information (the EVPI in this chapter and EVSI and ENBS in the next chap­ i tainty through probabilistic analysis. Although explicit valuation of additional

I
ter) can be compared with the expected costs of further investigation. These evidence through value-of-information analysis is not required, it is recom­
estimates can be used to identify those areas of research and the type of mended in order to inform the research recommendations that NICE issues
research which will be most valuable and can inform the allocation of research alongside its guidance on technologies:
and development resources within and between broad clinical areas. In addi­
Candidate topics for future research can be identified on the basis of evidence gaps
tion, the value of information can be used to inform the allocation of resources identified by the systematic review and cost-effectiveness analysis. These may be best
between research and the provision of health care. prioritized by considering the value of additional information in reducing the degree
The use of value-of-information analysis has increased dramatically since of decision uncertainty.
2000, particularly in health care (Yokota and Thompson 2004). Two recent
pilot studies on the use of value-of-information analysis have been conducted in (NICE 2004)
the UK to demonstrate the benefits and feasibility of using this type of approach I n fact almost all the Technology Assessment Reports commissioned by
to inform policy decision within the UK. The UK National Co-ordinating NICE already include decision analysis and synthesis of evidence. A number
Centre for Health Technology Assessment (NCCHTA) which identifies have also included probabilistic analysis and value-of-information analysis.
research priorities and commissions research for the UK NBS, commissioned More recently, NICE commissioned a pilot study of value-of-information
a pilot study on the use of decision analysis to inform its research prioritization analysis to support the Institute's research recommendations made as part of
process (Claxton et al. 2004). NICE also commissioned a pilot study of value­ its Guidance to the NHS in England and Wales and to inform the deliberations
of-information analysis to support the institute's research recommendations of the NICE Research and Development Committee (Claxton et al., 200Sa).
made as part of its guidance to the NHS in England and Wales and to inform The purpose was to establish the feasibility and requirements of using value­
the deliberations of the NICE Research and Development Committee (Claxton of-information analysis to consider the possible implementation of this
et al. 200Sa). framework within the NICE processes.
The pilot study consisted of a series of six case studies based on a re-analysis
6.4. 1 . Recent policy experience
of recent Technology Assessment Reports completed for NICE. These
In many countries, decisions to adopt, reimburse or issue guidance on health included; screening for age-related macular degeneration (AMD); glycoprotein
technologies are increasingly based on explicit cost-effectiveness analysis using lIb/IlIa antagonists for acute coronary syndrome (GPAs); clopidogrel and
a decision analytic framework. A number of countries now request manufac­ dipyridamole in the secondary prevention of occlusive vascular events (CLO);
turers of health care technologies to provide evidence on cost-effectiveness in neurominidase inhibitors for the treatment of influenza (NIs); liquid based
support of applications for funding by the health care system. This approach cytology screening for cervical cancer (LBC); and beta interferon and glatiramer
was first used in Australia and Ontario for decisions about the funding of new acetate in the management of multiple sclerosis (MS).
medicines (CDHHCS 1992; Ontario Ministry of Health 1994). The approach Value-of-information analysis was successfully implemented for each of the
has also been used by several managed care groups in the USA (Langley et al. six case studies using the type of nonparametric methods described previ­
1999). All these agencies, including the Food and Drug Administration in the ously. For each case study, this allowed comparison of the potential benefits of
USA, consider whether the evidence used to support claims of cost-effectiveness further research (EVPI) with the costs of further investigation and an indica­
is sufficient basis for the decision to reimburse, issue guidance or allow claims tion of whether further research may be required to support guidance on use
to be made (Claxton et al. 200Sb).A prime example is the UK National Institute of the technology issued by NICE. The EVPI associated with the groups of
for Health and Clinical Excellence, where recent guidance on the methods of parameters indicated the type of evidence that would be most valuable and,
technology appraisal reflect the importance of representing decision problems therefore, the type of studies that should be recommended. The pilot study
 �
V
I
192 1 DECISION MAKING, UNCERTAINTY A N D THE VALUE OF I N FORMATION I D E N T I FYING RESEARCH PRIORITIES 1 1 93

also enabled comparisons to be made across different technology assessments on the stroke and myocardial infarction subgroups should be regarded as a

!
(decision problems) and was able to inform the prioritization of the range of priority, alth6ugh research on the TIA and PAD subgroups may also be worth­
research recommendations made by the Institute. This type of analysis also while. Similarly, the value of additional evidence for AMD differs by visual
enables some comparison between the value of investing resources in research
or other activities, such as the provision of health service. In this sense the I acuity (from £6.2m to £lS.3m), and suggests that additional research should
include those subgroups with lower starting visual acuity.
application of the methods described in this and earlier chapters, for a body
such as NICE, provides a unified and coherent framework for prioritization of
I The analysis also indicates which endpoints should be included in further
research. For example, in the GPA case study further research is valuable and
both research and the use of health care technologies. I should be regarded as a priority. It also indicated that it is RCT evidence of the
The results of the pilot study illustrated many of the issues discussed previ­
ously in this chapter. For example, the value of research differed substantially I effect of GPA as medical management and clopidogrel which is most valuable.
However, it also indicates that the mortality endpoint for patients with non­

I
across the six technology appraisals and ranged from £2.8m (LBCs) to £86Sm acute PCI should be the primary endpoint in any future trial.
(CLO). In some cases the analysis indicated that further research should not A number of case studies presented scenarios to explore alternative views of
be regarded as a priority, for example, the EVPI surrounding LBC following the relevant evidence. These included inclusion of related and 'unrelated
evidence from a previous pilot study was low (£2.8m) and no further research events' in the assessment of CLO and impact of restricting consideration of
was required. In other cases it indicated that additional research should be evidence at 6 months in GPAs; different structural assumptions regarding the
regarded as a priority, for example, the EVPI surrounding CLO for stroke mechanism of action, such as the additive nature of information gains during
patients was high (£86Sm). In other cases, the analysis refocuses the original screening for AMD; as well as the impact on value of information when rele­
research recommendations, such as in the AMD case study, although further vant alternatives may have been excluded from the original scope of the
research appears to be potentially worthwhile, it is additional evidence about appraisal, such as including the potential role of c10pidigrel in the GPA case
quality of life with and without photodynamic therapy, rather than the study. These examples demonstrated that even when this type of structural
performance of self-screening itself which is valuable. uncertainty has a limited impact on the estimates of cost-effectiveness, it can
The analysis also indicated which comparators should be included in future still have substantial impact on the value of information. This indicates the
research and also suggested other parameters that could be excluded. For importance of accounting for structural uncertainty and, if possible, repre­
example, the value of information for NIs was significant (£66.7m) but it is senting these issues as additional uncertain parameters within the model. The
further evidence about quality of life with influenza which is most important parameterization of structural uncertainty enables the analysis to address the
(£44.3m), rather than additional evidence about the effect on symptoms. question of the value of generating evidence that can resolve these uncertain
Although there is some value in further randomized control trials (RCTs) on assumptions (Carlin and Chib 1995).
the effect of oselativir and amantadine on symptoms (£0,43m and £O.23m A more detailed discussion of the results and the implications for
respectively), there is no value in further trials of zanamivir. Similarly in the research prioritization including the implications for the design of any
MS case study, although there is value in additional RCT evidence of the effect future research in terms of features such as the relevant patient groups and
on progression of the disease, it is the effect of copaxone and betaferon which comparators, and whether experimental design was likely to be required in
should be regarded as a priority (£14m and £1 3.6m respectively). In this case, each of the areas, can be found in full reports (Claxton et al. 2004, Claxton et al.
however, evidence about cost of care and relapse, and quality of life are also 200Sa). Both pilot studies also highlighted a number of methodological and
valuable. For these parameters, further research would not require experimental computation challenges faced when applying value-of-information methods
design and may be less costly to acquire (£10m and £6m respectively) so may to policy-relevant decision problems in a timely way. However, it should be
be regarded as priorities. recognized that the key challenges for this type of analysis are not primarily
Estimates of value of information for the decision problem and for groups the value-of-information methods themselves, but issues associated with
of parameters were also presented for relevant patient subgroups, for example, structuring decision problems, synthesis of evidence and the characterization
the value of information differed across the patient groups considered in the of uncertainty (required for estimating costs and effects as well as value of
CLO case study (from £8S6m to £240m). This suggests that further research information).
1 94 I DECISION MAKING, UNCERTAINTY AND T H E VALUE OF I N FORMATION
1 EXERCISE: CALCULATING EXPECTED VALUE O F PERFECT I N FORMATION

basis o f expected values i n order to minimize opportunity costs. The second,

I 195

Overall, the pilot studies demonstrated that the framework of decision

analysis and value-of-information analysis can be applied to policy-relevant requires consideration of the value of obtaining information and an upper
decisions in a timely way to provide results which can inform both the deci­ bound on the,letums to future research was established: the expected value of
sion of whether a technology should be adopted based on existing evidence perfect informat.ion. This was argued to be a necessary condition for funding
and whether more evidence is required to support that decision in the future. future research - in the next chapter we consider the sufficient condition for
The studies also illustrate that the amount and type of evidence needed to funding additional research.
inform decisions about health technologies is essentially an empirical ques­ I 6.6. Exercise: calculating expected value of perfect

I
tion and different amounts and types of evidence will be needed for different
technologies, applied to different patient groups in different circumstances. information for the total hip replacement model
I
, 6.6. 1 . Objective
6.4.2. Necessary but not sufficient condition
The EVPI discussed in this chapter and presented in these pilot studies only
I' The purpose of this exercise is to show how the model you have already devel­
oped can be adapted to calculate the expected value of perfect information
places an upper bound on the returns to further research. This provides a
(EVPI). As a first step, the overall EVPI for the model is calculated in order to
necessary condition for conducting further research where additional research
summarise the overall importance of uncertainty for decision making.
about the decision problem as a whole or research about particular endpoints
Secondly, the EVPI for individual parameters (or groups of parameters) can be
may be worthwhile if the EVPI or EVPPI exceeds the cost of conducting
calculated in order to examine the individual contribution of different param­
further research. However, to establish a sufficient condition: decide if further
eter (combinations) to the overall uncertainty.
research will be worthwhile and identify efficient research design; we need to
The step-by-step guide below will take you through the following stages for
consider the marginal benefits and marginal cost of sample information.
calculating expected value of information for the overall decision problem
The same framework of value-of-information analysis can be extended to
and for the parameters of the decision model:
establish the expected value of sample information for particular research
designs and to compare these marginal benefits of research with the marginal 1 . Calculating per patient EVPI from simulation output.
costs (Claxton 1999; Ades et al. 2004). This type of analysis provides a societal 2. Calculating the effective population and population EVPI.
payoff to alternative designs and can be used to establish optimal sample size, 3. Plotting EVPI as a function of the ceiling ratio.
optimal allocation of patients within a clinical trial, appropriate follow-up and 4. EVPI for a parameter (EVPPI).
which endpoints should be included in the design. Indeed, this framework can be
5. Plotting EVPPls.
used to identify a portfolio of different types of studies, which may be required
to provide evidence sufficient to support the use of a health care technology. There are two templates for this exercise. Begin with the file Exercise 6.6a ­
template.xls for parts 1-3 above. Once complete, you will be prompted to open
Exercise 6.6b - template.xls to continue the exercise.
6 . 5 . Summary
6.6 2 Step-by-step guide
This chapter has argued that the conventional rules of inference are inappropri­
1 . Calculating per patient EVPI from simulation output
ate for adopt/do not adopt decisions in health care decision making because of
the potential opportunity costs associated with failing to implement cost­ In the template file you will find that we have returned to the model from
effectiveness decisions. As an alternative, it was argued that there are in fact Exercise 5.7 - where there is one new prosthesis to compare with the standard
two related, but distinct decisions: firstly, whether to adopt an intervention prosthesis. Move to the <Simulation> worksheet, which you should recognize
now, given the current state of knowledge concerning an intervention's cost­ from the previous exercise. Recall that, under a decision theoretic approach, it
effectiveness; and secondly, whether additional information is required is the expected outcome that should guide the decision. That is, we will use
to support such adoption decisions in the future. The first type of decision the new prosthesis if it has an expected net-benefit that is greater than that of
(in the absence of serious concerns over reversibility) should be made on the the standard prosthesis. However, due to uncertainty, sometimes we will make
196 ! DECISION MAKING, UNCERTAINTY AND T H E VALUE Of INFORMAT!ON 1 EXERCISE: CALCULATING EXPECTED VALUE OF PERFECT I N FORMATION ! 197

the wrong decision and the losses associated with the wrong decision are the
I ll. Now plot these results on the figure template <EVPI>.
net benefits forgone. iii. Notice the peak in EVPI - where does this occur?
i. In row 4 of the net-benefit columns (X to 2), record the mean net­
benefit across the 1000 Monte Carlo simulations. 4. EVPI for a p'arameter (EVPPi)
ii. In the EVPI column (AK) calculate the maximum of the net benefits The overall EVPI for the model is a useful upper limit on returns to future
observed for standard or NPI for each of the 1000 simulations (if we research. However, of crucial importance is which particular parameters (or
think of each simulation if like a 'realization of a state of the world' groups of related parameters) are most important in terms of having the greatest
then this column represents net benefits under perfect information value of information. EVPPI approaches are designed to look at just that. They
'
where we always choose the prosthesis with the greatest net-benefit). work by looking at the value of information of the remaining parameters of the
iii. Now calculate the mean across the net benefits we would have under model if we assume perfect information for the parameter of interest.
perfect information in cell AK4. Begin by opening up the template file 'Exercise 6.6b - Template.xls'.
iv. Finally, in cell AKI, calculate the difference between the mean of In this exercise, the approach is illustrated using the effectiveness parameter
the maximized net benefits and the maximum of the mean values for the new prosthesis rrNP l . The exercise will get you to manually run
(X4 and Y4) - this gives the EVPI for the individual. through a looped exercise of obtaining estimates of net-benefit assuming that
the value of rrNPI is known. This may begin to appear repetitive - but the aim
2 . Calculating the effective population and population EVPI is to get a full appreciation of the approach before running a pre-written macro
Of course, there will be more than one patient eligible for a particular treat­ which will do the same thing but many more times (the aim is to get an under­
ment. Therefore, the individual patient EVPI must be inflated by the effective standing of the process rather than to give you practice in macro writing!)
population, which is a function of the eligible patients per annum and the i. Start by checking the cost-effectiveness threshold value in cell Zl is set
expected lifetime of the technology. For the hip replacement example, we are to 2200 (this value is chosen to give a high EVPI for this part of the
going to assume an effective technology life of 1 0 years with 40 000 new exercise).
patients eligible for treatment each year. ii. We wish to estimate the effect of certain knowledge of rrNPI on reduc­
i. In column AN link the cell showing the 40 000 patients per annum to the ing the expected cost of uncertainty. However, we are uncertain as to the
individual years of the technology lifetime and discount the population. true value of rrNP1, therefore the first step is to draw a value at random
ii. Sum the result in cell AN4 to give the effective population (discounted) from the distribution of rrNPI. Note that this can be done simply by
over the lifetime of the technology. copying the value in cell C27 of the <Parameters> worksheet and using
iii. Finally, multiply this effective population by the individual EVPI in the Paste special: values command to enter this value into cell B27.
cell AKI to give the EVPI for the population in cell API. iii. The model now has a constant value for rrNPI (drawn from its esti­
mated distribution). Now run the MCsimulation macro to estimate
3 . Plotting EVPI as a function of the cost-effectiveness threshold the effect of the joint uncertainty in all the remaining parameters of
As EVPI is calculated using net-benefit and net-benefit is a function of the the model.
unknown cost-effectiveness threshold it is important to estimate EVPI as a iv. Copy the mean net benefits for each prosthesis (cells X4:Y4) into the
function of the cost-effectiveness threshold. AS6:AT6 cells, remembering to use the Paste special: values command.
i. If you have calculated everything correctly so far, the population EVPI v. Now repeat steps (ii) to (v) nine more times to give ten mean net benefits
should automatically change when you enter a new value for the cost­ for different draws from the distribution of rrNPI. Record these values
effectiveness threshold in cell 21. Test this, and when you are satisfied in AS7:ATl5.
all is well, run the macro 'EVPI' which will use cost-effectiveness Hint: ifyour PC is slow, such that this is taking a long period of time,
thresholds from column AP and record the corresponding EVPI in reduce the number of times you do this looping before progressing to the
columnAQ. next step.
1 98 1 DECISION MAKING, UNCERTAINTY AND T H E VALUE OF I N FORMATION R EFER ENCES 1 199

vi. Now average across these ten evaluations of net-benefit to calculate computer - or employ someone with specific programming skills (we
the mean overall net-benefit over the draws from the distribution of . recognize our code suggestions are not the fastest possible) but we are
rrNPl . Record these values in cells AS4:AT4. Note that these values not Visual Basic experts!)
are equivalent to the mean net benefits over all uncertainty in all ii. When the macro has finished running, plot the results of the EVPPI
parameters (as uncertainty over rrNPI has been reintegrated). analysis in the <partialEVPI> worksheet template as simple bar chart.
vii. In cells AU6:AUIS, find the maximum of the net benefits in columns
AS and AT for the ten realizations you have obtained. This is the References
perfect information payoff given certainty over rrNPl . Calculate the Ades, A. E., Lu, G. and Claxton, K(2004) 'Expected value of sample information in medical
mean value in cell AU4. decision modelling', Medical Decision Making, 24: 207-227.
viii. The EVPPI for an individual patient is now calculated by subtracting Buries A.) Clark W., Preston c., et al. (2000) Is zanamivir effective for the treatment of
influenza in adults (supplement). London, National Institute for Clinical Excellence.
the expected payoff with uncertainty over all parameters (the maxi­
Buxton, M. and Hanney, S. (1997) 'Assessing payback from Department of Health Research
mum of cells AS4 and AT4) from the expected payoff with perfect
and Development: Second report'. HERG Research Report 24 . Uxbridge, Brunel
information about rrNPI (cell AU4). Enter this formula in cell AX7 University.
and multiply by the effective population (cell AN4) to give the EVPPI Carlin, B. P. and Chib, S. (199 5) 'Bayesian model choice via Markov chain Monte Carlo
on rNPI for the population. methods') Journal of the Royal Statistical Society. Series B, 57(3): 473-484 .
ix. Copy this EVPPI from cell AX7 and paste into the EVPPI table (cell Claxton, K. ( l999 a) 'The irrelevance of inference: a decision making approach to the
AXI l ) using the Paste special: values command. stochastic evaluation of health care technologies', Journal ofHealth Economics,18: 342-64.
Claxton) K.(1999 b) 'Bayesian approaches to the value of information; implications for the
x. The final step relates to 'tidying up' the model (this is crucial before regulation of new health care technologies'. Health Economics,8: 269-274.
moving on to the next section). Copy the formula from cell B26 on Claxton, K. and Thompson) K. A.(2001) 'A dynamic programming approach to efficient
the parameters worksheet down to cell 327. This returns the rrNPI clinical trial design. Journal of Health Economics', 20: 432-448 .
from a deterministic draw to the deterministic mean or probabilistic Claxton, K. and Posnett, J. (1996) 'An economic approach to clinical trial design and
value (depending on the value of the switch in cell D3). research priority setting', Health Economics, 5: 513- 524 .
Claxton, K) Neuman, P. J., Araki) S. S. and Weinstein, M. C.(2001) 'The value of information:
5. Plotting EVPPls an application to a policy model of Alzheimer's disease', International Journal of
Now that you understand the process behind calculating EVPPls, take a Technology Assessment in Health Care, 17: 38-55.

moment to examine the macros recorded for you in the spreadsheet by opening Claxton, K., Sculpher, M. and Drummond, M. (2002) 'A rational framework for decision
making by the National Institute for Clinical Excellence', Lancet, 360: 711-71 5.
the visual basic editor. You will see a macro for six different parameters!
Claxton, K., Ginnelly, L., Sculpher, M. J, Palmer S. and Philips, z. (2004) 'A pilot study on
groups of parameters (where parameters are related) such as utilities or the use of decision theory and value of information analysis as part of the NHS Health
survival parameters, it makes sense to consider their EVPPls together). An Technology Assessment programme', Health Technology Assessment,8(31): 1-103.
overall macro called 'partialEVPI' will run the full set of six EVPPI macros and Claxton, K.) Eggingtol1, S., Ginnelly, l., Griffin, S., McCabe, c., Philips, Z.} Tappenden, P.
record the results in the table AXI l:AXI6. and Wailoo, A. (200 5a) 'A pilot study of using value of information analysis to support
research recommendations for the National Institute for Clinical Excellence: Centre for
i. When you are happy with what the macros are doing, run the
Health Economics, Research Paper No 4 , University of York.
'partialEVPI' macro. Claxton K, Cohen J. T, and Neumann P. J. (200Sb) 'When is evidence sufficient?' Health
Note: this macro runs 1 00 Monte Carlo simulations by 100 draws Affairs,(24) , 1: 93-10 l .
from the parameter (group) of interest. It will therefore take a while Commonwealth Department o fHealth, Housing and Community Services (CDHHCS)
to complete - you may need to leave your computer running for (1992) Guidelines for the pharmaceutical industry on preparation of submissions to the
Pharmaceutical Benefits Advisory Committee. Canberra, AGPS.
some time! However - and this is very important - accurate estima­
Davies, L.) Drummond, M. F. and Papanikolaou, P.(2000) 'Prioritizing investments in
tion of EVPPIs takes many more runs than this - we would recom­
health technology assessment', Internationaljournal o/Technology Assessment in Health
mend at least 1000 inner and 1000 outer loops. Time to buy a faster Care, 16: 73-91.
200 ! DECISION MAKlNG, UNCE RTAINTY AND T H E VALUE OF I N FORMATION
�
I

Felli, J. C. and Hazen, G. B. ( 199 8) 'Sensitivity analysis and the expected value of perfect
information'. Medical Decision Making, 18: 95- 109.
Fenwick, E., Claxton, K and Sculpher, M. (200 1) 'Representing uncertainty: the role of
cost-effectiveness acceptability curves'. Health Economics, 10: 779- 89.
I
I
Chapter 7
"
, ,

Efficient research design

Ginnelly, L., Claxton, K, Sculpher, M. J. and Golder, S. (2005) 'Using value of information I

I
analysis to inform publicly funded research priorities'. Applied Health Economics and
Health Policy, 4: 37-46 .
Gross, C. P., Anderson, GF, NR P.( 1999) 'The relation between funding by the National
Institutes of Health and the burden of disease', New England Journal of Medicine,340:
1881- 1887 . In this chapter we establish a sufficient condition for deciding to conduct
Howard, R. A. ( 1966) 'Information value theory', IEEE Transactions on Systems Science and I further research and identify efficient research design. To do so we need to
Cybernetics, SSC-2 , 22-26 . I consider the expected benefits and costs of sample information. The value of
Langley, P. C.( 1999) 'Formulary submission guidelines for Blue Cross and Blue Shield of sample information and the expected net benefits of sampling are outlined in
Colorado and Nevada. Structure, application and manufacturer responsibilities', the following section. This section considers the case where incremental net­
PharmacoEconomics, 16: 2 11-224.
benefit is normally distributed, as well as the calculation of the expected value
Michaud C. M., Murray Cl, BR. B (200 1) 'Burden of disease - implications for future
of sample information for parameters or groups of model parameters based
research'. Journal of the American Medical Association; 2 85 .
Ministry of Health( 1994) Ontario guidelines for economic analysis of pharmaceutical
on conjugacy assumptions. These methods are then applied to the simple
products. Ontario, Ministry of Health. model of zanamivir, which was introduced in Chapter 6. This example
National Institute for Clinical Excellence (NICE) (2000) 'Guidance on the use of zanamivir demonstrates how meaSures of the societal payoff from research can be used
(Relenza) in the treatment of influenza', Technology Appraisals Guidance IS. London, to identify efficient research designs. Additional issues for research design that
NICE. can be addressed within the general framework offered by considering the
National Institute for Clinical Excellence (NICE) (2004 ) Guide to the methods of technology value of sample information are discussed in the final section.
appraisal. London: NICE.
O'Hagan, A. and Luce, B.(2003) A primer on Bayesian statistics in health economics and
outcomes research. Bethesda, MD, MedTap International. 7 . 1 . Expected value of sample information (EVSI)
Phelps, C. E. and Parente, S. T.( 1990) 'Priority setting in medical technology and medical In Chapter 6 we saw that the expected value of perfect information (EVPI)
practice assessment', Medical Care, 2 8: 703-723 .
places an upper bound on the returns to further research. This provides a
Pratt J., Raiffa, H. and Schlaifer, R. ( 1995) Statistical decision theory. Cambridge, :.viA,
necessary condition for conducting further research, where additional research
MIT Press.
Raiffa, H. and Schlaifer, R. ( 1959) Probability and statistics for business decisions. New York,
about the decision problem as a whole, or research about particular parameters
McGraw-Hill. or groups of parameters, may be worthwhile if the EVPI or expected value of
Sculpher, M. J. and Claxton, K.(2005) 'Establishing the cost-effectiveness of new perfect information for a parameter(s) (EVPPI) exceeds the cost of conducting
pharmaceuticals under conditions of uncertainty - when is there sufficient evidence?' further research. However) to establish a sufficient condition, decide if further
Value in Health, 8: 433-446. research will be worthwhile and identify efficient research design we need to
Thompson, K. M. and Evans, J. S. ( 1997) 'The value of improved national exposure consider the marginal benefits and marginal costs of sample information.
information for perchloroethylene (perc): a case study for dry cleaners: Risk Analysis,
The same framework of value-of-information analysis can be extended to
17: 253-27 l.
Thompson, K. M. and Graham, J. D.( 1996) 'Going beyond the single number: using
establish the expected value of sample information for a sample of n (EVSIIn),
probabilistic risk assessment to improve risk management', Human Ecology Risk for particular research designs. These expected benefits of research can be
Assessment,24: 100 8- 1034 . compared with the expected costs of sampling (Csln). The difference between
Townsend, J. and Buxton, M. ( 1997) 'Cost-effectiveness scenario analysis for a proposed the EVSIln and Csln is the expected net-benefit of sampling for a sample size
trial of hormone replacement therapy', Health Policy, 39: 181-94 . of n (ENBSln). The ENBSln can be regarded as the societal payoff to research
Yokota, F. and Thompson, K M.(2004) 'Value of information literature analysis: a review and can be calculated for a range of samples, sizes and for alternative designs.
of applications in health risk management', Medical Decision Making, 24: 2 87-29 8.
It can be used to establish optimal sample size, optimal allocation of patients

�"
 --,­
202 I EFFICIENT RESEARCH DESIGN
I EXPECTED VALUE OF SAMPLE I N FORMATION (EVSI) 203

I decision of current and future patients who are not enrolled in this
within a clinical trial, appropriate follow-up and which endpoints should be
included in the design. Indeed, this framework can be used to identify a port­ proposed trial. The value of sample information for the population of current
folio of different types of studies, which may be required to provide sufficient and future p.atients over the effective lifetime of the technology ( 1') can be
evidence to support the use of a health care technology. It is then possible to established based on estimates of the incidence of patients or episodes (I) in
answer questions such as whether an additional clinical trial is required; and if each period ( t) discounted at rate r.
so, should an economic evaluation be conducted alongside the new trial; and It should be noted that those patients who are enrolled in the trial in each
what is the optimal follow-up, sample size and allocation of entrants between period (nt) will not be able to benefit from the information generated by the
research because they will have already received treatment:

\
the arms of the proposed trial?
7. 1 . 1 . EVSI when net-benefit is normally distributed
Population EVSIln, n(2) = EVSIln, n(2)·2: (t 1, 2. . . . . T)(It - n,) 1 (1 + r)' .
=

The expected benefit of sample information can be regarded as the resulting I Therefore, increasing the sample size provides more information for the
reduction in the cost of uncertainty surrounding the choice between alterna­
tives j. For an individual patient, the expected benefits of a sample of n I population of patients (or episodes of illness) who are not enrolled in the trial,
but it also 'uses up' those who would otherwise benefit from the sample infor­
allocated between a new technology (alternative j = 2) and current practice, i mation. The resource cost of a sample of n with n(2) allocated to j = 2 (Csln,
j
(alternative j = 1 ) in a proposed trial where n(2) are allocated to j = 2 and n(2») includes the fixed costs of proposed research (Cf), the marginal (incre­
n-n(2) are allocated to j = 1, can be expressed as the expected value of sample mental) treatment costs compared with current clinical practice (c, - CI) and
information (EVSIln, n(2» ' If incremental net-benefit is normally distributed marginal reporting costs (Cr):
then we can use analytic solutions similar to those described in Chapter 6
when we are considering samples of incremental net-benefit (Raiffa and Csln, n(2) = Cf + (C2 - CI ) · n(2) + Cr.n.
Schlaifer 1959; Claxton and Posnett 1996; Claxton 1999): In this case the cost function is very simple with a fixed element and
EVSIln, n(2) = A.. �(Vln, n(2» O"O . L(Dln, n(2» ' constant marginal costs. However, more complex cost functions can be easily
incorporated. The societal payoff to proposed research or the expected net
where: benefits of sampling (ENBSln, n(2») is simply the difference between the
A. = cost-effectiveness threshold expected benefits and expected cost of sampling:
ENBSln, n(2) = EVSIln, n(2) - Csln, n(2)'
It is now possible to establish a sufficient condition for conducting further
1)0 = prior mean incremental net health benefit
research and decide whether an additional study would be efficien t. If the
ENBSln, n(2) > 0 for any sample size and allocation of the sample between the
Vlln, n(2) = 0"5 1 (0"5 + (0"[2) 1 n(2» + 0"[2) + (0"[1) 1 n - n(2» )) arms of the trial, then further experimental research will be efficient. As well as
O"� = prior variance of the incremental net-benefit (1)0)' providing a sufficient condition for deciding whether further evidence is
required to support a decision to adopt a technology, this type of analysis also
The O"(�) and 0"(;) are the population variances surrounding the net-benefit indicates what type of research is needed and how it should be designed. It
of alternative j::;;:: 2 and j::;;:: 1, respectively, so that as sample size increases Yin, n(2) provides a framework for the efficient design of clinical trials based on the
tends to 1 and the expression for EVSI approaches the EVPI. In other words, societal payoff to research. In this case the optimal sample size ( n') for this
perfect information can be regarded as an infinite sample and is the upper proposed research will be where ENBS reaches a maximum (ENBSln" n(2)')'
bound to EVSI. given that the trial entrants are allocated optimally (n(2)*) between the two
As with EVPI, the expected value of sample information also has public arms of the trial. In fact, ENBS can be used to choose between alternative
good characteristics (it is nonrival), and can be used to inform the treatment designs across a range of dimensions of research design including optimal
204 I EFFICIENT RESEARCH DESIGN EXPECTED VALUE OF SAMPLE I N FORMATION (EVS!) I 205

follow-up; which endpoints should be included in the trial, when to stop expected net benefits are averaged over the posterior distribution of the net­
a sequential trial; and, in principle, the optimal portfolio of different types of benefit of each treatment j given the sample result D :
study, which can inform the decision prohlem. This discussion has outlined
the conceptual issues and the use to which estimates of EVSI and ENBS can be
put. However, in decision models it is very unlikely that net benefits will be
However) we d o not know the actual results o f the sample i n advance
normally distributed and it is more useful to focus on the value of acquiring
(we do not know which value D will take). Therefore, the expected value
sample information about groups of particular model parameters rather than
of a decision taken with the sample information is found by averaging
directly on incremental net-benefit itself.
these maximum expected net benefits over the distribution of possible values
7 . 1 2 . EVSI for conjugate distributions of D; that is, the expectation over the predictive distribution of the sample
results D conditional on e, averaged over the prior distribution (possible
For the same reasons as outlined in Chapter 6, net-benefit is very unlikely to
be normally distributed and is unlikely to have a parametric distribution at all. values) of e:
The simple analytic framework outlined above is also based on sampling
incremental net-benefit directly rather than sampling the uncertain parameters
that determine net-benefit. In decision models) net-benefit is a function of many So the EVSI is simply the difference between the expected value of a
parameters and it is more useful to focus on the value of sample information decision made with sample information and the expected value with current
associated with particular parameters or groups of model parameters. The information:
approach outlined below is very similar to EVPPI and requires, for each proposed
research design, the prediction of possible sample results, the possible poste­
rior distributions which could result from these sample results and the associ­
However, models include more than one uncertain parameter and this
ated net benefits (Pratt et al. 1995: Berry and Stangl 1996: Ades et al. 2004).
approach can be extended to consider the value of sample information about
This approach to EVSI simply asks what the payoff (expected net-benefit)
a parameter or a subset of the parameters qJ with remaining uncertain param­
would be ifwe could base our decision on having additional sample information.
eters in the model 1fI. If rp and IfI are independent, then a sample of n on rp will
We can do this by predicting the possible sample results we would get from a provide a sample result D. If we knew D then expected net benefits are averaged
particular study with a sample size of n. We can combine our prior with each
over the prior distribution of IfI and the posterior distribution of rp given the
of these possible sample results to form a number of possihle predicted poste­
sample result D :
rior mean values (one for each predicted sample result) (Gelman et al. 1995;
O'Hagan and Luce 2003). For each of these possible posteriors we can calculate
the expected net-benefit of our decision. We do not know which of these
predicted posteriors will result from the sample of n so we average the net As we do not koow which value D will take we must take the expectation of
benefits across them. The difference between the expected net benefits of our the maximum expected net benefits over the predictive distribution of D
decision based on the predicted posteriors and the expected net benefits based conditional on <p) averaged over the prior distribution of cpo
on current information is the EVSI for a sample of n (Pratt et al. 1995; Ades
et al. 2004).
More formally) suppose we were interested in the value of sample information
As before, the expected value of sample information is the difference
about e (initially imagine that e is a single parameter and the only uncertain
between the expected value of a decision with sample information and the
parameter in the model). A sample of n on e will provide a sample result D.
expected value of a decision based on current information:
If we koew what the result of the sample would be in advance, then we could
choose the alternative with the maximum expected net-benefit when these
 �
, " uT
i ,
!
206 I EFFICIENT RESEARCH DESIGN
AN EXAMPLE OF EVSI ! 207

It should be clear from this discussion that EVSI requires predicted sample If the model is linear and if' and ",are independent, or if the model is multilinear
results to be combined with the prior information about a parameter to form and all the parameters (not just the subsets if' and "') are independent, then
predicted posteriors. For this reason we restrict discussion in this chapter to the inner loop can be simplified:
those situations where the likelihood for the data (in this case predicted
sample results) is conjugate with the prior so there is an analytic solution to
combining the prior with the predicted sample to form the predicted posterior
(Gelman et al. 1995; O'Hagan and Luce 2003). If the prior and likelihood are In this case the net-beneftt for a particular value of D can be calculated from
not conjugate then there is no simple analytic solution and the computational the predicted posterior mean values of the parameters if' and the prior mean
burden of using numerical methods to form predicted posteriors is considerable of the other parameters ",. There is no need to sample from the predicted
(Ades et al. 2004).
Even with conjugacy EVSI still requires intensive computation. This is
I
I.
posterior distribution of if' or the prior distribution of ",. A range of
algorithms suitable for different levels of linearity in net-beneftt is outlined in
I Ades et al. (2004).
I
because in a similar way to EVPPI we have an inner and outer loop of expecta­
tion. However, EVSI is more intensive than EVPPI because we must first
sample a value for D from the predictive distribution of D conditional on a 7 . 2 . An example of EVSI
value of if', then sample from the prior distribution of '" and the posterior The principles of using an analysis of EVSI can be illustrated using the simple
distribution of if' given a sample result D (the inner loop). Then we mLlst example of zanamivir, which was discussed in Chapter 6. The National Institute
sample a new value for D from the predictive distribution of D conditional on for Clinical Excellence (NICE) considered zanamivir as a treatment for
a new value of if' and rerun the inner loop. This must be repeated until we have influenza in adults in 2000 and issued guidance (NICE 2000) imposing a
sampled sufficiently from the distribution (or joint distribution) of if' (the number of restrictions and conditions on its use. In particular, treatment
outer loop). should only be offered when influenza is circulating in the community and for
Of course, the EVSI described above is for a single study design and only at-risk adults who present within 36 hours of the onset of influenza-like
one sample size. To establish the optimal sample size for a particular type of illness. At-risk adults were deftned as those over the age of 65 or those with
study, these calculations need to be repeated for a range of sample sizes to chronic respiratory disease) significant cardiovascular disease} irnmunocom­
identify the point at which ENBS reaches a maximum. It should be noted that promised and those with diabetes mellitus.
the EVPPI can provide an upper bound on the sample sizes that may be worth This guidance and the appraisal of zanamivir were based on the independent
considering for particular groups of parameters. The EVPPI net of the ftxed assessment report (Burles 2000). The assessment report identified nine
costs of research, divided by the marginal costs of sampling provides an upper randomized control trials but the evidence for at-risk adults is based on
bound on sample size. This is because the cost of sample sizes greater than this subgroup analysis of eight of the all-adult trials and one trial that only
will exceed the EVPPI and the returns to any research study. Of course, if there recruited adults with chronic respiratory disease. The trials reported time to
are a number of different types of study which could generate information alleviation of symptoms for the influenza-positive population) as well as the
about different groups of parameters or studies with different follow-up number of complications requiring antibiotics. The assessment report
periods or different allocations of sample between the alternatives, then these included fixed-effect meta-analysis for both these endpoints. None of the trials
calculations would be needed for each possible design and for the range of included effect on hospitalization, costs or quality of life. A simple decision tree
possible sample sizes. model was presented in the assessment report where the proportion of
influenza-positive patients (pip) presenting with influenza-like symptoms was
7 . 1 .3. EVSI for linear models
based on a sponsored submission, the baseline risk of complication (pes) and
The approach outlined above provides a general solution for nonlinear hospitalization (phs) was based on an observational data set. The reduction in
models given that the groups of parameters if' and ", are (prior) independent complications requiring antibiotics was based on the log odds ratio from the
(Ades et al. 2004). The computation for EVSI can be somewhat simplifted if the fIxed effect meta-analysis of the trials (LORe) and the same ratio was used to
model is either linear or multilinear in a similar way to EVPPI in Chapter 6. model reduction in hospitalizations (LORh). The reduction in symptom
 I
I
I I

I
208 EFFICIENT R E S EARCH D E S I G N AN EXAMPLE OF EV5! 209

days (rds) for the influenza-positive population was based on a fixed effect research, so we must calculate the mean of these expected net benefits across
meta-analysis, and the impact on quality of life of a symptom day (qsd) was all the values of pip'. The difference between the expected net benefits of

I
based on an assumed impact on all dimensions of EQ5D. Resource use was our decision.based on the predicted posteriors for pip (pip') and the expected
based on published unit costs, hospital length of stay, and the number of GP net benefits based on current information is the EVSI for a sample of n.
visits. I n a nonmultilinear model the expected net benefits for a particular
This example was first introduced in Chapter 6 and the cost-effectiveness predicted posterior requires evaluation of net-benefit by sampling from the
acceptability curve (CEAC) for this decision problem was illustrated in J predicted posterior distribution. For a multilinear model (as in this case),

I
Fig. 6.1. The analysis of the population EVPI was illustrated in Fig. 6.3 and however, the predicted posterior mean can be used to calculate the expected
indicated that at a cost-effectiveness threshold of £50 000, further research may net benefits, substantially reducing the computation.

i
be potentially cost-effective. The analysis of the EVPIs for model parameters The relationship between the sample sizes considered, the prior and
was illustrated in Figs 6.5 and 6.6 and suggested that further research about predicted posterior distribution and the EVSI are illustrated in Fig. 7.1 for the
pip, phs, rds, upa and LORh may be worthwhile. In the following section we beta binomial case. Based on current information (the prior distribution )
illustrate how the EVSI for sample information about each of these parameters
can be calculated and illustrate the EVSI and ENBS for an additional clinical I imagine that alternative j has the highest net-benefit. When a small sample size
of n ; 1 is considered, there are only two possible results from the binomial
trial which would provide information on all parameters. I likelihood and the two possible posterior distributions will be similar to the
! prior distribution because the prior will dominate the limited sample infor­
7.2. 1 . EVSI for beta binomial mation. It is, therefore, very unlikely that these possible posteriors will change
The first uncertain parameter in this simple model is the probability that a the decision based on current information and thus EVSI will be low. In
patient is influenza-positive (pip). A beta distribution was assigned to this Fig. 7.1 we continue to choose alternative j irrespective of the results of the
parameter in the probabilistic analysis. The analysis of the EVPPI in Chapter 6 sample and EVSI ; O. However, when the sample size is larger ( n ; 20) there
indicated that additional information about this parameter may be worth­ are a large number of possible sample results and the predicted posterior
while. Therefore, we could consider conducting an epidemiological study to distributions can be some distance from the prior because the prior no longer
estimate the probability of the patient being influenza-positive. To calculate
the EVSI of such a study we must sample from the prior distribution:
N=l NBj NBj N=20 NBk
pip - Beta( a, {3);

predict possible sample results (rip) using the binomial distribution:

\" I
rip - Bin(pip, n);

and then form predicted posterior distributions (pip'):

pip' - Beta« a+rip), ({3 +n-nrip» .

The pip' are the possible (predicted) posterior distributions for the probability
of a patient being influenza-positive following the study with a sample size of n.
With sample information we will be able to make a different decision for each
predicted posterior pip'. Therefore, the expected value with sam pIe infor­ 0.0 0.2 0.8 1 .0 0.0 0.2 0.4 NBjo.6 0.8 1.0
mation is found by cakulating the expected net-benefit for a particular pip' Probability influenza positive (pip) Probability influenza positive (pip)

and choosing the alternative with the maximum net-benefit. However, we do -Prior distribution - Predicted posterior distributons

not know which of the possible posteriors will actually be the result of the Fig. 1.1 Prior and predicted posterior distributions.

<>p'"
 �.
':>?i
I
I
210 I EFFICIENT RESEARCH DESIGN I AN EXAMPLE OF EVSI 21 1

dominates the more substantial sample information. In this sense, as sample

size increases we are more uncertain about where the posterior distribu­
tion might lie. It is now much more likely that the posterior will change the
decision. In Fig. 7.1 some of the predicted posteriors change net-benefit so
that we would choose alternative k rather than j. When we calculate the mean
I The probability of complication requiring hospitalization with standard
care was aiso:.,assigned a beta distribution and was associated with significant
EVPPI in Fig,;6.6. The EVSI for an epidemiological study of the probability of
hospitalization Jar patients with influenza can be calculated using the same
approach, exploiting the conjugate relationship between the beta prior, the
expected net-benefit over these predicted posteriors it will be higher than with binomial likelihood and the predicted posterior beta distributions.
current information. In other words, the EVSI will be positive and increases
with n. 7.2.2. EVSI for the normal distribution
In fact, as sample size becomes very large, the predicted posterior can There is also uncertainty in the effectiveness of zanamivir and a normal distri­
resolve anywhere across the prior distribution. The variance in the predicted bution was assigned to the reduction in symptom days (rsd) for patients who
I-

!
posterior means approaches the prior variance and the EVSI for the parameter are influenza-positive based on the results of a fb:ed effect meta-analysis of the
will approach the EVPPI - once again confirming our interpretation of EVPI clinical trials. The EVPPI associated with rsd in Fig. 6.6 suggested that further
and EVPPI. research may be worthwhile. We can use the same approach to consider the

j
The EVSI for an epidemiological study to estimate the probability that a value of conducting a further clinical trial to provide a more precise estimate
patient is influenza-positive is illustrated in Fig. 7.2. This demonstrates that of rsd. To calculate the EVSI for a trial of n patients who are allocated equally
the EVSI will rise with sample size and approaches the EVPPI for pip as between zanamivir and standard care, we must first sample from the prior
sample size becomes very large. Of course, the EVSI will also depend on the
i distribution of rsd:
cost-effectiveness threshold and, just like the EVPPl, there is no clear monotonic I

I
relationship. For example, the EVSI will be lower at cost-effectiveness threshold
of £40 000 and £60 000 reflecting the relationship between EVPPI and the
cost-effectiveness threshold that was illustrated in Fig. 6.6. For each sample from this prior distribution we must predict possible
sample results (l1s) which will be normally distributed:

£900.000
I's - N(rsd, (J'2 In).
is £800,000 Then combine the prior with the predicted sample results to form
�
E £700,000 a predicted posterior distribution which will also be normally distributed:
�
'w £600.000
'l5. rsd' - N(1" , 1'),
E
m £500,000
'0 where:
ID
�
£400,000
'iii = «l'o .Io + lls · Is )I Uo +ls))
� £300,000
"

�
Io = 1 / (J'l
I,
:�::::::
1I«(J'2 In)
! - ---- -- -
-
A = £60,000
- ----
- -- -
- -� ----
I' (n + m) / (J'2
__ ...- -
- --
--

. .
-

. . .. . . I, :::: £40,000 m (J'2 Io '

--::.� �:.... . . .. .... ..........
....

£0 . . . . - ....
.. ..... .
' ....
-�- --....

o 50 100 150 200 250 300 350 400 450 500 As with pip, the EVSI for rsd is simply the difference between the expected
Sample size (n) net benefits of decisions based on the predicted posteriors foJ' rsd and the
Fig. 7.2 Expected value of sample information for an epidemjological study of pip. expected net benefits based on current information.
212 I EFFICIENT RESEARCH D E S I G N
1
I AN EXAMPLE OF EVSI I 213

The quality of life associated with influenza symptoms was also normally
distributed and was also associated with significant EVPPI in Fig. 6.6. The
EVSI for a study of quality of life in influenza-positive patients can be calcu­
lated using the same approach, exploiting the conjugate relationship between
the prior, the likelihood and the predicted posterior distributions.
II �
c
o
�
E
£2,500,000

£2,000,000
,�'
rsd

qsd

I
'"
� £ 1 ,500,000
7.2.3. EVSI for log odds ratios '"
"'
'0
The effect of zanamivir on complications requiring hospitalization was associ­ � pip
£1 ,000,000
ated with significant EVPPI in Fig: 6.6, which suggests additional research on .. . .
� -
. . .. . . .... . . . . . . . . " . . .

I .. . .. ..

l
this endpoint may also be worthwhile. A combination of the normal distribution

'� �I!
I
and the beta binomial relationship enables the calculation EVSI for a clinical trial LORh

I
"-

of n patients, which would estimate the log odds ratio of hospitalization for Qj
those treated with zanamivir compared with standard care. As the baseline O�
· �����---------------------------,
o 0 o 0 o 0 g o
probability, phs, has a beta prior and binomial likelihood and the LORh has a o o 0 o
� N co � as 0 '"
normal prior and normal likelihood, we can simply exploit the same principles �

Sample size (n)

outlined above: (i) sample from the beta prior on phs and the normal prior on
Fig. 7.3 Expected value of sample information for each model parameter.
LORh; (ii) calculate the prior on phz based on the sampled prior values of phs
and LORh; (iii) use these sampled prior values of phz and phs to predict the
results of the study from their respective binomial likelihood; (iv) use these
results to calculate the predicted sample LORh; (v) combine the prior LORh In Fig. 7.3 when n = 100 the EVSI associated with rsd is lower than for pip or
with the predicted sample LORh to form a predicted posterior LORh; (vi) upa. This reflects that fact that there is already substantial prior information
propagate the predicted posterior LORh through the model to calculate the about rsd and that a further clinical trial with a small sample size is unlikely to
expected net-benefit for this particular posterior; (vii) repeat this process to change the decision and generate net-benefit (see Fig. 6. 1 ). The pip and upa
sample across the range of possible posteriors recording the expected net parameters are more uncertain) so that relatively small amounts of sample
benefits from each; (viii) calcnlate the mean across these expected net benefits information can have a substantial effect on their posterior distribution and
with the sample information. It should be noted that it is not possible to are more likely to have an impact on the decision and on expected net-benefit.
simply use the mean posterior LORh to calculate phz and therefore net-benefit The methods we have just outlined mean that we can now think about
in the model as this relationship is not linear. Therefore, either each predicted designing research in an efficient way, taking account of the benefits and costs
posterior distribution (not just the mean) must be propagated through the of different research designs. Indeed, we can calculate the EVSI for any param­
model or mathematical approximations (Taylor series expansion) can be used. eter that is conjugate and consider the expected net benefits of sampling for
More detail about EVSI for common data structures and mathematical studies that will inform individual parameters, for example, a epidemiological
approximations can be found in (Ades et al. 2004). study for pip, a survey of quality of life for upa and a clinical trial to estimate
rsd and LORh. However, many types of study will provide information about
7.2.4. EVSI and research design more than one parameter. For example, a clinical trial may include the rsd
The EVSI for each of the parameters associated with a significant EVPPI in endpoint, but with sufficient sample size and follow-up could also estimate
Chapter 6 and for which further research may be worthwhile are illustrated in the log odds of complications requiring hospitalization.
Fig. 7.3. This demonstrates that the EVSl increases with sample size and The EVSI for a clinical trial that includes all the model parameters as endpoints
approaches the EVPPI as sample size increases. Therefore, the EVSI will differ is illustrated in Fig. 7.4. The EVSI when updating groups of parameters together
in the same way that EVPPI differed across model parameters when sample is not simply the sum of the EVSI for the individual parameters (for the same
size is very large. However) at lower sample sizes this is not necessarily the case. reasons that the EVPPI for parameters will not sum to the decision EVPI).
 I
214 I EFFICIENT RESEARCH DESIGN
I AN EXAMPLE OF EVSI I 215

£3,500,000 I £3,500,000
§
iii £3,000,000 EVSI
A' £50,000 £3,000,QPO ---

E
-
--
- � --

o A ' £60,009_ _ _ _ _ _ _ _
:s £2,500,000 --

£2,500,000 ENBS
-
--

(f)
-
-
,

'"
-

co
-
-
--
-

'5, z
- -
,

� £2,000,000 LU
'0 £2,000,000
:ii
'"

'0
� £1 ,500,000 £1 ,500,000 '
'"

A' £40,000 o
�
. . . . . .. .

.
g!
. . , ...
. . . .. .
. . ...
. .
. . . ..

'8 .
....
.... . . .

£1 ,000,000 . . ..
.
.. ,
. .
.. . .

UJ £1 ,000,000
hl
.
.

Cs
. "f' " ..
� £500,000
'"

£500,000 i
I
i
£O �����----�------------�--�'--------c_-,
o
g o 8 § g � � � R � � g g � g � �
Sample size (n) Sample size (n)
Fig. 7.4 Expected value of sample information for the decision problem. Fig. 7.5 Expected net-benefit of sampling, expected value of sample information,
expected costs of sampling and optimal sample size.

But rather the difference between the expected net benefits of decisions based research then the proposed trial can be can be considered cost-effective. We
on the predi�ted posteriors for all the parameters and the expected net benefits have met the sufficient condition for conducting further research.
based on current information. For this design) which provides information on As EVSI depends on the cost-effectiveness threshold, we would expect
all the uncertain parameters, the EVSI will approach the decision EVPI as different ENBS and different optimal sample sizes for different thresholds,
sample size increases) confirming the interpretation that the perfect information This is demonstrated in Fig. 7,6, which illustrates the ENBS and optimal
is the upper limit on the value of sample information, Again, the relationship samples size for three cost-effectiveness thresholds. Clearly the ENBS reflects
between EVSI and the cost-effectiveness threshold reflects the relationship the same type of relationship to the cost-effectiveness threshold as the EVPI in
illustrated in Fig, 6,3 for EVPI and demonstrates that a higher monetary Fig, 6.3 and EVSI in Fig. 7,1, It shows that the value of research and the opti­
valuation of health outcome does not necessarily mean a higher valuation of mal sample size depends crucially on the cost-effectiveness threshold, It
sample information. demonstrates that economic arguments are central to fundamental design
So far this example has been used to illustrate how to estimate the benefit of decisions. In this example, despite the fact that the ENBS is lower for a thresh­
sample information, However, by comparing the EVSI with the costs of old of £60 000 compared with £5 0 000 the optimal sample size increases.
sampling, we can identify optimal sample size where the ENBS reaches a In general) however, there is no reason why there should be a monotonic
maximum, This is illustrated in Fig, 7.5 for a clinical trial that includes all the relationship between the threshold and optimal sample size.
parameters as endpoints, where the trial entrants are allocated equally between In this example we have only explored a small number of the possible
zanamivir and standard care (for a cost-effectiveness threshold of £50 000), research designs. So, although the ENBS at the optimal sample size in Fig. 7.5
The EVSI increases with sample size but ultimately at a declining rate, In this is positive) indicating that this particular research design would be cost-effective
example the marginal costs of sampling with equal allocation are constant and and that further research is needed, it does not indicate whether there are
the ENBS reaches a maximum of £2Am at an optimal sample size of 1 100, The other possible research designs that are more efficient (with higher ENBS), In
fixed costs of the trial have been excluded as they will not change the optimal fact) there are a large number of possible research designs) including different
sample size, If the ENBS of £2Am is greater than the fixed costs of this endpoints included in a clinical trial, alternative allocation of trial entrants
 .�
216 I EFFICIENT R E S EARCH DESIGN DESIGN DECISIONS I 217

£3,000,000 experience o f using the tradition power calculation know how it i s used in
practice: (i)·estimate potential grant funding for the project, take away the
'" £2,500,000 fixed costs and divide by marginal cost to find the sample size you can afford,
,
c
0. and/or estimate the patients you will be able to recruit; (ii) solve for the effect

I
,
,
E size which can be detected at 5 per cent significance and 80 per cent power. In
ill £2,000,000 ,
}, = £60,Oq<?- � _ _ - - - -- -f - - - - - , - - - - - - -
'0 _ �-
-- -
I I
this sense, the traditional power calculation is bankrupt and dishonest
� (Claxton and Posnett 1996).
iii £1 ,500,000
.0

� 7.3. 1 . Fixed sample design

i
,
il £1 ,000,000 , Previously we demonstrated that the optimal sample size will be where the
1:$ !
I ,.
i ENBS reaches a maximum. This was illustrated in Fig. 7.4 for a clinical trial
!
"', - - · t . . .. . . .
. .l
.
. . . . . ... -- . . I
·

I
£500,000 . . which included all the parameters in the simple decision model of zanamivir.
. .. " , ,
, , However, the estimates of EVSI and ENBS were based on an equal allocation
I I
of the sample (trial entrants) between the two arms of this proposed trial
o o 0 0 0 o
o (zanamivir and standard care). An equal allocation of sample between the
R §5 � g �

Sample size (n) arms of a proposed trial would be efficient if the marginal benefits and
Fig. 7.6 Expected net-benefit of sampling and optimal sample size. the marginal costs of allocating are the same across the arms of the trial. If
the benefits and costs of allocation to the alternative arms differ then equal
allocation will not be optimal.
The equal allocation of patients between experimental and control arms of
between the arms of the trial or a portfolio of different research designs
a trial is often used and is implicitly justified by assuming that the variance of
addressing different uncertain model parameters. In principle, we could use
the outcome of interest for the control arm of the trial is the same as the
this framework to choose the design that maximizes ENBS over a1l1he dimen­
experimental arm, so that the marginal benefits (reduction in sample variance)
sions of design space. These issues are discussed in the next section along with
of assigning an additional trial entrant to either arm of the trial will be the
some of the computational challenges involved.
same. However, there is little justification for this rule of precedent and there
is no reason why the marginal costs and benefits of sampling should be the
7.3. Design decisions same across the arms of the proposed trial. For example, the costs of allocating
The concepts of EVSI and ENBS enable us to address some important research trial entrants to the zanamivir arm will be higher due to the positive incre­
design issues, as well as providing a framework for considering the opportunity mental costs of this technology. In addition, different parameters will be rele­
cost of ethical concerns about proposed research. This contrasts sharply with vant to the zanamivir (e.g. RSD, LORe and LORh) and the standard care arms
the more traditional approach to research design based on the power calcula­ of the trial, so there is no reason to believe that the marginal benefits of allo­
tion. The effect size, which must be specified in the traditional power calculation, cating trial entrants will be the same either. Estimates of ENBS are therefore
is not the effect that is expected, but what you wish to detect. However, there required for the range of possible allocations of each sample size considered.
are no resource or opportunity costs of sampling specified. If asked what you The optimal allocation will be the one that maximizes ENBS. Optimal alloca­
wish to detect but at no cost, the only honest answer must be you wish to tion needs to be established for each sample size so that the optimal sample
detect anything, however small. But as effect size approaches zero, the sample size, given that it will be allocated optimally between the arms of the trial, can
size required becomes unbounded. In practice) therefore, positive effect sizes be identified. In other words, we can use ENBS to identify the efficient design
are chosen in an essentially arbitrary way to achieve finite and (manageable' over these two dimensions of design space (optimal allocation and sample
sample sizes. In this sense, the traditional power calculation leads to only two size) (Claxton 1999; Claxton and Thompson 2001). It should be recognized
sample sizes: infinite ones and arbitrary ones. In fact, most researchers with that optimal allocation of a given sample will depend on the cost-effectiveness
 I
I
218 I EFFICIENT RESEARCH DESIGN I DESIGN DECISIONS I 219

threshold, as this will determine the weight placed on differences in the should be stopped early on ethical grounds because one o f the interventions
marginal costs and marginal benefits of allocating trial entrants to the arms of has been demonstrated to be more effective. As should be clear from the
the proposed triat demonstrating once again that fundamental design issues previous dis<:l'ssion, the question of when to stop a trial can be addressed by
cannot be separated from economic concerns. calculating the ENBS given the accumulated evidence. If the ENBS is positive
then it is efficient to continue the trial; if it is negative then the trial should be
7.3.2. Sequential design stopped.
So far we have only considered a fixed sample design where the results of the Much of data monitoring and questions about the ethics of randomized
trial are available at the end of the study after all entrants have been allocated trials are concerned with the implications for those individuals enrolled in the
and the sample size has been determined. There is a body of literature that triaL However, the EVSI and ENBS calculations outlined above are primarily
considers the optimal allocation of trial entrants in sequential clinical trials concerned with the benefits of information to the population of current and
where the results of the trial accumulate over time and can be used to assign future patients. These approaches do not explicitly model the expected costs
entrants to the different arms (Armitage 1985). An example of th is type of and benefits to individual trial entrants. It is possible to model the costs and
approach is Bather's 'play the winner rule' where patients are assigned to the benefits of those enrolled in the trial, but this issue of ' individual ethical
arm of the trial which appears to be most effective given the accumulated trial concerns' may be more appropriately left for those who are responsible for
results (Bather 1 98 1 ) . This approach, as well as others addressing the same the ethical approval. This approach, however, will allow us to establish the
problem and more recent dynamic programming approaches to this problem opportunity cost to 'collective ethics' of holding particular ethical concerns
(Carlin et al. 1994), do not explicitly consider the marginal cost and benefits for trial entrants. For example, the ENBS of a proposed trial represents the
of sampling, and tend to focus on minimizing the potential health cost to opportunity cost to other current and future patients of failing to approve it
individuals enrolled in the trial. on ethical grounds. Similarly the ENBS following data monitoring represents
However, the framework of value-of-information analysis can also be the opportunity cost of failing to continue the trial due to concerns for those
usefully applied to sequential trial designs to establish the optimal allocation enrolled.
of each patient or group of patients and when to stop the trial. In principle an
7.3.4. Relevant alternatives
EVSI and ENBS calculation can be conducted after each result: the parameters
are updated and the EVSI and ENBS are recalculated based on the new posterior The framework offered by value-of-information analysis can also be used to
distributions. The next patient or group of patients should simply be allocated identify which alternatives should be compared in a proposed clinical trial.
to the arm that provides the highest ENBS. Patients should continue to be The example of zanamivir conveniently considered only two alternatives.
recruited and allocated in this way until the ENBS for the next patient or In most decision problems, however, there are potentially many possible
group of patients is zero. At this point it is no longer efficient to continue the strategies. Clearly, any economic evaluation should compare all relevant alter­
trial. The trial should be then be stopped and the accumulated sample size and natives. However, it is not often made clear what does and what does not
allocation will be optimal. In principle, this seems very straightforward and in constitute a relevant alternative. In practice, many feasible alternatives
many circumstances could lead to an optimal solution. However, it should be are ruled out as irrelevant during the design of prospective research. If the
recognized that in deciding how to allocate the next patient, or whether to identification of'relevant alternatives' to be included in prospective research is
stop the trial, the EVSI calculation should take account of all the ways that the either arbitrary, or uses implicit decision rules which are inconsistent with
future trial entrants could be allocated as well. This is computationally very those that will be applied when the study is complete, then there is a danger
demanding and may require dynamic programming (Bellman e t al. 1962; that research will revolve around an arbitrarily selected subset of strategies.
Claxton and Thompson 2001). There is a real possibility that the optimal strategy may have already been
ruled out of the analysis prematurely as an 'irrelevant alternative'. The evaluation
7.3.3. Data monitoring and ethics of a particular clinical decision problem should, at least initially, consider all
Even when trials are not designed to be sequential there is often an opportunity feasible alternatives, rather than focus only on those currently used or those
to monitor the accumulated results for safety and to identify whether the trial identified as of interest in some arbitrary way. The 'relevant alternatives' that
220 I EFFICIENT RES EARCH DESIGN
1 D E S I G N DECISIONS I 221

should be compared in prospective research should be identified explicitly and c

£1 ,800,000

consistently. g rsd, LORe, LORh

� £1 ,600';000 ,

o ',.)"
I
If alternatives cannot be ruled out a priori but the prospective evaluation of -�I·--- - - -
:s I
-----
_ - -- - - -
£1 .400;000 I
- - -t
- - - -qsd
all feasible alternatives is not possible or efficient, then this poses the question <l> r /
.
I
--_
---

� £1 ,200.000
.
I I
of which of the many competing alternative strategies should be considered ..
I

'relevant' and be compared in future research. We have already illustrated the I �

� £1 ,000,000 ' /
! I
I
I
I
I

use of EVPI in identifying a strategy which is not relevant and can be excluded I ,
I
I
I

I �c £800,000 :
... .... ..
t, .. , · · · .. · ······,,···
I pip
in Table 6.2. However, most strategies will contribute to the EVPI and we will
·,
, , r t

!
.,
<l> , "
I I I
.0 I
i
.. I I I
not be able to exclude them on these grounds. Optimal sample allocation Q5
£600,000 ,
, I I I
, c ,
I I I
,
based on ENBS enables relevant alternatives that should be compared I '0 I I I
I. <l>
£400,000 I
13 : ./ I I I
in prospective research to be identified explicitly and consistently. Relevant
I
' :
<l> I I I
£200,000 I:
"

I
"- I I I
alternatives can be defined as those where it will be efficient to allocate some ,lj I I I
of the sample in a subsequent evaluative trial. vI/hen it is not efficient to allocate £0 I I
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
patients to an arm of a trial, then that alternative can be safely ruled out as '" co " "' <0 .... <Xl 0> 0 '" co " "'

Sample size (n)

� � �

irrelevant (Claxton and Thompson 2001).

It is also worth noting the flaw in the notion that dominated programmes Fig. 7.1 Expected net-benefit of sampling and optimal sample size for alternative
(whether strongly or extendedly dominated) can be excluded from an analysis types of study.
and from subsequent research. It is possible that a dominated alternative will
be relevant in subsequent evaluation (i.e. may have a high probability of being
cost-effective), while a uondominated but not optimal alternative may not we should choose the study with the highest ENBS. In this case, this would
(i.e. may have a very low probability of being cost-effective). In these circum­ be the clinical trial that provided information on all parameters illustrated in
stances, the concept of dominance is not only unnecessary when using net­ Fig. 7.6. However, we could consider conducting a number of different studies
benefit, it is also potentially misleading. It is only by explicitly considering to inform different parameters, specifically, a clinical trial to inform rsd, LORe
the marginal costs and benefits of sample information that a consistent and LORh, and an epidemiological study to inform pip and a survey to inform
and rational definition of what constitutes the set of relevant and irrelevant upa. It should be recognized that the ENBS for this portfolio of studies will
alternatives is available (Claxton and Thompson 2001). not be the simple addition of the ENBS for each separate study in Fig. 7.7.
Assuming that these three studies will report at the same time, so that the
7.3.5. A portfolio of research designs
results of one cannot inform the results of the others, the ENBS would need to
The earlier discussion of EVSI and ENBS considered a single trial, which be calculated for each combination of sample sizes across the three studies.
would provide sample information on all parameters in the decision model. The optimal portfolio of research would be the combination of sample sizes
However, this is only one possible design. The EVSI and ENBS could be across the studies that maximized ENBS. This clearly requires substantial
considered for a trial that only included clinical endpoints such as rsd and computation, particularly if we also consider the optimal allocation of samples
LORe and LORh, which are particularly vulnerable to selection bias. In addition, within, as well as between, studies.
rather than establish EVSI and ENBS for a clinical trial, we could consider a A more realistic situation, however) is that the results of one study will be
simple epidemiological study which would provide information about pip or available before the other studies are completed. Indeed, it may be worth wait­
a survey of quality of life with influenza to provide additional information ing until one study reports before initiating an other. In other words, as well
about qsd. The ENBS and optimal sample size for these three additional as considering the optimal combination of studies, we can also consider the
potential study designs are illustrated in Fig. 7.7. optimal sequence of studies within this portfolio. Now the computation is
The ENBS and optimal sample size differ across these alternative studies. very demanding, as the EVSI for the first study depends on all the possible
If we were restricted to choosing only one of the four possible study designs future studies that could be conducted given all the possible results of the first.
222 I EFFICIENT RESEARCH DESIGN REFERENCES I 223

It should be clear from this discussion that although ENBS provides a powerful reimbursement rules, it can also be used to explore the type of regulatory
tool to identify optimal research design, the dimensions of this design space regimens that will generate the socially optimal level of evidence to support a
make this a computationally challenging task. technology{Hawkins and Claxton 2003).

7.3.6. Development decisions 7.4. Summary

The discussion around value of information has been restricted to evaluative
The purpose of this chapter has been to demonstrate how we can move from
research and we have not explicitly considered the role of value of information
the necessary condition that the expected value of perfect information exceeds
in informing development decisions. In principle, value�of-information
the cost of research to the sufficient condition that the marginal benefits of
analysis can be used to inform the whole life cycle of the development, evalua­
sampling exceed the marginal costs. The way in which the expected value of
tion and implementation of a health technology. Of course there will be
sample information can be calculated was then reviewed, drawing parallels
substantial uncertainty surrounding the expected costs and effects of a tech­
with the calculation of the expected value of perfect information from the
nology which is in development. Therefore} a decision to continue to invest in
previous chapter. Examples of how to undertake these calculations were then
developing a technology should be based not only on the expected payoff
given for different types of parameter distribution. Finally, it was demonstrated
(expected net-benefit from a societal perspective) but also the value of infor­
how the framework offered by the expected value of sample information
mation. For example, even if the expected payoff from a technology in devel­
approach could be brought to bear on a number of different design questions
opment is negative, if the value of information is high then it may still be
that naturally arise in health care evaluation research.
worthwhile conducting further research, as there is a chance that the technol­
ogy will turn out to be cost-effective. The development of a technology should
References
stop if the expected payoff is negative and the costs of further investigation
Ades, A. E., Lu, G. and Claxton, K. (2004) 'Expected value of sample information in medical
exceed the value of information (i.e. it does not appear that it will be cost­
decision modelling'. Medical Decision Making, 24: 207-227.
effective and further research is unlikely to change this conclusion). In this
Armitage, P. (1985) 'The search for optimality in clinical trials', International Statistical
way the framework of value-of-information analysis can be used to focus on Review, 53:15-24.
those technologies which are likely to be worth developing and identify the Bather, J. A. (1981) 'Randomised allocation of treatments in sequential experiments',
type of research that is needed within the development programme. Journal of the Royal Statistical Society B, 43: 265-292.
Throughout Chapters 6 and 7, the value of information has been considered Bellman, R. K and Dreyfus, S. E. (1962) Applied dynamic programming, Princeton
from a societal perspective, that is, based on maximizing net-benefit or equiv­ University Press.
alently maximizing health outcome subject to a resource constraint. However, Berry, D. A. and Stangl D. K. ( 1996) Bayesian biostatistics. New York: Marcel Dekker.
there is no reason why the same type of analysis cannot incorporate other Burl:s A., CI�k W., Preston c., et' ai. (2000) Is zanamivir effective for the treatment of
mjluenza In adults (supplement), London, National Institute for Clinical Excellence.
objectives} such as a commercial objective to maximize profit. For example, a
Carlin, B.P., Kadane, J. B., and Gelfand, A. E. (1992) �pproaches for optimal sequential
commercial payoff function could be based on the licensing decision, modelled decision analysis in clinical trials: Biometrics, 54: 964-975,
as a frequentist hypothesis test of efficacy; a reimbursement decision, based Claxton, K. ( 1999) 'The irrelevance of inference: a decision making approach to the stochastic
on expected cost-effectiveness; and sales (conditional on successful licensing evaluation of health care technologies: Journal of Health Economics, 18: 342-364.
and reimbursement), also a function of either expected effectiveness or cost­ Claxton, K. and Posnett, J. (1996) �n economic approach to clinical trial design and
effectiveness. Once a commercial payoff function has been specified, the research priority setting', Health Economics, 5, 5 13-524.
value-of-information calculations will provide the value of conducting further Claxton, K. and Thompson, K. A, (2001) 'A dynamic programming approach to efficient
clinical trial design', Journal of Health Economics, 20, 432-448.
research from this commercial perspective. This can inform stop-go decisions
Gelman, A., Carlin, J. B., Stern, H. S, and Rubin, D. B. (1995) Bayesian data analysis,
at Phase II, the design of a portfolio of research which may be needed to
London, Chapman and Hall.
support this decision and support licensing and reimbursement at Phase III. Hawkins,N. and Claxton, K. (2003) 'Optimal drug development programmes and efficient
As value of information can be estimated based on both commercial and licensing and reimbursement rules: the role of value-of-information analysis', Medical
societal payoffs to drug development decisions for a range of licensing and Decision Making, 23(6): 564.
 1
!
224 I EFFICIENT RESEARCH DESIGN

National Institute for Clinical Excellence (NICE) (2000) 'Guidance on the use of zanamivir Chapter 8
(Relenza) in the treatment of influenza', Technology Appraisals Guidance No 15. , ,

London, NICE.
O'Hagan, A. and Luce, B.(2003) A primer on Bayesiall statistics in health economics and future challenges for
outcomes research. Bethesda, MD, MedTap International.
Pratt, J., Raiffa, H. and Schlaifer, R. (1995) Statistical decision theory. Cambridge, MA, cost-effectiveness modelling
MIT Press.
Raiffa, H. and Schlaifer, R. ( 1959) Probability and statistics for business decisions. New York, of health care interventions
McGraw-Hill.

This book has attempted to introduce the reader to some of the more sophisti­
cated aspects of applying cost-effectiveness models to health care decision
problems relating to the implementation of new technologies. We started by
briefly reviewing the common types of decision model employed for this task
in Chapter 2 and went on to consider more sophisticated extensions to the
standard 'cohort' model in Chapter 3. Chapter 4 considered how parameter
uncertainty could be captured in our decision models through the use of para­
metric distributions, while Chapter 5 showed how this uncertainty in the input
parameters of the model could be analysed in terms of the uncertainty relating
to decision itself, principally through the use of CEACs. Chapter 6 argued that
traditional rules of inference or consideration of Bayesian posterior probabili­
ties have no role for decision making and that current adoption decisions
should be made on the basis of expected values with additional research
collected to support those decisions where it can be demonstrated that such
additional information is valuable. Chapter 7 extended the value-of-information
framework to consider the value of sample information and show that
this framework has the potential to answer not only the question of whether
additional information should be collected, but also the appropriate study
design for doing so.
While the intention in writing this book was never to take a direct philosoph­
ical stand on the issue of statistical analysis of data, it should be clear to the
reader that the methods and techniques we describe sit most comfortably in a
Bayesian framework, with parameters assigned distributions to reflect their
uncertainty, inferential decision rules abandoned and direct consideration
given to the probability of error together with the consequences of error.
Nevertheless, the typical reader we anticipated as the audience for this book
was probably schooled (like us) in classical statistics, and could (like some of us)
be working within departments dominated by the notions of ,evidence-based'
medicine where confidence intervals and significance tests reign supreme.
226 I FUTURE CHALLENGES FOR COST�EFFECTIVENESS MODELLING EVIDENCE SYNTHESIS A N D CHARACTERIZING U N C E RTAINTY I 227

Furthermore, the normative idea of (evidence�based' medicine is one that we assumptions may have limited impact on estimates of cost�effectiveness, they
would sign up to in that our task in developing probabilistic cost-effectiveness may have a more substantial impact on the value of information. This suggests
models of health care interventions is to accurately describe the current that uncertainty, and therefore evidence, about the structural relationship may
evidence that bears on a particular decision problem rather than to include be as valuable as-evidence about the model parameters. These types of uncer­
our own subjective priors, or those of our clinical colleagues. In the absence of tainty may be modelled as scenarios, or more usefully by assigning additional
hard data relating to a parameter, we feel most comfortable assigning a non­ uncertain parameters to represent alternative assumptions. For example, in
informative prior reflecting the absence of evidence. This recognizes that the Chapter 3, the use of standard survival analysis methods for calculating Markov
analysts/decision makers are acting not for themselves, but on behalf of society transition probabilities was introduced. V\There there is uncertainty over the
as a whole. Perhaps our particular stance of adopting an informal approach choice of functional form, then alternative functional forms could be included
to Bayesian methods may infuriate the purists, but our hope is that this will as scenarios in the model. Alternatively, a general form such as the Weibull
encourage more, rather than fewer} analysts to consider the merits that function could be specified, and uncertainty over the functional form could be
a Bayesian approach brings to the problem of health care evaluation. represented through the ancillary parameter. This approach enables this
In trying to bring together a text on cost-effectiveness modelling, represen­ source of uncertainty to be fully integrated so that the value of information
tation of uncertainty and value-of-information analysis that is essentially associated with these uncertainties can be established (Ginnelly and Claxton
practical in nature, there is an inherent danger that we give the impression 2005). Nevertheless, such an approach raises the issues associated with the
that all the problems and issues are resolved. In truth, this is far from the case general problem of Bayesian model averaging (Draper 1995), in particular the
and our intention in this final chapler is to reflect on the challenges that are issues associated with an infinite number of potential model structures.
likely to face analysts attempting to implement the techniques we describe and Clearly, more complex model structures and computational expensive
the likely areas of future development. Below we consider future challenges models make the characterization of uncertainty and value-of-information
under three broad headings: structuring of decision problems; evidence analysis more computationally demanding. As mentioned in Chapter 3,
synthesis and uncertainty; and value-of-information techniques themselves. a number of approaches can be taken to overcome the computational problems,
including the use of more efficient programming platforms, mathematical
approximations and the use of emulators (Stevenson et al. 2004; O'Hagan
8. 1 . Structuring decision problems ef af. 2005). This raises the issue whether it is more appropriate to specify a
Identifying an appropriate structure for a decision problem is, of course, crucially highly complex model structure, such that some form of approximation is
important for estimating cost-effectiveness. A single structure for a model does necessary to undertake the uncertainty and value�of�information analysis, or
not, exist, however, and different ways of structuring a model can have important whether it is justifiable to use simplifying assumptions to keep the model
consequences for the characterization of uncertainty and the subsequent value­ manageable.
of-information calculations. Indeed, the value of information will in general be
highly sensitive to model specification, and a number of particular issues have
8 . 2 . Evidence synthesis and characterizing uncertainty
been highlighted in recent applications of value of information.
Ensuring a sufficiently wide scope for the assessment to include all the Chapter 4 focused on fitting distributions to represent uncertainty in different
relevant alternative strategies is particularly important. This includes other types of parameters. Most of the simplified examples focused on fitting
technologies as well as different clinical policies (e.g. start and stop criteria) parameter distributions based on single data sources guiding estimation of the
for a single technology. As discussed in the previous two chapters, the exclu­ parameter. In reality, there will often be multiple sources of evidence and the
sion of alternative strategies may not change the overall guidance on use of a identification and synthesis of all the available evidence for all the parameters
technology, but in some cases it may have a substantial impact on the value of in the model is essential to appropriately estimate the model parameters and
information and on any research recommendations. characterize the uncertainty surrounding their values (Ades et af. 2005).
Exploring and reflecting the additional uncertainty surrounding alternative The methods for evidence synthesis of treatment effect are well developed
but credible structural assumptions is also required. Even when the alternative (Sutton et al. 2000) and in principle should be applied to all model parameters.
228 I FUTURE CHALLENGES FOR COST-EFFECTIVENESS MODELLING ISSUES SPECIFIC TO VALUE-OF-!NFORMATION ANALYSIS I 229

In practice, the effect measure is usually the focus of formal evidence synthesis the alternative with the greatest expected net-benefit, variability will not be
in most economic modelling exercises to date (Cooper et al. 2005). At several import�nt in"many models. In some nonlinear model structures, however,
points the need for comparison of all treatment alternatives has been empha­ variability as well as uncertainty will affect the expectations of cost, outcome
sized and this will often necessitate indirect comparisons (even where head­ and, therefore, net�benefit. In these circumstances the analysis must integrate
to-head trials are available). Such indirect comparisons are necessary for a full both the variability in the parameter and the uncertainty in our estimate of
comparison of all treatment options) but are always subject to an increased its value.
level of uncertainty, and will clearly effect estimates of uncertainty and the
value of information. The correlation between parameters which results from
this type of evidence synthesis can add to the computational burden of value­ 8.3. Issues specific to value-of-information analysis
of-information analysis and restricts the way parameters can be grouped in Although the key challenges are more general and relevant to estimating
EVPPI and EVSI calculations (see Chapters 6 and 7). cost, effect and decision uncertainty, a number of issues specific to value-of­
In principle, we should also reflect any additional uncertainty due to potential information analysis need to be addressed.
biases in the evidence, which may come from different types of study andlor Estimating the effective population that may benefit from additional
suffer from publication bias. Similarly we should also attempt to model the evidence requires an estimate of the time horizons for different technologies
exchangeability of the evidence available with the parameters required in the and clearly has a significant impact on the estimates of population EVPI and
modeL reflecting any additional uncertainty in our parameter estimates. population EVSI. A number of approaches to estimating the time horizon and
Scenarios have been· used to explore these issues of bias and exchangeability effective population have been taken in the value-of-information literature,
(see Chapter 6) in a very similar way to issues of structural uncertainty. including considering only the current population; arbitrary time horizons
However, as with structural uncertainty, it is more useful to reflect the addi­ (l0, 20 years, infinite) and threshold population values. In addition, empirically­
tional uncertainty in our parameter estimates based on elicitation of priors based estimates of the time horizons for decisions in health exist and can be
andlor estimating the possible bias or exchangeability. The practical challenges incorporated in the analysis. However, all these approaches implicitly use time
associated with explicitly modelling the potential for bias as part of decision horizon as a proxy for future changes in technologies, prices and information.
models is huge, and the role for experimental designs for minimizing bias, It is possible to explicitly model these future changes. Such analysis shows that
particularly in treatment effects, remains strong. the value of information for the decision problem may increase or fall over
A very similar issue to exchangeability is the inclusion or exclusion of what time, but the value of information for the group of parameters that can be
are regarded as unrelated events from the evidence and the potential role of evaluated by current research tends to decline. Therefore, finite and infinite
prior elicitation from experts. The cost-effectiveness of a treatment strategy time horizons for the decision problem represent special cases (e.g. significant
and the value of information can be significantly affected by the inclusion price shock or no changes respectively). By explicitly modelling future change
(exclusion) of unrelated (related) events (see Chapter 6). Often there is it is also possible to inform the timing of research, specifically, whether the
considerable uncertainty about how observed effects on events, which are not research should be conducted now or delayed if further technologies are
believed to be connected to the treatments under consideration, should be expected to be developed. Of course, modelling future uncertain changes is
dealt with. In such instances, elicitation of prior beliefs on the magnitude and challenging and it may be that the use of finite time horizons are a reasonable
extent of any potential treatment effects on these events may have an important proxy for this more complex process.
role in integrating this additional uncertainty into the analysis. The EVPI for each patient group is useful in identifying where research will
Finally, in some cases it will be important to deal simultaneously with be most valuable (see Chapter 6 ) and which type of patients should be
heterogeneity (variability by observed patient characteristics), variability enrolled in any future trial. However, it must be recognized that further
(variability by unobserved characteristic) and uncertainty. It should be recog­ evidence about one subgroup may inform other patient groups as well. The
nized that the analysis discussed in previous chapters has primarily focused on subgroup EVPI is therefore likely to be a lower bound on the value of conduct­
uncertainty and has appropriately dealt with heterogeneity by conditioning ing research on that subgroup alone. However, the summation of value of
the analysis on observed patient characteristics. As our purpose is to identify information across subgroups will overestimate the value of research for all
230 I FUTURE CHALLENGES FOR COST-EFFECTIVENESS MODELLING REFERENCES I 231

patient groups together. What is required is to model the exchangeability benefits (the expected sunk costs avoided). However, a more complete analysis
between subgroups so that additional information about one generates value of thi,s issue'maybe offered by real options analysis, While the current empha­
to others. This is also true of decision problems: additional evidence about sis of this 'approach has been on financial options, in the form of various
one decision problem may also be useful in informing other decision problems derivative secti:i'ities, the methods have been extended to the evaluation of
for patients with different indications or even different clinical areas (e.g. real assets (e.g, real estate, proposed environmental developments etc,), These
using the same technology for a different indication). This is very similar to have become known as real options, Despite its analytical appeal, the
the subgroup problem and we must recognize that the value of information techniques of real options analysis have yet to be applied in practice in
for one decision problem is likely to be the lower bound on the value of the evaluation of health care in any substantive manner (Palmer and Smith
research. In principle, the value of information can be viewed at a system level 2000).
(the whole allocation problem for the health care system). This means there is There are, of course, many challenges when attempting to estimate costs and
an overall cost of uncertainty (EVPI) surrounding the whole allocation problem, effects across a range of possible interventions, over a relevant time horizon
and the value of information about a particular technology can been seen as and for specific patient groups, while attempting to represent fully the uncer­
the EVPPI for the parameters relevant to that technology. In other words, tainty surrounding the decision. The issues of interpretation of evidence,
what we have described as EVPI for the decision problem is really an EVPPI synthesis, potential bias, exchangeability and appropriate model structure
calculation when we place it in the context of the whole system. have always been present in any informal and partial review of evidence.
Value-of-information analysis indicates the societal value of further research In fact, until quite recently, these challenging issues could be conveniently
when decisions based on current, perfect or sample i.nformation are fully ignored by both policy makers, clinicians and analysts, while decision making
implemented. Although it is important to distinguish the value of information was opaque and based on implicit criteria and unspecified 'weighing' of the
and the value of implementation (see Chapter 6), relatively little consideration evidence, These challenges must be faced as more explicit and transparent
has been given to incorporating the value of implementation within the same approaches to decision making are being taken. Indeed, one of the many
analytic framework. In a budget constrained health care system, the decision advantages of a more transparent and explicit approach to decision making
to invest in implementation strategies must he made alongside those regarding and the characterization of uncertainty, is that it exposes many important
investment in health care services and further research. It is possible to present methodological issues previously avoided by presenting partial analyses which
both the value of information and the value of implementation strategies do not directly address the decisions which must be made in any health care
separately but simultaneously. An upper bound on the value of adopting system,
implementation strategies (expected value of perfect implementation) can be
established with current information and with perfect information based on
the methods described in Chapter 6. This can give an indication of where References
priorities for research, implementation or some combination of the two Ades, A. E" Sculpher, M, J., Sutton, A" Abrams, A, Cooper. N., Welton, N. et al. (2006)
should lie (Fenwick et al. 2004). 'Bayesian methods for evidence synthesi& in cost·effectiveness analysis',
PharmacoEconomics, 24: 1-19.
The discussion throughout Chapter 6 and 7 has assumed that decisions to
Cooper, N. J Coyle, D., Abrams, K. R., Mugford, M. and Sutton, A, J. (in press) 'Use of
.•

adopt or reject a technology are reversible with no sunk costs. That is, we can
evidence in decision models: An appraisal of health technology assessments in the UK
adopt now, conduct further research and then, as necessary, reverse this deci­ to date'. Journal of Health Services Research and Policy, 10: 245-250.
sion in the light of the new information. As noted in Chapter 5, however, if Draper, D. (1995) 'Assessment and propagation of model uncertainty', Journal of the Royal
there are substantial sunk costs associated with adopting a technology, or Statistical Society, Series B, 57: 45-97.
even irreversibility of the decision, then it may be better to delay adoption Fenwick, E., Claxton, K. and Sculpher, M. J. (2004) 'The value of implementation and the
while additional information is generated through research. One reason value of information: Combined and uneven development: Medical Decision Making,
24: 6 (Abstract).
why we might anticipate sunk costs is the cost of implementation discussed
Ginnelly, 1. and Claxton, K. (2005) 'Characterising structural uncertainty in decision
above. It is possible to estimate the expected opportunity cost of delay analytic models: review and application of available methods: Medical Decision Making,
(net-benefit forgone) and compare these costs of delay with the expected 25: 6 (Abstract).
232 I FUTURE CHALLENGES FOR COST-EFFECTiV E N E S S MODELLING

O'Hagan, A., Stevenson, M. and Madan, J. S. (2005). 'Monte Carlo probabilistic sensitivity
analysis for patient-level simulation models'. Sheffield Statistics Research Report
561105. Sheffield, University of Sheffield.
Palmer, S. and Smith, P. C. (2000) 'Incorporating option values into the economic evaluation
of health care technologies: Journal of Health Economics, 19: 755-766.
Stevenson, M. D., Oakley, J. and Chilcott, J. B. (2004) 'Gaussian process modelling in
conjunction with individual patient simulation modelling: a case study describing the A adding time-dependency 46-56
acceptability curves 126-30 in combination with decision
calculation of cost-effectiveness ratios for the treatment of established osteoporosis',
ACE inhibitors tree 46
Medical Decision Making, 24: 89-100. relaxing Markov assumption 57-8
EUROPA study 134-44
. F. (2000) Methods for
Sutton, A. J.,Abrams, K. R., Jones, D. R., Sheldon, T. A. and Song, costs 140-2 vs patient-level simulations 58-65
meta�analysis in medical research. WIley & Sons. Markov model 135-6 cohort simulation 33-6
QALYs 137-40 life expectancy calculations 36
risk equations 136-7 conditional probability 20
regression methods to represent confidence intervals, distributional
heterogeneity 134 assumptions 84
treatment of coronary disease 97-100, coronary disease
134-44 ACE inhibitors 97-100, 134--44
acute coronary syndrome see coronary disease acute coronary syndrome
age, generalistic linear models decision tree and Markov model
(GLMs) 105-7 in combination 47
age-related macular degeneration, Weibull model 54--6
NICE case study 1 9 1-3 correlation
AIDS see HIVJAIDS model parameters 95-7
analysis of covariance (A.�COVA) rank order, generating 96-7
methods 130-2 cost-benefit analysis,
appendix, technical, probabilistic decision Pareto improvement 3
modelling 1 16-18 cost-effectiveness acceptability curve
AZT see zidovudine (CEAC) 122-30, 167
EUROPA study 146
B HIV/AIDS model 122-30
Bayesian decision theory 166-7 influenza model 167-8
binomial distribution multiple 133-4
beta distribution 86-7 representing uncertainty 132-4
fitting by method of moments 88-9 cost-effectiveness analysis
central limit theorem 85-6 decision rules 4-5
breast cancer study future challenges 225-31
CEA using Markov model 64--5 measures of health care 3-4
PSA using Markov model 64--5 QALYs 4-5
using Weibull model 55-6 social decision making 3
see also decision modelling for CEA
C cost-effectiveness frontier 168
central limit theorem, normal cost--effectiveness plane 122-6
distributions 85--6 HIV/AlDS model 127-S
cervical cancer, NICE case study 191-3 cost-effectiveness ratios, interval
Cholesky decomposition estimates 126-30
alternative 102 cost-minimjzation analysis 4
multivariate normal distribution 95-6 cost-utility data
choleHerol, and myocardial infarction, logistic analysis, use of QALYs 4
regreSSion/probabilities 97-100 choosing parameters 91-2
clinical trials, trial-based economic diabetes, regression models 104-6
evaluation 8-9 gamma distribution, fitting 91-2
dopidogre1, NICE case study 191-3 parameters, summary 108
cohort models 22-37 primary care, ordinary least squares
decision tree 23-7 regression of age 104 .
extensions 45-58 regression models 104-6
 l
234 INDEX I N DEX I 235

covariance matrices 99, 103 E F fixed sample design 217

Cox proportional hazards model 50 EBNS (expected net-benefit of sampling) first order unc�rtainty, variability 81 NICE case study 191-3
, / information, perfect 170-9
201, 214-16, 220-2
D economic evaluation G ',�:' input/output parameters 130
decision analysis alternative paradigms 2-3 gastrosophageal reflux disease (GORD)
defined 5, 6 treatment 25-,-'30 L
cost-benefit analysis 3
CEACs 1 51--4 life expectancy calculations, cohort
role in economic evaluation 5-9 social welfare 2--3
decision modelling for CEA 15--43 cost-effectiveness assessment 145-55 simulation 36
defined 2
case studies decision modelling for CEA, Markov loss function (EVPI) 171
methods, early use 1
GORD treatment 25-30 trial-based 8-9 models 25-30
HIV treatment 30-6 decision tree 25 M
ethics, research design 218-19
cohort models 22-37 Monte Carlo simulations 150 Markov assumption 36-7
EUROPA study (ACE inhibitors in treatment
extensions 45-58 outcomes 147-8 relaxing, extensions to cohort
of coronary disease) 134-44
comparators 7 probabilistic assessment models 57-8
cost-effectiveness acceptabmty curve
defined 6 of model 149-55 Markov models 27-37
(CEAC) 146
expected values 20-1 results from mode1 148-55 advantages 27-30
Monte Carlo simulations 146
further developments 45-76 Gaussian process 61 cohort simulation 33-6
parameter uncertainty 142--4
models vs trials 8-9 generalistic linear models (GLMs) 105-7 decision modelling for CEA
risk equations 138-9
parameters, choosing distributions probabilistic analYSis 107 GORD treatment 25-30
evidence synthesis of treatment 227-9
84-93 exercises glycoprotein antagonists, NICE case study HIV treatment 30--6
payoffs 21-2 information on 10-----1 1 191-3 exercise 38--42
potential for bias 80 tunnel state 58
constructing a Markov model of hip
probabilities 20 H Markov transition probabilities 31-3
replacement 67--75
role in economic evaluation 6-3 H2 receptor antagonists, deCision modelling, estimation of discrete time transition
analysing simulation results 155-60
stages of development 15-19 Markov model 27-30 probabilities 53
introducing a third prosthesis 160-2
defining/structuring 16-17 EVPI, hip replacement model 195-9 Helicobacter pylori infection, patient-level time-dependency 46-56
identifying evidence 17-18 simulation 62-5 survival analysis 50-6
probabilistic analysis
specifying problem 15 heterogeneity 81-2 time in model 48
hip replacement model 1 1 1-16
value of additional research 18-19 defined 19, 81-2 time in state 48-9
HIVIAIDS model 109-11
synthesis 7 modelling 132-3 migraine treatment, decision tree 24
replicating a Markov model 38--42
time horizon 7-8 regression models 97-107 Monte Carlo simulations 123--4
expected net-benefit of sampling (ENBS) 201,
uncertainty and heterogeneity 8, 19 hip replacement EUROPA trial 146
2 14-16, 220-2
see also cost-effectiveness expected value of perfect information exercises GORD treatment 150
analysis; probabilistic analysing simulation results 155-60 HIV/AIDS model 123-4, 127
(EVPII 170--9
decision modelling constructing a Markov model 67-75 multinomial distribution
calculation, alternative treatments 175
decision problems, structuring 226-7 introducing a third prosthesis 160-2 Dirichlet distribution 87-8
EVPP1 210-13
decision rules Weibull model of late prosthesis polytomous case 1 1 7
example 185-9
standard CEA failure 101-2 multiple sclerosis, NICE case study 1 9 1-3
linear models 182-3
extended dominance 5 HIV/AIDS model myocardial infarction
multiple treatment options 183-5
ICERs 5 ANCOVA analysis 130-2 logistic regression/probabilities 97-100
parameters 181-2
decision tree 23-7 exercise 195-9 cost-effectiveness acceptability curve Weibull model 54-5
in combination with cohort models 46 loss function 171 (CEAC) 122-30
examples 24 cost-effectiveness plane 127-8 N
NICE case studies 178, 18i'
GORD treatment 25 decision modelling for CEA, Markov National Institute for Health and CUnical
nonparametric approach 174-9
migraine treatment 24 models 30-6 Excellence (NICE)
for parameters 179-89
Monte Carlo simulations 123--4, 127 case studies, expected value of pedect
and Markov model in combination 47 research priorities 189-94
decision uncertainty 19 using normal distribution 172-3 probabilistic analysis, exercise 109-11 information (EVPI) 173, 187-8
see also uncertainty in value-of-information analysis 229-30
compared with Scottish Medicines
Consortium 80
decision modelling expected value of sample information
incremental cost-effectiveness ratio Technology Assessment Reports (six case
diabetes mellitus (EVSI) 201-16
(lCER) 5, 122-4 studies) 191-2
regression models to estimate for beta binomial 208-11
infections, static vs dynamic models 65--7 net-benefit, expected net-benefit of sampling
cost-utility data 104-6 conjugate distributions 204-6
inference 168-9 (ENBS) 201, 214-16
UKPDS Outcomes Model 59, 60 example (influenza model) 207-11
traditional rules 168 net-benefit framework 128-30
Dirichlet distribution linear models 206-7
influenza model net-benefit statistic 122
multinomial distribution 87-8 log odds ratios 212
cost-effectiveness acceptability curve NHS Economic Evaluation Database,
sampling 117-18 net-benefit normally distributed 202-4
(CEAC) 167-8 trial"based economic
discrete-event simulation 59 research design 212-23
EVP1 185-9 evaluation 8-9
dynamic models, vs static models, using normal distribution 2 1 1
EVSI 207-Il nonlinear transformation, expectation 78
infections 65-7 expected values 20-1
236 I INDEX INDEX I 237

normal distributions, and central limit Q survival analysis (Continued) uniform distribution, and cumulative
theorem 85-6 quality"adjusted life-years (QALYs) 4-5 hazard rates vs Markov transition distributions function 93-5
NSTACS (non-ST elevation acute coronary EUROPA study 137-40 probabilities 50 utilities see cost-utility data
syndrome) 47, 54-6 HRQoL 140 time-depend�ncy 50-6
lifetime cost per 133 V
o T value-of-information analysis 79, 165-200,
opportunity costs 5 R tamoxifen, breast cancer study using Markov 229-30
random number generator (RNG) 93 mode1 64-5 current and future decisions
p randomized controlled trials (ReI's), sole time-dependency 79-80, 194
parameter uncertainty 19, 82, 83--4 source of evidence 8-9 adding to Markov transition models 46-56 variability
probabilistic assessment 144 rank order correlation, generating 96-7 time in model 48 defined 81
parameters rates, vs probabilities, defmed 5 1 time in state 48-9 first order uncertainty 81
choosing distributions 84-93 regression models 83-4 time-to"event data, probabilities 101-4 varicella vaccination programme,
costs 91-2 heterogeneity 97-107 transition matrix 39-41 study 65-7
probability parameters 86-9 logistic regression (probabilities from transition probabilities see Markov transition Visual Basic Editor, simulation
relative risk parameters 89-90 binomial data) 97-100 probabilities output 124-6
triangular 92-3 relative risk triangular distributions 92-3
utilities 92 parameters 89-90 tunnel state, Markov models 58 W
correlating 95-7 summary 108 1\'10 by two table, relative risk 89 Weibull model 54
summary of common types 108 two by two table 89 acute coronary syndrome
Pareto improvement 2-3 research design 201-24, 2 12-23 U (NSTACS) 54-6
patient�level simulations 58-65 data monitoring and ethics 218-19 UKPDS acute myocardial infarction 54-5
examples and comparison with cohort design decisions 216-17 ordinary least squares regression of age breast cancer study 55-6
models 62-5 EVSI and EBNS 220-2 (cost of primary care) 104 correlating parameters 103
features 58--61 ftxed sample design 217-18 Outcomes Model 59, 60 hazard of failure 101-4
Helicobacter pylori infection 62-5 identifying priorities 189-94 uncertainty in decision modelling 8, 19, late hip prosthesis failure 101-2
vs cohort models 58-65 necessary but not sufficient condition 19 82-3, 165-200 Weibull regression, STATA output 55
payoffs 21-2 relevant alternatives 2 1 9-20 first/second order uncertainty 81, 82
polytomous case, multinomial distribution sequential design 2 1 8 need for additional evidence 166 z
117 trialMbased economic evaluation 8-9 nonlinear models 78 zanamivir
primary care, cost-utility data 104--6 see also expected value of sample parameter/model uncertainty 82 influenza model 167-8x, 185-9
probabilistic analysis information (EVSI) possible strategies 165-6 EVSI 207-11, 2 1 7
GLMs 107 risk, and outcome 169 role 77-8 zidovudine, decision modelling,
hip replacement model, exercise 1 1 1-16 value of information 79, 165-200 Markov model 30-6
HIV/AIDS model, exercise 109- 1 1 S
probabilistic decision modelling 77-120 Scottish Medicines Consortium (SMC),
appendix 116-18 compared with NICE 80
option values and policy screening, Helicobacter pylori infection,
decisions 79-80 patient-level simulation 62-5
parameters, choosing distributions 84-93 simulation output
role of uncertainty 77-8 analysis 144-5
simulation output 12 1-62 probabilistic decision
uncertainty and nonlinear models 78 modelling 121-62
value of information 79 two treatment alternatives 122-30
probabilistic sensitivity analysis (PSA) 61 Visual Basic Editor 124-6
breast cancer model 64-5 social decision making 3
probabilities social welfare theory 2-3
concepts 20 standard distributions, and central limit
from binomial data, logistic regression theorem 85-6
97-100 standard regression model 83-4
time-to-event data 1 0 1-4 standard two by two table, relative risk 89
vs rates, defmed 5 1 static vs dynamic models
probability parameters infections 65-7
distributions 86-93 varicella vaccination programme 65-7
generation 100 survival analysis
summary 108 estimating probabilities from timeMtoMevent
proton pump inhibitors (PPIs), decision data 1 0 1-4
modelling, Markov model 30-6 fundamental relationships 50-6