Decision Analysis of Surgical Treatment Indications For Metastatic Epidural Spinal Cord Compression

Decision Analysis of Surgical
Treatment Indications for Metastatic

Epidural Spinal Cord Compression
by
Markian Pahuta
Thesis submitted to the

Faculty of Graduate and Postdoctoral Studies
In partial fulfillment of the requirements
For the Ph.D. degree in
Epidemiology
School of Epidemiology and Public Health

Faculty of Medicine
University of Ottawa

c Markian Pahuta, Ottawa, Canada, 2019
Abstract
Metastatic epidural spinal cord compression (MESCC) occurs when tumour invades the
epidural space and compresses the spinal cord. Despite Level 1 evidence that surgery is
the most effective treatment for MESCC, there is controversy regarding the role of surgery
because of fear that patients who have a short survival will spend a large fraction of their
remaining life recovering from surgery and potential complications. This controversy could
be resolved by decision-analysis of MESCC treatments using quality-adjusted-life-years
(QALYs).
There have been two barriers to conducting decision-analysis of QALYs for MESCC:
(a) lack of utility data, and (b) skepticism regarding decision-analysis. The first four
research chapters in this thesis address these barriers. The final research chapter reports
a decision-analysis of QALYs on the role of surgery in MESCC.
Chapter 1 provides background information on the controversy regarding surgical treat-
ment for MESCC and the rationale for each of the subsequent chapters. Chapter 2 reports
a psychometric validation study of a web-based utility valuation module for MESCC. In
Chapter 3, application of this module to a general population utility valuation study with
a market research panel is described. In Chapter 4, the beneficial properties of Bayesian
statistical analysis to minimizing “arbitrariness” in probabilistic sensitivity analysis are
described in relation to prognostication for MESCC. Chapter 5 presents a strategy for
simplifying and enhancing the transparency of Markov cohort simulation. Finally, the
work presented in the research chapters is applied in Chapter 6 to conduct Markov cohort
simulation to determine if patients with short survival derive net health-related quality-of-
life benefit from surgery.
Pragmatic research around barriers to decision-analysis of QALYs for MESCC was
conducted to resolve the controversy regarding the role of surgery in the treatment of
MESCC. Under most circumstances, MESCC patients who can ambulate prior to treatment
derive net HRQoL benefit from surgery, even if prognosis is poor. Non-ambulatory patients
can derive net HRQoL benefit but only if the morbidity of surgery is relatively low. It is
my hope that the work used to address barriers to decision-analysis of QALYs will be
disseminated and applied in other clinical problems.
ii
Preface/Statement of Academic Achievement
This manuscript-based thesis consists of a series of articles prepared for publication in

a peer-reviewed journal.
The first paper, “The Psychometric Properties of a Self-Administered, Open-Source
Module for Valuing Metastatic Epidural Spinal Cord Compression Utilities,” was published
in PharmacoEconomics – Open in 2018.[167] I designed the website described in this paper
but retained a web programmer to program and implement the code. The second paper,
“A General Population Utility Valuation Study for Metastatic Epidural Spinal Cord Com-
pression Health States,” was accepted for publication by Spine in January 2019.[168] The
third paper, “Back to Bayesian: a Strategy to Enhance Prognostication of Metastatic Spine
Disease,” was accepted for publication by the International Journal of Clinical Practice in
February 2019.[169] The fourth paper, “A Technique for Estimating Approximate Tran-
sition Rates from Published Partitioned Survival Analyses,” was accepted for publication
in Cost Effectiveness and Resource Allocation in June 2019. I developed the strategy and
programmed the models. The fifth paper, “Decision Analysis of Prognostic Contraindi-
cations to Surgery for Metastatic Epidural Spinal Cord Compression,” will be submitted
to a peer-reviewed journal. I programmed the Markov cohort simulation model using the
statistical programming language R.
I am the first and corresponding author on all papers. I wrote all papers with guidance
from supervisors and other co-authors. I submitted all manuscripts for publication and
addressed all feedback during the peer-review process.
iii
Acknowledgements
First and foremost, I would like to thank my accomplished, intelligent and beautiful wife
Dr. Therese Kozminski for her love, unwavering support, patience and prodding during this
drawn-out thesis process. I am notorious for multitasking and so embarked on additional
“projects” during my time in the PhD program. Therese and I married in my first year in
the PhD program. She gave birth to our first son Julian in year two of the PhD. I obtained
my Royal College Fellowship in Orthopaedic Surgery in year three during which we also
moved from Ottawa to Toronto so that I could start my fellowship in Spine Surgery. That
same year, Therese gave birth to our first daughter Kalyna. During year four we emigrated
from Canada to the USA where I accepted an academic spine surgeon position. Therese
not only tolerated my relative absence from our family while I was completing my PhD,
but she sacrificed her career as a Family Physician while encouraging me, motivating me,
and making sure I had time to finish my thesis.
Julian and Kalyna, I regret having to frequently hide in the office to work on my thesis.
I hope that you will learn from my example and become more efficient and productive
than I am. I look forward to spending more time with you and not having to divide my
attention.
I would like to express my deepest appreciation to my supervisor Dr. Doug Coyle
and co-supervisor Dr. Carl van Walraven. Doug: thank you for your mentorship and
willingness to engage with me in esoteric discussions on health economics and decision
analysis. Carl: thank you for reviewing esoteric components of this thesis with a fine-
tooth comb between the hours of 5 and 6 am so that I could continue working on them
the same day. Doug and Carl: thank you for bailing me out when I was kicked-out of the
PhD program because I did not check my University of Ottawa e-mail account in a timely
manner. Your thoughtful insights and feedback impressed me and challenged me to grow
as a researcher. Your engagement has strengthened this thesis.
I am extremely grateful to the members of my thesis committee: Dr. Eugene Wai and
Dr. Joel Werier. You have been outstanding clinical mentors. It is because of working with
both of you that I pursued spine surgery with a focus in oncology. Your involvement has
been instrumental in making this thesis clinically relevant. You were both my champions
during residency, and I know you worked behind the scenes to ensure I was well supported
during my hiatus from clinical residency training. I value you as colleagues and hope we
can collaborate in the future.
iv
Table of Contents
Nomenclature viii
List of Tables ix
List of Figures xi
1 Introduction 1
1.1 History of MESCC Treatment . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Surgical Decision Making for MESCC . . . . . . . . . . . . . . . . . . . . . 3
1.3 Barriers to Applying Decision-Analysis of QALYs to MESCC . . . . . . . . 5
1.4 Overview of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4.1 Synthesis of Research Chapters . . . . . . . . . . . . . . . . . . . . 8
2 SOAP-MESCC Psychometric Evaluation 9

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3 General Population Utility Valuation Study 21

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Material and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.6 Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.7 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
v
4 Bayesian Prognostication 36
4.1 Clinical Scenario . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Dichotomous Prognostication . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.4 Continuous Prognostication . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5 Approximating Transition Rates 47

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.2 Transition Rate Approximation . . . . . . . . . . . . . . . . . . . . . . . . 52
5.3 MESCC Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.4 Simulation Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
6 Decision Analysis 62
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
7 Discussion and Conclusion 77

7.1 Summary of Methodological Findings and Contributions . . . . . . . . . . 77
7.2 Summary of Main Clinical Findings . . . . . . . . . . . . . . . . . . . . . . 79
7.3 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
7.4 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
7.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
APPENDICES 83
A Appendices for Chapter 2 84

A.1 SOAP MESCC Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
A.2 SOAP platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
A.3 MESCC Health States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
vi
B Appendices for Chapter 3 110
B.1 Regression Model for Utilities . . . . . . . . . . . . . . . . . . . . . . . . . 110
B.2 Quota Samples Based on 2016 Canadian Census . . . . . . . . . . . . . . . 112
B.3 Regression Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
C Appendices for Chapter 4 129

C.1 Confidence Intervals for Median Survival Time . . . . . . . . . . . . . . . . 129
C.2 Sensitivity Analysis Through Effective Sample Size Adjustment . . . . . . 131
D Appendices for Chapter 5 132

D.1 Transition Parameters for an Illness-Death Model . . . . . . . . . . . . . . 132
D.2 Multistate Estimation of Transition Rates . . . . . . . . . . . . . . . . . . 134
D.3 Worked Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
E Appendices for Chapter 6 145

E.1 Bayesian Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
E.1.1 Bayesian MCMC Estimation of Event Probabilities for the Treatment-
Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
E.1.2 Bayesian MCMC Estimation of Transition Rates . . . . . . . . . . . 153
E.1.3 Bayesian MCMC Estimation of a Mixed Effects Beta Regression
Model for Utilities . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
E.2 Exact Formulas for Markov Cohort Simulation . . . . . . . . . . . . . . . . 167
References 168
vii
Nomenclature
cS+RT laminectomy without stabilization followed by external beam radiotherapy
DOC docetaxel
EORTC European Organisation for Research and Treatment of Cancer
HRQoL health-related quality-of-life
ICC intra-class correlation coefficient
MCID minimally clinically important difference
MESCC metastatic epidural spinal cord compression
mS+RT modern surgery followed by external beam radiotherapy
NSCLC non-small cell lung cancer
OS overall survival
PFS progression-free survival
QALY quality-adjusted life year
RAM ramucirumab
RT-alone external beam radiotherapy alone
SBRT stereotactic body radiation therapy
SOAP Self-directed Online Assessment of Preferences
viii
List of Tables
2.1 MESCC Health State Attribute Phrasing . . . . . . . . . . . . . . . . . . . 14

2.2 Participant Characteristics Stratified by Response Pattern . . . . . . . . . 18
2.3 Agreement, Reliability, and Responsiveness Results . . . . . . . . . . . . . 18
3.1 Participant Demographics . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

3.2 Utilities for all MESCC Health States . . . . . . . . . . . . . . . . . . . . . 34
3.3 Selection Parameters for all Models . . . . . . . . . . . . . . . . . . . . . . 35
3.4 Summary of Beta Regression Analysis for Model 4 . . . . . . . . . . . . . . 35
5.1 Data Abstracted from MESCC RCT PFS and OS Analyses . . . . . . . . . 54

5.2 Comparison of True and Approximate Transition Rates and Hazard Ratios
for MESCC RCT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.3 Simulation Results for (λhi + λhd ) . . . . . . . . . . . . . . . . . . . . . . . 59
5.4 Simulation Results for ρ . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.5 Simulation Results for λid . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.1 Selection Criteria used in the MESCC RCT[171] . . . . . . . . . . . . . . 67
B.1 Quotas Based on the 2016 Canadian Census . . . . . . . . . . . . . . . . . 112

B.2 Functional Form of Linear Predictor Formulas for Beta Regression . . . . 116
B.3 Selection Parameters for all Models . . . . . . . . . . . . . . . . . . . . . . 116
B.4 Calibration and Agreement Parameters for all Candidate Models . . . . . . 117
B.5 Mean Utilities by Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
B.6 Linear Predictor Formulas with Coefficients . . . . . . . . . . . . . . . . . 119
B.7 Utilities for all MESCC Health States using Full Dataset . . . . . . . . . . 128
C.1 Parameters used for Interpolation of var [S (tm )] . . . . . . . . . . . . . . . 130
ix
D.2 Comparison of Transition Rates and Hazard Ratios Estimated with Multi-
state Techniques under Short and Long Simulation Conditions . . . . . . . 143
E.1 Comparison of results from short and long simulation conditions for multi-
state estimated transition rates and hazard ratios . . . . . . . . . . . . . . 148
E.3 Comparison of Transition Rates and Hazard Ratios Estimated with Multi-
state Techniques under Short and Long Simulation Conditions . . . . . . . 162
E.4 Comparison of results from short and long simulation conditions for beta
regression coefficients for utilities . . . . . . . . . . . . . . . . . . . . . . . 164
x
List of Figures
3.1 Flow of participants. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.2 Calibration Plot for Model 4. . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.1 State-transition diagram for illness-death model . . . . . . . . . . . . . . . 50

5.2 Comparison of state-membership fractions, RT-alone arm . . . . . . . . . . 57
5.3 Comparison of state-membership fractions, mS+RT arm . . . . . . . . . . 58
6.1 State-transition diagram for decision model . . . . . . . . . . . . . . . . . . 68

6.2 Utilities by number of dysfunctions . . . . . . . . . . . . . . . . . . . . . . 70
6.3 Individual simulation ∆QALYs versus number of dysfunctions . . . . . . . 72
6.4 Mean ∆QALYs versus number of dysfunctions . . . . . . . . . . . . . . . . 73
A.1 MESCC SOAP: page 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

A.2 MESCC SOAP: page 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
A.3 MESCC SOAP: page 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
A.4 MESCC SOAP: perfect health . . . . . . . . . . . . . . . . . . . . . . . . . 89
A.5 MESCC SOAP: page 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
A.6 MESCC SOAP: probe health state . . . . . . . . . . . . . . . . . . . . . . 91
A.7 MESCC SOAP: standard gamble interface . . . . . . . . . . . . . . . . . . 92
A.8 MESCC SOAP: perfect health pop-out . . . . . . . . . . . . . . . . . . . . 93
A.9 MESCC SOAP: probe health state pop-out . . . . . . . . . . . . . . . . . . 94
A.10 MESCC SOAP: logic check . . . . . . . . . . . . . . . . . . . . . . . . . . 95
B.1 Variance function check. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

B.2 Link function check. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
xi
Chapter 1
Introduction
Over one-third of patients with carcinoma or hematological malignancy will develop spinal
metastases.[73, 163] A particularly disabling consequence of these metastases occurs when
tumour invades the space between the vertebra and spinal cord (epidural space) and com-
presses the spinal cord (metastatic epidural spinal cord compression, MESCC).[14, 62] The
clinical presentation of MESCC is variable owing to the complex functional anatomy of the
spinal cord. Depending on the axial location of compression, patients may experience a
motor and/or sensory disturbance. Furthermore, depending on the axial and cranio-caudal
location of compression, the neurologic deficit may affect any combination of extremities
and the trunk.[216]
A motor deficit can manifest as isolated weakness in a group of muscles, paraplegia,
hemiplegia, or quadriplegia. Sensory impairment can present as numbness, pain, or loss
of spatial sense in any limb. In concert, a sensory and motor dysfunction can cause loss
of hand function, ambulation, urinary and/or faecal continence. These consequences of
MESCC can be quite distressing to patients and may lead to clinical depression.[195]
Aggravating these functional losses is the loss of independence.
Patients must adapt to these physical and functional limitations. With lower extremity
dysfunction, patients may not be able to independently transfer from bed and will require
a lift or assistance from others.[158] Furthermore, immobility predisposes patients to skin
ulcers and therefore prophylactic skin care must be performed.[34] Sphincter dysfunction
is managed by diapers, urinary catheters and bowel regimes.[27, 44] If upper limbs are
affected, the manual dexterity to perform basic daily activities such as dressing, toileting
and bathing can be lost.[208] Unfortunately, patients with general functional decline from
cancer may not have the stamina to learn new adaptive skills.[218] MESCC patients with
severe neurologic deficits may be placed in long-term care or hospice institutions if there
are insufficient resources to care for them at home. MESCC can have a marked negative
impact on health-related quality-of-life (HRQoL).
1
1.1 History of MESCC Treatment
Surgery for MESCC is a controversial topic. Perceptions on the role of surgery were shaped
by high-quality comparative studies published in the late 1970s and early 1980s that found
surgery followed by external beam radiotherapy to be inferior to conventional external
beam radiotherapy alone (RT-alone) for maintenance of ambulation and survival.[83, 255]
Consequently initial surgical indications for MESCC were quite limited:[220, 42]
1. For tissue diagnosis
2. Following deterioration during or relapse after RT-alone
3. Inability to administer RT-alone due to spinal cord volume-dose limit having been
reached
It is important to recognize that the surgical technique used in these early studies – laminec-
tomy without stabilization – is not routinely used in the modern surgical era. Herein,
classical surgical treatment will refer to laminectomy without stabilization followed by
conventional radiotherapy and abbreviated as “cS+RT.”
Laminectomy is often ineffective for MESCC because the majority of spinal metastases
originate in the vertebral body (which is in the anterior portion of the spinal column).[83]
Therefore, for the majority of patients, a laminectomy will not effectively decompress
the spinal cord because laminectomy only directly decompresses the posterior aspect of
the spinal cord. Despite this limitation, in cancer patients who are often medically frail,
laminectomy was initially preferred over direct decompression of the vertebral body be-
cause it avoided the morbidity of formal anterior approaches.[83, 245, 85] It was only after
the mid-1980s that less morbid techniques for direct decompression of the vertebral body
through an extra-cavitary approach were popularized.[220]
Even if the neural elements are successfully decompressed with a laminectomy, pa-
tients who have no stabilization can experience late neurologic deterioration. During a
laminectomy, stabilizing ligaments and joints are removed.[184] When the vertebral body
is weakened by tumor, the destabilized posteriorly decompressed segment can result in
kyphosis and cause neurologic injury.[70, 103, 221] Due to better understanding of spinal
stability, and improved instrumentation, modern surgical techniques for MESCC routinely
include stabilization.[14, 71, 118, 204]
Modern surgical treatment for MESCC with direct decompression and stabilization fol-
lowed by conventional external beam radiotherapy addresses the key limitations of cS+RT.
Therefore, beliefs on the ineffectiveness of surgical treatment may not be applicable to the
modern surgical era. Herein, “mS+RT” will be used to denote the modern surgical treat-
ment approach. mS+RT has been subject to a randomized controlled trial and found to be
superior to RT-alone for maintenance of ambulation and survival.[171] Large observational
studies on mS+RT have found HRQoL benefit as well.[65]
More recently, treatment paradigms have changed based on advances in radiation
therapy.[14] In contrast to conventional external beam radiotherapy, stereotactic body
2
radiation therapy (SBRT) more precisely delivers higher ablative doses of radiation.[106]
Although radiation toxicity is lower with SBRT, high focal doses of radiation can predispose
to compression fractures in otherwise stable lesions.[197] With higher focal doses of radia-
tion, direct decompression may not be necessary because SBRT-alone can effectively treat
even moderate-grade symptomatic MESCC.[196] However, surgery does still play a role
in downgrading compression to facilitate SBRT and optimize local control.[190] Surgery
with this intent is termed “separation surgery” and consists of minimal decompression to
create a margin of cerebrospinal fluid (CSF) between the spinal cord and tumor. No Level
1 comparative studies of mS+RT to SBRT have been conducted.
1.2 Surgical Decision Making for MESCC

Although mS+RT is effective for MESCC, surgeons make this recommendation cautiously
due to prognostic uncertainty. It is generally recommended that appropriate surgical can-
didates have an estimated survival of 3 – 6 months.[42, 229, 232, 230, 133] Although
frequently used in the literature, there is no data to support this prognostic threshold
which is based on expert opinion alone.
In a 1986 review article, Sundaresan and colleagues first proposed a 6-month survival
threshold.[220] They noted that although neurologic response rates are equivalent with
cS+RT and RT-alone, cS+RT conferred a more durable response. Since the majority of
patients undergoing RT-alone relapsed within 6 months, they reasoned that patients with
predicted survival exceeding 6 months should be offered cS+RT to protect against late
neurologic deterioration. This logic may not apply to mS+RT because, in contrast to
cS+RT, neurologic response rates are higher with mS+RT than with RT-alone.[255, 171,
189]
Several prognostic indices have been developed and validated for estimating expected
survival of patients with metastatic spine disease.[172, 131, 211] The Tokuhashi Score, a
widely used survival index for these patients, considers performance status, extra-spinal
bone metastases, vertebral metastases, visceral metastases, pathologic diagnosis, and neu-
rologic function.[230] The prognostic value of the Tokuhashi score for predicting 6-month
survival has been studied in a recent meta-analysis.[133] The Tokuhashi score had a good
discriminative performance (area under the receiver operator curve of 0.748) between pa-
tients who die prior to, or survive beyond, 6 months.[64]
Systemic chemotherapy is an important component of treatment for all MESCC pa-
tients regardless of treatment offered for spinal cord dysfunction (RT-alone, cS+RT, or
mS+RT). Whereas the goal of RT-alone, cS+RT, and mS+RT is to improve physical func-
tion; systemic chemotherapy is administered to prolong survival. Due to rapid advances
in systemic therapies, some authors have pointed out that existing prognostic scores used
to guide surgical decision making may no longer be useful – in particular for non-small
cell lung cancer (NSCLC). For example, the Tokuhashi Score was developed in 1990 and
revised in 2005.[229, 230] During this time-period, only cytotoxic agents were available for
the treatment of NSCLC and, consequently, patients with advanced cancer had a median
3
overall survival of 4 – 6 months. Currently, patients with advanced NSCLC may be offered
targeted therapy such as anti-angiogenic drugs which has been shown to improve over-
all survival in NSCLC patients.[185] Prognostic uncertainty poses a significant challenge
for the surgeon when the complication profile of the two treatment options, mS+RT and
RT-alone, is considered.
Patients undergoing RT-alone may experience nausea, dermatitis, and symptoms from
inflammation of the gastrointestinal tract (pharyngitis, esophagitis, and gastritis) within a
month of receiving spinal external beam radiotherapy. Less than 10% of patients will report
these symptoms to be severe.[146, 145, 237] Early toxicity may be even less frequent with
the newer technique of stereotactic radiosurgery because radiation is delivered to a smaller
field.[107] Infrequently, patients may experience late toxicities months after treatment,
including myelopathy and vertebral compression fractures.[144] When compared to those
from mS+RT, complications from RT-alone are less common and relatively mild.
Analysis of 30-day National Surgical Quality Improvement Program data found that
14.4% of patients undergoing mS+RT for primary and metastatic spinal tumors expe-
rience a major complication with 5.3% undergoing a re-operation and 10.2% being re-
admitted.[121] Prospectively collected data from the AOSpine MESCC study reported the
most common complications to be urinary tract (25.0%) and wound infections (10.4%).[65]
Adverse events were significantly higher in a single-center cohort of patients undergoing
emergency surgery for spinal metastases when complications were studied as the primary
outcome.[47] In this study, 76.2% of patients experienced at least one complication, and the
mortality rate during the admission was 10.9%. Intraoperative adverse events occurred in
31.7% of patients and included: blood loss >2L (16.8%), incidental durotomy (9.9%), mal-
positioned implants requiring revision (5.9%), nerve root injury (2%), and cardiac events
(1%). Postoperative adverse events occurred in 20.8% of patients and included: delir-
ium (20.8%), transient neurological deterioration (5.9%), pressure sores (4.0%), deep vein
thrombosis (4.0%), early construct failure (2.0%), electrolyte imbalance (11.0%), arrhyth-
mia (4.0%), dysphagia (5.0%). The risks of mS+RT are substantial and can negatively
affect HRQoL. However, patients experiencing complications can still derive overall net
benefit from mS+RT.[47, 3]
The decision to offer mS+RT to a patient with MESCC is difficult because surgical
recovery and potential complications must be considered vis-à-vis an uncertain survival
time. Some clinicians may take a conservative approach and only offer RT-alone, and not
mS+RT, to MESCC patients out of concern for the “worst-case” survival scenario.[196, 136]
This is because recovery from surgery and potential complications will impact on a greater
fraction of remaining life in patients having short survival, thereby negating net HRQoL
benefit. A decision-analysis of quality-adjusted life years (QALYs) could help patients and
clinicians jointly assess the trade-offs between survival benefits, HRQoL benefits, recovery
and potential complications to reach an optimal treatment decision.[55, 123] This is because
QALY concurrently quantify morbidity and mortality within a single parameter. To date,
such an analysis has not been conducted for MESCC.
4
1.3 Barriers to Applying Decision-Analysis of QALYs
to MESCC
Barrier 1: Lack of Utility Data

QALYs are calculated using utilities, or health-related HRQoL weights, which can be
obtained by direct valuation or from generic health status measures.[249] Direct valuation
is the classical approach in which individuals rate hypothetical health state descriptions
using the time-trade-off or standard gamble procedures.[76] These procedures can be used
to measure utilities for very specific and uncommon health states. However, it can be
cumbersome to develop valid health state descriptions for particular diseases. Alternatively,
techniques have been developed to convert generic health status measures (e.g. EuroQol-
5D, Short Form-6D, or Health Utilities Index Mark 3) to utilities.[55] Conversion of generic
health state measures is advantageous because custom health state descriptions are not
required. However, utilities can only be obtained for health states actually observed in a
cohort of patients completing a generic health survey. For MESCC, few high-quality studies
compared mS+RT and RT-alone using generic health status measures, and no study has
conducted direct valuation. Therefore, with a paucity of utility data, the ability to apply
decision-analysis of QALYs to MESCC has been limited.
Barrier 2: Skepticism Regarding Decision Analysis

Once utility data are available, QALYs can be estimated directly from a study or calculated
using simulation-based techniques. Study-based QALY estimates can be prohibitive due
to large sample size requirements.[23] Furthermore, study-based decision-analysis is not a
viable option when few high-quality comparative studies have collected HRQoL data (as
in MESCC). Unfortunately, the alternative, simulation-based QALY estimates, are often
viewed with skepticism.[58]
Markov cohort simulation is the most frequently used technique for estimating QALYs
through simulation.[22] This technique first involves simulating disease history from a the-
oretical model of the disease process composed of distinct and clinically-relevant health
states. The fraction of simulated patients in each health state over time is simulated using
transition probabilities derived from the literature. QALYs are then calculated by weigh-
ing the fraction of simulated patients in each health state by the utility of those health
states, and then summing the weighed state-membership fractions over the time-horizon
of interest.
One criticism of Markov cohort simulation stems from current recommendations for
probabilistic rather than deterministic analysis.[21, 36] Deterministic analysis refers to
running the simulation model with only one “best” value for each of the model inputs. On
the other hand, probabilistic analysis involves randomly selecting many plausible values
for each model input and re-running the analysis with each random draw, the goal of
which being the examination of the sensitivity of conclusions to the uncertainty in model
5
inputs. The random draws are taken from probability density functions derived from
confidence intervals reported in the literature for each of the model inputs.[22] It is crucial to
appreciate that confidence intervals do not quantify the uncertainty in the model inputs.[28]
Furthermore, converting a confidence interval to a probability density function can be
arbitrary and subjective. Hence, Markov cohort simulation can be criticized because the
analysis can be biased by the choice of probability density functions.[21]
Many studies do not report sufficient data to properly compute transition probabilities
which are needed for Markov cohort simulation. Many studies will only report progression-
free survival (PFS) and overall survival (OS) analyses.[251] PFS is the probability of re-
maining healthy to time t; while OS is the probability of being alive (either healthy or ill)
at time t. Notably, the PFS and OS do not directly analyze ill patients. Multistate estima-
tion of transition rates is needed to properly calculate transition probabilities.[105, 4, 181]
Unfortunately sufficient data for multistate estimation is infrequently reported. Despite
this, transition probabilities are often opaquely and improperly derived from PFS and OS
analyses.[253]
Not surprisingly, another criticism of Markov cohort simulation is that the techinque is
complex and poorly reported in the literature.[57, 63, 63, 13, 56, 176, 192, 29] Perhaps due
to the resulting lack of transparency in modeling, some authors characterize simulation-
base techniques as a “black box.”[29] To be useful for treatment decision making, clinicians
and patients must have faith in, and understanding of, these models.[16]
1.4 Overview of Thesis

This thesis consists of seven chapters: an introduction and background chapter; five re-
search chapters; and a discussion chapter. The first four research chapters, Chapters 2, 3, 4,
and 5, address the barriers to decision-analysis in MESCC. In the final research chapter,
Chapter 6, I present a QALY decision-analysis of mS+RT and RT-alone for MESCC.

Given that few high-quality studies have compared mS+RT and RT-alone for MESCC
using generic health status measures, direct utility valuation is the most efficient way
to collect utility data. Best practices in economic evaluation involve recruiting a sam-
ple of healthy individuals from the general population for direct utility valuation (ex ante
utilities).[248, 198] It is important to appreciate that ex ante MESCC utilities from healthy
individuals may not be equivalent to ex post utilities from patients who have experienced
MESCC.[15, 256, 46] Although ex ante utilities are theoretically restricted to system policy
decisions,[187] ex ante utilities have become the de facto standard for individual patient de-
cisions. This is because utilities obtained from generic health surveys such as the EuroQol-
5D, Short Form-6D, and Health Utilities Index Mark 3 are actually ex ante valuations.[55]
Therefore, I sought to elicit ex ante MESCC utilities to conform with conventions in the
literature.
6
Traditional general population utility valuation are conducted using face-to-face inter-
views, phone interviews, or postal surveys.[59] These forms of survey administration are
time-intensive and costly. As a result, web-based surveys are increasingly used.[60, 61,
180, 101, 143, 201, 239, 84, 109, 240, 6, 117, 159, 247] Web-based studies are typically
conducted using proprietary software which limit application to other disease contexts.
Furthermore, the psychometric properties of these propriety software programs have not
been assessed.[223]
It is important to determine whether web-based utility valuation has acceptable psycho-
metric properties. If this mode of administration has acceptable psychometric properties,
it would be beneficial and efficient for other investigators to be able to build disease-specific
modules on a common platform that has been used to develop modules with acceptable
psychometric properties. To meet this need, in Chapter 2 “The Psychometric Properties
of a Self-Administered, Open-Source Module for Valuing Metastatic Epidural Spinal Cord
Compression Utilities,” I describe the development of a new open-source (non-proprietary)
web-based self-directed utility valuation platform, called the Self-directed Online Assess-
ment of Preferences (SOAP), that can be used on major computer systems (including
touch-screen devices). In Chapter 2, I determine the validity, reproducibility, and respon-
siveness of a SOAP module for MESCC.
Chapter 3 “A General Population Utility Valuation Study for Metastatic Epidural
Spinal Cord Compression Health States” presents a general population utility valuation
study using the SOAP MESCC module. For this study, a sample of 1138 adult Canadians
was recruited using a market research company. I report utility valuations for 31 unique
MESCC health states and explore the relative importance of MESCC-related consequences
on HRQoL.
Barrier 2: Skepticism Regarding Decision-Analysis

Regarding the concern that probabilistic analysis is arbitrary and prone to manipulation, I
propose that Bayesian statistics offer a solution. This branch of statistics was developed in
the 18th century and predates the more commonly used Frequentist statistics which were
developed in the early 20th century.[68] This type of analysis provides a rigorous and non-
arbitrary framework for analyzing data. The output of a Bayesian analysis is a posterior
distribution which can be used in place of user defined distributions in economic analysis.
The purpose of Chapter 4 “Back to Bayesian: a Strategy to Enhance Prognostication of
Metastatic Spine Disease” is to introduce the concepts of Bayesian statistics to clinicians.
This chapter also reviews the difficulties in treatment decision making for MESCC.
Credibility of Markov cohort simulation suffers when models are improperly populated
with the results of partitioned-survival analysis.[251] However, the correct technique for
these situations, partitioned-survival modeling, does not yield as informative of results as
state-transition modeling. Chapter 5 “A Technique for Approximating Transition Rates
from Published Survival Analyses” presents a solution to this problem. In this chap-
ter, I develop and validate a technique to derive transition probabilities from summary
partitioned-survival analysis data.
7
1.4.1 Synthesis of Research Chapters
The ultimate purpose of this thesis is to determine whether MESCC patients derive net
HRQoL benefit from mS+RT compared to RT-alone. This work is motivated by contro-
versy regarding the role of mS+RT despite Level 1 evidence that this treatment is more
effective than RT-alone for MESCC.[196, 136]
I will try to resolve the controversy regarding the role of mS+RT by determining if
patients with short survival derive net HRQoL with mS+RT relative to RT-alone. If
patients benefit from surgery even under the “worst-case” survival scenario, clinicians and
patients could make treatment decisions with greater comfort. In Chapter 6 “Decision
Analysis of Prognostic Contraindications to Surgery for Metastatic Epidural Spinal Cord
Compression,” I determine the conditions under which patients with poor prognosis benefit
from surgery using decision-analysis of QALYs.
The first four research chapters set the foundation for Chapter 6. In this final chapter,
I conduct Markov cohort simulation using a simple and transparent multistate MESCC
disease model developed in Chapter 5. QALYs are calculated using the utilities valued in
Chapter 3. In an effort to minimize perceptions of “arbitrariness,” the analysis is run using
posterior Bayesian distributions (Chapter 4).
8
Chapter 2
The Psychometric Properties of a

Self-Administered, Open-Source
Module for Valuing Metastatic
Utilities
Markian A. Pahuta1 , Aaron Frombach2 , Emile Hashem3 , Stewart Spence3 , Christina Sun3 ,
Eugene K. Wai4,5 , Joel Werier4 , Carl van Walraven5,6,7 , Doug Coyle5
1 – Divisions of Orthopaedic & Neurological Surgery, Henry Ford Health System, Detroit,
Michigan, USA
2 – Postgraduate Medical Education, The University of Ottawa, Ottawa, Ontario, Canada
3 – Undergraduate Medical Education, The University of Ottawa, Ottawa, Ontario,

Canada
4 – Division of Orthopaedic Surgery, The University of Ottawa, Ottawa, Ontario, Canada
5 – School of Epidemiology and Public Health, The University of Ottawa, Ottawa, On-
tario, Canada
6 – The Ottawa Hospital Research Institute, Ottawa, Ontario, Canada
7 – The Institute for Clinical Evaluative Sciences, Ottawa, Ontario, Canada
9
Context
In this chapter, I describe the development of a new open-source (non-proprietary) web-
based self-directed utility valuation platform, called the Self-directed Online Assessment of
Preferences (SOAP), that can be used on major computer systems (including touch-screen
devices). I also determine the validity, reproducibility, and responsiveness of a SOAP
module for MESCC. The purpose of this chapter is to present a valid tool that will be used
to address the lack of utility data for MESCC.
10
Abstract
Objectives
Web surveys are often used for utility valuation. Typically, custom utility valuation tools
that have not undergone psychometric evaluation are used.
This study aims to determine the psychometric properties of a metastatic epidural
spinal cord compression (MESCC) module run on a customizable open-source internet-
based self-directed utility valuation platform (Self-directed Online Assessment of Prefer-
ences; SOAP).
Methods
Individuals accompanying patients to the emergency department waiting room in Ottawa,
Canada were recruited. Participants made SOAP MESCC health state valuations in the
waiting room, and 48 hours later at home. Validity, agreement reliability, and responsive-
ness were measured by logical consistency of responses, Smallest Detectable Change, the
Interclass Correlation Coefficient, and Guyatt’s Responsiveness Index respectively.
Results
Of 285 participants who completed utility valuations, only 113 (39.6%) completed the re-
test. Of these 113 participants, 92 (81.4%) provided valid responses on the first test, and
75 (66.4%) provided valid responses on the test and re-test. Agreement for all groups
of health states was adequate since their Smallest Detectable Change was less than the
Minimally Clinically Important Difference. The mean Interclass Correlation Coefficient s
for all health states were greater than 0.8 indicating at least substantial reliability. Guyatt’s
Responsiveness Indices all exceeded 0.80, indicating high level of responsiveness.
Conclusions
To our knowledge, this is the first validated open-source, web-based, self-directed util-
ity valuation module. We have demonstrated the SOAP MESCC module to be a valid,
reproducible and responsive for obtaining ex ante utilities. Considering the successful
psychometric validation of the SOAP MESCC module, other investigators can consider
developing modules for other diseases where direct utility valuation is needed.
11
2.1 Introduction
Quality-adjusted-life-years (QALYs) are used to concurrently quantify morbidity and mor-
tality within a single parameter.[55] For this reason, QALYs can facilitate the discussion
of risks and benefits during patient counselling regarding treatment options.[123] To help
make funding decisions, policy makers may also combine QALYs with cost estimates to
calculate the incremental cost effectiveness ratio.[37] QALYs are calculated using “utili-
ties,” or health-related quality of life weights, which are obtained by direct valuation or
from generic health status measures.[249]
The choice of utility valuation approach is driven by available data. Direct valuation
is the classical approach in which individuals rate hypothetical health state descriptions
using the time-trade-off or standard gamble procedures.[76] These procedures can be used
to measure utilities for very specific and uncommon health states. However, it can be
cumbersome to develop valid health state descriptions for particular diseases. Alternatively,
techniques have been developed to covert generic health status measures (e.g. EuroQol-5D,
Short Form-6D, or Health Utilities Index Mark 3) to utilities.[55] Conversion of generic
health state measures is advantageous because custom health state descriptions are not
required. However, utilities can only be obtained for health states actually observed in a
cohort of patients involved in the generic health survey.
Unfortunately, generic health scores have not been collected for many diseases, thereby
making direct valuation a necessary option for measuring utilities. Best practices in eco-
nomic evaluation are to recruit a sample of healthy individuals from the general population
for utility valuation.[198, 248] Traditionally, general population utility valuation was con-
ducted using face-to-face interviews, phone interviews, or postal surveys.[59] These forms
of survey administration are time-intensive and costly, and as a result web-based surveys
are increasingly used. [60, 61, 6, 117, 160, 247, 179, 101, 143, 201, 238, 84, 109, 241]
Typically these studies are conducted using proprietary software which limits application
to other disease contexts. Furthermore, the psychometric properties of these propriety
software programs have not been assessed.[223] [84]
It is important to determine whether web-based utility valuation has acceptable psy-
chometric properties. If this mode of administration has acceptable psychometric proper-
ties, rather than building custom software for new utility valuation studies, it would be
beneficial and efficient for investigators to be able to build disease-specific modules on a
common platform that has been used to develop modules with acceptable psychometric
properties. To meet this need, we developed a new open-source (non-proprietary) web-
based self-directed utility valuation platform usable on major computer systems (including
touch-screen devices) called the Self-directed Online Assessment of Preferences (SOAP)
(Appendix A.1 and Appendix A.2). SOAP was designed with flexibility in mind and can
accept new health state descriptions (modules) with minimal programming.
We decided to first create a SOAP module for metastatic epidural spinal cord com-
pression (MESCC), a condition for which there is limited health-related quality of life
data. MESCC can be treated with surgery or radiotherapy. But few high-quality studies
compare these interventions using generic health status measures for patients. However,
12
surgery and/or radiotherapy outcomes could be compared using utilities obtained by di-
rect valuation of hypothetical probe health state descriptions. The European Organisation
for Research and Treatment of Cancer (EORTC) MESCC working group has developed a
health-related quality of life questionnaire for MESCC.[151] Items from this questionnaire
could be used to generate health state descriptions for a SOAP module.
The objective of this study was to determine if the SOAP platform can be used to
develop a valid, reproducible, and responsive module for MESCC. For this first application
of the SOAP platform, we developed a MESCC module based on the work done by the
EORTC and measured psychometric properties in a general population sample.
2.2 Methods
SOAP Platform
Electronic utility valuation protocols are distinguished by the form of health state de-
scriptions, assessment approach, navigation rules, and auxiliary functions.[138] A detailed
description of these elements for the SOAP MESCC module are provided in Appendix A.1
and Appendix A.2.
MESCC Module
EORTC Phase 1 development of a MESCC questionnaire in Canada found that patients
and health care providers felt that ambulation, urinary continence, pain, and independence
were important health-related quality of life issues for MESCC. Since Phase 1 development
was restricted to health-related quality of life, and did not specifically consider treatment
effects and adverse events, we reviewed prospective studies on MESCC to identify reported
outcomes and adverse events.[145, 171, 147, 65] The EORTC items captured all treatment
outcomes identified in our review. However, the review identified a large and disparate
set of adverse effects. To develop a manageable decision analytic model, all adverse effects
were grouped as an “other symptoms” attribute.
A tabular (point-form) presentation of health states was chosen as it is preferred by
participants, is believed to decrease cognitive burden compared to the narrative format,
and produces similar results to the narrative format.[142, 199] Therefore, we presented
health states as a point-form list of five dysfunctional attributes: non-ambulatory (N),
incontinent of urine (I), pain (P), dependent (D), and “other symptoms” (S). To reduce
the number of potential health states, EORTC items were collapsed to indicate the pres-
ence (+) or absence (-) of the dysfunctional attribute, producing 32 discrete health states
(Appendix A.3).
When possible, the phrasing for presence or absence of dysfunctional attributes was
created using same EORTC items identified in the MESCC module development process
(Table 2.1). Items were rephrased to the second person and restructured as declarative sen-
tences. Items describing feelings or worries were not utilized as it was desired to make the
13
Table 2.1: MESCC Health State Attribute Phrasing
Attribute Symbol Item EORTC Code
Dependence D- You do not need help with eating, C05, BM49

dressing, washing yourself or using the
toilet. You are not dependent on others.
D+ You need help with eating, dressing, C05, BM49
washing yourself or using the toilet.
You are dependent on others.
Non-ambulation N- You walk without assistance. NA
N+ You have lost mobility because of weakness BM48, BN48,
in your legs. You need to stay in bed or a C04
chair (or wheelchair) during the day.
Incontinence I- You have no trouble controlling your BN50
bladder.
I+ You need to insert a tube in your bladder BN50,
or you need to wear a diaper full time, to BLM44
control your bladder.
Pain P- You have back pain relieved by pain BM31, BM38
medications.
P+ You have back pain not relieved by pain BM31, BM38
medications.
Other Symptoms S- You have no other uncomfortable NA
symptoms.
S+ You have one or more of: nausea, C14, C15,
vomiting, shortness of breath, lack of C08, C17
appetite, diarrhea.
The symbol (+) indicates the presence of a dysfunctional attribute, and (-) indicates
absence. The European Organisation for Research and Treatment of Cancer (EORTC)
codes refer to identifiers in the EORTC Item Bank.
14
health states descriptions as objective as possible. The rationale for the specific attribute
formulation was:
1. Dependence (D). The two items identified by the MESCC working group were
combined into one attribute to highlight the implications of loss of independence.
The qualifiers “do” and “do not” were added to indicate complete function and
dysfunction.
2. Lack of Ambulation (N). The MESCC working group developed a new item which
was used as the functional level. Again, two items were combined to highlight the
implications of loss of mobility.
3. Incontinence of Urine (I). The item identified by the MESCC working group with
a qualifier was used as the functional level. An item from the EORTC bladder cancer
module (BLM44) was used to highlight the implications of loss of bladder control.
4. Pain (P). As MESCC can only occur in the cervical spine to the thoracolumbar
junction, pain was not differentiated by the terms “upper” and “lower” back as was
identified by the MESCC working group. As most patients with spine metastasis
will have some element of pain, the functional state had patients requiring pain
medications. Use of pain medication served as a qualifier, and was taken from the
EORTC bone metastasis module (BM38).
5. Other Symptoms (S). Again, to maintain efficiency, all adverse effects were char-
acterized by several common adverse symptoms. These items were all taken from the
core EORTC questionnaire.
Evaluations were obtained using standard gamble method using a ping-pong search
algorithm. Typically in the standard gamble, success is framed as “perfect” health for an
undermined period of time. In this context, this can be inferred to be the absence of any
dysfunctions. Therefore the fully functional health state (D-,N-,I-,P-,S-) was chosen as the
success anchor. To eliminate confusion around life expectancy, all scenarios were framed
as having a certain life expectancy of five years; that is, for both the probe health scenario
and success health scenario, participants were told their life expectancy would certainly be
five years. Five years was the maximum survival reported in a randomized controlled trial
on treatments for MESCC.[171] Probe health states were presented in a random order.
The MESCC module was pilot tested in a sample of 40 participants to assess accept-
ability and ease of use. Participants were asked to rate the SOAP MESCC module using
a five-point Likert rating for the statement “[t]his website is easy to use.” 92.5% of par-
ticipants strongly agreed or agreed with the statement.
Subjects
To be compliant with best practices in economic evaluation, we sought to conduct a direct
utility valuation study with a sample of the general population who have not experienced
15
MESCC using the SOAP MESCC module (ex ante valuation).[248] Prior to this general
population direct valuation study, psychometric properties of the SOAP MESCC module
had to be evaluated. To approximate a general population sample for this psychomet-
ric validation study, participants were recruited from the emergency department waiting
rooms at The Ottawa Hospital, an academic hospital in Ottawa, Ontario, Canada. Only
patients’ family members or friends (i.e. individuals accompanying patients) at least 18
years of age were eligible to participate. Participants were required to be able to read
English and have access to the internet outside of the hospital. A minimum sample size of
50 participants has been recommended in published guidelines for reliability and respon-
siveness evaluations.[223] To ensure robust results, we set the sample size for this study at
75.
Survey Procedures
Participants completed the first survey in the emergency department using a touch screen
device. Investigators did not assist participants in navigating or completing the survey.
Each participant valued the health state D+N+I+P+S+, one randomly selected singly
dysfunctional health state, and another triply dysfunctional health state. Dysfunctional
elements were nested to ensure a logical ordering of utilities for the three health states. For
example, if the singly dysfunctional health state was D-N-I+P-S-, the triply dysfunctional
health state could include incontinence and two of: dependence, lack of ambulation, pain,
or other symptoms.
Two days after the initial survey, investigators contacted participants through e-mail
and/or phone with information to access the re-test. Participants completed the second
survey using their personal device. For the retest, participants were presented with the
same probe health states as they completed in the emergency room, but the states were
presented in a random order (and not necessarily identical to the initial survey).
Statistical Analysis
Validity refers to whether a tool under investigation measures what it is supposed to
measure.[152] Specifically, construct validity concerns whether results obtained using the
tool under investigation are consistent with a priori hypotheses.[152] We hypothesized that
utility valuations should follow the logical ordering of health states with utility valuations
following the relationship: singly ≥ triply ≥ fully dysfunctional. We considered singly =
triply = fully a valid response because we could not exclude the possibility of a ceiling
effect with one dysfunction. Participant responses were deemed “valid” if their utilities
followed this order. The proportion of participants providing valid responses on the test
and re-test was computed.
Reproducibility concerns the stability of participants’ responses on repeated testing,
and can be characterized by agreement and reliability.[152] Agreement quantifies the ab-
solute differences in participants’ repeated responses. We assessed agreement using the
16
Smallest Detectable Change.[223] We classified agreement as adequate if the Smallest De-
tectable Change was less than the Minimally Clinically Important Difference (MCID).[223]
By anchoring to Eastern Cancer Oncology Group functional levels, an MCID of 0.05 for
cancer utilities obtained by the EuroQol-5D-3L has been proposed.[175] This MCID has
also been used for direct utility valuation by the standard gamble and time-trade-off of
EuroQol-5D-3L health states.[246] The precision of the standard gamble algorithm used in
our study was also 0.05. Therefore, we use an MCID of 0.05 in this study. Systematic dif-
ferences between the test and re-test sessions were quantified using the Smallest Detectable
Change calculation. Reliability concerns the fraction of pooled study variance across the
repeated tests attributable to differences between participants (participant variance) and
individual test-retest variability (noise).[152] If responses are stable, the ratio of noise to
participant variance should be small, and the ratio of participant variance to variance for
the pooled results from test and re-test should be high. Reliability accounting for system-
atic differences between the test and re-test, stratified for the number of dysfunctions in the
health state, was quantified using the Intra-Class Correlation Coefficient (ICC) using the
following categories: < 0.21, slight reliability 0.21–0.40, fair reliability; 0.41–0.60, moder-
ate reliability; 0.61–0.80, substantial reliability; > 0.80, almost perfect reliability.[129] An
ICC ≥ 0.70 was considered adequate.[223]
Responsiveness reflects the ability of a tool to detect clinically important changes and
can be quantified using Guyatt’s Responsiveness Index.[91] The Guyatt’s Responsiveness
Index is proportional to the ratio of the MCID to the root mean squared error of the
difference between the test and re-test value. If test-retest variability is small relative
to the MCID, the tool is deemed responsive because meaningful changes are of greater
magnitude than test-retest fluctuation.[90] Values of 0.20, 0.50, and 0.80 were interpreted
as small, moderate and large levels of responsiveness, respectively.[108]
Statistical analysis was performed using the statistical programming language R.[183]
The distribution of age (Kruskal-Wallis test) and gender (Chi-squared test) was compared
between participants providing valid and invalid responses on the test and re-test. We
considered logically ordered responses to be valid, that is decreasing utilities assigned to
the singly, triply, and fully dysfunctional states. Age was assessed using a one-way ANOVA.
Gender was assessed using the chi-squared test. Reproducibility, agreement, reliability, and
responsiveness were only measured on participants providing valid responses on both the
test and re-test. Since the SOAP tool is intended for measuring average utilities from the
general public, average measures (rather than individual measures) of Smallest Detectable
Change, ICCs, and Guyatt’s Responsiveness Indices were calculated.[45]
2.3 Results
Of 285 participants who completed utility valuations in the emergency department, only
113 (39.6%) completed the re-test. Of these 113 participants, 92 (81.4%) provided valid
responses on the first test, and 75 (66.4%) provided valid responses on the test and re-test
(Table 2.2). The response validity pattern was not associated with age (p = 0.2336) or
sex (p = 0.971) (Table 2.2). Only data from the participants providing valid responses
17
on both the test and re-test were used for reproducibility and responsiveness analysis.
Seven respondents skipped at least one scenario during the test and were classified as
providing invalid responses. Only one respondent skipped one question during the re-test,
and the responses were also classified as invalid. Agreement for all groups of health
Table 2.2: Participant Characteristics Stratified by Response Pattern
Valid Valid Invalid Invalid

p-value
to Valid to Invalid to Valid to Invalid
no. (%) 75 (66.4) 17 (15.0) 17 (15.0) 4 (3.6)
Age – yr† 53.0 43.0 51.0 43.0 0.2336
(33.5, 61.0) (33.0, 48.0) (38.0, 62.0) (33.25, 52.0)
Sex – no. (%) 0.971
Female 47 (62.7) 11 (64.7) 11 (58.8) 3 (75.0)
Male 28 (37.3) 6 (35.3) 6 (41.2) 1 (25.0)
†
median and interquartile range in parentheses.
states was adequate since their Smallest Detectable Change was less than the MCID of
0.05 (Table 2.3). Mean ICCs were all greater than 0.8 indicating substantial reliability,
and all ICCs were significantly greater than the pre-specified threshold of 0.7 (Table 2.3).
Guyatt’s Responsiveness Indices all exceeded 0.80, indicating large responsiveness for the
utility evaluation (Table 2.3).[90]
Table 2.3: Agreement, Reliability, and Responsiveness Results
Number Smallest ICC (95% CI)† Guyatt’s

of dysfunctions detectable change responsiveness
index
Single 0.036 0.879 (0.808, 0.923) 1.45
Triple 0.035 0.898 (0.838, 0.935) 1.47
Full 0.042 0.826 (0.724, 0.890) 1.24
†
Intra-Class Correlation Coefficient (ICC) and 95% Confidence Interval
(95% CI).
2.4 Discussion
Traditional utility valuation studies have been conducted using face-to-face interviews,
phone interviews, or postal surveys. These modes of administration have undergone psy-
chometric validation. There is increasing use of web surveys for utility valuation which
18
are usually conducted using custom and proprietary valuation tools which have not been
psychometrically validated. It would be beneficial and efficient for investigators to be able
to build disease-specific modules on a common platform that has been used to develop
modules with acceptable psychometric properties.
We developed a new platform called the Self-directed Online Assessment of Preferences
(SOAP) (Appendix A.1 and Appendix A.2). For the first application of this platform,
we developed a module for MESCC health states. The SOAP platform met published
benchmarks for reproducibility (both agreement and reliability) and responsiveness for
utility measurement. This study demonstrates that the SOAP platform can be used to
develop modules with acceptable psychometric properties.
81.4% of participants provided valid responses on the first test, and 66.4% of partic-
ipants provided valid responses on both the test and re-test. These results should be
considered in the context of other ex ante valuation studies reported in the literature. We
classified a participant’s responses valid if their utility valuations decreased with increas-
ing dysfunctional attributes in the health state. For example, if a participant valued the
fully dysfunctional health state higher than the single dysfunctional health state, their re-
sponses were classified as invalid. This definition of validity is termed “logical consistency”
and has been used in traditional general population ex ante utility valuation studies of
EuroQol-5D-3L health states.
Logical consistency rates for face-to face valuations have been reported for the UK and
Netherlands.[51, 128] In the UK study, 12 pairs of health states per participant, could be
evaluated for logical consistency. The median rate of logical consistency, per participant,
ranged from 83.8% to 91.7%. In the Dutch study, 87.6% of participants provided at
least one pair of logically inconsistent valuations. Postal surveys conducted in the US
and New Zealand reported at least one logically inconsistent pairing in 88% and 79% of
participants respectively.[113, 49] With 81.4% of participants providing a valid response
(28.6% providing a logically inconsistent response), the logical consistency rate for the
SOAP MESCC module is similar to traditional population studies. Logical consistency
and has also been assessed for other self-administered general population ex ante utility
valuation studies of EuroQol-5D-3L health states over the internet.[6, 12, 7] Each study
reported a logical consistency rate below 70%.
Compared to the SOAP MESCC module, the face-to-face, postal and web-based EuroQol-
5D-3L utility valuation studies required greater cognitive effort because participants rated a
greater number of health states (between 5-10) that were also more complex (5 attributes
and 3 levels of dysfunction). Furthermore, these studies did not provide error checking
whereas the SOAP MESCC module notified participants of a logical error if they rejected
a lottery with 100% of success. Considering these differences, a logical consistency rate of
81.4% on the first test with the SOAP MESCC module is consistent with the literature.
Valuing MESCC health states using the classical standard gamble is problematic for two
reasons. The first, the classical standard gamble uses perfect health as a top anchor which
is an unrealistic outcome for metastatic cancer. The second, the classical standard gamble
considers timeless (ie. perpetual) health states which are incongruent with the metastatic
cancer disease process. To make the standard gamble more realistic, we characterised
19
perfect health as the absence of dysfunctions, and restricted all health states (including
the top anchor) to a survival period of five years. These modifications may impact on the
interpretation of our results relative to classic utility assessment.
Utilities are typically estimated for specific health states and are used to weight the
time in such health states. Consequently, a utility value for a specific state is typically
considered “timeless,” that is utilities are usually assumed not to change with time spent in
a health state.[10] As a reflection of this, the duration of time spent in a probe health state
is not specified in the classical standard gamble.[76] For MESCC health states, we were
concerned that the most severe health states would connote poor survival, and therefore
confound the measurement of health-related quality of life using the standard gamble with
quantity-of-life. To alleviate this difficulty, we explicitly stated a five-year duration for
each health state which was the longest survival observed in a randomized controlled trial
of treatments for MESCC.[171] This approach has also been used in other utility valuation
studies for cancer health states.[214] This modification to health state descriptions should
not affect results because the standard gamble (and all other utility elicitation methods)
relies on the utility independence assumption.[177] Under this assumption, if a health state
has a utility of x, the utility of this health state for 5 years should still be x. Unfortunately,
a systematic review concluded that individuals tend not to satisfy the utility independence
assumption with no consistent pattern of violation.[234] We are unaware of any algorithm to
convert utilities for fixed period of time to “timeless” utilities. Consequently, the utilities
measured in this study may not be directly comparable to utilities obtained using the
classical standard gamble.
A strength of our study is that we built on the work conducted by the EORTC MESCC
working group to ensure that the attributes in the MESCC module were appropriate and
representative of the MESCC disease process. A limitation of our study is that we did
not assess criterion validity by comparing utilities obtained by SOAP MESCC and a “gold
standard.[152] This could be done by having MESCC patients to value their own health
using the SOAP MESCC module and comparing these utility valuations to those derived
from a generic health questionnaire. We did not have the resources to conduct such a study.
Furthermore, measures of logical validity, reproducibility and responsiveness are more rel-
evant than MESCC criterion validity to investigators considering developing modules for
new diseases.
To our knowledge, this is the first validated open-source, web-based, self-directed utility
valuation module. For the first application of the SOAP platform, we developed a module
for MESCC health states. We have demonstrated the SOAP MESCC module to be a valid,
reproducible and responsive for obtaining ex ante utilities. Considering the successful
psychometric validation of the SOAP MESCC module, other investigators can consider
developing modules for other diseases where direct utility valuation is needed.
20
Chapter 3
A General Population Utility

Valuation Study for Metastatic
Health States
Markian A. Pahuta1 , Eugene K. Wai2,3 , Joel Werier2 , Carl van Walraven3,4,5 , Doug Coyle3
Michigan, USA
tario, Canada
21
Context
In this chapter, I present a general population utility valuation study using the validated
SOAP MESCC module described in Chapter 2. This is the first comprehensive set of ex
ante utility estimates for MESCC health states. These data will be used to conduct QALY
decision analysis in Chapter 6.
22
Abstract
Study Design
General population utility valuation study.
Objectives
This study obtained utility valuations from a Canadian general population perspective
for 31 unique Metastatic Epidural Spinal Cord Compression (MESCC) health states and
determined the relative importance of MESCC-related consequences on quality-of-life.
Summary of Background Data

Few prospective studies on the treatment of MESCC have collected quality-adjusted-life-
year weights (termed “utilities”). Utilities are an important summative measure which
distills health outcomes to a single number that can assist healthcare providers, patients,
and policy makers in decision making.
Methods
We recruited a sample of 1138 adult Canadians using a market research company. Quota
sampling was used to ensure that the participants were representative of the Canadian
population in terms of age, gender, and province of residence. Using the validated MESCC
module for the “Self-administered Online Assessment of Preferences” (SOAP) tool, par-
ticipants were asked to rate 6 of the 31 MESCC health states, each of which presented
varying severities of 5 MESCC-related dysfunctions (dependent; non-ambulatory; inconti-
nent; pain; other symptoms).
Results
Participants equally valued all MESCC-related dysfunctions which followed a pattern of
diminishing marginal disutility (each additional consequence resulted in a smaller incre-
mental decrease in utility than the previous). These results demonstrate that the general
population values physical function equal to other facets of quality-of-life.
Conclusions
We provide a comprehensive set of ex ante utility estimates for MESCC health states that
can be used to help inform decision making. This is the first study reporting direct utility
valuation for a spinal disorder. Our methodology offers a feasible solution for obtaining
23
quality-of-life data without collecting generic health status questionnaire responses from
patients.
24
3.1 Introduction
The decision to offer a patient with metastatic epidural spinal cord compression (MESCC)
surgery followed by radiotherapy (S+RT) or radiotherapy alone (RT) is difficult because
survival, physical function, and complications must be considered jointly. Quality-adjusted-
life-year (QALY) analysis could allow patients and clinicians to determine the relative
weight of these factors and reach an optimal decision. QALYs are calculated using “util-
ities,” or health-related quality of life weights, which are usually derived from patient re-
sponses to generic health status questionnaires (e.g. EuroQol-5D, Short Form-6D, or Health
Utilities Index 3).[55] Unfortunately there is a paucity of quality-of-life data for MESCC
as few high-quality studies compare interventions using generic health status measures.
When generic health status measures are not available, utilities can be derived by direct
valuation. Direct valuation is the classical approach in which individuals rate hypothetical
health state descriptions using the time-trade-off or standard gamble procedures.[76] These
procedures can be used to measure utilities for very specific and uncommon health states.
Best practices in economic evaluation are to recruit a sample of healthy individuals from
the general population for direct utility valuation.[248, 198]
Based on the work of the European Organisation for Research and Treatment of Cancer
(EORTC),[151] our group has developed a comprehensive set of 31 unique MESCC health
state descriptions.[167] We have also developed an online utility valuation module for
MESCC using the “Self-administered Online Assessment of Preferences”(SOAP) tool. This
module was found to be valid, reproducible and responsive in a sample of individuals how
have not experienced MESCC.[167]
The primary objective of this study is to obtain general population utility valuations for
the 31 unique MESCC health states. The secondary objective of this study is to determine
the relative importance of various aspects to quality-of-life in MESCC.
3.2 Material and Methods
Subjects
Recruiting a general population probability sample is a time-intensive and costly under-
taking. Typically this is done by visiting homes, using random-digit-dialing, and selecting
random phone numbers. In Canada, with over 80% of Canadians age 16 years and older
having access to the internet,[94] web recruitment offers a practical alternative to tradi-
tional strategies. As has been done by several other investigators, we recruited participants
from a proprietary market research panel (Toluna Group Ltd) for this utility elicitation
study for MESCC.[6, 12, 7]
The market research panel was composed of over 80,000 individuals across Canada
recruited by the company into a panel of potential survey participants through random-
digit-dialing, internet banner advertisements, and partnerships with corporations to be-
come panel members.[231] Panel members agree to be contacted about new surveys. We
25
did not provide an incentive for participating in our study. However, the market research
company managing the panel does award monthly prizes to panel members based on the
number and length of surveys completed. Quota sampling was used to ensure that the
sample of the market research panel represented the general Canadian population in terms
of region of residence, gender, and age based on the 2016 Canadian Census (Table B.1).
The market research company sent panel members an e-mail invitation to participate in
our study. Interested panel members were redirected to a secure website hosting the utility
valuation exercise.
Survey Procedures
Participants were asked to value six health states in the online SOAP MESSC module.
The SOAP MESCC module has previously been described.[167] Briefly, the first three
pages of the module explained the utility valuation task, and provided an overview of
MESCC. Following this, participants completed the standard gamble health state valuation
exercises.[167]
Health states were derived from the EORTC item bank and were presented as a point-
form description of five dysfunctional attributes: dependent (D), non-ambulatory (N),
incontinent of urine (I), pain (P), and “other symptoms” (S). Each attribute was char-
acterized by the presence (+) or absence (-) of the dysfunctional attribute. Dependence
was described as “you need help with eating, dressing, washing yourself or using the toilet.
You are dependent on others.” Other symptoms were described as “You have one or more
of: nausea, vomiting, shortness of breath, lack of appetite, diarrhea.” Valuations were
obtained using standard gamble method.[243]
In this study, “perfect” health is represented by the fully functional health state (D-,
N-, I-, P-, S-). To eliminate any bias introduced by respondents assuming different life ex-
pectancies for each scenario, all scenarios were framed as having a certain life expectancy of
five years; that is, for both the probe health scenario, and perfect health, participants were
told their life expectancy would certainly be five years.(13) Five years was the maximum
survival reported in a randomized controlled trial on treatments for MESCC.[171]
Participants were asked to value six health states. Two health states formed a “test
pair.” These pairs shared non-dysfunctional attributes (eg. D-, N+, I-, P-, S- and D-, N+,
I-, P- , S+). The other four health states were selected at random.
Statistical Analysis
We deemed participants to have misunderstood the task, or not engaged, if: they provided
an illogical valuation for the test pair; or provided the same valuation for more than 4
health states. Such participants were excluded from the analysis. A sensitivity analysis
including all participants is provided in Appendix B.3.
Utility values lie between 0 and 1 with 0 representing death and 1 perfect health.
Therefore beta regression analysis was chosen to explore the relationship between health
26
attributes and utility valuation (Appendix B.1).[67] Regression was performed on the mean
(µ) of the beta model. The dispersion parameter (σ) was treated as a constant.[67] To
account for the presence of six observations per participant (indexed by i), we incorporated
a random intercept term for each participant (i ) in the model for µ.[78]
The SOAP MESCC module is designed for ex ante (without experience) utility valua-
tion from a general population sample. These utilities should reflect population preferences
for the health states and were used to compute expected (mean) quality-of-life changes in
economic analysis.
Within this context, the objective of the analysis was to estimate mean utilities for the
population, not predict individual utility valuations. Consequently, goodness-of-fit mea-
sures such as the R2 statistic or root mean squared (RMSE) are not appropriate because
they quantify how well the model fits individual observations.[174] To instead quantify the
performance of the regression model in estimating the mean utility for the 31 health states
relative to direct estimation of mean utilities, we used the absolute agreement intraclass
correlation coefficient (ICC). For this ICC calculation, health states were targets and util-
ities were ratings.[203] A two-way model was used because ratings can only be obtained
by direct estimation or regression estimation.
In an effort to strengthen the generalizability of the regression analysis, we implemented
internal validation by allocating participants to a test set and validation set in a 1 :
1 ratio.[213] Regression models were fit using the test set, and the optimal model was
identified by jointly the Akaike Information Criterion (AIC) and Bayesian Information
Criterion (BIC).[125] The optimal regression model was used to compute mean utilities
for each health state (regression estimation of means). Mean utilities for each health state
were then directly estimated using the validation set (direct estimation of means). The
performance of the optimal regression model was quantified using the ICC by comparing
mean utilities obtained from the regression model to mean utilities obtained by direct
estimation.
Four regression models were considered. In Model 1, each dysfunctional attribute was
coded as a categorical variable. In Model 2, the number of dysfunctional attributes was
coded as nominal categorical variables. This strategy was used because it avoided assump-
tions of linear or extra-linear effects. Model 3 incorporated all first-order interactions in
Model 1. Model 4 combined Model 1 and Model 2.
Before proceeding with model selection, we graphically checked all four models for
misspecification of the variance and link function.[149] We attempted to simplify the op-
timal model using standard regression model building procedures.[100] These procedures
ultimately led us to consider two additional models. In Model 5, the square root of the
number dysfunctional attributes (num dys) was included as a continuous predictor. In
Model 6, the natural logarithm of the number dysfunctional attributes (num dys) was in-
cluded as a continuous predictor. The formula for the beta regression linear predictor for
each model is outlined in Table B.2.
27
Sample Size Calculation
The sample size for the regression analysis was based on a commonly cited rule-of-thumb
that linear regression requires 10 to 20 subjects per variable.[86] Our regression models
considered a total of 21 variables: four numerical variables, five categorical variables, ten
interaction terms, and two intercepts. Following this rule of thumb, we would require
15 × 21 = 315 subjects. As we planned to split into a test and validation set in a 1 : 1
ratio, we required 630 subjects. However, to ensure at least 2 participants in each quota,
650 participants were required (Table B.1).
3.3 Results
1138 panel members logged into the SOAP MESCC module and provided consent. How-
ever, 488 were excluded for not valuing all health states, providing an illogical valuation for
the test pair, or providing invariant responses (Figure 3.1). Of the 1029 participants who
valued six health states, 379 (36.8%) were excluded for not having understood or engaged
in the task.
The socio-demographic characteristics of 650 included and 488 excluded participants
are shown in Table 3.1. There were no statistically or qualitatively significant differences
between the groups in terms of sex, age, or province of residence.
Mean valuations for the 31 health states tended to decrease with an increasing number
of dysfunctions (Table 3.2). For example, the mean utility valuation for 1, 2, 3, 4, and
5 dysfunctions was 0.691, 0.581, 0.471, 0.364, and 0.333 respectively (p-value for trend
<0.0001).
The functional form of each model and example mean utility calculations are provided
in Appendix B.3. Calibration and agreement parameters for all candidate models are
provided in Table B.4 and Figures B.4, B.5, B.6, 3.2, and B.7. The model selection
procedure identified the square root of the number of dysfunctions to be the optimal
predictor of mean utility based on both the AIC and BIC criteria (Table B.3).
The coefficients for Model 4 are given in Table 3.4. This model was well calibrated,
estimating mean utilities in the validation set with a mean absolute error of 0.047 and
mean signed difference of -0.02 (both of which are less than the MCID of 0.05 for utili-
ties)(Figure 3.2).[175, 246] There was excellent agreement between mean utilities obtained
from this regression model and mean utilities obtained by direct estimation with an ICC
of 0.936 (95% CI: 0.86, 0.97).[129]
Mean and median utility valuations were not sensitive to the inclusion of all participants
(Table B.7).
Formulas for the linear predictor in the beta regression are provided in Table B.2.
Formulas for mean utility valuation for a given MESCC health state using the fitted models
are provided in Table B.6. To illustrate the use of these formulas, we will calculate mean
utilities for the health state D-, N+, I-, P+, S- using the optimal Model 4. This health
28
state is coded as: num dys=2. Therefore using Model 4 the mean utility valuation is
calculated as: √
inverse logit 1.71 − 1.11 × 2 = 0.53.
3.4 Discussion
In this paper we report directly valued utilities for MESCC health states using the validated
SOAP MESCC module.[167] The study sample was representative of the population in all
English-speaking Canadian provinces. Our findings are valid with 63.2% of participants
having understood and engaged in the task. This rate is superior to validity rates reported
in general population direct valuation studies for the EuroQol-5D.[12, 128, 114, 49]
The regression model building exercise revealed that members of the general popu-
lation value all dysfunctions characterizing MESCC health states equally. Furthermore,
dysfunction follows a pattern of diminishing marginal disutility. That is, each additional
dysfunction effects a smaller incremental change in utility than the previous dysfunction.
These results demonstrate that from the societal perspective, physical function is valued
equal to other facets of well-being.
It must be recognized that ex post utilities are not equivalent to ex ante utilities ob-
tained from patients who have experienced the health states.[156] In part due to adapta-
tion, patients tend to provide higher valuations for health states which predominantly affect
physical health than the general population for the same health state. The ex ante utilities
collected in this study are highly appropriate for facilitating healthcare decision making
and can be used to conduct decision analysis and cost-utility analysis for MESCC.[54]
Although it may seem that applying lower ex ante utilities may infringe on patient
autonomy and deny care, healthcare system decision making impacts patients with various
conditions. If the objective of healthcare decision making is to maximize the benefit of
all patients, utilities across different disease must be comparable in order to set priori-
ties. Rawles argues that ex ante utilities can be used ethically if valued under a “veil of
ignorance.”[187] If we assume that the general population providing ex ante utility valu-
ations may eventually develop the condition of interest, out of self-interest, they should
provide fair valuations. Although ex ante utilities are theoretically restricted to system
policy decisions, ex ante utilities have become the de facto standard for individual patient
decisions. Utilities obtained from generic health surveys such as the EuroQol-5D, Short
Form-6D, and Health Utilities Index 3 are actually ex ante valuations.[55] Therefore we
have chosen to evaluate ex ante SOAP MESCC module to conform with conventions in
the literature.
Utility valuations for a single health state were highly variable across participants. This
was evidenced by wide Inter-quartile ranges. However, since our objective was to measure
general population ex ante utilities for health policy decision making, the expected values
and the underlying uncertainty represented by the 95% confidence intervals are pertinent.
Health policy decisions concern the allocation of scarce resources to health programs
with superior cost-effectiveness ratios at the expense of those with inferior cost-effectiveness
29
ratios.[55] If cost-effectiveness ratios are uncertain, there is a risk that a wrong policy
decision will be made (ie. funding the program with an inferior cost-effectiveness ratio).
Even if the risk of a wrong policy decision is large (ie. great uncertainty in cost-effectiveness
ratio), from both ethical and economic perspective a decision should still be made.[35] From
an ethical perspective, deferring a decision (ie. not funding any health program under
consideration) denies care to several groups of patients. From the economic perspective,
convincing arguments have been given to defend the practise of expected value decision
making – i.e. the irrelevance of inference with respect to public decision making.[5] For
these reasons, unbiased estimates of costs and effects are more important than precise
estimates.
The results of the regression model building exercise have relevance for clinicians coun-
seling MESCC patients regarding treatment. As dysfunctions follow a pattern of dimin-
ishing marginal disutility, each dysfunction is valued equal to the others – it is the total
number of dysfunctions that drive quality-of-life. Surgeons should be cognizant that ambu-
lation and continence, which are dysfunctions addressed by surgery, are no more important
than other attributes (pain, other symptoms, and level of independence). Furthermore,
attempting to reverse a single dysfunction in a patient with high functional status will lead
to a greater increase in quality-of-life relative to a patient with low functional status.
To our knowledge, this is the first study reporting direct utility valuation for a spinal
disorder. Utilities for spine disease specific instruments such as the NDI, ODI, and SRS-
22 have been developed using an indirect “cross-walk” protocol.[193, 31, 32, 19] Patient
responses are collected using both the disease specific instrument and a generic health mea-
sure (eg. EuroQol-5D, Short Form-6D, and Health Utilities Index 3). Next, a regression
model is developed to relate the disease specific score to the generic health measure. Then
the regression model estimated generic health measure score can be used to compute a util-
ity. However, utilities obtained from generic health measure scores are actually computed
from another regression model relating the generic health measure score to directly valued
utilities.[55] Our study demonstrates that disease specific direct utility valuation is feasi-
ble and valid. Investigators may consider applying our study protocol and the validated
SOAP tool to other disease contexts. This approach may eliminate the potential error of
the cross-walk approach introduced by the need for two regression models.
3.5 Conclusion
We provide a comprehensive set of ex ante utility estimates for MESCC health states. The
utility values derived from this study can be used to help inform population level healthcare
decision making, such as allocation of limited resources for specific treatments. The results
of this study can also help clinicians counsel MESCC patients regarding treatment.
This is the first study reporting direct utility valuation for a spinal disorder. We
demonstrate that direct utility valuation over the internet is a feasible solution for obtaining
quality-of-life data when generic health status questionnaire data is lacking. Investigators
may consider applying our approach to other disease contexts.
30
3.6 Figures
Logged into website

n = 1138
excluded Skipped one question

n = 109
Completed all 6 gambles

n = 1029
excluded Provided illogical rating

n = 144
Passed logic test

n = 885
excluded Invariant for 6 scenarios

n = 182
≥ 1 pair of 6 scenarios variant

n = 703
excluded Invariant for final 5 scenarios

n = 53
≥ 1 pair of final 5 scenarios variant

n = 650
Figure 3.1: Flow of participants.
31
Figure 3.2: Calibration Plot for Model 4.
32
3.7 Tables
Table 3.1: Participant Demographics
Included Excluded
N = 650 N = 488
Sex – no. (%)

Female 331 (51) 234 (48)
Male 319 (49) 254 (52)
Age – no. (%)
18 – 24 yrs 78 (12) 41 (8)
25 – 34 yrs 113 (17) 76 (16)
35 – 44 yrs 106 (16) 83 (17)
45 – 54 yrs 118 (18) 91 (19)
55 – 64 yrs 107 (17) 79 (16)
≥ 65 yrs 128 (20) 118 (24)
Province – no. (%)
Maritimes 57 (9) 40 (8)
Ontario 326 (50) 265 (54)
Manitoba 30 (5) 27 (6)
Saskatchewan 26 (4) 9 (2)
Alberta 98 (15) 65 (13)
British Columbia 113 (17) 82 (17)
33
Table 3.2: Utilities for all MESCC Health States
Health State Number of N Mean (95% CI) Median (IQR)

Dysfunctions
D+N-I-P-S- 1 181 0.676 (0.628, 0.725) 0.75 (0.45, 0.95)
D-N+I-P-S- 1 185 0.717 (0.671, 0.763) 0.85 (0.55, 0.95)
D-N-I+P-S- 1 187 0.681 (0.632, 0.73) 0.85 (0.45, 0.95)
D-N-I-P+S- 1 182 0.695 (0.647, 0.743) 0.85 (0.512, 0.95)
D-N-I-P-S+ 1 202 0.685 (0.639, 0.732) 0.85 (0.462, 0.95)
D+N+I-P-S- 2 88 0.548 (0.474, 0.623) 0.65 (0.225, 0.9)
D+N-I+P-S- 2 108 0.619 (0.554, 0.684) 0.75 (0.4, 0.95)
D+N-I-P+S- 2 106 0.542 (0.475, 0.608) 0.575 (0.25, 0.85)
D+N-I-P-S+ 2 112 0.6 (0.534, 0.666) 0.7 (0.35, 0.95)
D-N+I+P-S- 2 88 0.659 (0.589, 0.728) 0.75 (0.45, 0.95)
D-N+I-P+S- 2 107 0.57 (0.505, 0.635) 0.65 (0.325, 0.875)
D-N+I-P-S+ 2 105 0.513 (0.44, 0.586) 0.6 (0.1, 0.9)
D-N-I+P+S- 2 100 0.583 (0.514, 0.652) 0.65 (0.35, 0.9)
D-N-I+P-S+ 2 89 0.601 (0.525, 0.676) 0.65 (0.3, 0.95)
D-N-I-P+S+ 2 113 0.582 (0.514, 0.65) 0.65 (0.25, 0.95)
D+N+I+P-S- 3 81 0.512 (0.428, 0.596) 0.55 (0.1, 0.9)
D+N+I-P+S- 3 76 0.443 (0.363, 0.524) 0.45 (0.1, 0.75)
D+N+I-P-S+ 3 80 0.492 (0.413, 0.572) 0.5 (0.075, 0.75)
D+N-I+P+S- 3 88 0.445 (0.367, 0.524) 0.45 (0, 0.75)
D+N-I+P-S+ 3 96 0.491 (0.415, 0.566) 0.5 (0.1, 0.85)
D+N-I-P+S+ 3 88 0.479 (0.398, 0.56) 0.6 (0, 0.813)
D-N+I+P+S- 3 86 0.478 (0.401, 0.554) 0.525 (0.1, 0.75)
D-N+I+P-S+ 3 81 0.407 (0.329, 0.485) 0.45 (0, 0.75)
D-N-I+P+S+ 3 91 0.501 (0.426, 0.576) 0.5 (0.15, 0.85)
D-N+I-P+S+ 3 81 0.457 (0.374, 0.539) 0.45 (0, 0.8)
D+N+I+P-S+ 4 192 0.408 (0.357, 0.46) 0.35 (0, 0.75)
D+N+I+P+S- 4 199 0.359 (0.308, 0.41) 0.25 (0, 0.65)
D+N+I-P+S+ 4 191 0.342 (0.292, 0.391) 0.25 (0, 0.65)
D+N-I+P+S+ 4 173 0.357 (0.303, 0.411) 0.25 (0, 0.65)
D-N+I+P+S+ 4 170 0.351 (0.297, 0.404) 0.25 (0, 0.65)
D+N+I+P+S+ 5 174 0.333 (0.281, 0.384) 0.2 (0, 0.638)
95% Confidence Interval (95% CI). Interquartile range (IQR).
34
Table 3.3: Selection Parameters for all Models
Description AIC BIC
Model 1 Attributes -2750.24 -899.148

Model 2 Number of dysfunctions (nominal variable) -2751.78 -906.259
Model 3 Attributes and second-order interactions -2747.42 -840.572
Model 4 Square root of number of dysfunctions -2756.85 -928.061
(continuous variable)
Model 5 Logarithm of number of dysfunctions -2754.9 -926.11
Table 3.4: Summary of Beta Regression Analysis for Model 4
Variable Coefficient SE t-value p-value
Mean (µ) Intercept 1.7058 0.11594 14.71 <0.0001

coefficients
Square root of -1.11247 0.07175 -15.51 <0.0001
number of dysfunctions
Dispersion (σ) Intercept 0.74727 0.02139 34.94 <0.0001
coefficient
35
Chapter 4
Back to Bayesian: a Strategy to

Enhance Prognostication of
Metastatic Spine Disease
Markian A. Pahuta1 , Joel Werier2 , Eugene K. Wai2,3 , Carl van Walraven3,4,5 , Doug Coyle3
Michigan, USA
tario, Canada
36
Context
This chapter addresses skepticism regarding decision-analysis. In this chapter I review
Bayesian statistics which I believe can address the concern that probabilistic analysis is
arbitrary and prone to manipulation. This is because the output of a Bayesian analysis is a
posterior distribution which can be used in place of user defined distributions in economic
analysis. Bayesian analysis will be used to conduct QALY decision analysis in Chapter 6.
37
Abstract
Clinicians must consider prognosis when offering treatment to patients with spine metas-
tases. Although several prognostic indices have been developed and validated for this
purpose, they may not be applicable in the current era of targeted systemic therapies.
Even before the introduction of targeted therapies, these prognostic indices should not
have been directly used for individual patient decision making without contextualizing with
other sources of data. By contextualizing, we mean that prognostic estimates should not
be based on these scores alone and should formally incorporate clinically relevant factors
not part of prognostic indices. Contextualization requires the use of Bayesian statistics
which may be unfamiliar to many readers.
In this paper we show readers how to correctly apply prognostic scores to individual
patients using Bayesian statistics. Through Bayesian analysis, we explore the impact of
new targeted therapies on prognostic estimates obtained using the Tokuhashi score.
We provide a worked calculation for the probability a patient surviving to 6 months
using dichotomous prognostication. We then demonstrate how to calculate a patient’s
expected survival using continuous prognostication. Sensitivity of the posterior distribution
to prior assumptions is illustrated through effective sample size adjustment.
When the predicted prognosis from the Tokuhashi score is contextualized with data
on contemporary systemic treatments, patients previously deemed non-surgical candidates
may be eligible for surgery.
Bayesian prognostication generates intuitive results and allows multiple data points
to be synthesized transparently. These techniques can extend the usefulness of existing
prognostic scores in the era of targeted systemic therapies.
38
4.1 Clinical Scenario
Tom is a 60 year-old man with Stage IV non-small-cell lung cancer (NSCLC). He has
metastases to four vertebral bodies with no epidural disease. He is treated with platinum-
based chemotherapy cisplatin and radiotherapy. Four months into treatment, the disease
at T5 progresses and Tom develops an incomplete spinal cord injury from epidural disease.
As a result, Tom experiences a functional decline in his Karnofsky Performance Score to
70 and is admitted to hospital.
The consulting spine surgeon, radiation oncologist, and medical oncologist all agree that
direct decompressive surgery and stabilization would be the most effective treatment for
neurologic dysfunction.[171] However, Tom’s Tokuhashi Score is 6, which is associated with
an expected survival of less than 6 months.[228] Furthermore, neither the spine surgeon
nor the radiation oncologist has seen a patient like Tom (with stage IV NSCLC refractory
to first-line treatment) survive beyond 3 months. Therefore, both the spine surgeon and
radiation oncologist recommend against surgery because Tom’s prognosis is poor.
However, contemporary trials have reported improvement in overall survival for patients
with advanced NSCLC refractory to first-line treatment with targeted therapies. Therefore,
the medical oncologist disagrees with the spine surgeon and medical oncologist because
she/he plans to give Tom targeted therapy and therefore expects Tom to survive longer
than 6 months.
The three clinicians plan to meet to resolve their disagreement.
4.2 Introduction
Surgical treatment of metastatic spine disease is effective for maintaining and improving
patients’ physical function and health-related quality-of-life.[171, 65] However, benefit from
a surgical intervention requires that patients survive longer than the time needed to recover
and rehabilitate from surgery. Surgeons typically believe that patients should have an
expected survival of at least 3 – 6 months to be surgical candidates.[227]
Several prognostic indices have been developed and validated for estimating expected
survival of patients with metastatic spine disease.[172, 132, 211] The Tokuhashi Score,
widely used survival index for these patients, considers performance status, extra-spinal
bone metastases, vertebral metastases, visceral metastases, pathologic diagnosis, and neu-
rologic function.[228] The Tokuhashi score has good performance in discriminating between
patients who die prior to or survive beyond 6-months.[64]
However, some authors have pointed out that existing prognostic scores are not useful
in the era of targeted therapies, in particular for non-small cell lung cancer (NSCLC).
The Tokuhashi Score, for example, was developed in 1990 and revised in 2005.[229, 228]
During this time-period, only cytotoxic agents were available for the treatment of NSCLC,
and consequently patients with advanced cancer had a median overall survival was 4 –
6 months. In the contemporary period, patients with advanced NSCLC may be offered
39
targeted therapy such as antiangiogenic drugs which has been shown to improve overall
survival in NSCLC patients.[185]
However, even before the introduction of targeted therapies, the Tokuhashi Score should
not have been directly used for individual patient decision making without contextualizing
with other sources of data. By contextualizing, we mean that estimates of prognosis should
not be based on the Tokuhashi score alone: prognosis should be estimated from multiple
data points. This is because the Tokuhashi Score does not consider all clinically important
prognostic factors.
Failing to contextualize statistical results, such as the Tokuhashi Score, is a common
mistake.[41] This is because contextualization requires Bayesian statistics – a form of statis-
tics that may be unfamiliar to many readers. This form of statistics that actually predates
the more commonly used Frequentist statistics (the branch of statistics underpinning con-
fidence intervals and p-values) and is distinguished by the fact that it incorporates multiple
sources of information into statistical analysis.[68] Although the term “Bayesian” may be
new to many readers, its concepts are likely familiar to many readers as they are also
used for diagnostic test interpretation. In essence, the Bayesian statistical approach that
integrates new information with previously known information.[68]
In this paper we aim to show readers how to correctly apply the Tokuhashi Score
to individual patients using Bayesian statistics. We illustrate the impact of new targeted
therapies on prognostic estimates. In Section 4.3, we will determine the probability of Tom
surviving to 6 months (dichotomous prognostication). In Section 4.4, we will determine
Tom’s expected survival using continuous prognostication. We then conclude the paper by
reviewing the debate between Frequentist and Bayesian statistics and its implications on
clinical decision making.
4.3 Dichotomous Prognostication

Returning to the clinical scenario, the medical oncologist decides to argue her/his point
by estimating the probability of Tom surviving ≥6-months with targeted therapy (do-
cetaxel combined with ramucirumab, DOC+RAM) versus standard therapy (docetaxel
alone, DOC). This is an example of dichotomous prognostication because there are only
two mutually exclusive outcomes: ≥6-month survival and ¡6-month survival. For Tom, di-
chotomous prognostication is the process of updating beliefs on the probability of ≥6-month
survival with the different treatments (prior probability, PriorPr) with the Tokuhashi Score
(converted to a likelihood ratio, LR) to yield a new belief on the probability of ≥6-month
survival (posterior probability, PostPr).[2]
A search of the literature reveals that no studies have recruited patients that exactly
match Tom’s pathologic diagnosis and disease stage. However, the REVEL investigators
performed a planned subgroup analysis in patients like Tom: those with refractory NSCLC
within 9-months of starting first-line treatment.[188] Therefore this analysis has the best
external validity for the clinical scenario at hand.[186] Six-month survival probabilities
for this subgroup are not reported in the main paper; however, we can determine this
40
probability by digitizing the survival curve shown in supplementary Figure S1a.[92] From
the digitized survival curve, we determine that the 6-month survival in the DOC+RAM
group was 68%, while in the DOC group 6-month survival was 56
Prognostic performance of the Tokuhashi Score have been reported in several cohort
studies and a meta-analysis conducted by Lee and colleagues.[133] We should use the
highest level of evidence provide by the meta-analysis reported by Lee et al. The authors
reported that a Tokuhashi Score of ≥8 has a sensitivity and specificity of 0.657 and 0.676,
respectively, for ≥6-month survival. Tom has a negative test result because his Tokuhashi
Score is ¡8. Therefore, a negative likelihood-ratio (LR)1 is required:
1 − Sensitivity
LR = [89]
Specificity
1 − 0.676
=
0.657
= 0.493.
The LR quantifies how much more likely a Tokuhashi Score <8 is found in patients who
survive ≥6-months compared to those who survive <6-months. We must emphasize that
the LR does not quantify how much more likely a ≥6-month survival is found in patients
with a Tokuhashi Score <8 compared to those with a Tokuhashi score of ≥8. Therefore
the LR provides information in a way that is “backwards” to what is needed for clinical
decision making – the spine surgeon and oncologist require the likelihood of ≥6-month
survival, not the likelihood of a Tokuhashi Score of <8.
The posterior probability can be calculated from the prior probability and LR using
the formula
PriorPr × LR
PostPr =
1 − PriorPr + PriorPr × LR
For DOC only treatment we compute
0.56 × 0.493
PostPr =
1 − 0.56 + 0.56 × 0.493
= 0.386.
Repeating the same steps for DOC+RAM treatment generates a posterior probability of
0.512.
These calculations indicate that while a Tokuhashi Score of <8 generally predicts a
<6-month survival, the score should be integrated with other pieces of information more
generalizable to the clinical problem at hand. Given the effectiveness of DOC+RAM,
a negative Tokuhashi Score does not translate to a 6-month survival probability <50%.
In contrast, DOC is less effective and a negative Tokuhashi Score results in a survival
probability less than 50%.
1
If Tom had a positive test result, we would instead compute the positive likelihood ratio LR =
Sensitivity
1−Specificity
41
Based on this analysis, the medical oncologist concludes that if treated with DOC+RAM,
Tom is a surgical candidate because he has a greater than 50% probability of surviving
≥6-months. The spine surgeon and radiation oncologist are not entirely convinced and
request more analysis.
4.4 Continuous Prognostication

Tom, like many cancer patients, has difficulty interpreting and applying probabilities.[79,
74, 141] Like many cancer patients, Tom has also expressed that he wishes to know the
typical survival experience for patients like him, that is, he wants to know the worst
and best case scenarios.[93, 215] The medical oncologist would like to present Tom with
his expected median survival time because it is a less biased measure than mean survival
time.[202] To communicate the best and worst case scenarios, the medical oncologist would
like to present range of values which she/he is 95% confident the true median survival lies.
The medical oncologist also believes this type of analysis will satisfy the spine surgeon and
radiation oncologist.
The medical oncologist now wishes to conduct continuous prognostication. Recall that
the outcome of interest for dichotomous prognostication could only take on two values
(≥6-month survival: yes or no). In contrast the median survival is a continuous variable,
that is, it can theoretically take on an infinite number of values between 0 and infinity. We
can apply the same studies used for dichotomous prognostication (the REVEL trial and
the Lee et al. meta-analysis).[188, 133]
To perform continuous prognostication for Tom, beliefs on the median survival time
reported by the REVEL trial need to be updated with the median survival time associ-
ated with a Tokuhashi Score of ¡8 to yield a new belief on the median survival time. A
prior probability is needed for each potential median survival time; this collection of prior
probabilities is called a prior distribution. A likelihood ratio is also needed for each of the
potential median survival times; this information is contained in a likelihood distribution.
The updated beliefs on median survival time are contained in the posterior distribution.
To develop a prior distribution from the REVEL subgroup analysis, we require an esti-
mate of the median survival time and standard deviation for each treatment group. From
supplementary Figure S1a, we find the median survival time was 7.0 months (95% CI, 6.1 –
8.5 months) for the DOC group; while for the DOC+RAM group, median survival was 9.3
months (95% CI, 8.4 – 10.1 months). These confidence intervals are asymmetric, therefore
additional calculations are needed to calculate standard deviation (Appendix C.1).[25] The
standard deviation for the DOC group is 0.68, while the standard error for the DOC+RAM
group is 0.42.
To develop a likelihood distribution for Tokuhashi Score ¡8, we require an estimate of
the median survival time and standard deviation for this group of patients. This analysis
was not reported in the publication, but was provided by the authors on request.[133] Using
individual patient analysis, Lee and colleagues determined that the median survival and
42
95% CI for a Tokuhashi Score ¡8 was 5.0 ± 0.6 months. Therefore the standard deviation
is 0.6 ÷ 1.96 = 0.33.
It is important for readers to appreciate that the meaning of confidence intervals from
the REVEL trial and Lee et al. meta-analysis. These intervals ensure that 95% of similarly
computed intervals computed in an infinite number of times would contain the true median
survival.[200] The qualifier 95% applies to the technique of computing confidence intervals
and not the true expected survival.[148] Confidence intervals do not indicate the range
within which we are 95% confident the true median survival lies – such an interval is a
credible interval. Much in the same way that the likelihood ratio provides information in
a way that is “backwards” to what is needed for clinical decision making, the confidence
interval provides information in a “backwards” to a credible interval. Bayesian continuous
prognostication is required to compute a credible interval for the median survival.
The posterior distribution can be easily derived if we assume the estimates of median
survival time follow a normal (or bell-curve) distribution.[209] We denote the median sur-
vival time and standard deviation from the REVEL data as m and s respectively. We
denote the median survival time and standard deviation from the Lee et al. meta-analysis
as µ and σ respectively. Through exact mathematical analysis it can be shown that the
mean of the posterior distribution for median survival time is
m µ
s2
+ σ2
1 1
s2
+ σ2
and the standard deviation is s

1
1 1 .
s2
+ σ2
Substituting the data for the DOC treatment arm and Tokuhashi Score ¡ 8 yields a
posterior distribution with mean
7.0 5.0
0.682
+ 0.33 2
1 1
0.682
+ 0.332
and standard deviation s
1
1 1 .
0.682
+ 0.332
We can calculate a 95% credible interval as 5.38 ± (0.09 times1.96) months. In summary,
Tom’s estimated median survival with DOC alone is 5.38 months with a 95% credible
interval from 5.20 to 5.56 months. By repeating the same calculations for DOC+RAM
we determine that Tom’s expected median survival is 6.64 months with a 95% credible
interval from 6.50 to 6.78 months
The typical, worst and best case scenarios are all under 6 months if Tom is treated with
DOC alone. But if treated with DOC+RAM, Tom’s typcial, worst and best case scenarios
are all over 6 months. Based on the expectation that recovery from surgery can take over
3 months, the medical oncologist concludes that if treated with DOC+RAM, surgery is a
43
reasonable proposition because even in the worst-case scenario, expected survival is over 6
months.
However, the spine surgeon and radiation oncologist are not convinced. They have
reservations regarding the generalizability of the REVEL trial to Tom’s individual case
because the REVEL investigators did not report the number of patients with vertebral
metastases in the study sample. The medical oncologist suggests that a sensitivity analysis
would demonstrate the robustness of conclusions based on the REVEL trial data, and would
address concerns regarding generalizability.
By questioning the the generalizability of the REVEL trial, the spine surgeon and
radiation oncologist are implicitly discounting the REVEL trial to the degree that the
upper limit of the resulting 95% credible interval is less than and does not include 6
months. To make their discounting explicit, the medical oncologist conducts a “what-
if” analysis by re-calculating the posterior distribution for the DOC+RAM treatment
multiple times with decreasing effective sizes, but keeping all other parameters constant
(Appendix C.2). She/he determines that only if the REVEL subgroup analysis for the
DOC+RAM treatment arms consisted of ≤64 patients would the 95% credible interval for
Tom’s prognosis not include 6 months.
The medical oncologist explains that this exercise formally quantified the degree to
which the other two clinicians are implicitly discounting the generalizability of the REVEL
trial. In other words, the spine surgeon and radiation oncologist discounted the REVEL
trial by 83%. Although the REVEL trial may not be fully generalizable to Tom’s case, the
spine surgeon and radiation oncologist do not believe the REVEL trial should be discounted
by 83%.2 Consequently, all three clinicians agree that Tom should be offered surgery if he
receives DOC+RAM treatment.
4.5 Discussion
The motivation for this article was to enhance prognostication for patients with metastatic
spine disease in an era of more effective systemic treatments. We introduced Bayesian
statistics and sought to demystify this statistical approach. In this article, we focused on
the principles and application of Bayesian analysis rather than the mechanics of Bayesian
computation. Without knowledge of detailed computational steps, some readers may
regard Bayesian statistics as a “black box.” Interested readers are directed to excellent
textbooks on the subject.[209, 153, 30, 124] Dichotomous prognostication requires several
data-points:(i) a prior survival probability at a time of interest; and (ii) the sensitivity
and specificity of the prognostic factor on interest. If a only a survival probability, but not
sensitivity and specificity for a prognostic factor of interest are reported, beta-binomial
conjugate analysis can be used.[135] Continuous prognostication requires:(i) a prior be-
lief on median survival time and confidence interval; and (ii) median survival time and
confidence interval for a prognostic factor of interest.
2 64
1− 384 = 0.8333
44
We took care to highlight that confidence intervals do not quantify the uncertainty
in a value such as median survival time. Although confidence intervals are often used
for this purpose, this is because confidence intervals are frequently misinterpreted by the
“confidence-level misconception” to be Bayesian credible intervals.[28] Under this miscon-
ception, a 95% confidence interval is interpreted to contain the true value with a probability
of 95%.[41] Cumming et al. asked 263 researchers in psychology, behavioural neuroscience,
or medicine to interpret confidence intervals and found that over 40% suffered from the
confidence-level misconception.[41] This misconception was actually anticipated in 1937
by the developers of the confidence interval.[155] As even practicing researchers have dif-
ficulty interpreting confidence intervals, the journal Basic and Applied Social Psychology
recently “banned” the use of confidence intervals in new submissions.[233] The prevalence
of the confidence-level misconception suggests that for many individuals the Bayesian cred-
ible interval is more intuitive construct for expressing the uncertainty in a reported value.
Therefore, we believe Bayesian statistics can enhance clinician-patient communication.
In this article, we focused on the principles and application of Bayesian analysis rather
than the mechanics of Bayesian computation. Without knowledge of the computational
steps, some readers may regard Bayesian statistics as a “black box.” Interested readers are
directed to Appendix C.1 and excellent textbooks on the subject.[209, 153, 30, 124] Without
delving into Bayesian computation, we believe readers can gain sufficient understanding
of Bayesian analysis through the analogy with diagnostic testing. It is well accepted that
diagnosis should be based on the post-test probability of disease rather than the sensitivity
and specificity of a test.[110] For similar statistical considerations, credible intervals should
be used instead of confidence intervals. Diagnostic test interpretation will be incorrect if
the pre-test probability is inappropriate, similarly, credible intervals will be incorrect in
the prior probability is inappropriate.
The sensitivity of the posterior distribution to the choice of prior distribution is an
important reason for the lack of uptake of Bayesian methods. R.A. Fisher, one of the
founders of modern Frequentist statistical methods, opposed Bayesian statistics and char-
acterized the approach as: “. . . extremely arbitrary. . . by evolving a vitally important piece
of knowledge, that of the exact form of the distribution of [the value of interest], out of
an assumption of complete ignorance. . . ”[72] We exploited the sensitivity of the poste-
rior distribution to choice of prior distribution in a positive way for effective sample size
sensitivity analysis. However, arbitrary and/or improper selection of prior distributions
negatively impacts on Bayesian analysis.
Lack of public trust in science illustrates these issues regarding the prior distributions.
Despite scientific evidence on issues such as safety of childhood vaccines, safety of ge-
netically modified foods, and effectiveness of sexual health education these topics remain
controversial.[96] Although individual psychology may predispose individuals to embracing
controversy, access to “prior distributions” is an important determinant.[139] Controversy
is stoked by consumption of selective and biased media.[24] Whereas controversy is less-
ened when scientific consensus is communicated in a coherent and unified manner.[140]
Just as the public is cautioned to consume media critically and identify “fake news,”[150]
readers should critically appraise prior distributions before accepting a Bayesian analy-
sis. To mitigate manipulation of the posterior distribution, it has been suggested that
45
researchers repeat their analysis with several different prior distributions that capture the
spectrum of reasonable clinical beliefs.[104] Consequently, readers must assess the sensitiv-
ity of results to prior distribution choice, and make a judgment on whether the results are
sufficiently robust to make conclusions. Sensitivity analysis was illustrated in Section 4.4
and Appendix C.2 when the effective sample size was adjusted.
We argue that for clinical scenarios with sparse evidence, such as prognostication in
metastatic spine disease in the era of targeted therapies, a non-Bayesian approach may lead
to suboptimal decision making. A nihilistic course of action based on non-Bayesian analysis
would be to ignore recent advances in systemic treatment because their effectiveness for
patients with metastatic spine disease has not been formally studied. An overly optimistic
course of action base on non-Bayesian analysis would be to assume the results of oncology
trials enrolling few patients with metastatic spine disease fully apply to patients with
metastatic disease. In this paper we demonstrated how effective sample size sensitivity
analysis quantifies the impact of clinicians’ assumptions and biases on decision making, and
therefore, can make decisions more transparent. In this manner, different – and potentially
not fully compatible – sources of evidence can be synthesized for decision making in a
rational and transparent way.
It is not known whether Bayesian prognostication for metastatic spine disease is su-
perior to informal clinical decision making.[33] However, it has been shown that informal
prognostication for metastatic spine disease is superior to prognostication based on the
Tokuhashi Score alone.[126] While for the diagnosis of pulmonary embolism, integrating
clinical gestalt with a clinical prediction rule (a Bayesian strategy) is superior to relying
solely on a clinical prediction rule.[173] Given the effectiveness of Bayesian clinical decision
making in other clinical contexts, and the limitations of existing decision making strategies,
Bayesian prognostication for metastatic spine disease should be formally studied.
In this article we present the correct statistical procedures for prognostication. This ap-
proach uses Bayesian statistics which is intuitive and can transparently synthesize evidence.
Bayesian prognostication can enhance counseling for metastatic spine disease because clin-
icians’ assumptions and biases are formally incorporated into analysis. Future studies
should compare Bayesian prognostication with traditional prognostication in a cohort of
patients with metastatic spine disease.
46
Chapter 5
A Technique for Approximating

Transition Rates from Published
Survival Analyses
Markian A. Pahuta1 , Joel Werier2 , Eugene K. Wai2,3 , Roy A. Patchell4 , Doug Coyle3
Michigan, USA
tario, Canada
4 – Department of Neurosurgery, University of Kentucky Medical Center, Lexington,

Kentucky, USA
47
Context
This chapter addresses skepticism regarding decision-analysis. Credibility of Markov cohort
simulation suffers when models are improperly populated with the results of partitioned-
survival analysis. However, the correct technique for these situations, partitioned-survival
modeling, does not yield as informative of results as Markov cohort simulation. In this
chapter I present a solution to this problem. I develop and validate a technique to ap-
proximate transition rates from published partitioned-survival analysis. Formal multistate
techniques, as described in this chapter, will be used to conduct QALY decision analysis
in Chapter 6.
48
Abstract
Background
Quality-adjusted-life-years (QALYs) are used to concurrently quantify morbidity and mor-
tality within a single parameter. For this reason, QALYs can facilitate the discussion of
risks and benefits during patient counseling regarding treatment options.
QALYs are often calculated using partitioned-survival modelling. Alternatively, QALYs
can be calculated using more flexible and informative state-transition models populated
with transition rates estimated using multistate modelling (MSM) techniques. Unfortu-
nately the latter approach is considered not possible when only progression-free survival
(PFS) and overall survival (OS) analyses are reported.
Methods and Results

We have developed a method that, under particular conditions, can be used to estimate ap-
proximate transition rates from published PFS and OS analyses (we will refer to transition
rates estimated using full multistate methods as true transition rates).
Conclusions
In this article we present the basis for and use of the transition rate approximation method.
We then apply the method to a case study and evaluate the method in a simulation study.
49
5.1 Introduction
Chronic, progressive, and non-communicable diseases (such as cancer, diabetes, cardiovas-
cular disease and chronic respiratory disorders) are now the leading cause of morbidity
and mortality around the world. More than 60% of global deaths are attributable to these
types of diseases;[224] consequently these diseases now account for up to 50% of the total
healthcare budget in some countries.[154] Many of these diseases can be conceptualized as
consisting of three health states: healthy (h), ill (i), or dead (d) (Figure 5.1).
λ id
d
λh
h i
λhi
Figure 5.1: State-transition diagram for an illness-death model. The model consists of
three health states: healthy (h), ill (i), and dead (d). Variable names adjacent to the
solid arrows are transition rates (λhi , λhd , and λid ). This model is said to be “progressive”
because transitions are irreversible (i.e. unidirectional). The curved arrows indicate that
individuals can remain in a particular state over time. See text for more details.
Treatment decisions for chronic, progressive, and non-communicable diseases are diffi-
cult because interventions can have distinct, and sometimes opposite, influences on the
probability that a patient experiences a given health state. For example, a therapy
(e.g. high-risk cancer surgery) may decrease the risk of death (by controlling cancer) but
may increase the risk of becoming ill (if a post-operative complication occurs). Quality-
adjusted-life-years (QALYs) can be used to concurrently quantify morbidity and mortal-
ity within a single parameter.[55] For this reason, QALYs may facilitate the discussion
of risks and benefits during patient counseling regarding treatment options.[123] QALY
calculation requires knowledge of state-membership fractions. These are the proportion
of patients from a defined cohort that are in a given health state at a given time t.
State-membership fractions can be calculated using partitioned-survival modelling or state-
transition modelling.[226, 253, 251]
Partitioned-survival modelling uses data abstracted from progression-free survival (PFS)
curves and overall survival (OS) curves reported in the literature.[251] PFS curves show
the fraction of the cohort that is healthy over time t (P F S(t); OS curves show the fraction
of the cohort that is alive (either healthy or ill) over time t (OS(t)). Since OS curves
show the fraction of alive patients, the fraction of dead patients is simply 1 − OS(t). The
fraction of ill (but alive) patients is the difference between the fraction of alive and healthy
50
patients OS(t) − P F S(t). We will refer to state-membership fractions calculated in this
way as partitioned-survival fractions.[226, 253, 251] In contrast, state-transition modelling
applies the results of a multistate analysis. For the disease shown in Figure 5.1, these
techniques would be used to estimate the transition rate (i.e. the instantaneous risk (or
hazard) of moving from one state to another) from health to illness (h → i), from health
to death (h → d), and from illness to death (i → d).[253] Transition rates can be used
compute transition probability matrices to calculate state-membership fractions (“multi-
state fractions”). It is important to recognize that state-transition modelling is based on a
set of mutually exclusive health states (health, illness, death), whereas partitioned survival
modelling is based on non-mutually exclusive health states (health and illness or death for
the PFS curve, and alive and dead for the OS curve). Partitioned-survival modelling is
used when sufficient data for state-transition modelling is unavailable.
QALY calculations based on partitioned-survival fractions can suffer from two impor-
tant limitations that result from the fact that (i) the OS analysis does not consider the
survival of ill patients separate from healthy patients, and (ii) the risk of progressing to
illness rather than death for healthy patients cannot be determined from PFS analysis.
The first limitation of partitioned-survival fractions stems from the difficultly of extrapo-
lating partitioned-survival fractions beyond the study’s observation period.[253] This is a
significant deficiency because clinical studies often have a limited observation period that is
of insufficient duration to characterize long-term clinical outcomes.[102, 9, 130, 88, 38, 222]
The second limitation of partitioned-survival fractions is that computed QALYs are not
generalizable to patient cohorts whose baseline fractions of healthy, ill and dead patients
differs from those of the study cohort.[253] This is because the OS curve is a weighted aver-
age of OS curves for healthy and ill patients; therefore, the shape of the curve will change if
the baseline ratio of healthy to ill patients differs. These two limitations restrict the use of
partitioned survival fractions for decision analysis. These limitations can be avoided by cal-
culating QALYs using multistate fractions. Because they are based on granular analyses of
all transitions, multistate fractions have several advantages over partitioned-survival frac-
tions. First, they can be reliably extrapolated beyond the study observation period.[253]
Second, they can be used for decision analysis in cohorts with baseline characteristics that
differ from the original study cohort.[4]
Unfortunately, one cannot usually calculate transition rates using data abstracted from
PFS and OS analyses.[253, 251] Given the limitations of partitioned-survival fractions and
the advantages of multistate fractions, it would be helpful to obtain transition rates and
calculate the latter when one only has access to PFS and OS analyses. We have developed
a method that, under particular conditions, can be used to estimate approximate transition
rates from published PFS and OS analyses (we will refer to transition rates estimated using
full multistate methods as true transition rates).
This article organized as follows. We first present the basis for and use of the transition
rate approximation method. A case study is then reported in which we apply transition
rate approximation to data from a randomized controlled trial (RCT) of treatments for
metastatic epidural spinal cord compression (MESCC). We then report a simulation study
evaluating the accuracy of the approximation method. In the last section we summarize
and discuss our findings.
51
5.2 Transition Rate Approximation
The approximation technique is restricted to three-state progressive, time-homogenous
Markov disease processes such as the one shown in Figure 5.1.[105] Progressive means that
transitions are irreversible (i.e. cannot return to health from illness). Time-homogenous,
means that transition rates do not change over time. Markov means that transition rates
do not depend on disease history; in other words, the probability that a patient transitions
from state x to state y during a particular time period is independent of their previous
health state.
The data needed to use the approximation technique can be abstracted from most
articles reporting PFS and OS analyses. The number of patients experiencing an event
e c e c
and number of censored patients in both the PFS (Npf s and Npf s ) and OS (Nos and Nos )
analyses can be determined from the article text or patients-at-risk risk table. To obtain
the remaining data points, PFS and OS KM curves need to be digitized. Digitized KM
curves can then be used to reconstruct individual patient data using validated algorithms to
determine the event times in the PFS and OS analyses.[244] The approximation technique
requires that we make note of the maximum observation time (event or censoring) in the
PFS and OS analyses (τpf s and τos respectively). The area under the PFS and OS curves
(AU Cpf s and AU Cos respectively) are calculated by summing the area under each step of
the KM curve.
We denote h → i, h → d, and i → d transition rates as λhi , λhd , and λid . For the
time-homogenous disease processes (i.e. constant transition rates), exit times from the (i)
healthy state (i.e. h → i or h → d transition) and (ii) ill state (i.e. i → d transition) are
exponentially distributed. Furthermore, once a patient exits health, the probability that
they make an h → d transition is
λhd
ρ=
λhi + λhd
We will refer to ρ as the risk of death for healthy patients. As there are only two possible
transitions out of health, the probability that a transition out of the health state is an
h → i transition is 1 − ρ.
The mean time of exit from the healthy state (i.e. mean progression-free survival
time) is a biased measure in the presence of right censoring.[43] Instead we calculate the
restricted mean progression free-survival time (RMPFST−τ ) which is interpreted as the
mean progression-free survival time if observation is restricted to a truncation time τ .[134]
Since the exit time from health is exponentially distributed, the RMPFST−τ can be cal-
culated as
−τ 1 − e−(λhi +λhd )τ
RMPFST = . (5.1)
λhi + λhd
By definition, the area under the PFS curve is equal to RMPFST−τ when τ is set to
the maximum observation time in the PFS analysis, τpf s .[194, 257] Using Formula 5.1, we
can then numerically solve for λhi + λhd using standard algorithmic methods.[18] Simulta-
neous events in the PFS and OS analyses indicate h → d transitions. Therefore, we can
52
approximate the risk of death for healthy patients as
Nsimul
ρ≈ e
. (5.2)
Npf s
To approximate λid we need to use information gathered from the OS analysis. It is

more challenging to define an exact formula for the restricted mean overall survival time
(RMOST−τ ) than form the RMPFST−τ because exit from the alive state (i.e. healthy or
ill) is defined by a mixture of two exponential distributions: exit from health and exit from
illness. However, if we know the death times oei and censoring times ocj for a cohort of
e c
alive patients, Nos who had an observed event, and Nos who were right censored, we can
−τ −τos
approximate RMOST truncated to τos , RMOST , using inverse probability weighting
e e
Nos c
Nos
!
c
X
−τos Nos + Nos e
X
c 1
RMOST ≈ e
oi + oj e + Nc
.[95] (5.3)
Nos i=1 j=1
Nos os
Next, we determine the total person-time of observation in the OS analysis

Nose c
Nos
X X
Eos = oei + ocj . (5.4)
i=1 j=1
If censoring times are not denoted on the OS curve, it is not possible to determine ocj .
However, we can rearrange Formula 5.3 to yield
c
Nos e
Nos e c

X
−τos
X Nos + Nos
ocj ≈ RMOST e
(Nos + c
Nos ) − oei e
.
j=1 i=1
Nos
If we substitute this relationship into Formula 5.4 we obtain

Nos e e
Nos
e c

X
−τos Nos + Nos X
Eos ≈ + RMOST oei + − e
(Nos c
Nos
e
) oei
i=1 i=1
Nos
e
Nos e c

−τos e c
X
e Nos + Nos
≈ RMOST (Nos + Nos ) + oi 1 − e
(5.5)
i=1
Nos
We can repeat the same calculations using the corresponding data from the PFS anal-
ysis to approximate total person-time of observation in the OS analysis, Epf s . We then
approximate the total person-time of observation in the ill state as
Eill ≈ Eos − Epf s . (5.6)
If we make the assumption that the number of i → d transitions is
e
Nid ≈ Nos − Nsimul , (5.7)
we can compute
Nid
λid ≈ .[250] (5.8)
Eill
53
5.3 MESCC Case Study
To evaluate whether the approximation method can generate reasonable results, we com-
pared approximate transitions rates against a gold standard of true transition rates esti-
mated from real study data.
Table 5.1: Data Abstracted from MESCC RCT PFS and OS Analyses
Description RT-alone Arm mS+RT Arm
PFS Analysis
e
Npf s total number of PFS events 17 31
c
Npf s total number of patients 5 6
censored from PFS analysis
e
Npf s
pei
P
person-time of PFS observation 7.02 years 27.22 years
i=1
τpf s maximum observation time in 2.97 years 5.25 years
the PFS analysis
AU Cpf s Area under PFS curve 0.63 1.16
OS Analysis
e
Nos total number of OS events 44 45
c
Nos total number of patients 1 3
censored from OS analysis
τos maximum observation time in 2.99 years 5.25 years
the OS analysis
N e
os
oei
P
person-time of OS observation 24.17 years 35.89 years
i=1
AU Cos Area under OS curve 0.62 0.98
Synthesis of PFS and OS Analyses
Nsimul total number of simultaneous 10 19
events in PFS and OS curves
Patchell et al. conducted a randomized controlled trial (RCT) comparing modern

surgery and radiotherapy (mS+RT) versus radiotherapy alone (RT-alone) for the treatment
of metastatic epidural spinal cord compression (MESCC).[171] MESCC occurs when cancer
metastasizes to the spine which and can lead to loss of ambulation from paralysis. MESCC
can be modelled as in Figure 5.1 if we consider ability to ambulate as the healthy state
h and the inability to ambulate due to neurologic dysfunction as the ill state i. True
transition rates were estimated using individual patient data provided by the study authors.
To eliminate the potential for transcription error and inaccuracy in individual patient
data reconstruction, we used actual individual patient data to generate the data listed
54
in Table 5.1. We estimated true transition rates using the Bayesian modeling language
Stan,[212] run through the statistical programming language R (Appendix D).[183] The
effect of mS+RT was parametrized as a log-hazard ratio for each RT-alone transition rate.
Prior to comparing true and approximate transition rates, we conducted non-parametric
multistate analysis to assess whether our assumed model (progressive, time-homogenous
and Markov) was appropriate for MESCC. Non-parametric multistate fractions were esti-
mated from individual patient data from the MESCC RCT using the etm library[1] run
through the statistical programming language R.[183] We compared non-parametric multi-
state fractions and multistate fractions calculated from true transition rates. Goodness-of-
fit tests for true multistate analysis of data observed with exact transition times affected
by right censoring have not been developed.[226] We therefore used informal graphical
methods.
Plots comparing proper non-parametric multistate and proper parametric multistate
fractions showed good agreement, and no evidence of systematic deviation (Figures 5.2
and 5.3). Therefore, a progressive time-homogenous three-state Markov model is appro-
priate for the MESCC RCT data and true transition rates can serve as an appropriate
comparator to evaluate approximate transition rates. Calculations for the mS+RT arm
are shown in Appendix D.3.
The true transition rates shown in Table 5.2 provides useful insights into the impact
of treatment. mS+RT prolongs ambulation with a statistically significant hazard ratio
of 0.53 (95% CrI: 0.30, 0.94) on the total transition rate for exit from the ambulatory
state (λhi + λhd ). For patients making a transition out of the ambulatory state, the risk of
death was similar with both treatments: relative risk 1.07 (95% CrI, 0.65 – 1.75). mS+RT
tended to increase the mortality rate for non-ambulatory patients, hazard ratio for λid of
1.61 (95% CrI: 0.89, 2.66), but this effect was not statistically significant.
All approximate transition rates lay within the 95% credible intervals for true transition
rates. There was no consistent direction of error indicating the approximation method does
not consistently under-or over-estimate true transition rates.
5.4 Simulation Study

To assess the validity of the approximation strategy in a wider set of conditions, we con-
ducted a simulation study to assess the impact of censoring on the accuracy of the approx-
imation method for (λhi + λhd ), ρ, and λid .
Data were generated randomly for a three-state progressive, time-homogenous Markov
disease process with parameters similar to those for the mS+RT arm from the MESCC
trial. A simulated cohort of 100 patients, 75 of which were healthy at baseline, was created
with λhi = 0.33, λhd = 0.53, λid = 3.28. Events times were independently censored using a
uniform distribution to achieve all combinations of 0, 2, 5, and 10 patients censored from
the OS and PFS analysis. 100 000 replications were generated for each set of simulation
conditions.
55
Table 5.2: Comparison of True and Approximate Transition Rates and Hazard Ratios for MESCC RCT
Transition Rates Hazard Ratios

True (95% CrI) Approximate Error % Error True (95% CrI) Approximate Error % Error
RT-alone group
λhi 0.66 (0.27, 1.20) 0.65 −0.01 −1.52
λhd 1.00 (0.50, 1.67) 0.93 +0.07 +7.53
56
λid 2.12 (1.40, 2.98) 2.24 +0.12 +5.66
mS+RT group
λhi 0.33 (0.17, 0.54) 0.33 0.00 0.00 0.50 (0.21, 1.40) 0.51 +0.01 +0.02
λhd 0.53 (0.32, 0.80) 0.52 −0.01 −1.89 0.55 (0.26, 1.16) 0.56 +0.04 +0.07
λid 3.28 (2.09, 4.69) 3.63 +0.35 +10.67 1.61 (0.89, 2.66) 1.62 +0.01 +0.01
CrI, Bayesian credible interval. Error, Approximate − True. % Error, (Approximate − True) ÷ True.
Figure 5.2: State-membership fractions for RT-alone arm. Non-parametric multistate
fractions, solid black line. Parametric multistate fractions based on true transition rates,
small dashed red line.
57
Figure 5.3: State-membership fractions for mS+RT arm. Non-parametric multistate
fractions, solid black line. Parametric multistate fractions based on true transition rates,
small dashed red line.
58
We calculated the mean error (ME), mean absolute error (MAE), mean percentage error
(%ME), and mean absolute percentage error (%MAE) for each set of simulation conditions
(Tables 5.3,5.4,and 5.5). ME and %ME are a measure of the direction of bias (systematic
over- or underestimation). MAE and %MAE are a measure of the magnitude of error,
regardless of direction.
The approximation method tended to underestimate (λhi + λhd ) and ρ as the censoring
rate increased, however the bias was small with %ME under 3% in all censoring conditions.
Even under no censoring, the approximation method was imprecise with a relatively high
MAE and %MAE; increasing censoring did not significantly decrease precision.
The approximation method tended to underestimate λid as the censoring rate increased,
however the bias was small with %ME under 3% in all censoring conditions. Even under
no censoring, the approximation method was imprecise with a relatively high MAE and
%MAE; increasing censoring did not significantly decrease precision.
The approximation method tended to underestimate λid as the censoring rate increased,
however the bias was small with %ME under 3% in all censoring conditions. Even under
no censoring, the approximation method was imprecise with a relatively high MAE and
%MAE; increasing censoring did not significantly decrease precision.
Table 5.3: Simulation Results for (λhi + λhd )
# Censored OS # Censored PFS ME MAE %ME % MAE
0 0 0.00 0.08 0.11 9.34

5 8 -0.06 0.10 -7.33 11.19
5 15 -0.10 0.12 -12.18 14.09
5 30 -0.21 0.21 -24.69 24.86
10 15 -0.10 0.12 -11.50 13.70
10 30 -0.20 0.21 -23.69 23.93
20 15 -0.08 0.11 -9.85 12.74
20 30 -0.19 0.19 -22.32 22.65
40 30 -0.17 0.17 -19.30 20.01
5.5 Discussion
Although chronic, progressive, and non-communicable diseases chronic diseases affect both
patients’ survival and quality-of-life, interventions may impact on these two outcomes
differentially. QALYs can simplify decision-making and counselling regarding treatment
options.[123] For clinicians and decision makers, QALYs calculated using multistate frac-
tions are useful because they can be used to extrapolate long-term quality-of-life and to
59
Table 5.4: Simulation Results for ρ
0 0 -0.00 0.04 -0.01 7.27

5 8 -0.00 0.05 -0.66 7.90
5 15 -0.00 0.05 -0.62 8.24
5 30 -0.00 0.06 -0.58 9.50
10 15 -0.01 0.05 -1.42 8.30
10 30 -0.01 0.06 -1.33 9.52
20 15 -0.02 0.05 -3.75 8.70
20 30 -0.02 0.06 -3.05 9.74
40 30 -0.06 0.08 -9.38 12.21
conduct rich decision analysis. Unfortunately, one cannot usually calculate multistate
fractions from PFS and OS curves.[226, 253, 251]
In this paper, we presented a technique for approximating transition rates, which can be
used to calculate multistate fractions, from PFS and OS analysis. Our technique requires
that three elements be abstracted from each of the PFS and OS analyses: (i) total number
of events, (ii) total number of censored patients, and (iii) event times.
Approximate transition rates provide a reasonable estimate of true transition rates
estimated using full multistate methods. For the MESCC RCT case study, all approximate
transition rates lay within the 95% Bayesian credible intervals for true transition rates.
The simulation study indicates that the approximation method is relatively unbiased and
precise for estimating the transition rate out of health (λhi + λhd ) and the risk of death for
healthy patients ρ.
It is important to recognize that our techniques only apply to a time-homogenous
progressive three-state irreversible disease process. Time-homogeneity is violated if the
transition rates change with time (i.e. any parametric model aside from the exponen-
tial) or depend on the amount of time spent in the preceding health state (non-Markov
phenomenon).[120] Irreversibility is violated if patients can become healthy after being
ill.[105] Our approximation approach can be scaled-up to more complex (e.g. reversible
transitions, >3 health states) disease models, however, the formulas will become more
complex. Furthermore, as was done in this article, it would be necessary to validate the
scaled-up approximation approach to evaluate for bias.
In this paper, we have demonstrated that transition rates can be approximated from
published PFS and OS analyses. The approximation method is more accurate for estimat-
ing the transition rates out of health than the transition rate out of illness. The method
tends to under-estimate true transition rates as censoring increases; therefore, approximate
transition rates are not a substitute for true transition rates estimated with full multistate
60
Table 5.5: Simulation Results for λid
0 0 0.07 0.41 2.25 12.42

5 8 -0.41 0.56 -12.59 17.19
5 15 -0.55 0.72 -16.81 21.92
5 30 -69.35 71.28 -2114.42 2173.07
10 15 -0.52 0.70 -15.96 21.35
10 30 -0.73 2.36 -22.13 72.08
20 15 -4.54 4.60 -138.32 140.30
20 30 0.79 2.75 24.22 83.88
40 30 -1.89 1.90 -57.67 57.88
methods. However, when proper multistate analysis is not available, approximate transi-
tion rates can guide probabilistic modeling and enhance QALY analysis if one considers
and accounts for the limitations of the approximation method.
61
Chapter 6
Decision Analysis of Prognostic

Contraindications to Surgery for
Metastatic Epidural Spinal Cord
Compression
Markian A. Pahuta1 , Joel Werier2 , Eugene K. Wai2,3 , Roy A. Patchell4 , Carl van Walraven3,5,6 ,
Doug Coyle3
Michigan, USA
tario, Canada
4 – Department of Neurosurgery, University of Kentucky Medical Center, Lexington,

Kentucky, USA
62
Context
The ultimate purpose of this thesis is to determine whether MESCC patients derive net
HRQoL benefit from mS+RT compared to RT-alone. This work is motivated by contro-
versy regarding the role of mS+RT despite Level 1 evidence that this treatment is more
effective than RT-alone for MESCC.
In this chapter I will try to resolve the controversy regarding the role of mS+RT by
determining if patients with short survival derive net HRQoL with mS+RT relative to
RT-alone. If patients benefit from surgery even under the worst-case survival scenario,
clinicians and patients could make treatment decisions with greater comfort.
The first four research chapters set the foundation for this final chapter. I conduct
Markov cohort simulation using a simple and transparent multistate MESCC disease model
developed in Chapter 5. QALYs are calculated using the utilities valued in Chapter 3. In an
effort to minimize perceptions of arbitrariness, the analysis is run using posterior Bayesian
distributions (Chapter 4).
63
Abstract
Objectives
To determine whether patients with metastatic epidural spinal cord compression (MESCC)
and poor prognosis can derive net health-related quality-of-life (HRQoL) benefit from mod-
ern surgery followed by radiotherapy (mS+RT).
Summary of Background Data

To benefit from mS+RT, patients should survive longer than the time needed to recover
and rehabilitate. Many clinicians consider an expected survival less than 3-6 months to be
a contraindication to surgery and instead offer radiotherapy alone (RT-alone).
Treatment decisions are particularly challenging for MESCC patients who have uncer-
tain prognoses. If survival is short, recovery from mS+RT and possible operative compli-
cations will impact on a greater proportion of their remaining life.
Methods
Decision analytic and Markov simulation methods were used to measure changes in incre-
mental quality-of-life-years (∆QALYs) in patients with a poor prognosis. Poor prognosis
was defined as median survival ≤3. The model was populated using outcomes from a ran-
domized controlled trial of mS+RT versus RT-alone for MESCC, and utility values from a
general population utility valuation study for MESCC. To model the belief that mS+RT
is more morbid than RT-alone, we implemented a “surgical penalty.”
Results
On average, baseline non-ambulatory and ambulatory patients with poor prognosis benefit
from mS+RT if the incremental disutility of mS+RT vs RT-alone on an average day (sur-
gical penalty) is less than 4.02 and 4.50 dysfunctions respectively. We are 97.5% confident
that mS+RT provides net HRQoL benefit for non-ambulatory patients with poor progno-
sis when the surgical penalty is at less than 1.78, and for ambulatory patients with poor
prognosis when the surgical penalty is less than 3.10.
Conclusions
This is the first study to critically examine prognostic contraindication to surgery for
MESCC treatment. We have demonstrated that MESCC patients with poor predicted
prognosis can derive HRQoL benefit for S+RT. Given the inaccuracy of prognostication
for MESCC, blanket survival indications/contraindications to surgery should be avoided.
64
6.1 Introduction
Contemporary surgery followed by radiotherapy (direct decompression and stabilization,
mS+RT) for symptomatic Metastatic Epidural Spinal Cord Compression (MESCC) pro-
vides sustained improvements to patients’ physical function and health-related quality-of-
life (HRQoL).[171, 65] However, patients should survive longer than the time needed to
recover and rehabilitate from surgery to benefit from that surgical intervention. Many sur-
geons regard an expected survival less than 3 – 6 months to be contraindication to surgery
for MESCC.[42, 229, 232, 228, 227] Patients with poor prognosis are offered radiotherapy
alone (RT-alone).
Although several prognostic indices have been developed and validated for estimating
expected survival of patients with spinal metastases, their use in surgical decision making
has been limited.[172, 132, 211] The most commonly used indices, the Tokuhashi and
Tomita scores, have only “fair” discriminative capacity for 6-month survival.[133, 99, 98]
Furthermore, the Tokuhashi and Tomita scores may not be applicable to patients treated
with modern chemotherapy. This is because these scores were developed in the 1990s and
2000s prior to the introduction of targeted therapies.[87, 236] Prognostic uncertainty poses
a significant challenge for the surgeon when the complication profile of the two treatment
options, mS+RT and RT-alone, is considered.
Patients undergoing RT-alone may experience nausea, dermatitis, and symptoms from
inflammation of the gastrointestinal tract (pharyngitis, esophagitis, and gastritis) within a
month of receiving spinal external beam radiotherapy. Less than 10% of patients will report
these symptoms to be severe.[146, 145, 237] Early toxicity may be even less frequent with
the newer technique of stereotactic radiosurgery because radiation is delivered to a smaller
field.[107] Infrequently, patients may experience late toxicities months after treatment,
including myelopathy and vertebral compression fractures.[144] When compared to those
from mS+RT, complications from RT-alone are less common and relatively mild.
Analysis of 30-day National Surgical Quality Improvement Program data found that
14.4% of patients undergoing mS+RT for primary and metastatic spinal tumors expe-
rience a major complication with 5.3% undergoing a re-operation and 10.2% being re-
admitted.[121] Prospectively collected data from the AOSpine MESCC study reported the
most common complications to be urinary tract (25.0%) and wound infections (10.4%).[65]
Adverse events were significantly higher in a single-center cohort of patients undergoing
emergency surgery for spinal metastases when complications were studied as the primary
outcome.[47] In this study, 76.2% of patients experienced at least one complication, and the
mortality rate during the admission was 10.9%. Intraoperative adverse events occurred in
31.7% of patients and included: blood loss >2L (16.8%), incidental durotomy (9.9%), mal-
positioned implants requiring revision (5.9%), nerve root injury (2%), and cardiac events
(1%). Postoperative adverse events occurred in 20.8% of patients and included: delir-
ium (20.8%), transient neurological deterioration (5.9%), pressure sores (4.0%), deep vein
thrombosis (4.0%), early construct failure (2.0%), electrolyte imbalance (11.0%), arrhyth-
mia (4.0%), dysphagia (5.0%). The risks of mS+RT are substantial and can negatively
affect HRQoL. However, patients experiencing complications can still derive overall net
benefit from mS+RT.[47, 3]
65
Given prognostic uncertainty and the risks of mS+RT, some clinicians may take a
conservative approach and not offer surgery out of concern for the worst-case scenario.[169,
196, 136] This is because recovery from mS+RT and complications will impact on a greater
fraction of remaining life in patients having short survival thereby negating net HRQoL
benefit. This conservative approach may deny patients with uncertain prognosis the more
effective treatment.
To help guide treatment decision making in the context of uncertain prognosis in
MESCC, it would be useful to determine whether such patients with poor prognosis can
derive net HRQoL benefit from mS+RT. In this paper we study this question using clinical
decision analytic methods combined with Markov simulation using data from a randomized
controlled trial comparing mS+RT and RT-alone for MESCC[171] and a general population
utility evaluation study for MESCC health states[168].
6.2 Methods
Markov simulation was used to measure changes in the incremental quality-of-life-years
(∆QALYs) for patients with poor prognosis undergoing mS+RT versus those undergoing
RT-alone. Poor prognosis was defined as median survival ≤3 months because 3 months is
the lower limit for which patients are generally considered to be mS+RT candidates.[227]
Measurement of QALYs was done from the societal perspective.
Estimation of Effectiveness and Model Structure

Effectiveness parameters were estimated by re-analysis of data from a randomized con-
trolled trial (RCT) by Patchell et al.[171] The trial compared radiotherapy alone (30 Gy in
ten fractions) to direct decompression and stabilization followed by radiotherapy (30 Gy in
ten fractions) for the treatment of symptomatic MESCC (Table 6.1). In this trial, patients’
health state (ambulatory, a; non-ambulatory, n; or dead, d) was assessed immediately be-
fore treatment and within one day of completion of treatment. Time-to-event data was
not collected during the treatment phase, while it was collected during the post-treatment
phase. Due to differences in data collection, effectiveness parameters were estimated in a
different way for the treatment phase and post-treatment phase.
Without time-to-event analysis, the treatment phase data could only be used to com-
pute the event probabilities being in the a, n or d state immediately after treatment. Time-
to-event data permitted calculation of transition rates and probabilities for a multistate
model. As transitions from the non-ambulatory to ambulatory state were not analyzed in
the MESCC RCT, we were unable to compute the rate and probability of this transition.
Event probabilities for the treatment-phase (θa , θn , and θd ) were estimated separately
for (1) baseline ambulatory patients undergoing RT-alone, (2) baseline non-ambulatory
patients undergoing RT-alone, (3) baseline ambulatory patients undergoing mS+RT, and
66
Table 6.1: Selection Criteria used in the MESCC RCT[171]
Inclusion Criteria Exclusion Criteria
-Adult patients (≥18 years old) -Multifocal MESCC on MRI

-MESCC -Primary spinal tumour (sarcoma,
-Tissue proven diagnosis of cancer chondrosarcoma, chordoma, meningioma)
-Radiologic evidence of spinal cord
displacement by tumor
-At least one neurologic sign attributable
to spinal cord compression (including pain)
(4) baseline non-ambulatory patients undergoing mS+RT. Treatment-phase event proba-

bilities were estimated using Bayesian Markov Chain Monte Carlo (MCMC) methods with
non-informative prior distributions (Appendix E.1.1).
We estimated post-treatment transition probabilities (λan , λad , and λnd ) separately for
each treatment group. A time-homogenous progressive three-state irreversible multistate
model fitted using Bayesian MCMC methods with non-informative prior distributions (Ap-
pendix E.1.2).
The state-transition diagram for the Markov simulation model is shown in Figure 6.1.
The model consists of three health states: ambulatory (a), non-ambulatory (n), and dead
(d). Due to the structure of the effectiveness data reported in the MESCC RCT, the sim-
ulation was divided into a treatment-phase and post-treatment phase. As in the MESCC
RCT, patients begin the simulation in either the a or n state. During the treatment phase,
patients are permitted to transition to any of the three states. After treatment, patients en-
ter a progressive time-homogenous Markov model meaning that transitions between states
are irreversible and occur at a constant rate. Furthermore, patients in the n state cannot
regain the ability to ambulate.
Population
Immediate post-treatment event probabilities and transition probabilities have not been
reported for MESCC patients with poor prognosis. We defined such a cohort by restricted
Monte Carlo sampling of the posterior distributions for immediate post-treatment event
probabilities. We used samples for which baseline ambulatory patients undergoing RT-
alone were expected to have a median survival of ≤2.25 months in the post-treatment
phase. This threshold was chosen because the treatment phase in the MESCC RCT was 3
weeks, thus the threshold corresponds to ≤3 months from the start of treatment.
Valuing Health States

For the MESCC health states shown in Figure 6.1, utilities have been estimated in a general
population valuation study and modeled using mixed effects beta regression.[168] This
67
a or n
θd
θa θn
d
λ nd
d
λa
a n
λan
Figure 6.1: State-transition diagram for the Markov simulation. The model consists of
three health states: ambulatory (a), non-ambulatory (n), and dead (d). Patients enter the
simulation in either state a or n. During the treatment-phase, patients can transition to any
of the three states; this transition is denoted by dashed arrows. Variable names adjacent to
the dashed arrows are treatment-phase event probabilities (θa , θn , and θd ). After treatment,
patients enter a progressive time-homogenous Markov model of the MESCC disease process.
Variable names adjacent to the solid arrows are post-treatment transition rates (λan , λad ,
and λnd ). See text for more details.
68
model was re-fit using Bayesian MCMC methods with non-informative prior distributions
(Appendix E.1.3).
However, the health states shown in Figure 6.1 do not account for a longer recovery
from mS+RT (compared to RT-alone) and the more severe complications associated with
this treatment. Consider two identical patients having the same health state; compared
to the patient receiving RT-alone, the patient receiving mS+RT is expected to have a
lower utility (at least initially). Therefore we would anticipate an initial penalty on utility
if a patient undergoes S+RT. The surgical penalty is time-dependent, in that it resolves
after (and if) the patient recovers from mS+RT. In some patients, the surgical penalty
may persist if a permanent complication occurs. The intensity of the surgical penalty also
varies. Unfortunately, the temporal course, frequency, and severity of the surgical penalty
of MESCC patients undergoing spine surgery has not been reported in the literature.
Because our analysis uses cohort simulation to compute mean ∆QALYs, granular data
on all characteristics of the surgical penalty are not necessary. We apply the surgical
penalty as a constant over the entire lifespan to all simulated patients. Therefore, the
surgical penalty can be viewed as the average decrease in utility attributable to mS+RT vs
RT-alone on an average day for the average patient. Since the surgical penalty is applied
as a constant over the entire lifespan, it is not necessary to model the time-course. Since
the surgical penalty is applied to all simulated patients, it is not necessary to model the
frequency.
As constructed, the impact of the surgical penalty on ∆QALYs also depends on the ex-
pected survival of a simulated cohort. For example, a low surgical penalty in a simulation
with long survival could result in the same ∆QALYs as a high surgical penalty in a simu-
lation with a short survival. We note that the objective of our analysis is to examine the
benefit of mS+RT in patient with poor prognosis in relation to the morbidity of this treat-
ment. Therefore, survival is a nuisance parameter which is addressed through probabilistic
Monte Carlo parameter sampling of posterior distributions for immediate post-treatment
event probabilities. Thus in our analyses, the confounding effect of expected survival on
∆QALYs is addressed by averaging out over expected survival periods consistent with a
cohort with poor prognosis.
The results of our general population MESCC utility valuation study can be used
to model the severity of the surgical penalty.[168] In this study, respondents were asked
to value a set of dysfunctions that characterize recovery from mS+RT and its potential
complications (Table 2.1). Standard gamble utility elicitation revealed that respondents
applied equal weight to all dysfunctions and therefore utility was related simply to the
number of dysfunctions. Thus we can quantify the average severity of the surgical penalty
by specifying the number of dysfunctions listed in Table 2.1.
A uniform distribution on the interval 0 to 4 was placed on the number of dysfunctions
in the surgical penalty. Posterior distributions for mean utility valuations and Bayesian
95% credible intervals for the number of dysfunctions are shown in Figure 6.2.
69
Figure 6.2: Histograms of posterior distributions of MESCC utility values for health
states with (a) no dysfunctions, (b) one dysfunction, (c) two dysfunctions, (d) three dys-
functions, (e) four dysfunctions, and (f) five dysfunctions. Mean and 95% credible intervals
reported in panel titles.
70
Analytical Methods, Parameter Uncertainty and Heterogeneity
We derived exact formulas for the results of Markov simulation for a time-homogenous pro-
gressive disease process using matrix algebra and integral calculus (Appendix E.2). Since
these formulas were derived using exact mathematical analysis, computational considera-
tions such as the cohort size, number of cycles, cycle length, and half-cycle correction are
not relevant.[22]
A probabilistic model was implemented through Monte Carlo sampling of the poste-
rior distributions for effectiveness parameters and utilities. Since Bayesian analysis yields
probability distributions jointly for all parameters, it was not necessary to directly spec-
ify covariance between parameters. The surgical penalty was modeled using a uniform
distribution on the number-of-dysfunctions. The model was run using 50 000 replications.
Separate analyses were conducted for baseline ambulatory and non-ambulatory pa-
tients. The relationship between mean and surgical penalty was examined using linear
regression ∆QALYs. Quantile regression was used to relate the 95% credible interval for
∆QALYs and the surgical penalty.[164]
6.3 Results
When the surgical penalty was between 0 and 1 dysfunctions, the mean ∆QALY (95% cred-
ible interval) benefit was 0.35 (0.09, 0.66) and 0.46 (0.21, 0.80) for baseline non-ambulatory
and ambulatory patients respectively. For both groups of patients, greater surgical penalty
was associated with a greater frequency of simulations with ∆QALYs<0 (Figure 6.3).
For baseline non-ambulatory patients we found that, on average, mS+RT provides
net HRQoL benefit if the surgical penalty was less than 4.02 dysfunctions (extrapolated).
However, when accounting for uncertainty in the model, we could only be 97.5% confident
that mS+RT provides net HRQoL benefit if the number of dysfunctions was less than 1.78
(Figure 6.4a). On average, ambulatory patients also derived HRQoL benefit from mS+RT
but with a threshold above 4.50 dysfunctions (extrapolated). We can only be 97.5% con-
fident of HRQoL benefit in non-ambulatory patients to a threshold of 3.10 dysfunctions
(Figure 6.4b).
6.4 Discussion
Although mS+RT is the most effective treatment for maintaining neurologic function in
MESCC patients, surgeons must make this recommendation cautiously due to prognostic
uncertainty.[169] It is generally recommended that appropriate surgical candidates have an
estimated survival of 3 – 6 months.[42, 229, 232, 228, 227] Although frequently used in the
literature, support and rationale for this prognostic threshold is usually not provided.
71
Figure 6.3: ∆QALYs for (a) baseline non-ambulatory patients, and (b) baseline am-
bulatory patients. Each point represents on replication of the simulation. Blue points
represent ∆QALYs > 0 indicating superiority of S+RT. Red points represent ∆QALYs <
0 indicating superiority of RT. Black points represent ∆QALYs = 0 indicating equivalence.
72
Figure 6.4: Mean ∆QALYs versus surgical penalty (a) baseline non-ambulatory patients,
and (b) baseline ambulatory patients. Shaded area represents 95% credible interval. X-
axis intercept is 4.02 and 4.50 (extrapolated) dysfunctions for baseline non-ambulatory and
ambulatory patients respectively. The lower limit of the credible interval is 1.78 and 3.10
dysfunctions for baseline non-ambulatory and ambulatory patients respectively.
73
Sundaresan and colleagues first proposed a 6-month threshold in a review article in
1986.[220] They noted that although neurologic response rates are equivalent with de-
compressive laminectomy and radiotherapy alone, decompressive laminectomy conferred a
more durable response. Since the majority of patients undergoing radiotherapy alone re-
lapsed within 6 months, they reasoned that patients with predicted survival greater than 6
months be offered surgery. This logic may not apply to the modern era because, in contrast
to decompressive laminectomy, neurologic response rates are higher with circumferential
decompression and stabilization (modern surgery) than with radiotherapy alone.[255, 171]
In this study, we used Markov simulation to evaluate the HRQoL benefit of mS+RT in
MESCC patients with poor prognosis (median survival ≤3 months). Contrary to widely
used guidelines, we found that both ambulatory and non-ambulatory patients with poor
prognosis can derive HRQoL benefit from mS+RT but this conclusion depends upon the
surgical penalty (measured by number of dysfunctions, Table 2.1)with ambulatory patients
able to withstand a larger surgical penalty than non-ambulatory patients.
Our methods permitted a more sophisticated decision analysis compared to previous re-
ports using Patchell et al.’s MESCC RCT data.[225, 75] To mitigate the effects of distribu-
tional assumptions on results, we employed posterior distributions obtained from Bayesian
analysis with non-informative prior distributions.[21] Previous MESCC decision analyses
used the results of partitioned survival analysis, whereas we used the results of multistate
analysis. Multistate analysis is advantageous because it permits extrapolation of study
data beyond the study observation period and application of study data to new patient
subgroups.[253] We populated the model using the only reported set of comprehensive of
MESCC utilities.[168]
It is important to consider our results in the context of new treatment paradigms
based on advances in radiation therapy techniques.[14] In contrast to conventional exter-
nal beam radiotherapy (cEBRT) used in the MESCC RCT, stereotactic body radiation
therapy (SBRT) more precisely delivers higher ablative doses of radiation.[106] Although
radiation toxicity is lower with SBRT compared with cEBRT, high focal doses of radiation
can predispose to compression fractures in otherwise stable lesions.[197] These competing
issues have not been studied and thus the relative treatment disutlity of cEBRT versus
SBRT is unknown. With higher focal doses of radiation, circumferential decompression
used in the MESCC RCT may not be necessary because SBRT-alone can effectively treat
even moderate-grade symptomatic MESCC.[196] However, surgery does still play a role in
downgrading compression to facilitate SBRT and optimize local control.[190] Surgery with
this intent is termed “separation surgery” and consists of minimal decompression to create
a margin of CSF between the spinal cord and tumor which has a lower surgical penalty
than circumferential decompression used in the MESCC RCT. No Level 1 comparative
studies of separation surgery versus SBRT-alone have been conducted. It is expected that
our decision analysis of the new treatment paradigm would be influenced by differences in
the (i) incremental disutility of separation surgery compared to SBRT-alone and (ii) the
impact of improved local control on maintenance of ambulation.
The main limitation of our study relates to the lack of published data on the surgical
penalty. We defined the surgical penalty as average decrease in utility attributable to
74
mS+RT vs RT-alone on an average day for the average patient. We can quantified the
average severity of the surgical penalty by specifying the number of dysfunctions listed
in Table 2.1. Rather than make a distributional assumption for the surgical penalty, we
conducted sensitivity analysis. We are 97.5% confident that mS+RT provides net HRQoL
benefit for patients with poor prognosis when the surgical penalty is less than 1.78 and
3.10 dysfunctions for baseline non-ambulatory and ambulatory patients respectively.
We made several assumptions in this study which may impact on results. First, pa-
tients could not regain the ability to ambulate after the initial treatment phase. This
assumption likely decreases overall QALYs in each treatment arm. This assumption could
also potentially impact on the ∆QALYs if neurologic recovery is tends to be delayed with
a treatment. Conceivably, recovery could be delayed from mS+RT due to need to recover
from the surgical insult. Second, we assumed that transition rates are constant, and did
verify this assumption by performing diagnostics for multistate models. Third, ambulation
was defined as a binary phenomenon. If a more granular definition of ambulation was used,
overall QALYs would likely be increased.
It is important to appreciate that the MESCC utilities used in this study were valued
by healthy individuals (ex ante utilities).[168] ex ante MESCC utilities from healthy in-
dividuals may not be equivalent to ex post utilities from patients who have experienced
MESCC. It has been reported that healthy individuals provide lower utility valuations
than patients who have experienced living with a colostomy, menopausal symptoms, and
hemodialysis.[15, 256, 46] However, the impact of experience may be transient as evi-
denced by former colostomy patients providing utility valuations equivalent to healthy
individuals.[206] Furthermore, in the case of depression, healthy individuals provide higher
utility valuations than patients.[182] Currently the understanding of differences between
ex ante and ex post utilities is insufficient to specify general conversion rules.[52]
Rawls argues that ex ante utilities can be used ethically if valued under a “veil of ig-
norance.” If we assume that the general population providing ex ante utility valuations
may eventually develop the condition of interest, out of self-interest, they should provide
fair valuations.[187] Although ex ante utilities are theoretically restricted to system policy
decisions, ex ante utilities have become the de facto standard for individual patient deci-
sions. Utilities obtained from generic health surveys such as the EQ-5D, HUI and SF-6D
are actually ex ante valuations.[55] Therefore our use of ex ante utilities conforms with
conventions in the literature.
These results can help patients and clinicians make decisions in the face of prognostic
uncertainty. By applying the results of this study, there is no need to rely on the worst-case
scenario, but instead rely on the most likely scenario identified through simulation methods
in this work. Our results indicate that both ambulatory and non-ambulatory patients can
gain net HRQoL benefit from mS+RT. However, it is important to only apply these results
to patients fitting the recruitment criteria for the MESCC RCT. Notably, patients with
multifocal disease and prolonged neurology were not recruited into the study.
75
6.5 Conclusion
This is the first study to critically examine prognostic contraindications to surgery for
MESCC. We have demonstrated that MESCC patients with poor predicted prognosis can
derive HRQoL benefit for mS+RT. On average, baseline non-ambulatory and ambula-
tory patients with poor prognosis benefit from mS+RT if the incremental disutility of
mS+RT vs RT-alone on an average day (surgical penalty) is less than 4.02 and 4.50 dys-
functions respectively. We are 97.5% confident that mS+RT provides net HRQoL benefit
for non-ambulatory patients with poor prognosis when the surgical penalty is at less than
1.78, and for ambulatory patients with poor prognosis when the surgical penalty is less
than 3.10. Given the inaccuracy of prognostication for MESCC, blanket survival indica-
tions/contraindications to surgery should be avoided.
76
Chapter 7
Discussion and Conclusion
Despite Level 1 evidence that surgery is the most effective treatment for metastatic epidu-
ral spinal cord compression (MESCC), there is controversy regarding the role of surgery
because of fear that patients who have a short survival will spend a large fraction of their
remaining life recovering from surgery and potential complications. In this thesis, I sought
to resolve this controversy by quality-adjusted life year (QALY) analysis of MESCC treat-
ments.
7.1 Summary of Methodological Findings and Con-

tributions
Prior to addressing the clinical problem, I tackled two methodological barriers: (a) lack of
MESCC utility data, and (b) skepticism regarding decision-analysis.

QALYs are calculated using utilities, or health-related quality-of-life (HRQoL) weights,
which can be obtained by direct valuation or from generic health status measures. For
MESCC, few high-quality studies compared modern surgery followed by external beam
radiotherapy (mS+RT) and external beam radiotherapy (RT-alone) using generic health
status measures, and no study has conducted a direct valuation. Therefore, with a paucity
of utility data, the ability to apply decision-analysis of QALYs to MESCC has been limited.
Given that few high-quality studies have compared mS+RT and RT-alone for MESCC
using generic health status measures, direct utility valuation was the most efficient way
to collect utility data. Best practices in economic evaluation involve recruiting a sample
of healthy individuals from the general population for direct utility valuation (ex ante
utilities).[248, 198]
Traditional utility valuation studies have been conducted using face-to-face interviews,
phone interviews, or postal surveys. These modes of administration have undergone psy-
chometric validation. However, there is increasing use of web surveys for utility valuation
77
which are usually conducted using custom and proprietary valuation tools which have not
been psychometrically validated. I felt it would be beneficial and efficient for investigators
to be able to build disease-specific modules on a common platform that has been used to
develop modules with acceptable psychometric properties.
In Chapter 2, I presented a new platform called the Self-directed Online Assessment of
Preferences (SOAP). For the first application of this platform, I developed a module for
MESCC health states. The SOAP platform met published benchmarks for reproducibil-
ity (both agreement and reliability) and responsiveness for utility measurement. To our
knowledge, this is the first validated open-source, web-based, self-directed utility valuation
module. This study demonstrated that the SOAP platform can be used to develop modules
which have acceptable psychometric properties
In Chapter 3, I reported the first general population direct utility valuation study for
a spinal disorder in the literature. This is in contrast to existing utilities for spine disease-
specific instruments (such as the NDI, the ODI, and the SRS-22) which have been derived
using an indirect “cross-walk” protocol.[193, 31, 32, 19] The “cross-walk” protocol involves:
1. collecting responses with a disease specific instrument and a generic health measure
(eg. EuroQol-5D, Short Form-6D, or Health Utilities Index Mark 3).
2. developing a regression model to relate the disease-specific score to the generic health
measure.
3. estimating a generic health measure score with the regression model.
4. converting generic health measure score to a utility value using another regression
model.[55]
The cross-walk protocol is complex as it requires two regression models. Chapter 2 demon-
strates that disease-specific direct utility valuation is feasible and valid. Investigators may
consider applying our study protocol and the validated SOAP tool to other disease con-
texts.
Barrier 2: Skepticism Regarding Decision-Analysis

Once utility data are available, QALYs can be estimated directly from a study or calculated
using simulation-based techniques. Study-based QALY estimates can be prohibitive due to
large sample size requirements.[20] Furthermore, study-based decision analysis of QALYs
is not a viable option when few high-quality comparative studies have collected HRQoL
data (as in MESCC). Unfortunately, the alternative (simulation-based QALY estimates)
are often viewed with skepticism.[58]
Markov cohort simulation is the most frequently used technique for estimating QALYs
through simulation.[22] This technique first involves simulating disease history from a the-
oretical model of the disease process composed of distinct and clinically-relevant health
states. The fraction of simulated patients in each health state over time is simulated using
78
transition probabilities derived from the literature. QALYs are then calculated by weigh-
ing the fraction of simulated patients in each health state by the utility of those health
states, and then summing the weighed state-membership fractions over the time-horizon
of interest.
One criticism of Markov cohort simulation relates to implementation of probabilistic
sensitivity analysis.[21, 36] The random draws are taken from probability density functions
derived from confidence intervals reported in the literature for each of the model inputs.[22]
In Chapter 4, I took care to highlight that confidence intervals do not actually quantify
the uncertainty in a value such as median survival time. Although confidence intervals are
often used for this purpose, this is done because confidence intervals are frequently mis-
interpreted by the “confidence-level misconception” to be Bayesian credible intervals.[28]
Under this misconception, a 95% confidence interval is interpreted to contain the true
value with a probability of 95%.[41] The prevalence of such “confidence-level misconcep-
tion” suggests that, for many individuals, the Bayesian credible interval is a more intuitive
construct for expressing the uncertainty in a reported value. An added benefit is that the
posterior distribution, from which a Bayesian credible interval is derived, can be used to
directly populate Markov cohort simulation. Therefore posterior distributions obtained
from Bayesian analysis with non-informative prior distributions can mitigate the effects of
distributional assumptions in Markov cohort simulation.
Another criticism of Markov cohort simulation is that the methods are complex and
poorly reported in the literature.[57, 63, 13, 56, 176, 192, 29] Due to the resulting lack of
transparency, some authors characterize simulation-base techniques as a “black box.”[29]
To be useful for treatment decision making, clinicians and patients must have faith in,
and an understanding of, these models.[16] This perception is aggravated by the fact that
transition probabilities are often improperly derived from partitioned-survival analysis.
In Chapter 5, I developed and evaluated a technique to approximate transition rates,
from PFS and OS analysis. Approximate transition rates reasonably estimate transition
rates estimated using full multistate methods. Approximation requires that three elements
be abstracted from each of the PFS and OS analyses: (i) total number of events, (ii) total
number of censored patients, and (iii) event times. It is my hope that other researchers
will use this approximation technique and abandon improper derivation of transition prob-
abilities from partitioned-survival analyses.
7.2 Summary of Main Clinical Findings

In Chapter 3, I reported directly valued utilities for MESCC health states using the vali-
dated SOAP MESCC module.[166] The study sample was representative of the population
in all English-speaking Canadian provinces. Our findings are valid with 63.2% of partic-
ipants having understood and engaged in the task. This rate is superior to validity rates
reported in general population direct valuation studies for the EuroQol-5D.[12, 128, 113, 49]
The regression model building exercise revealed that members of the general popu-
lation value all dysfunctions characterizing MESCC health states equally. Furthermore,
79
dysfunction follows a pattern of diminishing marginal disutility. That is, each additional
dysfunction effects a smaller incremental change in utility than the previous dysfunction.
These results demonstrate that, from the societal perspective, physical function is valued
equal to other facets of well-being. Ambulation and continence (which are dysfunctions
that are addressed by surgery) are no more important to overall health than other at-
tributes (including pain, other symptoms, and level of independence).
In Chapter 4, I applied Bayesian analytical methods to prognostication for MESCC
in an era of more effective systemic treatments. I explored the impact of new targeted
therapies on prognostic estimates obtained using the Tokuhashi score. I provide a worked
calculation for the probability of a patient surviving to 6 months using dichotomous prog-
nostication. Continuous prognostication was performed using conjugate analysis. When
the predicted prognosis from the Tokuhashi score is contextualized with data on contem-
porary systemic treatments, patients previously deemed non-surgical candidates may be
eligible for surgery.
Some clinicians may be reluctant to adopt Bayesian techniques based on R.A. Fisher’s
criticism that the Bayesian approach is “. . . extremely arbitrary... by evolving a vitally
important piece of knowledge, that of the exact form of the distribution of [the value of
interest], out of an assumption of complete ignorance . . . ”[72] This is a fair criticism, but
it should be considered in context of the non-Bayesian alternative courses of action.
A nihilistic course of action based on non-Bayesian analysis would be to ignore recent
advances in systemic treatment because their effectiveness for patients with metastatic
spine disease has not been formally studied. An overly optimistic course of action based
on non-Bayesian analysis would be to assume that the results of oncology trials enrolling
few patients with metastatic spine disease fully apply to patients with metastatic disease.
In Chapter 4, I demonstrated how effective sample size sensitivity analysis quantifies the
impact of clinicians’ assumptions and biases on decision making and, therefore, can make
decisions more transparent. In this manner, different – and potentially not fully compatible
– sources of evidence can be synthesized for decision making in a rational and transparent
way. Therefore, Bayesian prognostication can enhance counseling for metastatic spine
disease because clinicians’ assumptions and biases are formally incorporated into analysis.
In Chapter 6, I applied the work from the previous four research chapters and used
Markov simulation to evaluate the HRQoL benefit of mS+RT in MESCC patients with
poor prognosis (median survival ≤3 months). Contrary to widely used guidelines, I found
that both ambulatory and non-ambulatory patients with poor prognosis can derive HRQoL
benefit from mS+RT but this conclusion depends upon the surgical penalty (as measured
by number of dysfunctions resulting from surgery, Table 2.1) with non-ambulatory patients
able to withstand a smaller surgical penalty than ambulatory patients.
7.3 Limitations
It is important to appreciate that the MESCC utilities used in this study were valued
by healthy individuals (ex ante utilities).[168] ex ante MESCC utilities from healthy in-
80
dividuals may not be equivalent to ex post utilities from patients who have experienced
I made several assumptions in Chapter 6 which may impact on results. First, patients
could not regain the ability to ambulate after the initial treatment phase. This assump-
tion likely decreases overall QALYs in each treatment arm. This assumption could also
potentially impact on the ∆QALYs if neurologic recovery is tends to be delayed with a
treatment. Conceivably, recovery could be delayed from mS+RT due to need to recover
from the surgical insult. Second, I assumed that transition rates are constant, and did
verify this assumption by performing diagnostics for multistate models. Third, ambulation
was defined as a binary phenomenon. If a more granular definition of ambulation was used,
overall QALYs would likely be increased.
The main limitation of Chapter 6 relates to the lack of published data on the surgical
penalty. I defined the surgical penalty as average decrease in utility attributable to mS+RT
vs RT-alone on an average day for the average patient. I then quantified the average severity
of the surgical penalty by specifying the number of dysfunctions listed in Table 2.1. Rather
than make a distributional assumption for the surgical penalty, I conducted sensitivity
analysis. I am 97.5% confident that mS+RT provides net HRQoL benefit for patients
with poor prognosis when the surgical penalty is less than 1.78 dysfunctions for baseline
non-ambulatory patients and 3.10 dysfunctions for ambulatory patients.
7.4 Future Work

The surgical penalty is an abstract concept. Decision-analysis of QALYs could be better
conducted through a high-quality comparative study of mS+RT and RT-alone that collects
patient-reported HRQoL outcomes. Unfortunately, the utility data obtained through the
vignette study described in Chapter 3 has limited application for this purpose.[252] There
now exist better tools for this purpose.
The Spine Oncology Study Group (SOSG) has developed a spine oncology-specific
outcome questionnaire (SOSGOQ).[217] The original questionnaire consisted of 27 items
belonging to five domains: physical function, neurologic function, pain, mental health and
social function. The tool has face validity based on expert opinion.[217] SOSGOQ also
has demonstrated content validity through high correlation International Classification of
Functioning and EQ-5D.[217, 111] A revised version of the tool, SOSGOQ2.0, has recently
undergone successful psychometric validation.[242] However, there currently exists no map-
ping of this tool to utilities. I am collaborating with the AOSpine Knowledge Forum Tumor
to create a SOAP module for a shortened version of the SOSGOQ2.0, SOSGOQ-8D with
81
the aim of conducting a general population valuation study. More recently, discrete choice
experiements (DCEs) have been used for online utility valuation because cognitive burden
is lessened.[157, 112] We will utilize the EuroQol EQ-VT protocol for DCE, implemented
on the SOAP platform, to measure utilities for SOSGOQ-8D.[161, 162]
It is important to appreciate that the MESCC utilities reported in Chapter3 were
valued by healthy individuals (ex ante utilities). MESCC utilities from healthy indi-
viduals may not be equivalent to ex post utilities from patients who have experienced
Rawls argues that ex ante utilities can be used ethically if valued under a “veil of ig-
norance.” If we assume that the general population providing ex ante utility valuations
may eventually develop the condition of interest, out of self-interest, they should provide
fair valuations.[187] Although ex ante utilities are theoretically restricted to system pol-
icy decisions, ex ante utilities have become the de facto standard for individual patient
decisions. Utilities obtained from generic health surveys such as the EuroQol-5D, Short
Form-6D, and Health Utilities Index Mark 3 are actually ex ante valuations.[55] Therefore
our use of ex ante utilities conforms with conventions in the literature. However, it would
be informative to conduct a utility valuation study with MESCC patients and derive ex
post utilities. These utilities could be used to create a true patient decision aid.
7.5 Conclusion
In summary, my thesis explored prognostic contraindications to surgery for MESCC and
determined that existing guidelines based on expert opinion are overly restrictive. As
part of this thesis, I presented solutions to two barriers to decision-analysis of QALYs in
MESCC: (a) lack of utility data and (b) skepticism regarding decision-analysis. These
barriers are not unique to MESCC. It is my hope that this thesis and the methodological
contributions will be disseminated to facilitate decision-analysis of QALYs in other disease
contexts.
82
APPENDICES
83
Appendix A
Appendices for Chapter 2
A.1 SOAP MESCC Module

After logging-in, respondents are first presented with a description of MESCC and its func-
tional consequences (Figure A.1). Next, the basic premise of the standard gamble task is
outlined (Figure A.2). The best outcome (perfect health) is then outlined (Figure A.3), and
this outcome is encapsulated as “Chris” whose characteristics are easily accessible in the
remainder of the survey (Figure A.4). The flow the website is then outlined (Figure A.5).
The standard gamble task begins with a tabular description of the probe health state
(Figure A.6). Participants then move to a screen with the probability wheel (Figure A.7).
Rejecting a lottery is indicated by “Reject the Pill,” accepting the lottery by “Take the
Pill,” and indifference to a lottery by “I Can’t Decide if the Risk is Worth it.” A description
of the best outcome is accessible by clicking on “Chris” (Figure A.8). The probe health
state is also accessible (Figure A.9).
The search algorithm presents the following sequence of probabilities for the best and
worst outcome: 100-0, 10-90, 90-10, 20-80, 80-20,. . . 50-50. The first iteration always
proposes a lottery with a 100% probability of the best outcome. If the participant does
not accept the lottery, a pop-up window appears which highlights the potential logical
error with an explanation (Figure A.10). If the participant still refuses the lottery, the
probe health state is assigned a utility of 1.
The search algorithm stops presenting lotteries when a lottery with a probability of the
best outcome above 50% is rejected, or a lottery with the probability of the best outcome
below 50% is accepted. The algorithm also stops if participants are indifferent to a lottery.
If the participant rejects a lottery above the 50% probability, the utility of the probe
health state is imputed to be:
probability best outcome
+ 0.05
100
If the participant accepts a lottery below the 50% probability, the utility of the probe
84
health state is imputed to be:
probability best outcome
− 0.05
100
If the participant is indifferent to a lottery, the utility of the probe health is imputed to
be the probability of the best outcome.
85
Figure A.1: Description of impairments caused by MESCC.
86
Figure A.2: Explanation of Standard Gamble task.
87
Figure A.3: Caveats to and nuances of best possible outcome.
88
Figure A.4: Description of perfect health.
89
Figure A.5: Instructions on navigating webpages.
90
Figure A.6: Probe health state in tabular form.
91
Figure A.7: Probability wheel.
92
Figure A.8: Ability to review definition of perfect health.
93
Figure A.9: Ability to review probe health state.
94
Figure A.10: Logic check.
95
A.2 SOAP platform
Assessment Approach
Scaling Method
The time-trade-off and standard gamble can be used to directly elicit utilities. However,
it was felt that framing health scenarios in the time-trade-off would be difficult as the
life-expectancy of patient with MESCC is less than five years. Therefore, the standard
gamble was used for the SOAP MESCC module.
Frame
Typically, in the standard gamble success is framed as “perfect” health for an undermined
period of time. This anchor is not appropriate for MESCC because metastatic cancer
cannot be cured, and life expectancy is relatively short. Therefore the fully functional
health state, D-N-I-P-S- was chosen as the success anchor. To eliminate confusion around
life expectancy, all scenarios were framed as having a certain life expectancy of five years;
that is for both the probe health scenario, and success health scenario, participants were
told their life expectancy would certainly be five years.
Search Procedure
For each health state, it is necessary to vary the probabilities of the gamble until partic-
ipants are indifferent to the lottery. Three “search procedures” have been described for
varying probabilities:
1. Ping-Pong Variant (PPV). This is the classic search algorithm proposed by von
Neumann and Morgenstern.[243] In contemporary applications, the gambles oscillate
between extremes of probability: 100-0, 10-90, 90-10, 20-80, 80-20,. . . 50-50. The
algorithm stops when a gamble with success above 50% is rejected, or a gamble
with success below 50% is accepted. The algorithm also stops if participants are
indifferent to a gamble. With 10% decrements, the algorithm iterates a maximum of
11 times.[76]
2. Titration Up Variants. The Titration Up Variant (TUV) algorithm was first

described in a transportation economics paper.[116] Here the probability of success
progressively decreases: 100-0, 90-10, 80-20,... 0-100. In the Titration Down Variant
(TDV), the probability of success increases: 0-100, 10-90, 20-80,. . . 100-0.
3. Bisection Variant (BV). This algorithm was first applied in U-titer, a comput-
erized utility elicitation tool.[219] The algorithm is applied by offering 100-0, 0-100
and then successively bisecting the difference between the probe gamble and either
100 or 0, depending on whether the gamble was accepted or rejected. For example,
96
if 50-50 was rejected, the next gamble would be 75-25; while 25-75 would be posed
if the gamble was accepted. An arbitrary limit is usually place on the number of
iterations. To obtain a precise result, it is necessary to use at least six iterations;
consequently, respondent fatigue can be an issue with this procedure.
Unfortunately, there are no guidelines for choosing among the four algorithms. More-
over, three studies have demonstrated that the titration variants and PPB do not produce
equivalent results (the bisecting algorithm has not yet been compared to other methods).
In a computer utility elicitation study, precision and utility estimates were found to be
higher (mean difference between 0.1 to 0.15) with the BUV compared to the PPV.[137]
Comparing interviewer-administered PPV to self-administered paper-based TDV, Ham-
merschmidt found PPV to generate higher utility estimates than TDV.[97] In the same
study, no significant difference was found between estimates using the TDV and BUV.
Another study compared interviewer led PPV and TDV, and reported utility estimates to
be higher with the TDV.[17] These three papers are heterogeneous in methods and results,
making it difficult to synthesize a coherent recommendation. Brazier and Dolan, however,
offer a sensible conclusion on the differences between the TDV and PPV.[17] They suggest
that based on the starting-point, the TDV anchors participants to the upper end of the
utility scale. Alternatively, they suggest the PPV may confuse participants and encourage
risk taking. However, the alternative conclusion is not supported by their results as inter-
nal consistency was equivalent in both the TDV and PPV groups. Anchoring is further
supported by Lenert’s results as PPV produced higher utility estimates than the BUV.
Although Hammerschmidt’s results counter the presence of anchoring, they are likely con-
founded by the mode of administration. In summary, there is weak evidence for anchoring
with titration which is plausible given the ordering of risks posed by the search procedures.
We selected the PPV because it is the classical search procedure and may protect against
anchoring.
Navigation Rules
Sequence Specification
The SOAP tool is flexible by allowing researchers to vary the number and order of health
scenarios administered to participants. Each scenario asks participants to rate up to ten
different combinations of probabilities.[76]
Error Checking
Participants may go back one iteration to correct an error. The first iteration always
proposes a lottery with 100% chance of success. If participants do not choose to take
treatment, a pop-up window appears which identifies the logical error, and explains why
this choice was not logical.
97
Auxiliary Functions
Questionnaire Administration
Traditionally, direct valuation administered face-to-face by an interviewer using props to

illustrate probabilities.[76] Clearly, this method is time-intensive, and presents logistical
difficulties in recruiting a representative general population sample. Consequently, there is
a growing body of literature utilizing developed self-directed standard gamble elicitation
exercises over the internet.
Data Management
SOAP has been developed to work with Microsoft Internet Explorer, Google Chrome,
Mozilla Firefox, Apple Safari, and the Blackberry Browser on Windows, Apple OS, iOS,
Linux and Blackberry desktop and touch-screen configurations. The website can be run
with the most restrictive security settings, and does not generate any security errors.
Participants login with a secure password. The first set of web pages prepare the participant
for the task. The second set of pages present the standard gambles. The last section elicits
demographic information. SOAP is fully customizable, and the ordering can be easily
changed. SOAP was programmed in the open-source PHP and JavaScript language. These
languages were chosen as it is unlikely they will be depreciated in the future. Results are
stored in an open-source mySQL database, however, most major databases are supported.
Training
The first set of web pages prepare the participant for the task.
98
A.3 MESCC Health States
D-N-I-P-S-
In Scenario X you are 60 years old, you have spine cancer and...
• You do not need help with eating, dressing, washing yourself, or getting on the toilet
• You walk without assistance.
• You have no trouble controlling your bladder.
• You have back pain relieved by pain medications.
• You have no other uncomfortable symptoms.
D-N-I-P-S+
• Everyday, you have one or more of: nausea, vomiting, shortness of breath, lack of
appetite, or diarrhea.
D-N-I-P+S-
• You have back pain not relieved by pain medications.
99
D-N-I-P+S+
D-N-I+P-S-
• You need to insert a tube in your bladder to control your bladder.
D-N-I+P-S+
100
D-N-I+P+S-
D-N-I+P+S+
D-N+I-P-S-
• You have lost mobility because of weakness in your legs. You need to stay in bed or
a chair (or wheelchair) during the day.
101
D-N+I-P-S+
D-N+I-P+S-
D-N+I-P+S+
102
D-N+I+P-S-
D-N+I+P-S+
D-N+I+P+S-
103
D-N+I+P+S+
D+N-I-P-S-
• You need help with eating, dressing, washing yourself, or getting on the toilet.
D+N-I-P-S+
104
D+N-I-P+S-
D+N-I-P+S+
D+N-I+P-S-
105
D+N-I+P-S+
D+N-I+P+S-
D+N-I+P+S+
106
D+N+I-P-S-
D+N+I-P-S+
D+N+I-P+S-
107
D+N+I-P+S+
D+N+I+P-S-
D+N+I+P-S+
108
D+N+I+P+S-
D+N+I+P+S+
109
Appendix B
B.1 Regression Model for Utilities

We wish to relate utility values to a set of predictors using regression. A general linear
model (GLM) is inappropriate for two reasons. The first, utilities can only take on values
in the closed interval [0, 1] whereas a GLM supports (−∞, ∞). The second, the GLM
assumes homoskedacticity which is not a reasonable assumption for utilities.[50] By virtue
of the fact that utilities cannot be greater than 1 or less than 0, dispersion in valuations for
health states with mean 0.5 will be greater than dispersion in valuations for health states
at the extremes near 0 and 1.
Generalized Linear Models (GzLMs) based on the beta distribution, which supports
the open interval (0, 1), offer a potential solution to the limitations of GLMs for utilties.
However, adjustments are required as utilities can in include the values 0 and 1. One
solution is to transform utilities to the closed unit interval:[207]
y (n − 1) + 0.5
y∗ = .
n
An alternative strategy is to use a beta-inflated model which is a mixture model of
three separate regression models: the standard beta model and binomial models for values
of 0 and 1.[165] We elected to use the simpler and more readily interpretable beta model
with utilities transformed to the closed unit interval.
The Beta Distribution

The beta probability density function (pdf) is an equation which describes the contour
of the histogram for a beta distributed random variable Y .[67, 115, 149] The relative
frequencies generated by a pdf, f (y), are termed probability densities which for the beta
distribution satisfy Z 1
f (y)dy = 1
0
110
Using this definition, the probability that Y ≤ y can be computed using the cumulative
distribution function (cdf) F (y)
Z y
Pr (Y ≤ y) = F (y) = f (y)dy
0
The beta pdf is a function of two shape parameters: p ∈ R+ and q ∈ R+ . Both parameters
influence the mean (E [Y ]), variance (Var [Y ]), skewness, and kurtosis (peakedness and tail
weight) of the beta pdf.[11, 170] When p = q the beta pdf is symmetric about Y = 0.5,
and larger values of the equivalent shape parameters increase peakedness while decreasing
tail-weight. When p < q, the beta pdf is right-skewed, and left-skewed when p > q. When
p 6= q, p tends to control the mode, and q tends to control the tail weight.
The mean of a beta distributed variable Y is
p
E [Y ] = µ =
p+q
The variance of a beta distributed variable Y is
pq
Var [Y ] = 2
(p + q) (p + q + 1)
For exponential family of distributions, such as the beta, the variance can be partitioned
as
Var [Y ] = V (µ) σ
where V is the variance function and σ is the precision parameter. For exponential-family
distributions, V is a function of the mean µ, and for the beta distribution is
V (µ) σ = µ (1 − µ) .
It is through the variance function that heteroscedasticity can be modeled with the beta
distribution because variance is smaller at the upper and lower bounds of the distribution.
While the precision parameter
σ =1+p+q
controls the relative dispersion for a given mean.
Beta Regression
The beta GzLM[67, 149] relates the mean for a set of n covariates xi to a linear predictor
using an inverse link function g −1 :
n
!
X
−1
E [Y ] = µ = g β i xi
i=1
The logit function is often used for g −1 for beta regression. The dispersion parameter σ is
often treated as a constant for beta regression.
111
B.2 Quota Samples Based on 2016 Canadian Census
Table B.1: Quotas Based on the 2016 Canadian Census
Region Gender Age Pop. % Cnd. Pop. Quota

Maritimes Female 18-24 98041 0.444064 3
Maritimes Female 65+ 242201 1.097017 7
Maritimes Male 18-24 104383 0.472789 3
Maritimes Male 25-34 142078 0.643523 4
Maritimes Male 35-44 145154 0.657456 4
Maritimes Male 45-54 178200 0.807133 5
Maritimes Male 55-64 178502 0.808501 5
Maritimes Male 65+ 203558 0.921989 6
Ontario Female 18-24 655894 2.970785 19
Ontario Female 25-34 953313 4.317904 28
Ontario Female 35-44 923760 4.184048 27
Ontario Female 45-54 1033308 4.680231 31
Ontario Female 55-64 930168 4.213072 27
Ontario Female 65+ 1219557 5.523821 36
Ontario Male 18-24 689380 3.122455 20
Ontario Male 25-34 921938 4.175795 27
Ontario Male 35-44 882531 3.997307 26
Ontario Male 45-54 1023430 4.63549 30
Ontario Male 55-64 892852 4.044054 26
Ontario Male 65+ 991595 4.491298 29
Manitoba Female 18-24 62942 0.285087 2
Manitoba Female 25-34 90091 0.408055 3
Manitoba Female 35-44 83631 0.378795 2
Manitoba Female 45-54 86261 0.390708 3
Manitoba Female 55-64 79745 0.361194 2
Manitoba Female 65+ 105817 0.479284 3
Manitoba Male 18-24 65881 0.298399 2
Manitoba Male 25-34 91173 0.412956 3
Manitoba Male 35-44 80922 0.366525 2
Manitoba Male 45-54 88596 0.401284 3
Manitoba Male 55-64 81220 0.367875 2
Manitoba Male 65+ 86100 0.389978 3
Saskatchewan Female 18-24 51435 0.232968 2
Continued on next page
112
Table B.1 – Continued from previous page
Region Gender Age Pop. % Cnd. Pop. Quota
Saskatchewan Female 65+ 90911 0.411769 3
Saskatchewan Male 18-24 57438 0.260158 2
Saskatchewan Male 65+ 74983 0.339626 2
Alberta Female 18-24 190216 0.861558 6
Alberta Female 25-34 347060 1.571962 10
Alberta Female 35-44 297495 1.347464 9
Alberta Female 45-54 276512 1.252424 8
Alberta Female 55-64 244383 1.1069 7
Alberta Female 65+ 263549 1.19371 8
Alberta Male 18-24 203940 0.923719 6
Alberta Male 25-34 364870 1.65263 11
Alberta Male 35-44 318403 1.442164 9
Alberta Male 45-54 296531 1.343098 9
Alberta Male 55-64 256490 1.161737 8
Alberta Male 65+ 223618 1.012848 7
British Columbia Female 18-24 212117 0.960756 6
British Columbia Female 65+ 434352 1.96734 13
British Columbia Male 18-24 231772 1.04978 7
British Columbia Male 65+ 383712 1.737972 11
113
B.3 Regression Modeling
Fitting
We considered three base-case regression models for the linear predictor (Model 1, Model
2 and Model 3) (Table B.2). A random intercept for each participant i was incorporated
in the models. We considered φ a fixed effect and constant in each model.
Prior to comparing models, we checked for misspecification of the variance and/or
or link function. The variance function was checked by comparing normalized quantile
residuals against fitted values (Figure B.1). As utilities lie in the closed interval [0, 1],
residuals also lie in an interval which for fitted values of 0 is (0, 1) and for fitted values of 1
is (−1, 0). The smoothed line indicated that the mean residual for all fitted values was 0 and
there was no trend. Therefore, we concluded the variance function was properly specified.
The link function was checked by comparing the adjusted dependent variable against linear
predictor for µ (Figure B.2). Each plot demonstrated a straight line supporting the choice
of link function. Of the three base-case regression models, Model 2 was optimal as it had
the lowest AIC and BIC (Table B.3).
A plot of these nominal coefficients against the number of dysfunctional attributes
revealed a non-linear concave up relationship (Figure B.3(a)). Tukey’s and Mosteller’s
Bulging Rule guided that a square root (Model 4) or logarithmic transformation (Model
5) of the number of dysfunctional attributes could linearize the relationship (Figure B.3(b)
and (c)).[235] The beta regression model was re-fitted using Model 4 and Model 5. Model
4 was optimal yielding the lowest AIC and BIC (Table B.3).
Formulas for the linear predictor in the beta regression are provided in Table B.6. Mean
utilities for each health state estimated by each regression model are shown in Table B.5.
Calibration and agreement parameters for all candidate models are provided in Table B.4
and Figures B.4, B.5, B.6, 3.2, and B.7. We will calculate mean utilities for the health
state D-N+I-P+S- to illustrate the use of the formulas listed in Table B.6.
For Model 1, this health states is coded as: D= 0, N= 1, I= 0, P= 1, S= 0. The mean
utility valuation is calculated as
inverse logit (0.90 − 0.32(0) − 0.39(1) − 0.26(0) − 0.40(1) − 0.43(0)) = 0.53.
For Model 2, this health states is coded as: two dys= 1, three dys= 0, four dys= 0, five
dys= 0. The mean utility valuation is calculated as:
inverse logit (0.59 − 0.44(1) − 0.84(0) − 1.12(0) − 1.27(0)) = 0.54.
For Model 3, this health states is coded as: D= 0, N= 1, I= 0, P= 1, S= 0, D×N= 0,

D×I= 0, D×P= 0, D×S= 0, N×I= 0, N×P= 1, N×S= 0, I×P= 0, I×S= 0, P×S= 0. The
mean utility valuation is calculated as:
inverse logit (1.18 − 0.58(0) − 0.59(1) − 0.64(0) − 0.45(1) − 0.62(0)
+ 0.17(0) + 0.16(0) − 0.03(0) + 0.21(0) + 0.32(0)
− 0.01(1) − 0.12(0) + 0.05(0) + 0.20(0) +0.05(0)) = 0.53.
114
For both Model 4 and Model 5 the health state is coded as: num dys= 2. Using Model
4 the mean utility valuation is calculated as:
√
inverse logit 1.71 − 1.11 × 2 = 0.53.
Using Model 5 the mean utility valuation is calculated as:
inverse logit (0.64 − 0.81 × ln 2) = 0.52.
115
Table B.2: Functional Form of Linear Predictor Formulas for Beta Regression
Functional form
Model 1 β0 + β1 (D) + β2 (N) + β3 (I) + β4 (P) + β5 (S) + i

Model 2 β0 + β1 (two dys) + β2 (three dys) + β3 (four dys) + β4 (five dys) + i
Model 3 β0 + β1 (D) + β2 (N) + β3 (I) + β4 (P) + β5 (S) + β6 (D × N) + β7 (D × I) +
β8 (D × P) + β9 (D × S) + β10 (N × I) + β11 (N × P) + β12 (N × S) +
β13 (I × P) + β14 (I × S) + β15 (P × S) + i
√
Model 4 β0 + β1 num dys + i
Model 5 β0 + β1 ln (num dys) + i
Table B.3: Selection Parameters for all Models
Description AIC BIC
Model 1 Attributes -2750.24 -899.148

Model 2 Number of dysfunctions (nominal variable) -2751.78 -906.259
Model 3 Attributes and second-order interactions -2747.42 -840.572
Model 4 Square root of number of dysfunctions -2756.85 -928.061
Model 5 Logarithm of number of dysfunctions -2754.9 -926.11
116
Table B.4: Calibration and Agreement Parameters for all Candidate Models
Mean Absolute Mean Signed Intraclass Correlation

Error Difference Coefficient (95% CI)
Model 1 0.046036 -0.0161 0.931 (0.857, 0.967)
Model 2 0.046548 -0.01967 0.937 (0.863, 0.97)
Model 3 0.053878 -0.03416 0.917 (0.766, 0.965)
Model 4 0.046582 -0.01934 0.936 (0.86, 0.97)
Model 5 0.049875 -0.02303 0.927 (0.834, 0.966)
117
Table B.5: Comparison of Model-Based Mean Utilities for all MESCC Health States
Health State Observed Mean Model 1 Model 2 Model 3 Model 4 Model 5

D+N-I-P-S- 0.67 0.64 0.644 0.645 0.644 0.654
D-N+I-P-S- 0.73 0.624 0.644 0.644 0.644 0.654
D-N-I+P-S- 0.7 0.655 0.644 0.633 0.644 0.654
D-N-I-P+S- 0.71 0.622 0.644 0.676 0.644 0.654
D-N-I-P-S+ 0.71 0.615 0.644 0.636 0.644 0.654
D+N+I-P-S- 0.5 0.546 0.538 0.543 0.533 0.519
D+N-I+P-S- 0.6 0.579 0.538 0.531 0.533 0.519
D+N-I-P+S- 0.56 0.544 0.538 0.529 0.533 0.519
D+N-I-P-S+ 0.58 0.537 0.538 0.493 0.533 0.519
D-N+I+P-S- 0.68 0.562 0.538 0.567 0.533 0.519
D-N+I-P+S- 0.54 0.527 0.538 0.533 0.533 0.519
D-N+I-P-S+ 0.52 0.519 0.538 0.461 0.533 0.519
D-N-I+P+S- 0.56 0.56 0.538 0.538 0.533 0.519
D-N-I+P-S+ 0.61 0.553 0.538 0.531 0.533 0.519
D-N-I-P+S+ 0.55 0.518 0.538 0.539 0.533 0.519
D+N+I+P-S- 0.47 0.482 0.438 0.504 0.445 0.437
D+N+I-P+S- 0.41 0.447 0.438 0.421 0.445 0.437
D+N+I-P-S+ 0.51 0.439 0.438 0.36 0.445 0.437
D+N-I+P+S- 0.36 0.48 0.438 0.425 0.445 0.437
D+N-I+P-S+ 0.51 0.472 0.438 0.426 0.445 0.437
D+N-I-P+S+ 0.4 0.437 0.438 0.387 0.445 0.437
D-N+I+P+S- 0.5 0.463 0.438 0.466 0.445 0.437
D-N+I+P-S+ 0.39 0.455 0.438 0.431 0.445 0.437
D-N-I+P+S+ 0.48 0.453 0.438 0.444 0.445 0.437
D-N+I-P+S+ 0.49 0.42 0.438 0.362 0.445 0.437
D+N+I+P-S+ 0.37 0.377 0.37 0.37 0.373 0.38
D+N+I+P+S- 0.29 0.384 0.37 0.396 0.373 0.38
D+N+I-P+S+ 0.39 0.345 0.37 0.265 0.373 0.38
D+N-I+P+S+ 0.35 0.375 0.37 0.337 0.373 0.38
D-N+I+P+S+ 0.34 0.359 0.37 0.347 0.373 0.38
D+N+I+P+S+ 0.3 0.289 0.337 0.285 0.314 0.339
118
Table B.6: Linear Predictor Formulas with Coefficients
Equation
Model 1 0.90 − 0.32 (D) − 0.39 (N) − 0.26 (I) − 0.40 (P) − 0.43 (S)
Model 2 0.59 − 0.44 (two dys) − 0.84 (three dys) − 1.12 (four dys) − 1.27 (five dys)
Model 3 1.18 + 0.85 (D) − 0.59 (N) − 0.64 (I) − 0.45 (P) − 0.62 (S) + 0.17 (D × N) +
0.16 (D × I) − 0.03 (D × P) + 0.21 (D × S) + 0.32 (N × I) − 0.01 (N × P) −
0.12 (N × S) + 0.05 (I × P) + 0.20 (I × S) + 0.05 (P × S)
√
Model 4 1.71 − 1.11 num dys
Model 5 0.64 − 0.81 ln (num dys)
119
Figure B.1: Variance function check.
120
Figure B.2: Link function check.
121
Figure B.3: Plots of nominal coefficients against the number of dysfunctions. No transfor-
mation on number of dysfunctions (a), square root (b), and logarithmic (c) transformations.
122
Figure B.4: Calibration Plot for Model 1.
123
124
125
126
Sensitivity analysis
Our analysis excluded participants deemed to have not have understood or not have en-
gaged in the utility valuation task. Mean and median utility valuations for each health
state including these participants are provided in Supplemental Table B.7. These valua-
tions are quite similar to the results obtained using the more restricted dataset (Table 3.2).
Therefore, we deemed our analysis not to be sensitive to the inclusion criteria defined a
priori.
127
Table B.7: Utilities for all MESCC Health States using Full Dataset
Health State Number of N Mean (95% CI) Median (IQR)

Dysfunctions
D+N-I-P-S- 1 300 0.597 (0.553, 0.641) 0.75 (0.15, 0.95)
D-N+I-P-S- 1 303 0.641 (0.597, 0.682) 0.8 (0.35, 0.95)
D-N-I+P-S- 1 290 0.596 (0.551, 0.638) 0.75 (0.163, 0.95)
D-N-I-P+S- 1 302 0.605 (0.558, 0.649) 0.75 (0.15, 0.95)
D-N-I-P-S+ 1 321 0.603 (0.56, 0.646) 0.75 (0.15, 0.95)
D+N+I-P-S- 2 142 0.539 (0.475, 0.601) 0.65 (0.15, 0.95)
D+N-I+P-S- 2 179 0.569 (0.511, 0.624) 0.65 (0.1, 0.95)
D+N-I-P+S- 2 179 0.478 (0.418, 0.534) 0.45 (0, 0.85)
D+N-I-P-S+ 2 179 0.545 (0.487, 0.601) 0.65 (0.1, 0.95)
D-N+I+P-S- 2 134 0.591 (0.523, 0.654) 0.675 (0.15, 0.95)
D-N+I-P+S- 2 169 0.522 (0.463, 0.579) 0.6 (0.1, 0.95)
D-N+I-P-S+ 2 169 0.495 (0.433, 0.555) 0.5 (0, 0.95)
D-N-I+P+S- 2 163 0.562 (0.502, 0.618) 0.65 (0.125, 0.95)
D-N-I+P-S+ 2 152 0.514 (0.451, 0.577) 0.575 (0, 0.95)
D-N-I-P+S+ 2 170 0.561 (0.5, 0.619) 0.65 (0.15, 0.95)
D+N+I+P-S- 3 137 0.473 (0.405, 0.54) 0.45 (0, 0.95)
D+N+I-P+S- 3 132 0.481 (0.413, 0.548) 0.45 (0, 0.95)
D+N+I-P-S+ 3 142 0.457 (0.391, 0.525) 0.45 (0, 0.938)
D+N-I+P+S- 3 151 0.494 (0.43, 0.556) 0.55 (0, 0.925)
D+N-I+P-S+ 3 153 0.521 (0.459, 0.584) 0.55 (0, 0.95)
D+N-I-P+S+ 3 158 0.487 (0.424, 0.55) 0.6 (0, 0.95)
D-N+I+P+S- 3 148 0.467 (0.404, 0.528) 0.45 (0, 0.85)
D-N+I+P-S+ 3 135 0.433 (0.367, 0.499) 0.45 (0, 0.825)
D-N+I-P+S+ 3 130 0.505 (0.433, 0.572) 0.5 (0, 0.95)
D-N-I+P+S+ 3 154 0.496 (0.432, 0.559) 0.5 (0, 0.95)
D+N+I+P+S- 4 325 0.432 (0.388, 0.477) 0.45 (0, 0.85)
D+N+I+P-S+ 4 308 0.451 (0.407, 0.495) 0.45 (0, 0.85)
D+N+I-P+S+ 4 311 0.415 (0.371, 0.46) 0.35 (0, 0.85)
D+N-I+P+S+ 4 280 0.409 (0.364, 0.456) 0.35 (0, 0.75)
D-N+I+P+S+ 4 270 0.413 (0.367, 0.46) 0.35 (0, 0.85)
D+N+I+P+S+ 5 289 0.402 (0.358, 0.449) 0.35 (0, 0.75)
95% Confidence Interval (95% CI). Interquartile range (IQR).
128
Appendix C
C.1 Confidence Intervals for Median Survival Time

Bayesian analysis of median survival time requires knowledge of the mean and standard
deviation of the median survival time (tm ) for each group of the REVEL study and for
each stratum in the Lee et al. meta-analysis.
For the Lee et al. meta-analysis this can be easily done because they report symmetric
asymptotic confidence intervals for tm . These confidence intervals are based on the assump-
tion that tm follows a normal distribution. Symmetric asymptotic confidence intervals for
tm are based on the relationship:[191]
var [S (tm )]
var [tm ] = . (C.1)
f (tm )2
In the original report, tm was not stratified by Tokuhashi score. We requested this
analysis from the study authors:
• a Tokuhashi Score <8 was associated with a median survival of 5 months (95% CI,
4.356 – 5.644 months); thus µ=5, and σ=(5.644-5)÷1.96=0.332.
• a Tokuhashi Score ≥ 8 was associated with a median survival of 14 months (95% CI,
12.666 – 15.334 months); thus µ=14, and σ=(15.334-14)÷1.96=0.681.
Many studies, such as the REVEL trial, report asymmetric confidence intervals for
tm . A commonly used technique for computing asymmetric confidence intervals is that
described by Brookmeyer and Crowley and was used to analyze the REVEL trial.[25]
Brookmeyer and Crowley confidence intervals contain the times t in an α-level confidence
interval if they satisfy the relationship:
cα var [S (t)] ≥ (S(t) − 0.5)2 .
129
Table C.1: Parameters used for Interpolation of var [S (tm )]
Lower Confidence Limit for tm Upper Confidence Limit for tm
DOC DOC+RAM DOC DOC+RAM
t 6.1 8.4 8.5 10.1

S (t) 0.54 0.55 0.44 0.45
†
var [S (t)] 0.00042 0.00065 0.00094 0.00065
†
calculated using Formula C.2.
Where S(t) is the survival fraction at time t; var [S (t)] is the Greenwood variance; and
cα is the critical value of the χ2 (1) distribution with a right tail area of α. Within the
Brookmeyer and Crowley interval,
(S(t) − 0.5)2
var [S (t)] ≥ .
cα
Thus outside the interval,
(S(t) − 0.5)2
var [S (t)] < .
cα
We therefore approximate the variance of the survival fraction at the confidence limits by,
(S(t) − 0.5)2
var [S (t)] ≈ . (C.2)
cα
Greenwood variances computed from the limits of the 95% confidence intervals for tm
reported in the REVEL study are shown in Table C.1. Datapoints for the DOC and
DOC+RAM arms were obtained from Figure S1a of the REVEL publication.[188] We
the use trilinear interpolation of t, S (t) and var [S (t)] between the confidence limits to
estimate var [S (tm )] which was 0.00074 and 0.00065 for the DOC and DOC+RAM groups
respectively.
We must still compute the probability density function f (tm ) in order to compute
var [tm ] based on Formula C.1. A variety of strategies have been proposed to compute
f (tm ).[254] We calculated f (tm ) using the upper and lower confidence limits for tm ,tlm and
tum , respectively
S tlm − S (tum )
f (tm ) =
tum − tlm
f (tm ) was -0.04 and -0.06 for the DOC and DOC+RAM groups respectively.
Substituting var [S (tm )] and f (tm ) into Formula C.1 yields a var [tm ] of 0.46 and 0.18
for the DOC and DOC+RAM groups respectively.
130
C.2 Sensitivity Analysis Through Effective Sample Size
Adjustment
It is helpful to consider the effective sample instead of the variance. Larger studies will
have a smaller variance and larger precision. The Cutler-Ederer effective sample size (n0 )
is a measure of the number of individuals who survived and contibuted to S(t), and is
related to the Greenwood variance (var [S (t)]) through:[53]
S (t) (1 − S (t))
n0 = (C.3)
var [S (t)]
Since S(tm ) = 0.5, the effective sample size at tm is

0.25
n0tm = .
var [S (tm )]
Therefore
0.25
var [S (tm )] = . (C.4)
n0tm
Substituting Formula C.4 into Formula C.1 yields an expression for the asymptotic variance
of median survival time
0.25
var [tm ] = 0 . (C.5)
ntm (f (tm ))2
For the REVEL trial, we can calculate a baseline effective sample size for each arm
because we calculated f (tm ) and var [tm ] (Appendix C.1). The baseline effective samples
sizes for the DOC and DOC+RAM arms are 337 and 384 respectively. By keeping f (tm )
and var [tm ] constant while changing n0tm , we can explore the impact of the effective sample
size on the posterior distribution.
131
Appendix D
D.1 Transition Parameters for an Illness-Death Model

The transition rates between states can be organized in a transition rate matrix R where
rows correspond to the “from” state, and the columns to the “to” state. States in the
illness-death model are healthy (h), ill (i), and dead (d). Disallowed transitions are given
a value of 0, and all rows must sum to 0. We model the MESCC disease process using
time-homogenous transition rates: that is transition rates do not change over time and
thus R is constant over time. Disallowed transitions are i to h, and d to either h or i. The
transition rate matrix is
 
− (λhi + λhd ) λhi λhd
 0 −λid λid  .[40]
0 0 0
The transition rate matrix can be converted to a transition probability matrix P. This
matrix records the probability of making transitions in the time interval from t = 0 to the
time t. Using the transition rate matrix R, the transition probability matrix is computed
as
P (t) = eRt
 
phh (t) phi (t) phd (t)
= 0 pid (t) pid (t) 
0 0 0

λan e id t −1+e−(λhi +λhd −λid )t
 −λ

−(λhi +λhd )t
 e − λhi +λhd −λid
1 − phh (t) − phi (t) 
= 0 e −λid t
1 − pii (t)  .[40]
0 0 0
The fraction of patients in each state at time t is recorded in the state membership
vector m (t). The initial state membership fractions are m (0) = [θh , θi , θd ]. m (t) is
computed as the product of m (0) and the transition probability matrix P (t),
m (t) = P (0) · P (t) .
132
The first element of m (t) is H (t), the fraction of healthy patients
H (t) = θh e−(λhi +λhd )t .
The second element of m (t) is I (t), the fraction of ill patients
θh λhi e−λid t −1 + e−(λhi +λhd −λid )t

−λid t
I (t) = θi e − .
λhi + λhd − λid
The third element of m (t) is D (t), the fraction of dead patients
θh λhi e−λid t −1 + e−(λhi +λhd −λid )t

−(λhi +λhd )t −λid t
D (t) = 1 − θh e − θi e + .
λhi + λhd − λid
133
D.2 Multistate Estimation of Transition Rates
Methods
Likelihood
From Patchell et al.’s trial data, patients’ health state trajectories can be categorized into
seven different sequences (Table E.2). For some patients, the state was unknown at some
time intervals. The time spent in the h, i and unknown state are denoted th , ti , and tu
respectively. The transition rate and transition probability matrices are used to calculate
the likelihood contribution for each trajectory.[122, 119, 66]
Table D.1: Health state trajectories and corresponding likelihood contributions
Sequence of states Likelihood contribution
h→i→d λhi phh (th ) × λid pii (ti )

h→d λhd phh (th )
i→d λid pii (ti )
h → censored phh (th )
i → censored pii (ti )
h → unknown → d phh (th ) × (λhd phh (tu ) + λid pii (tu ) + λid phi (tu ))
unknown → d λhd phh (tu ) + λid pii (tu ) + λid phi (tu )
Prior Distributions
Non-informative prior distributions were used for multistate model estimation
log λ ∼ Normal (mean = 0, variance = 10) ,

β ∼ Normal (mean = 0, variance = 10) .[178]
An upper truncation limit of 10 was applied to all prior distributions.[178]
Implementation
The model was estimated using the Bayesian modelling language Stan[212] run through
the statistical programming language R.[183] Four Markov chains were implemented with
different random initial values for each chain.
Initially all chains were run for 5000 iterations without thinning; 2500 burn-in iterations
were discarded. Post-burn-in traceplots for each parameter were examined to ensure that
134
all four chains reached a similar mean and sampled similar regions of the distribution.[80]
Iterations were increased until (1) the potential scale reduction factor, R̂, was < 1.1; and
(2) histograms of the posterior distribution were smooth and without gaps.[39, 26, 48]
To mitigate autocorrelation, thinning was increased so that effective sample size, Nef f ,was
similar for all parameters, and the effective sample size as ≥ 20% of the number of sampling
iterations.[80] Post-burn in iterations were increased until Monte Carlo (MC) error for each
parameter was < 5% of the sample standard deviation.[210] The final simulation conditions
were termed the short conditions.
To ensure that global convergence was achieved (rather than local convergence), models
was re-run with iterations doubled.[48] Simulation conditions were adjusted to ensure con-
vergence and mitigate autocorrelation. These simulation conditions were termed the long
conditions. Percent relative deviation for the estimates from the long conditions to the
short conditions was computed. Global convergence was indicated if the percent relative
deviation was less than ±2% for each parameter.[48]
135
STAN Code
functions{
real MarkovKnownlpdf(real t, real [] lam, real censind, int state, int trans){
//log-likelihood function when state known
real lamstay [3]; real LL;
lamstay[1]=-(lam[1]+lam[2]);
lamstay[2]=-lam[3];
LL = (lamstay[state])*t + log(lam[trans])*censind;
return LL;
}
real MarkovUnknownlpdf(real t, real [] lam){
//log-likelihood function when state unknown
real LL; real LikeA; real LikeB; real LikeC;
LikeA=exp((-lam[1] - lam[2])*t)*lam[2];
LikeB=(lam[1]/(lam[1]+lam[2]-lam[3]))*(exp(-lam[3]*t)-exp((-lam[1] - lam[2])*t))*lam[3];
LikeC=exp(-lam[3]*t)*lam[3];
LL = log(LikeA + LikeB +LikeC);
return LL;
}
}
data {
int <lower=0> Nknown; int <lower=0> Nunknown; vector [Nknown] tknown; vector [Nunknown] tunknown;
int <lower=0> stateknown [Nknown]; int <lower=0> transknown [Nknown]; vector [Nknown] censindknown;
vector [Nknown] grpknown; vector [Nunknown] grpunknown;
}
parameters {
real llam [3]; real bet [3];
}
transformed parameters {
real lam [3]; real slam [3];
for (k in 1:3){
//transition rates for RT-alone group
lam[k]=exp(llam[k]);
//transition rates for mS+RT group
slam[k]=lam[k]*exp(bet[k]);
}
}
model {
real tempknown [3]; real tempunknown [3];
//likelihood for transitions where from and to state known
for (i in 1:Nknown){
tempknown[1]=lam[1]*exp(bet[1]*grpknown[i]);
target += MarkovKnownlpdf(tknown[i], tempknown, censindknown[i], stateknown[i], transknown[i]);
}
//likelihood for transitions where either from or to state unknown
for (j in 1:Nunknown){
tempunknown[1]=lam[1]*exp(bet[1]*grpunknown[j]);
target += MarkovUnknownlpdf(tunknown[j], tempunknown);
}
//Prior Distributions
llam ~ normal(0, 10)T[,10];
bet ~ normal(0, 10)T[,10];
}
136
Results
To meet pre-specified benchmarks for autocorrelation, thinning was increased to 10 to
ensure that effective sample size as ≥ 20% of the number of sampling iterations (Table D.2).
Traceplots for λ, and β showed a stable mean and variance under both short (Fig-
ure D.1)and long (Figure D.2) conditions. R̂ was < 1.1 for all parameters under all con-
ditions; and percent relative deviation was < ±2% for all parameters (Table D.2). We
conclude that global convergence was reached.
Histograms of the posterior distribution under both short (Figure D.3) and long (Fig-
ure D.4) simulation conditions show smooth histograms without empty bins. Under both
short and long simulation conditions, Monte Carlo error was < 5% of the sample stan-
dard deviation for each parameter (Table D.2). We conclude that both the short and long
simulation conditions were sufficient to characterize the shape of the posterior distribution.
137
Figure D.1: Post-burn-in traceplots for transition rates and hazard ratios under short
simulation conditions.
138
Figure D.2: Post-burn-in traceplots for transition rates and hazard ratios under long
139
Figure D.3: Histograms of posterior distributions for transition rates and hazard ratios
under short simulation conditions.
140
Figure D.4: Histograms of posterior distributions for transition rates and hazard ratios
141
D.3 Worked Example
In this appendix we perform transition rate approximation for the mS+RT arm in the
MESCC RCT using the formulas developed in Chapter 5.
Table 5.1 is used to populate Formula 5.1 with AU Cpf s = RMPFST−τpf s ,
1 − e−(λhi +λhd )τpf s

RMPFST−τpf s =
λhi + λhd
1 − e−(λhi +λhd )5.25
1.159 = .
λhi + λhd
This equation can be solved using uniroot function in the statistical programming
language R with lambdaS as the variable for (λhi + λhd )
lambdaS_solve<-function(lambdaS, tau_pfs, RMPFSTtau_pfs)(1-exp(-lambdaS*tau_pfs))/lambdaS-RMPFSTtau_pfs
uniroot(lambdaS_solve,c(0.00001,10000),5.25,1.16)
Using this code we obtain λhi + λhd = 0.85.
Table 5.1 is also used to populate Formula 5.2,
Nsimul
ρ≈ e
Npf s
19
≈
31
= 0.61.
We can now compute
λhi = ρ (λhi + λhd )

= 0.61 × 0.85
≈ 0.52,
and
λhd = (λhi + λhd ) − λhi

≈ 0.85 − 0.52
≈ 0.33.
We populate Formula 5.5 with values from Table 5.1 to yield

e
Nos e c

−τos e c
X Nos + Nos
Eos ≈ RMOST (Nos + Nos ) + oei 1− e
Nos
i=1
45 + 3
≈ 0.98 + 35.89 1 −
45
≈ 45.27.
142
Table D.2: Comparison of Transition Rates and Hazard Ratios Estimated with Multistate
Techniques under Short and Long Simulation Conditions
Short Conditions Long Conditions % Rel

Mean SDev MCSE R̂ Nef f Mean SDev MCSE R̂ Nef f Dev
λhi 0.66 0.25 0.00 10094 1.00 0.66 0.24 0.00 19751 1.00 0.80
λhd 0.99 0.30 0.00 9956 1.00 1.00 0.30 0.00 19084 1.00 -0.50
143
λid 2.12 0.40 0.00 10039 1.00 2.12 0.40 0.00 19627 1.00 0.00
βhi -0.66 0.49 0.00 10303 1.00 -0.66 0.48 0.00 19906 1.00 0.30
βhd -0.60 0.39 0.00 9665 1.00 -0.61 0.38 0.00 19387 1.00 -1.00
βid 0.44 0.28 0.00 9805 1.00 0.44 0.28 0.00 19750 1.00 0.00
SDev, standard deviation. MCSE, Monte Carlo standard error. % Rel. Dev., percent relative
deviation
We adapt Formula 5.5 for the PFS analysis and substitute relevant values from Table 5.1
to yield
e
Npf !
s e c
X Npf s + Npf s
Epf s ≈ RMPFT−τpf s Npf
e c
pei

s + Npf s + 1− e
i=1
Npf s

31 + 6
≈ 1.16 + 27.22 1 −
31
≈ 37.65.
We can then use Eos and Epf s in Formula 5.6
Eill ≈ Eos − Epf s

≈ 45.27 − 35.94
≈ 7.00.
We then substitute values from Table 5.1 in Formula 5.7

e
Nid ≈ Nos − Nsimul
≈ 45 − 19
≈ 26
Nid and Eill are used in Formula 5.8 to obtain

Nid
λid ≈
Eill
26
≈
7.00
≈ 3.71.
144
Appendix E
E.1 Bayesian Estimation

This section describes the methods for and results of Bayesian MCMC estimation for
• treatment-phase event probabilities

• a mixed effects beta regression model for utilities
Details on the likelihood function and prior distribution for each application of Bayesian
MCMC estimation are provided in the following subsections. However, a common analytic
approach was used for each application.
Each model was estimated using the Bayesian modelling language Stan[212] run through
Initially all chains were run for 10 000 iterations without thinning; 5000 burn-in it-
erations were discarded. Post-burn-in traceplots for each parameter were examined to
ensure that all four chains reached a similar mean and sampled similar regions of the
distribution.[80] Iterations were increased until (1) the potential scale reduction factor,
R̂, was < 1.1; and (2) histograms of the posterior distribution were smooth and without
gaps.[39, 26, 48] To mitigate autocorrelation, thinning was increased so that effective sam-
ple size, Nef f ,was similar for all parameters, and the effective sample size as ≥ 0.1% of the
number of sampling iterations.[80] The final simulation conditions were termed the short
conditions.
deviation was less than ±1% for each parameter.[48] Results for long conditions were used
to run the decision model (Section 6.3).
145
E.1.1 Bayesian MCMC Estimation of Event Probabilities for the
Treatment-Phase
Likelihood
In Patchell et al.’s RCT, patients were evaluated at t = 0 and immediately after treat-
ment. At t = 0 patients could be ambulatory (a) or non-ambulatory (n). After treatment
patients could be a, n or dead (d). If we stratify by treatment group and health state at
t = 0, the health state immediately after treatment can be modeled using a multinomial
distribution.[81]
The multinomial distribution is a generalization of the binomial distribution. The multi-
nomial distribution models the probability of observing a vector of counts after repeated
k
P
trials. This distribution is parametrized by θ1 . . . θk event probabilities with θi = 1.[170]
i=1
Concretely for the Patchell et al.’s RCT, the vector of counts is the number of ambulatory,
non-ambulatory and dead patients, [Na , Nn , Nd ], for a given treatment arm stratified by
baseline ambulatory status at is distributed as:
[Na , Nn , Nd ] ∼ Multinomial (θa , θn , θd ))
Prior Distributions
We utilized non-informative prior distributions to form a hierarchical multinomial-Dirichlet

model.[8] Event probabilities were organized in the vector Θ = [θa , θn , θd ]. θ was stratified
for treatment arm and baseline ambulatory status. Stratum was listed in the subscript
on θx,y with x denoting the treatment arm, and y denoting baseline ambulatory status.
Consequently four prior distributions were specified:
Θrt,a ∼ Dirichlet (τa Art )

Θrt,n ∼ Dirichlet (τn Art )
Θs+rt,a ∼ Dirichlet (τa As+rt )
Θs+rt,n ∼ Dirichlet (τn As+rt ) .
Note that treatment arms share a common concentration parameter vector A (subscripted
for treatment arm). In this way, Θ vectors for the same arm are dependent. A non-
informative hyper-prior distribution was placed on A
A ∼ Dirichlet (1, 1, 1) .
The informativeness of each prior distribution was controlled by an equivalent sample size
parameter τ . To make Θ vectors for baseline ambulatory status strata dependent, τ was
common for baseline ambulatory status. A non-informative hyper-prior distribution was
placed on τ
τ ∼ Uniform (0, 100) .
146
Results

ensure that effective sample size as ≥ 20% of the number of sampling iterations (Table E.1).
Traceplots for Θrt,a , Θrt,n , Θs+rt,a , and Θs+rt,n showed a stable mean and variance
under both short (Figure E.1)and long (Figure E.2) conditions. R̂ was < 1.1 for all param-
eters under all conditions; and percent relative deviation was < ±2% for all parameters
(Table E.1). We conclude that global convergence was reached.
Histograms of the posterior distribution under both short (Figure E.3) and long (Fig-
ure E.4) simulation conditions show smooth histograms without empty bins. Under both
dard deviation for each parameter (Table E.1). We conclude that both the short and long
147
Table E.1: Comparison of results from short and long simulation conditions for multistate estimated
transition rates and hazard ratios

θa,rt,a 0.54 0.08 0.00 9703 1.00 0.54 0.07 0.00 20000 1.00 -0.40
θn,rt,a 0.32 0.07 0.00 9695 1.00 0.32 0.07 0.00 19790 1.00 0.00
θd,rt,a 0.14 0.05 0.00 10000 1.00 0.14 0.05 0.00 20000 1.00 0.00
θa,rt,n 0.41 0.10 0.00 9723 1.00 0.42 0.10 0.00 20000 1.00 -0.50
θn,rt,n 0.41 0.09 0.00 9961 1.00 0.41 0.09 0.00 20000 1.00 0.20
148
θd,rt,n 0.17 0.07 0.00 9964 1.00 0.17 0.07 0.00 20000 1.00 0.60
θa,s+rt,a 0.77 0.07 0.00 9703 1.00 0.77 0.07 0.00 20000 1.00 0.00
θn,s+rt,a 0.18 0.06 0.00 9695 1.00 0.18 0.06 0.00 20000 1.00 -0.60
θd,s+rt,a 0.05 0.03 0.00 10000 1.00 0.05 0.03 0.00 19538 1.00 0.00
θa,s+rt,n 0.63 0.09 0.00 9723 1.00 0.63 0.09 0.00 19823 1.00 0.20
θn,s+rt,n 0.30 0.08 0.00 9961 1.00 0.30 0.09 0.00 20000 1.00 -0.70
θd,s+rt,n 0.07 0.05 0.00 9964 1.00 0.07 0.05 0.00 18866 1.00 0.00
deviation
Figure E.1: Post-burn-in traceplots for event probabilities under short simulation condi-
tions.
149
Figure E.2: Post-burn-in traceplots for event probabilities under long simulation condi-
tions.
150
Figure E.3: Histograms of posterior distributions for event probabilities under short
151
Figure E.4: Histograms of posterior distributions for event probabilities under long sim-
ulation conditions.
152
E.1.2 Bayesian MCMC Estimation of Transition Rates
Likelihood
From Patchell et al.’s trial data, patients’ health state trajectories can be categorized into
seven different sequences (Table E.2). For some patients, the state was unknown at some
time intervals. The time spent in the a, n and unknown state are denoted ta , tn , and tu
respectively. The transition rate and transition probability matrices are used to calculate
the likelihood contribution for each trajectory.[122, 119, 66]
Table E.2: Health state trajectories and corresponding likelihood contributions
Sequence of states Likelihood contribution
a→n→d λan paa (ta ) × λnd pnn (tn )

a→d λad paa (ta )
n→d λnd pnn (tn )
a → censored paa (ta )
n → censored pnn (tn )
a → unknown → d paa (ta ) × (λad paa (tu ) + λnd pnn (tu ) + λnd pan (tu ))
unknown → d λad paa (tu ) + λnd pnn (tu ) + λnd pan (tu )
Prior Distributions
Non-informative prior distributions were used for multistate model estimation
log λ ∼ Normal (mean = 0, variance = 10) ,

β ∼ Normal (mean = 0, variance = 10) .[178]
An upper truncation limit of 10 was applied to all prior distributions.[178]
Implementation
The model was estimated using the Bayesian modelling language Stan[212] run through
Initially all chains were run for 5000 iterations without thinning; 2500 burn-in iterations
were discarded. Post-burn-in traceplots for each parameter were examined to ensure that
all four chains reached a similar mean and sampled similar regions of the distribution.[80]
Iterations were increased until (1) the potential scale reduction factor, R̂, was < 1.1; and
(2) histograms of the posterior distribution were smooth and without gaps.[39, 26, 48]
153
To mitigate autocorrelation, thinning was increased so that effective sample size, Nef f ,was
similar for all parameters, and the effective sample size as ≥ 20% of the number of sampling
iterations.[80] Post-burn in iterations were increased until Monte Carlo (MC) error for each
parameter was < 5% of the sample standard deviation.[210] The final simulation conditions
were termed the short conditions.
deviation was less than ±2% for each parameter.[48]
154
STAN Code
functions{
real MarkovKnownlpdf(real t, real [] lam, real censind, int state, int trans){
//log-likelihood function when state known
real lamstay [3]; real LL;
lamstay[1]=-(lam[1]+lam[2]);
lamstay[2]=-lam[3];
LL = (lamstay[state])*t + log(lam[trans])*censind;
return LL;
}
real MarkovUnknownlpdf(real t, real [] lam){
//log-likelihood function when state unknown
real LL; real LikeA; real LikeB; real LikeC;
LikeA=exp((-lam[1] - lam[2])*t)*lam[2];
LikeB=(lam[1]/(lam[1]+lam[2]-lam[3]))*(exp(-lam[3]*t)-exp((-lam[1] - lam[2])*t))*lam[3];
LikeC=exp(-lam[3]*t)*lam[3];
LL = log(LikeA + LikeB +LikeC);
return LL;
}
}
data {
int <lower=0> Nknown; int <lower=0> Nunknown; vector [Nknown] tknown; vector [Nunknown] tunknown;
int <lower=0> stateknown [Nknown]; int <lower=0> transknown [Nknown]; vector [Nknown] censindknown;
vector [Nknown] grpknown; vector [Nunknown] grpunknown;
}
parameters {
real llam [3]; real bet [3];
}
transformed parameters {
real lam [3]; real slam [3];
for (k in 1:3){
//transition rates for RT-alone group
lam[k]=exp(llam[k]);
//transition rates for mS+RT group
slam[k]=lam[k]*exp(bet[k]);
}
}
model {
real tempknown [3]; real tempunknown [3];
//likelihood for transitions where from and to state known
for (i in 1:Nknown){
target += MarkovKnownlpdf(tknown[i], tempknown, censindknown[i], stateknown[i], transknown[i]);
}
//likelihood for transitions where either from or to state unknown
for (j in 1:Nunknown){
target += MarkovUnknownlpdf(tunknown[j], tempunknown);
}
//Prior Distributions
llam ~ normal(0, 10)T[,10];
bet ~ normal(0, 10)T[,10];
}
155
Results

ensure that effective sample size as ≥ 20% of the number of sampling iterations (Table E.3).
Traceplots for λ, and β showed a stable mean and variance under both short (Fig-
ure E.5)and long (Figure E.6) conditions. R̂ was < 1.1 for all parameters under all con-
ditions; and percent relative deviation was < ±2% for all parameters (Table E.3). We
156
Figure E.5: Post-burn-in traceplots for transition rates and hazard ratios under short
157
Figure E.6: Post-burn-in traceplots for transition rates and hazard ratios under long
158
Figure E.7: Histograms of posterior distributions for transition rates and hazard ratios
159
Figure E.8: Histograms of posterior distributions for transition rates and hazard ratios
160
E.1.3 Bayesian MCMC Estimation of a Mixed Effects Beta Re-
gression Model for Utilities
Model
In previous work, a beta regression model with the square root of the number of dys-
functional attributes in a health state was found to be well calibrated for estimating mean
utility.[168] For the purposes of decision analysis, we re-fit this model using Bayesian meth-
ods.
Beta regression was used because utilities are restricted to the interval [0, 1]. The beta
distribution is typically parametrized using two shape parameters: p ∈ R+ and q ∈ R+ .[115]
For regression estimation, it is more useful to parametrize with the mean (µ) and precision
(φ).[67] This alternative parametrization of the beta distribution is related to the typical
parametrization by p = µφ and q = (1 − µ) φ.[67]
Mixed beta regression model for the mean, µ, was implemented as described by Figueroa-
Zúñiga and colleagues.[69] It was assumed that participant i’s j-th utility valuation, Yij ,
is drawn from a beta distribution
Yij ∼ Beta (µij φ, (1 − µij ) φ) .
The inverse logit function was used to map linear predictors to the [0, 1] interval.[149] Based
on rigorous model selection, the covariate for participant i’s j-th utility valuation was the
square root of the number of dysfunctions. β0 and β1 are fixed effect regression coefficients.
To account for six responses per participant a random intercept term i was added for each
participant i.[78] The precision parameter was common to all utility valuations.[67]
The expectation of participant i’s j-th utility valuation is the mean of then beta dis-
tribution (µij ) served as the dependent variable of the regression model:
E [Yij ] = µij = inverse logit (β0 + xij β1 + i ) .
Prior Distributions
Non-informative prior distributions were used
β0 ∼ Normal (mean = 0, standard deviation = 10) ,

β1 ∼ Normal (mean = 0, standard deviation = 10) ,
i ∼ Normal (mean = 0, standard deviation = 1) ,
φ ∼ Gamma (0.1, 0.01) .[127, 82, 77]
Results
To meet pre-specified benchmarks for autocorrelation, thinning was increased to 3 to ensure

that effective sample size as ≥ 20% of the number of sampling iterations (Table E.4).
161
Table E.3: Comparison of Transition Rates and Hazard Ratios Estimated with Multistate
Techniques under Short and Long Simulation Conditions

λan 0.66 0.25 0.00 10094 1.00 0.66 0.24 0.00 19751 1.00 0.80
λad 0.99 0.30 0.00 9956 1.00 1.00 0.30 0.00 19084 1.00 -0.50
162
λnd 2.12 0.40 0.00 10039 1.00 2.12 0.40 0.00 19627 1.00 0.00
βan -0.66 0.49 0.00 10303 1.00 -0.66 0.48 0.00 19906 1.00 0.30
βad -0.60 0.39 0.00 9665 1.00 -0.61 0.38 0.00 19387 1.00 -1.00
βnd 0.44 0.28 0.00 9805 1.00 0.44 0.28 0.00 19750 1.00 0.00
deviation
Traceplots for β0 , β1 and φ showed a stable mean and variance under both short (Fig-
ure E.10)and long (Figure E.10) conditions. R̂ was < 1.1 for all parameters under all
conditions; and percent relative deviation was < ±2% for all parameters (Table E.4). We
163
Table E.4: Comparison of results from short and long simulation conditions for beta regression
coefficients for utilities

164
β0 1.73 0.10 0.00 9655 1.00 1.72 0.10 0.00 19892 1.00 0.10
β1 -1.15 0.06 0.00 10378 1.00 -1.15 0.05 0.00 20695 1.00 0.00
φ 0.96 0.02 0.00 10081 1.00 0.96 0.02 0.00 21283 1.00 0.10
SDev, standard deviation. MCSE, Monte Carlo standard error. % Rel. Dev., percent
relative deviation
Figure E.9: Post-burn-in traceplots for regression coefficients under short simulation
conditions.
Figure E.10: Post-burn-in traceplots for regression coefficients under long simulation
conditions.
Figure E.11: Histograms of posterior distributions for regression coefficients under short
165
Figure E.12: Histograms of posterior distributions for regression coefficients under short
166
E.2 Exact Formulas for Markov Cohort Simulation
Markov cohort simulation is used to determine the incremental quality-of-life-years (∆QALYs)
gained/lost by patients undergoing treatment y versus treatment x. Markov cohort simu-
lation involves three steps for each treatment arm:[22]
1. simulate state membership fractions over a time-horizon of interest
2. weigh state membership fractions by their utility
3. sum the weighed state membership fractions over the time-horizon of interest
The formulas presented in the remainder of this section apply to either the RT-alone or
mS+RT treatment arm. θ will be used to denote event probabilities for any treatment arm
and baseline ambulatory status stratum (Appendix E.1.1). θ should be replaced by the
appropriate strata-specific parameter in the formulas. λ will be used to denote transition
rates. For the mS+RT arm, λ should be replaced by λβ.
The simulation model (Figure 6.1) consists of three health states: ambulatory (a), non-
ambulatory (n), and dead (d). The fraction of simulated patients in each health state at
time t is recorded in the state membership vector m (t). The state membership fractions
at the start of simulation are the event probabilities for the appropriate treatment arm and
baseline ambulatory status stratum (Section E.1.1); m (0) = [θa , θn , θd ]. m (t) is computed
as the product of m (0) and the transition probability matrix P (t).[40]
The first element of m (t) is A (t), the fraction of ambulatory patients
A (t) = θa e−(λan +λad )t .
The second element of m (t) is N (t), the fraction of non-ambulatory patients

−λnd t −(λan +λad −λnd )t

θa λ an e −1 + e
N (t) = θn e−λnd t − .
λan + λad − λnd
If the utility of the ambulatory state is Wa ; the utility of the non-ambulatory state is
Wn ; and the utility of the dead state is 0, the total QALYs gained/lost by a simulated
cohort can be expressed as a sum of integrals,
Z ∞ Z ∞
QALYs = Wa A (t) dt + Wn N (t) dt
0 0
θa (Wa λnd + Wi λan ) + θn Wn (λan + λad )
= .
(λan + λad ) λnd
167
References
[1] A. Allignol, M. Schumacher, and J. Beyersmann. Empirical Transition Matrix of

Multi-State Models: The etm Package. J. Stat. Softw., 38(4), 2011.
[2] D. G. Altman and J. M. Bland. Diagnostic tests 2: Predictive values. BMJ,

309(6947):102, jul 1994.
[3] A. Amelot, L. Balabaud, D. Choi, Z. Fox, H. A. Crockard, T. Albert, C. M. Arts,

J. M. Buchowski, C. Bunger, C. K. Chung, M. H. Coppes, B. Depreitere, M. G.
Fehlings, J. Harrop, N. Kawahara, E. S. Kim, C. S. Lee, Y. Leung, Z. J. Liu, J. A.
Martin-Benlloch, E. M. Massicotte, B. Meyer, F. C. Oner, W. Peul, N. Quraishi,
Y. Tokuhashi, K. Tomita, C. Ulbricht, J. J. Verlaan, M. Wang, and C. Mazel. Surgery
for metastatic spine tumors in the elderly. Advanced age is not a contraindication to
surgery! Spine J., 17(6):759–767, jul 2017.
[4] P. K. Andersen, S. Z. Abildstrom, and S. Rosthøj. Competing risks as a multi-state

model. Stat. Methods Med. Res., 11(2):203–15, apr 2002.
[5] K. J. Arrow and R. C. Lind. Uncertainty and the Evaluation Decisions of Public
Investment. Am. Econ. Rev., 60(3):364–378, 1970.
[6] L. A. Augestad, K. Stavem, I. S. Kristiansen, C. H. Samuelsen, and K. Rand-

Hendriksen. Influenced from the start: anchoring bias in time trade-off valuations.
Qual. Life Res., 25(9):2179–2191, 2016.
[7] F. Augustovski, L. Rey-Ares, V. Irazola, M. Oppe, and N. J. Devlin. Lead versus

lag-time trade-off variants: does it make any difference?. Eur. J. Health Econ., 14
Suppl 1:S25–31, 2013.
[8] L. Azzimonti, G. Corani, and M. Zaffalon. Hierarchical Multinomial-Dirichlet model

for the estimation of conditional probability tables, aug 2017.
[9] A. Bagust and S. Beale. Survival analysis and extrapolation modeling of time-to-
event clinical trial data for economic evaluation: an alternative approach. Med. Decis.
Making, 34(3):343–51, 2014.
[10] M. V. Bala, L. L. Wood, G. A. Zarkin, E. C. Norton, A. Gafni, and B. J. O’Brien.

Are health states ”timeless”? The case of the standard gamble method. J. Clin.
Epidemiol., 52(11):1047–53, nov 1999.
168
[11] K. P. Balanda and H. L. Macgillivray. Kurtosis : A Critical Review. Am. Stat.,
42(2):111–119, 1988.
[12] N. Bansback, A. Tsuchiya, J. Brazier, and A. Anis. Canadian Valuation of EQ-
5D Health States: Preliminary Value Set and Considerations for Future Valuation
Studies. PLoS One, 7(2):e31115, feb 2012.
[13] C. M. Bell, D. R. Urbach, J. G. Ray, A. Bayoumi, A. B. Rosen, D. Greenberg, and
P. J. Neumann. Bias in published cost effectiveness studies: systematic review. BMJ,
332(7543):699–703, mar 2006.
[14] M. H. Bilsky, I. Laufer, and S. Burch. Shifting Paradigms in the Treatment of
Metastatic Spine Disease. Spine (Phila. Pa. 1976)., 34(Supplement):S101–S107, oct
2009.
[15] N. F. Boyd, H. J. Sutherland, K. Z. Heasman, D. L. Tritchler, and B. J. Cummings.
Whose utilities for decision analysis? Med. Decis. Making, 10(1):58–67, 1990.
[16] C. A. Brauer, P. J. Neumann, and A. B. Rosen. Trends in Cost Effectiveness Analyses
in Orthopedic Surgery. Clin. Orthop. Relat. Res., PAP, jan 2007.
[17] J. Brazier and P. Dolan. Evidence of preference construction in a comparison of
variants of the standard gamble method. 2005.
[18] R. P. Brent. Algorithms for minimization without derivatives. Pretince-Hall, Engle-
wood Cliffs, N.J., 1973.
[19] K. H. Bridwell, W. Cats-Baril, J. Harrast, S. Berven, S. Glassman, J.-P. Farcy,
W. C. Horton, L. G. Lenke, C. Baldus, and T. Radake. The validity of the SRS-22
instrument in an adult spinal deformity population compared with the Oswestry and
SF-12: a study of response distribution, concurrent validity, internal consistency, and
reliability. Spine (Phila. Pa. 1976)., 30(4):455–61, feb 2005.
[20] A. Briggs. Economic evaluation and clinical trials: size matters. BMJ,
321(7273):1362–3, 2000.
[21] A. Briggs. Probabilistic analysis of cost-effectiveness models: statistical representa-
tion of parameter uncertainty.[comment]. Value Heal., 8(1):1–2, 2005.
[22] A. Briggs, K. Claxton, and M. Sculpher. Decision Modelling for Health Economic
Evaluation (Handbooks in Health Economic Evaluation). Oxford University Press,
USA, 2006.
[23] A. H. Briggs. Handling uncertainty in cost-effectiveness models. Pharmacoeconomics,
17(5):479–500, may 2000.
[24] D. A. Broniatowski, A. M. Jamison, S. Qi, L. AlKulaib, T. Chen, A. Benton, S. C.
Quinn, and M. Dredze. Weaponized Health Communication: Twitter Bots and
Russian Trolls Amplify the Vaccine Debate. Am. J. Public Health, 108(10):1378–
1384, oct 2018.
169
[25] R. Brookmeyer and J. Crowley. A Confidence Interval for the Median Survival Time.
Biometrics, 38(1):29, mar 1982.
[26] S. P. Brooks and A. Gelman. General methods for monitoring convergence of iterative
simulations. J. Comput. Graph. Stat., 7(4):434–455, dec 2013.
[27] A. S. Burns, D. A. Rivas, and J. F. Ditunno. The management of neurogenic bladder

and sexual dysfunction after spinal cord injury. Spine (Phila. Pa. 1976)., 26(24
Suppl):S129–36, dec 2001.
[28] P. R. Burton. Helping doctors to draw appropriate inferences from the analysis of
medical studies. Stat. Med., 13(17):1699–1713, 1994.
[29] M. J. Buxton, M. F. Drummond, B. A. Van Hout, R. L. Prince, T. A. Sheldon,

T. Szucs, and M. Vray. Modelling in economic evaluation: An unavoidable fact of
life. Health Econ., 6(3):217–227, 1997.
[30] B. P. Carlin and T. A. Louis. Bayesian methods for data analysis. CRC Press, Boca
Raton, 3rd edition, 2009.
[31] L. Y. Carreon, P. a. Anderson, C. M. McDonough, M. Djurasovic, and S. D. Glass-

man. Predicting SF-6D utility scores from the neck disability index and numeric
rating scales for neck and arm pain. Spine (Phila. Pa. 1976)., 36(6):490–494, 2011.
[32] L. Y. Carreon, S. D. Glassman, C. M. McDonough, R. Rampersaud, S. Berven, and

M. Shainline. Predicting SF-6D utility scores from the Oswestry disability index and
numeric rating scales for back and leg pain. Spine (Phila. Pa. 1976)., 34(19):2085–9,
sep 2009.
[33] G. Cervellin, L. Borghi, and G. Lippi. Do clinicians decide relying primarily on

Bayesians principles or on Gestalt perception? Some pearls and pitfalls of Gestalt
perception in medicine. Intern. Emerg. Med., 9(5):513–519, 2014.
[34] Y. Chen, M. J. Devivo, and A. B. Jackson. Pressure ulcer prevalence in people

with spinal cord injury: age-period-duration effects. Arch. Phys. Med. Rehabil.,
86(6):1208–13, jun 2005.
[35] K. Claxton. The irrelevance of inference: a decision-making approach to the stochas-

tic evaluation of health care technologies. J. Health Econ., 18(3):341–364, jun 1999.
[36] K. Claxton, M. Sculpher, C. McCabe, A. Briggs, R. Akehurst, M. Buxton, J. Brazier,

and T. O’Hagan. Probabilistic sensitivity analysis for NICE technology assessment:
not an optional extra. Health Econ., 14(4):339–347, apr 2005.
[37] I. Cleemput, M. Neyt, N. Thiry, C. De Laet, and M. Leys. Using threshold values
for cost per quality-adjusted life-year gained in healthcare decisions. Int. J. Technol.
Assess. Health Care, 27(1):71–6, jan 2011.
170
[38] M. Connock, C. Hyde, and D. Moore. Cautions regarding the fitting and interpre-
tation of survival curves: examples from NICE single technology appraisals of drugs
for cancer. Pharmacoeconomics, 29(10):827–37, oct 2011.
[39] M. K. Cowles and B. P. Carlin. Markov Chain Monte Carlo Convergence Diagnostics
: A Comparative Review. J. Am. Stat. Assoc., 91(434):883–904, 1996.
[40] D. R. Cox and H. D. Miller. The theory of stochastic processes. Wiley, New York,
1965.
[41] G. Cumming, J. Williams, and F. Fidler. Replication and Researchers’ Understanding

of Confidence Intervals and Standard Error Bars. Underst. Stat., 3(4):299–311, 2004.
[42] G. R. Cybulski. Methods of surgical stabilization for metastatic disease of the spine.
Neurosurgery, 25(2):240–52, aug 1989.
[43] S. Datta. Estimating the mean life time using right censored data. Stat. Methodol.,
2(1):65–69, mar 2005.
[44] D. De Looze, M. Van Laere, M. De Muynck, R. Beke, and A. Elewaut. Constipation

and other chronic gastrointestinal problems in spinal cord injury patients. Spinal
Cord, 36(1):63–6, jan 1998.
[45] H. C. de Vet, L. M. Bouter, P. D. Bezemer, and a. J. Beurskens. Reproducibility and

responsiveness of evaluative outcome measures. Theoretical considerations illustrated
by an empirical example. Int. J. Technol. Assess. Health Care, 17(4):479–487, 2001.
[46] G. A. De Wit, J. J. V. Busschbach, and F. T. De Charro. Sensitivity and perspective

in the valuation of health status: Whose values count? Health Econ., 9(2):109–126,
2000.
[47] N. Dea, A. Versteeg, C. Fisher, A. Kelly, D. Hartig, M. Boyd, S. Paquette, B. K.

Kwon, M. Dvorak, and J. Street. Adverse events in emergency oncological spine
surgery: a prospective analysis. J. Neurosurg. Spine, 21(5):698–703, 2014.
[48] S. Depaoli and R. van de Schoot. Improving transparency and replication in Bayesian
statistics: The WAMBS-Checklist. Psychol. Methods, 22(2):240–261, 2017.
[49] N. J. Devlin, P. Hansen, P. Kind, and A. Williams. Logical inconsistencies in sur-

vey respondents’ health state valuations - A methodological challenge for estimating
social tariffs. Health Econ., 12(7):529–544, 2003.
[50] P. Dolan, C. Gudex, P. Kind, and A. Williams. A social tariff for EuroQol: results
from a UK general population survey. 1995.
[51] P. Dolan, C. Gudex, P. Kind, and A. Williams. Valuing health states: a comparison
of methods. J. Health Econ., 15(2):209–31, apr 1996.
171
[52] M. G. T. Dolders, M. P. A. Zeegers, W. Groot, and A. Ament. A meta-analysis
demonstrates no significant differences between patient and population preferences.
J. Clin. Epidemiol., 59(7):653–64, jul 2006.
[53] F. J. Dorey and E. L. Korn. Effective sample sizes for confidence intervals for survival
probabilities. Stat. Med., 6(6):679–687, 1987.
[54] M. Drummond, D. Brixner, M. Gold, P. Kind, A. McGuire, and E. Nord. Toward a

Consensus on the QALY. Value Heal., 12:S31–S35, mar 2009.
[55] M. F. Drummond, M. J. Sculpher, G. W. Torrance, B. J. O’Brien, and G. L. Stoddart.

Methods for the Economic Evaluation of Health Care Programmes. Oxford University
Press, USA, 2005.
[56] D. M. Eddy. Accuracy versus transparency in pharmacoeconomic modelling: finding

the right balance. Pharmacoeconomics, 24(9):837–44, 2006.
[57] D. M. Eddy, W. Hollingworth, J. J. Caro, J. Tsevat, K. M. Mcdonald, and J. B.

Wong. Model transparency and validation: A Report of the ISPOR-SMDM Good
Research Practices Task Force -7. Value Heal., 15:843–850, 2012.
[58] N. C. Edwards, A. C. Skelly, J. E. Ziewacz, K. Cahill, and M. J. McGirt. The role of

decision analytic modeling in the health economic assessment of spinal intervention.
Spine (Phila. Pa. 1976)., 39(22 Suppl 1):S16–42, 2014.
[59] L. Engel, N. Bansback, S. Bryan, M. M. Doyle-Waters, and D. G. Whitehurst. Ex-

clusion Criteria in National Health State Valuation Studies: A Systematic Review.
Med. Decis. Mak., 36(7):798–810, 2016.
[60] M. Evans, H. H. Jensen, M. Bogelund, J. Gundgaard, B. Chubb, and K. Khunti.

Flexible insulin dosing improves health-related quality-of-life (HRQoL): a time trade-
off survey. J. Med. Econ., 16(11):1357–1365, 2013.
[61] M. Evans, K. Khunti, M. Mamdani, C. B. Galbo-Jorgensen, J. Gundgaard, M. Bo-

gelund, and S. Harris. Health-related quality of life associated with daytime and
nocturnal hypoglycaemic events: a time trade-off survey in five countries. Health
Qual. Life Outcomes, 11:90, 2013.
[62] Expert Panel on Radiation Oncology-Bone Metastases, S. S.-M. Lo, S. Ryu, E. L.

Chang, N. Galanopoulos, J. Jones, E. Y. Kim, C. D. Kubicky, C. P. Lee, P. S. Rose,
A. Sahgal, A. E. Sloan, B. S. Teh, B. J. Traughber, C. Van Poznak, and A. D. Vassil.
ACR Appropriateness Criteria R Metastatic Epidural Spinal Cord Compression and
Recurrent Spinal Metastasis. J. Palliat. Med., 18(7):573–84, jul 2015.
[63] K. Fairman and F. Curtiss. It’s Only a Pharmacoeconomic Model–Believe It or Not.

J Manag Care Pharm, 14(1):83–85, 2008.
172
[64] J. Fan, S. Upadhye, and A. Worster. Understanding receiver operating characteristic
(ROC) curves. CJEM Can. J. Emerg. Med. care = JCMU J. Can. soins medicaux
d’urgence, 8(1):19–20, 2006.
[65] M. G. Fehlings, A. Nater, L. Tetreault, B. Kopjar, P. Arnold, M. Dekutoski, J. Finkel-

stein, C. Fisher, J. France, Z. Gokaslan, E. Massicotte, L. Rhines, P. Rose, A. Sahgal,
J. Schuster, and A. Vaccaro. Survival and Clinical Outcomes in Surgically Treated
Patients With Metastatic Epidural Spinal Cord Compression: Results of the Prospec-
tive Multicenter AOSpine Study. J. Clin. Oncol., 34(3):268–276, nov 2015.
[66] P. D. Feigin. Maximum likelihood estimation for continuous-time stochastic pro-

cesses. Adv. Appl. Probab., 8(04):712–736, dec 1976.
[67] S. Ferrari and F. Cribari-Neto. Beta Regression for Modelling Rates and Proportions.
J. Appl. Stat., 31(7):799–815, aug 2004.
[68] S. E. Fienberg. When did Bayesian inference become ”Bayesian”? Bayesian Anal.,
1(1):1–40, mar 2006.
[69] J. I. Figueroa-Zúñiga, R. B. Arellano-Valle, and S. L. Ferrari. Mixed beta regression:

A Bayesian perspective. Comput. Stat. Data Anal., 61:137–147, may 2013.
[70] G. F. Findlay. The role of vertebral body collapse in the management of malignant
spinal cord compression. J. Neurol. Neurosurg. Psychiatry, 50(2):151–4, feb 1987.
[71] L. H. Fischer, J. M. Abzug, A. L. Osterman, P. J. Stern, and J. Chang. Complications

of common hand and wrist surgery procedures: flexor and extensor tendon surgery.
Instr. Course Lect., 63:97–103, jan 2014.
[72] R. A. Fisher. On the Mathematical Foundations of Theoretical Statistics. Philos.

Trans. R. Soc. A Math. Phys. Eng. Sci., 222(594-604):309–368, jan 1922.
[73] V. L. Fornasier and J. G. Horne. Metastases to the vertebral column. Cancer,

36(2):590–4, aug 1975.
[74] R. Fuller, N. Dudley, and J. Blacktop. How informed is consent? Understanding of

pictorial and verbal probability information by medical inpatients. Postgrad. Med.
J., 78(923):543–4, sep 2002.
[75] J. C. Furlan, K. K.-W. Chan, G. a. Sandoval, K. C. K. Lam, C. a. Klinger, R. A.

Patchell, A. Laporte, and M. G. Fehlings. The combined use of surgery and radio-
therapy to treat patients with epidural cord compression due to metastatic disease:
a cost-utility analysis. Neuro. Oncol., 14(5):631–40, may 2012.
[76] W. Furlong, D. Feeny, G. W. Torrance, R. Barr, and J. Horsman. Guide to design

and development of health-state utility instrumentation. 1990.
[77] D. M. Galvis, D. Bandyopadhyay, and V. H. Lachos. Augmented mixed beta regres-

sion models for periodontal proportion data. Stat. Med., 33(21):3759–3771, 2014.
173
[78] J. C. Gardiner, Z. Luo, and L. A. Roman. Fixed effects, random effects and GEE:
what are the differences? Stat. Med., 28(2):221–39, jan 2009.
[79] M. Gattellari, P. N. Butow, M. H. N. Tattersall, S. M. Dunn, and C. A. MacLeod.

Misunderstanding in cancer patients: Why shoot the messenger? Ann. Oncol.,
10(1):39–46, 1999.
[80] A. Gelman, B. P. Carlin, H. S. Stern, and D. B. Rubin. Bayesian Data Analysis.

Chapman and Hall/CRC, Boca Raton, second edition, 2003.
[81] A. Gelman, J. B. Carlin, H. S. Stern, D. B. Dunson, A. Vehtari, and D. B. R. Rubin.

Bayesian Data Analysis. CRC Press, Boca Raton, third edition, 2014.
[82] M. Gheorghe, S. Picavet, M. Verschuren, W. B. Brouwer, and P. H. van Baal. Health

losses at the end of life: a Bayesian mixed beta regression approach. J. R. Stat. Soc.
Ser. A Stat. Soc., 180(3):723–749, 2017.
[83] R. W. Gilbert, J. H. Kim, and J. B. Posner. Epidural spinal cord compression from
metastatic tumor: diagnosis and treatment. Ann. Neurol., 3(1):40–51, jan 1978.
[84] E. Goodwin, C. Green, and A. Spencer. Estimating a Preference-Based Index for an

Eight-Dimensional Health State Classification System for Multiple Sclerosis. Value
Health, 18(8):1025–1036, 2015.
[85] E. J. Graham, L. G. Lenke, T. G. Lowe, R. R. Betz, K. H. Bridwell, Y. Kong, and

K. Blanke. Prospective pulmonary function evaluation following open thoracotomy
for anterior spinal fusion in adolescent idiopathic scoliosis. Spine (Phila. Pa. 1976).,
25(18):2319–25, sep 2000.
[86] S. B. Green. How Many Subjects Does It Take To Do A Regression Analysis. Mul-
tivariate Behav. Res., 26(3):499–510, 1991.
[87] T. M. Gregory, R. Coriat, and O. Mir. Prognostic scoring systems for spinal metas-
tases in the era of anti-VEGF therapies. Spine (Phila. Pa. 1976)., 38(11):965–966,
2013.
[88] R. Grieve, N. Hawkins, and M. Pennington. Extrapolation of survival data in cost-

effectiveness analyses: improving the current state of play. Med. Decis. Making,
33(6):740–2, 2013.
[89] D. a. Grimes and K. F. Schulz. Refining clinical diagnosis with likelihood ratios.
Lancet, 365(panel 1):1500–1505, 2005.
[90] G. Guyatt, S. Walter, and G. Norman. Measuring Change Over Time- Aseessing the
Usefulness of Evaluative Instruments. J. Chronic Dis., 40(2):171–178, 1987.
[91] G. H. Guyatt, R. A. Deyo, M. Charlson, M. N. Levine, and A. Mitchell. Responsive-

ness and validity in health status measurement: A clarification. J. Clin. Epidemiol.,
42(5):403–408, jan 1989.
174
[92] P. Guyot, A. E. Ades, M. J. N. M. Ouwens, and N. J. Welton. Enhanced secondary
analysis of survival data: reconstructing the data from published Kaplan-Meier sur-
vival curves. BMC Med. Res. Methodol., 12(1):9, feb 2012.
[93] R. G. Hagerty, P. N. Butow, P. A. Ellis, E. A. Lobb, S. Pendlebury, N. Leighl,

D. Goldstein, S. K. Lo, and M. H. N. Tattersall. Cancer patient preferences for
communication of prognosis in the metastatic setting. J. Clin. Oncol., 22(9):1721–
1730, 2004.
[94] M. Haight, A. Quan-Haase, and B. a. Corbett. Revisiting the digital divide in

Canada: The impact of demographic factors on access to the Internet, level of online
activity, and social networking site usage. Information, Commun. Soc., 17(4):503–
519, 2014.
[95] E. F. Halpern. Behind the numbers: inverse probability weighting. Radiology,

271(3):625–8, jun 2014.
[96] L. C. Hamilton, J. Hartter, and K. Saito. Trust in Scientists on Climate Change and
Vaccines. SAGE Open, 5(3), 2015.
[97] T. Hammerschmidt, H.-P. Zeitler, M. Gulich, and R. Leidl. A comparison of different

strategies to collect standard gamble utilities. Med. Decis. Making, 24(5):493–503,
2004.
[98] J. A. Hanley. Receiver operating characteristic (ROC) methodology: the state of the
art. Crit. Rev. Diagn. Imaging, 29(3):307–35, 1989.
[99] J. a. Hanley and B. J. McNeil. The meaning and use of the area under a receiver
operating characteristic (ROC) curve. Radiology, 143(1):29–36, apr 1982.
[100] F. E. Harrell. Regression Modeling Strategies. Springer Series in Statistics. Springer

International Publishing, Cham, 2 edition, 2015.
[101] S. Harris, M. Mamdani, C. B. Galbo-Jorgensen, M. Bogelund, J. Gundgaard, and

D. Groleau. The effect of hypoglycemia on health-related quality of life: Canadian
results from a multinational time trade-off survey. Can. J. diabetes, 38(1):45–52,
2014.
[102] N. Hawkins and R. Grieve. Extrapolation of Survival Data in Cost-effectiveness

Analyses: The Need for Causal Clarity. Med. Decis. Making, 37(4):337–339, 2017.
[103] J. M. Herman and V. K. Sonntag. Cervical corpectomy and plate fixation for post-
laminectomy kyphosis. J. Neurosurg., 80(6):963–70, jun 1994.
[104] J. P. T. Higgins and D. J. Spiegelhalter. Being sceptical about meta-analyses: a

Bayesian perspective on magnesium trials in myocardial infarction. Int. J. Epi-
demiol., 31(1):96–104, 2002.
[105] P. Hougaard. Multi-state Models: A Review. Lifetime Data Anal., 5(3), 1999.
175
[106] M. Huo, A. Sahgal, D. Pryor, K. Redmond, S. Lo, and M. Foote. Stereotactic spine
radiosurgery: Review of safety and efficacy with respect to dose and fractionation.
Surg. Neurol. Int., 8(1):30, 2017.
[107] M. Huo, A. Sahgal, D. Pryor, K. Redmond, S. Lo, and M. Foote. Stereotactic spine
radiosurgery: Review of safety and efficacy with respect to dose and fractionation.
Surg. Neurol. Int., 8:30, 2017.
[108] J. A. Husted, R. J. Cook, V. T. Farewell, and D. D. Gladman. Methods for assessing
responsiveness. J. Clin. Epidemiol., 53(5):459–468, 2000.
[109] R. Hutchins, A. J. Viera, S. L. Sheridan, and M. P. Pignone. Quantifying the utility
of taking pills for cardiovascular prevention. Circ. Cardiovasc. Qual. Outcomes,
8(2):155–163, 2015.
[110] R. Jaeschke, G. H. Guyatt, and D. L. Sackett. Users’ guides to the medical literature.
III. How to use an article about a diagnostic test. B. What are the results and will
they help me in caring for my patients? The Evidence-Based Medicine Working
Group. JAMA, 271(9):703–7, mar 1994.
[111] S. J. Janssen, T. Teunis, E. van Dijk, M. L. Ferrone, J. H. Shin, F. Hornicek, and J. H.
Schwab. Validation of the Spine Oncology Study Group—Outcomes Questionnaire
to assess quality of life in patients with metastatic spine disease. Spine J., 17(6):768–
776, 2017.
[112] F. R. Johnson, E. Lancsar, D. Marshall, V. Kilambi, A. Mühlbacher, D. A. Regier,
B. W. Bresnahan, B. Kanninen, and J. F. Bridges. Constructing experimental designs
for discrete-choice experiments: Report of the ISPOR conjoint analysis experimental
design good research practices task force. Value Heal., 16(1):3–13, 2013.
[113] J. A. Johnson, N. Luo, J. W. Shaw, P. Kind, and S. J. Coons. Valuations of EQ-5D
health states: are the United States and United Kingdom different? Med Care,
43(3):221–228, 2005.
[114] N. L. Johnson, A. W. Kemp, and S. Kotz. Univariate Discrete Distributions. Wiley,
third edition, 2005.
[115] N. L. Johnson, S. Kotz, and N. Balakrishnan. Continuous Univariate Distributions,
Volume 2. Wiley, second edition, 1994.
[116] M. Jones-Lee, G. Loomes, D. O’Reilly, and P. Philips. The Value of Preventing
Non Fatal Road Injuries: Findings of a Willingness to Pay National Sample Survey.
Technical report, Transport Research Laboratory, Wokingham, United Kingdom,
1993.
[117] T. R. Jorgensen, C. Emborg, K. Dahlen, M. Bogelund, and A. Carlborg. The effect
of the medicine administration route on health-related quality of life: Results from a
time trade-off survey in patients with bipolar disorder or schizophrenia in 2 Nordic
countries. BMC Psychiatry, 16:244, 2016.
176
[118] M. B. Kabins and J. N. Weinstein. The History of Vertebral Screw and Pedicle Screw
Fixation. Iowa Orthop. J., 11:127–136, 1991.
[119] J. D. Kalbfleisch and J. F. Lawless. The Analysis of Panel Data under a Markov
Assumption. J. Am. Stat. Assoc., 80(392):863–871, dec 1985.
[120] J. D. Kalbfleisch and R. L. Pretince. The Statistical Analysis of Failure Time Data.
Wiley-Interscience, Hoboken, NJ, 2nd ed. edition, 2002.
[121] A. Karhade, V. Vasudeva, H. Dasenbrock, Y. Lu, W. Gormley, and W. Groff. Thirty-
day readmission and reoperation after surgery for spinal tumors: a National Surgical
Quality Improvement Program analysis. Neurosurg Focus, 41(2):E5, 2016.
[122] R. Kay. A Markov model for analysing cancer markers and disease states in survival
studies. Biometrics, 42(4):855–65, dec 1986.
[123] P. Kind, J. E. Lafata, K. Matuszewski, and D. Raisch. The use of QALYs in clinical
and patient decision-making: issues and prospects. Value Health, 12 Suppl 1:S27–30,
mar 2009.
[124] J. R. Kruschke. Doing bayesian data analysis: a tutorial with r and bugs. Academic
Press Inc., Oxford, 2011.
[125] J. Kuha. AIC and BIC: Comparisons of Assumptions and Performance. Sociol.
Methods Res., 33(2):188–229, nov 2004.
[126] K. Y. H. Kwan, T. C. Lam, H. C. W. Choi, H. Y. Koh, and K. M. C. Cheung.
Prediction of survival in patients with symptomatic spinal metastases: Comparison
between the Tokuhashi score and expert oncologists. Surg. Oncol., 27(1):7–10, 2018.
[127] B. M. Lagos-Álvarez, R. Fustos-Toribio, J. Figueroa-Zúñiga, and J. Mateu. Geo-
statistical mixed beta regression: a Bayesian approach. Stoch. Environ. Res. Risk
Assess., 31(2):571–584, 2017.
[128] L. M. Lamers, P. F. M. Stalmeier, P. F. M. Krabbe, and J. J. V. Busschbach. In-
consistencies in TTO and VAS values for EQ-5D health states. Med. Decis. Mak.,
26(2):173–181, 2006.
[129] J. R. Landis and G. G. Koch. The measurement of observer agreement for categorical
data. Biometrics, 33(1):159–174, 1977.
[130] N. R. Latimer. Survival analysis for economic evaluations alongside clinical trials–
extrapolation with patient-level data: inconsistencies, limitations, and a practical
guide. Med. Decis. Making, 33(6):743–54, aug 2013.
[131] I. Laufer, J. B. Iorgulescu, T. Chapman, E. Lis, W. Shi, Z. Zhang, B. W. Cox,
Y. Yamada, and M. H. Bilsky. Local disease control for spinal metastases follow-
ing ”separation surgery” and adjuvant hypofractionated or high-dose single-fraction
stereotactic radiosurgery: outcome analysis in 186 patients. J. Neurosurg. Spine,
18(3):207–14, mar 2013.
177
[132] I. Laufer, D. G. Rubin, E. Lis, B. W. Cox, M. D. Stubblefield, Y. Yamada, and
M. H. Bilsky. The NOMS framework: approach to the treatment of spinal metastatic
tumors. Oncologist, 18(6):744–51, jun 2013.
[133] C.-H. Lee, C. K. Chung, T.-A. Jahng, K.-j. Kim, C. H. Kim, S.-J. Hyun, H.-J. Kim,
S. R. Jeon, U.-K. Chang, S.-H. Lee, S.-H. Moon, H. Majeed, D. Zhang, G. Gravis,
C. Wibmer, N. Kumar, K. Y. Moon, J. H. Park, E. Tabouret, and S. Fuentes. Which
one is a valuable surrogate for predicting survival between Tomita and Tokuhashi
scores in patients with spinal metastases? A meta-analysis for diagnostic test ac-
curacy and individual participant data analysis. J. Neurooncol., 123(2):267–75, jun
2015.
[134] C. H. Lee, J. Ning, and Y. Shen. Analysis of restricted mean survival time for
length-biased data. Biometrics, 74(2):575–583, 2018.
[135] J. C. Lee and D. J. Sabavala. Bayesian Estimation and Prediction for the Beta-
Binomial Model. J. Bus. Econ. Stat., 5(3):357–367, jul 1987.
[136] J. S. Lee, K. T. Suh, J. I. Kim, and T. S. Goh. Analysis of Sagittal Balance of Anky-
losing Spondylitis Using Spinopelvic Parameters. J. Spinal Disord. Tech., 27(3):E94–
E98, 2014.
[137] L. A. Lenert, D. J. Cher, M. K. Goldstein, M. R. Bergen, and A. Garber. The effect

of search procedures on utility elicitations. Med. Decis. Making, 18(1):76–83, 1998.
[138] L. A. Lenert and A. E. Sturley. Use of the internet to study the utility values of the
public. Proc. AMIA Symp., pages 440–444, 2002.
[139] S. Lewandowsky, G. E. Gignac, and K. Oberauer. The Role of Conspiracist Ideation

and Worldviews in Predicting Rejection of Science. PLoS One, 8(10):e75637, oct
2013.
[140] S. Lewandowsky, G. E. Gignac, and S. Vaughan. The pivotal role of perceived

scientific consensus in acceptance of science. Nat. Clim. Chang., 3(4):399–404, apr
2013.
[141] I. M. Lipkus. Numeric, verbal, and visual formats of conveying health risks: Sug-
gested best practices and future recommendations. Med. Decis. Mak., 27(5):696–713,
2007.
[142] H. Llewellyn-Thomas, H. J. Sutherland, R. Tibshirani, A. Ciampi, J. E. Till, and

N. F. Boyd. Describing health states. Methodologic issues in obtaining values for
health states. Med. Care, 22(6):543–52, jun 1984.
[143] J. C. Lloyd, T. Yen, R. Pietrobon, J. S. Wiener, S. S. Ross, P. J. Kokorowski, C. P.

Nelson, and J. C. Routh. Estimating utility values for vesicoureteral reflux in the
general public using an online tool. J. Pediatr. Urol., 10(6):1026–1031, 2014.
178
[144] S. Lutz, T. Balboni, J. Jones, S. Lo, J. Petit, S. E. Rich, R. Wong, and C. Hahn.
Palliative radiation therapy for bone metastases: Update of an ASTRO Evidence-
Based Guideline. Pract. Radiat. Oncol., 7(1):4–12, 2017.
[145] E. Maranzano, R. Bellavita, R. Rossi, V. De Angelis, A. Frattegiani, R. Bagnoli,

M. Mignogna, S. Beneventi, M. Lupattelli, P. Ponticelli, G. P. Biti, and P. Latini.
Short-course versus split-course radiotherapy in metastatic spinal cord compression:
results of a phase III, randomized, multicenter trial. J. Clin. Oncol., 23(15):3358–65,
may 2005.
[146] E. Maranzano, P. Latini, E. Perrucci, S. Beneventi, M. Lupattelli, and E. Corgna.

Short-course radiotherapy (8 Gy x 2) in metastatic spinal cord compression: an
effective and feasible treatment. Int. J. Radiat. Oncol. Biol. Phys., 38(5):1037–44,
jul 1997.
[147] E. Maranzano, F. Trippa, M. Casale, S. Costantini, M. Lupattelli, R. Bellavita,

L. Marafioti, S. Pergolizzi, A. Santacaterina, M. Mignogna, G. Silvano, and V. Fusco.
8Gy single-dose radiotherapy is effective in metastatic spinal cord compression: re-
sults of a phase III randomized multicentre Italian trial. Radiother. Oncol., 93(2):174–
9, nov 2009.
[148] D. Mayo. In Defense of the Neyman-Pearson Theory of Confidence Intervals. Philos.

Sci., 48(2):269–280, 1981.
[149] P. McCullagh and J. Nelder. Generalized Linear Models. Chapman & Hall/CRC,
London, 2 edition, 1989.
[150] R. M. Merchant and D. A. Asch. Protecting the Value of Medical Science in the Age
of Social Media and ”Fake News”. JAMA, nov 2018.
[151] G. Mitera, A. Loblaw, A. Sahgal, and B. Danielson. Canadian-led International

Development of a European Organization for Research and Treatment of Cancer
Quality of Life Module for Malignant Spinal Cord Compression: Results of Phase I.
Int. J. Radiat. Oncol., 78(3):S604, nov 2010.
[152] L. B. Mokkink, C. B. Terwee, D. L. Patrick, J. Alonso, P. W. Stratford, D. L. Knol,

L. M. Bouter, and H. C. W. de Vet. The COSMIN study reached international
consensus on taxonomy, terminology, and definitions of measurement properties for
health-related patient-reported outcomes. J. Clin. Epidemiol., 63(7):737–745, 2010.
[153] L. A. Moye. Elementary bayesian biostatistics. Chapman & Hall/CRC, Boca Raton,
2008.
[154] T. Muka, D. Imo, L. Jaspers, V. Colpani, L. Chaker, S. J. van der Lee, S. Mendis,
R. Chowdhury, W. M. Bramer, A. Falla, R. Pazoki, and O. H. Franco. The global
impact of non-communicable diseases on healthcare spending and national income:
a systematic review. Eur. J. Epidemiol., 30(4):251–277, 2015.
179
[155] J. Neyman. Outline of a Theory of Statistical Estimation Based on the Classicial
Theory of Probability. Philos. Trans. R. Soc. London. Ser. A, Math. Phys. Sci.,
236(767):333–380, 1937.
[156] E. Nord, N. Daniels, and M. Kamlet. QALYs: Some Challenges. Value Heal.,
12(10):S10–S15, mar 2009.
[157] R. Norman, R. Viney, N. K. Aaronson, J. E. Brazier, D. Cella, D. S. J. Costa,

P. M. Fayers, G. Kemmler, S. Peacock, A. S. Pickard, D. Rowen, D. J. Street,
G. Velikova, T. A. Young, and M. T. King. Using a discrete choice experiment to
value the QLU-C10D: feasibility and sensitivity to presentation format. Qual. Life
Res., 25(3):637–49, mar 2016.
[158] J. Nyland, P. Quigley, C. Huang, J. Lloyd, J. Harrow, and A. Nelson. Preserv-

ing transfer independence among individuals with spinal cord injury. Spinal Cord,
38(11):649–57, nov 2000.
[159] S. Olofsson, H. Norrlid, and U. Persson. Preferences for improvements in attributes

associated with basal insulin: a time trade-off and willingness-to-pay survey of a
diabetic and non-diabetic population in Sweden. J. Med. Econ., 19(10):945–958,
2016.
[160] S. Olofsson, H. Norrlid, and U. Persson. Preferences for improvements in attributes

associated with basal insulin: a time trade-off and willingness-to-pay survey of a
diabetic and non-diabetic population in Sweden. J. Med. Econ., 19(10):945–58, oct
2016.
[161] M. Oppe, N. J. Devlin, B. van Hout, P. F. M. Krabbe, and F. de Charro. A program

of methodological research to arrive at the new international EQ-5D-5L valuation
protocol. Value Health, 17(4):445–53, jun 2014.
[162] M. Oppe, K. Rand-Hendriksen, K. Shah, J. M. Ramos-Goñi, and N. Luo. EuroQol

Protocols for Time Trade-Off Valuation of Health Outcomes. Pharmacoeconomics,
34(10):993–1004, 2016.
[163] J. A. Ortiz Gómez. The incidence of vertebral body metastases. Int. Orthop.,
19(5):309–11, 1995.
[164] M. Osborne. An effective method for computing regression quantiles. IMA J. Numer.
Anal., 12(2):151–166, 1992.
[165] R. Ospina and S. L. P. Ferrari. Inflated beta distributions. Stat. Pap., 51(1):111–126,
jan 2010.
[166] M. Pahuta, A. Formbach, G. Mitera, D. Coyle, J. Werier, and E. Wai. Validation

of the self-administered online assessment of preferences (SOAP) utility elicitation
tool. Can. J. Surg., 59(3 Suppl 2):S40, 2016.
180
[167] M. Pahuta, A. Frombach, E. Hashem, S. Spence, C. Sun, E. K. Wai, J. Werier, C. van
Walraven, and D. Coyle. The Psychometric Properties of a Self-Administered, Open-
Source Module for Valuing Metastatic Epidural Spinal Cord Compression Utilities.
PharmacoEconomics - open, 3(2):197–204, jun 2019.
[168] M. A. Pahuta, E. K. Wai, J. Werier, C. van Walraven, and D. Coyle. A General Pop-
ulation Utility Valuation Study For Metastatic Epidural Spinal Cord Compression
Health States. Spine (Phila. Pa. 1976)., 44(13):943–950, 2019.
[169] M. A. Pahuta, J. Werier, E. K. Wai, C. van Walraven, and D. Coyle. Back to

Bayesian: A strategy to enhance prognostication of metastatic spine disease. Int. J.
Clin. Pract., 73(4):e13322, apr 2019.
[170] A. Papoulis. Probability, Random Variables, and Stochastic Processes. McGraw-Hill

Inc., New York, third edition, 1991.
[171] R. a. Patchell, P. a. Tibbs, W. F. Regine, R. Payne, S. Saris, R. J. Kryscio, M. Mo-

hiuddin, and B. Young. Direct decompressive surgical resection in the treatment of
spinal cord compression caused by metastatic cancer: a randomised trial. Lancet
(London, England), 366(9486):643–8, aug 2005.
[172] G. R. Paton, E. Frangou, and D. R. Fourney. Contemporary treatment strategy for

spinal metastasis: the ”LMNOP” system. Can. J. Neurol. Sci., 38(3):396–403, may
2011.
[173] A. Penaloza, F. Verschuren, G. Meyer, S. Quentin-Georget, C. Soulie, F. Thys, and

P.-M. Roy. Comparison of the unstructured clinician gestalt, the wells score, and
the revised Geneva score to estimate pretest probability for suspected pulmonary
embolism. Ann. Emerg. Med., 62(2):117–124.e2, aug 2013.
[174] R. R. Picard and R. D. Cook. Cross-Validation of Regression Models. J. Am. Stat.

Assoc., 79(387):575–583, sep 1984.
[175] A. S. Pickard, M. P. Neary, and D. Cella. Estimation of minimally important dif-

ferences in EQ-5D utility and VAS scores in cancer. Health Qual. Life Outcomes,
5(1):70, nov 2007.
[176] M. Pignone, S. Saha, T. Hoerger, K. N. Lohr, S. Teutsch, and J. Mandelblatt. Chal-

lenges in systematic reviews of economic analyses. Ann. Intern. Med., 142(12 Pt
2):1073–9, jun 2005.
[177] J. Pliskin, D. Shepard, and M. Weinstein. Utility Functions for Life Years and Health
Status. Oper. Res., 28(1):206–224, 1980.
[178] M. J. Price, N. J. Welton, and A. E. Ades. Parameterization of treatment effects for

meta-analysis in multi-state Markov models. Stat. Med., 30(2):140–51, jan 2011.
181
[179] G. H. Prosser, P. J. Yates, D. J. Wood, S. E. Graves, R. N. de Steiger, and L. N.
Miller. Outcome of primary resurfacing hip replacement: evaluation of risk factors
for early revision. Acta Orthop., 81(1):66–71, 2010.
[180] L. A. Prosser, K. Payne, D. Rusinak, P. Shi, and M. Messonnier. Using a discrete

choice experiment to elicit time trade-off and willingness-to-pay amounts for influenza
health-related quality of life at different ages. Pharmacoeconomics, 31(4):305–315,
2013.
[181] H. Putter, M. Fiocco, and R. B. Geskus. Tutorial in biostatistics: competing risks

and multi-state models. Stat. Med., 26(11):2389–430, may 2007.
[182] J. M. Pyne, J. C. Fortney, S. Tripathi, D. Feeny, P. Ubel, and J. Brazier. How Bad
Is Depression? Preference Score Estimates from Depressed Patients and the General
Population. Health Serv. Res., 44(4):1406–1423, aug 2009.
[183] R Core Team. R: A Language and Environment for Statistical Computing, 2018.
[184] W. Rahman, C. J. Richards, and C. P. Duncan. Surgical nuances to minimize muscle

damage during the anterolateral intermuscular approach in minimally invasive hip
replacement. Instr Course Lect, 57:243–247, 2008.
[185] J. Raphael, K. Chan, S. Karim, R. Kerbel, H. Lam, K. delos Santos, R. Saluja, and
S. Verma. Antiangiogenic Therapy in Advanced Non–small-cell Lung Cancer: A
Meta-analysis of Phase III Randomized Trials. Clin. Lung Cancer, 18(4):345–353.e5,
2017.
[186] M. Rawlins. De Testimonio: On the evidence for decisions about the use of thera-
peutic interventions. Clin. Med. J. R. Coll. Physicians London, 8(6):579–588, 2008.
[187] J. Rawls. A Theory of Justice: Revised Edition. Belknap Press, Cambridge, Mas-
sachusetts, United States, 2 edition, 1999.
[188] M. Reck, L. Paz-Ares, P. Bidoli, F. Cappuzzo, S. Dakhil, D. Moro-Sibilot,

H. Borghaei, M. Johnson, R. Jotte, N. A. Pennell, F. A. Shepherd, A. Tsao,
M. Thomas, G. C. Carter, F. Chan-Diehl, E. Alexandris, P. Lee, A. Zimmermann,
A. Sashegyi, and M. Pérol. Outcomes in patients with aggressive or refractory dis-
ease from REVEL: A randomized phase III study of docetaxel with ramucirumab
or placebo for second-line treatment of stage IV non-small-cell lung cancer. Lung
Cancer, 112(March 2017):181–187, 2017.
[189] K. J. Redmond, S. S. Lo, C. Fisher, and A. Sahgal. Postoperative Stereotactic

Body Radiation Therapy (SBRT) for Spine Metastases: A Critical Review to Guide
Practice. Int. J. Radiat. Oncol. Biol. Phys., 95(5):1414–28, 2016.
[190] K. J. Redmond, S. S. Lo, C. Fisher, and A. Sahgal. Postoperative Stereotactic

Body Radiation Therapy (SBRT) for Spine Metastases: A Critical Review to Guide
Practice. Int. J. Radiat. Oncol. Biol. Phys., 95(5):1414–1428, 2016.
182
[191] N. Reid. Estimating the Median Survival Time. Biometrika, 68(3):601, dec 1981.
[192] D. Rennie and H. S. Luft. Pharmacoeconomic analyses: making them transparent,

making them credible. JAMA, 283(16):2158–60, apr 2000.
[193] S. S. Richardson and S. Berven. The development of a model for translation of the
Neck Disability Index to utility scores for cost-utility analysis in cervical disorders.
Spine J., 12(1):55–62, jan 2012.
[194] P. Royston and M. K. Parmar. Restricted mean survival time: an alternative to the
hazard ratio for the design and analysis of randomized trials with a time-to-event
outcome. BMC Med. Res. Methodol., 13(1):152, dec 2013.
[195] R. L. Ruff, V. W. Adamson, S. S. Ruff, and X. Wang. Directed rehabilitation

reduces pain and depression while increasing independence and satisfaction with life
for patients with paraplegia due to epidural metastatic spinal cord compression. J.
Rehabil. Res. Dev., 44(1):1–10, 2007.
[196] S. Ryu, J. Rock, R. Jain, M. Lu, J. Anderson, J. Y. Jin, M. Rosenblum, B. Movsas,

and J. H. Kim. Radiosurgical decompression of metastatic epidural compression.
Cancer, 116(9):2250–2257, 2010.
[197] A. Sahgal, C. M. Whyne, L. Ma, D. A. Larson, and M. G. Fehlings. Vertebral com-

pression fracture after stereotactic body radiotherapy for spinal metastases. Lancet
Oncol., 14(8):e310–e320, jul 2013.
[198] G. D. Sanders, P. J. Neumann, A. Basu, D. W. Brock, D. Feeny, M. Krahn, K. M.

Kuntz, D. O. Meltzer, D. K. Owens, L. A. Prosser, J. A. Salomon, M. J. Sculpher,
T. A. Trikalinos, L. B. Russell, J. E. Siegel, and T. G. Ganiats. Recommendations
for conduct, methodological practices, and reporting of cost-effectiveness analyses:
Second panel on cost-effectiveness in health and medicine. JAMA - J. Am. Med.
Assoc., 316(10):1093–1103, 2016.
[199] H. J. Schünemann, E. Ståhl, P. Austin, E. Akl, D. Armstrong, and G. H. Guyatt. A

comparison of narrative and table formats for presenting hypothetical health states to
patients with gastrointestinal or pulmonary disease. Med. Decis. Making, 24(1):53–
60.
[200] Schuyler W Huck. Statistical Misconceptions. Taylor and Francis Group, New York,
2009.
[201] E. D. Sears, R. Shin, L. A. Prosser, and K. C. Chung. Economic analysis of re-

vision amputation and replantation treatment of finger amputation injuries. Plast.
Reconstr. Surg., 133(4):827–840, 2014.
[202] K. Seppä and T. Hakulinen. Mean and median survival times of cancer patients
should be corrected for informative censoring. J. Clin. Epidemiol., 62(10):1095–1102,
2009.
183
[203] P. E. Shrout and J. L. Fleiss. Intraclass correlations: Uses in assessing rater reliabil-
ity., 1979.
[204] R. M. Slone, M. MacMillan, W. J. Montgomery, and M. Heare. Spinal fixation.

Part 2. Fixation techniques and hardware for the thoracic and lumbosacral spine.
Radiographics, 13(3):521–43, may 1993.
[205] D. M. Smith, R. L. Sherriff, L. Damschroder, G. Loewenstein, and P. A. Ubel.

Misremembering colostomies? Former patients give lower utility ratings than do
current patients. Health Psychol., 25(6):688–695, 2006.
[206] W. R. Smith, Y.-J. Kim, B. H. Ziran, R. Ganz, J. O. Anglen, S. B. Murphy, P. F.

Stahel, R. L. Buly, and M. B. Millis. Hip joint-preserving surgery: beyond the classic
osteotomy. Instr. Course Lect., 55:145–158, jan 2006.
[207] M. Smithson and J. Verkuilen. A better lemon squeezer? Maximum-likelihood re-

gression with beta-distributed dependent variables. Psychol. Methods, 11(1):54–71,
mar 2006.
[208] G. J. Snoek, M. J. IJzerman, H. J. Hermens, D. Maxwell, and F. Biering-Sorensen.

Survey of the needs of patients with spinal cord injury: impact and priority for
improvement in hand function in tetraplegics. Spinal Cord, 42(9):526–32, sep 2004.
[209] D. J. Spiegelhalter, K. R. Abrams, and J. P. Myles. Bayesian Approaches to Clinical

Trials and Health-Care Evaluation. Wiley, Chischester, 2004.
[210] D. J. Spiegelhalter, A. Thomas, N. Best, and D. Lunn. WinBUGS User Manual:

version 1.4. Technical report, MRC Biostatistics Unit, Cambridge, UK, 2003.
[211] D. E. Spratt, W. H. Beeler, F. Y. de Moraes, L. D. Rhines, J. J. Gemmete, N. Chaud-

hary, D. B. Shultz, S. R. Smith, A. Berlin, M. Dahele, B. J. Slotman, K. C. Younge,
M. Bilsky, P. Park, and N. J. Szerlip. An integrated multidisciplinary algorithm for
the management of spinal metastases: an International Spine Oncology Consortium
report. Lancet Oncol., 18(12):e720–e730, 2017.
[212] Stan Development Team. The Stan Core Library, Version 2.17.0, 2017.
[213] E. W. Steyerberg, M. J. Eijkemans, F. E. Harrell Jr., and J. D. Habbema. Prognostic

modeling with logistic regression analysis: in search of a sensible strategy in small
data sets. Med.Decis.Making, 21(0272-989X (Print)):45–56, 2001.
[214] A. M. Stiggelbout, G. M. Kiebert, J. Kievit, J. W. Leer, G. Stoter, and J. C. de Haes.

Utility assessment in cancer patients: adjustment of time tradeoff scores for the utility
of life years and comparison with standard gamble scores. Med. Decis. Making,
14(1):82–90, 1994.
[215] M. R. Stockler, M. H. N. Tattersall, M. J. Boyer, S. J. Clarke, P. J. Beale, and

R. J. Simes. Disarming the guarded prognosis: predicting survival in newly referred
patients with incurable cancer. Br. J. Cancer, 94(2):208–12, jan 2006.
184
[216] J. Street, B. Lenehan, S. Berven, and C. Fisher. Introducing a new health-related
quality of life outcome tool for metastatic disease of the spine: Content validation
using the international classification of functioning, disability, and health; On behalf
of the spine oncology study group. Spine (Phila. Pa. 1976)., 35(14):1377–1386, 2010.
[217] J. Street, B. Lenehan, S. Berven, and C. Fisher. Introducing a new health-related

quality of life outcome tool for metastatic disease of the spine: content validation
using the International Classification of Functioning, Disability, and Health; on behalf
of the Spine Oncology Study Group. Spine (Phila. Pa. 1976)., 35(14):1377–86, jun
2010.
[218] M. Stubblefield. Barriers to Rehabilitation of the Malignant Spine. Top. Spinal Cord
Inj. Rehabil., 14(2):19–30, oct 2008.
[219] W. Sumner, R. Nease, and B. Littenberg. U-titer: a utility assessment tool. Proc.
Annu. Symp. Comput. Appl. Med. Care, pages 701–5, jan 1991.
[220] N. Sundaresan, G. V. Digiacinto, and J. E. Hughes. Surgical treatment of spinal

metastases. Clin. Neurosurg., 33:503–22, 1986.
[221] N. Sundaresan, A. A. Steinberger, F. Moore, V. P. Sachdev, G. Krol, L. Hough, and

K. Kelliher. Indications and results of combined anterior—posterior approaches for
spine tumor surgery. J. Neurosurg., 85(3):438–446, 1996.
[222] P. Tappenden, J. Chilcott, S. Ward, S. Eggington, D. Hind, and S. Hummel. Method-

ological issues in the economic analysis of cancer treatments. Eur. J. Cancer,
42(17):2867–75, nov 2006.
[223] C. B. Terwee, S. D. M. Bot, M. R. de Boer, D. van der Windt, D. L. Knol, J. Dekker,

L. M. Bouter, and H. C. W. de Vet. Quality criteria were proposed for measurement
properties of health status questionnaires. J. Clin. Epidemiol., 60(1):34–42, 2007.
[224] The World Health Organization. Global status report on noncommunicable diseases
2010. Geneva, 2010.
[225] K. C. Thomas, B. Nosyk, C. G. Fisher, M. Dvorak, R. A. Patchell, W. F. Regine,

A. Loblaw, N. Bansback, D. Guh, H. Sun, and A. Anis. Cost-effectiveness of surgery
plus radiotherapy versus radiotherapy alone for metastatic epidural spinal cord com-
pression. Int. J. Radiat. Oncol. Biol. Phys., 66(4):1212–8, nov 2006.
[226] A. C. Titman and L. D. Sharples. Model diagnostics for multi-state models. Stat.
Methods Med. Res., 19(6):621–51, dec 2010.
[227] Y. Tokuhashi, Y. Ajiro, and M. Oshima. Algorithms and planning in metastatic

spine tumors. Orthop. Clin. North Am., 40(1):37–46, v–vi, jan 2009.
[228] Y. Tokuhashi, H. Matsuzaki, H. Oda, M. Oshima, and J. Ryu. A revised scoring

system for preoperative evaluation of metastatic spine tumor prognosis. Spine (Phila.
Pa. 1976)., 30(19):2186–91, oct 2005.
185
[229] Y. Tokuhashi, H. Matsuzaki, S. Toriyama, H. Kawano, and S. Ohsaka. Scoring
system for the preoperative evaluation of metastatic spine tumor prognosis. Spine
(Phila. Pa. 1976)., 15(11):1110–3, nov 1990.
[230] Y. Tokuhashi, Y. Tokuhashi, H. Matsuzaki, H. Matsuzaki, H. Oda, H. Oda, M. Os-

hima, M. Oshima, J. Ryu, and J. Ryu. A revised scoring system for preoperative eval-
uation of metastatic spine tumor prognosis. Spine (Phila. Pa. 1976)., 30(19):2186–91,
oct 2005.
[231] Toluna Ltd. Esomar 26: 26 questions to help research buyers of online samples.
Technical report, 2013.
[232] K. Tomita, N. Kawahara, T. Kobayashi, A. Yoshida, H. Murakami, and T. Akamaru.

Surgical strategy for spinal metastases. Spine (Phila. Pa. 1976)., 26(3):298–306, feb
2001.
[233] D. Trafimow and M. Marks. Editorial. Basic Appl. Soc. Psych., 37(1):1–2, 2015.
[234] A. Tsuchiya and P. Dolan. The QALY model and individual preferences for health
states and health profiles over time: a systematic review of the literature. Med.
Decis. Making, 25(4):460–7, 2005.
[235] J. Tukey. Exploratory Data Analysis. Addison-Wesley, Reading, PA, 1977.
[236] H. Uei, Y. Tokuhashi, and M. Maseda. Treatment Outcome of Metastatic Spine

Tumor in Lung Cancer Patients. Spine (Phila. Pa. 1976)., 42(24):E1446–E1451,
2017.
[237] Y. M. van der Linden, S. P. D. S. Dijkstra, E. J. A. Vonk, C. A. M. Marijnen,

J. W. H. Leer, and Dutch Bone Metastasis Study Group. Prediction of survival in
patients with metastases in the spinal column: results based on a randomized trial
of radiotherapy. Cancer, 103(2):320–8, jan 2005.
[238] F. E. van Nooten, X. Koolman, and W. B. F. Brouwer. The influence of subjective life
expectancy on health state valuations using a 10 year TTO. Health Econ., 18(5):549–
558, 2009.
[239] F. E. van Nooten, X. Koolman, J. J. V. Busschbach, and W. B. F. Brouwer. Thirty

down, only ten to go?! Awareness and influence of a 10-year time frame in TTO.
Qual. Life Res., 23(2):377–384, 2014.
[240] F. E. van Nooten, N. J. A. van Exel, X. Koolman, and W. B. F. Brouwer. ”Married

with children” the influence of significant others in TTO exercises. Health Qual. Life
Outcomes, 13:94, 2015.
[241] F. E. van Nooten, N. J. A. van Exel, X. Koolman, and W. B. F. Brouwer. ”Married

with children” the influence of significant others in TTO exercises. Health Qual. Life
Outcomes, 13:94, jul 2015.
186
[242] A. L. Versteeg, A. Sahgal, L. D. Rhines, D. M. Sciubba, J. M. Schuster, M. H.
Weber, P. P. Varga, S. Boriani, C. Bettegowda, M. G. Fehlings, M. J. Clarke, P. M.
Arnold, Z. L. Gokaslan, and C. G. Fisher. Psychometric evaluation and adaptation of
the Spine Oncology Study Group Outcomes Questionnaire to evaluate health-related
quality of life in patients with spinal metastases. Cancer, 124(8):1828–1838, 2018.
[243] J. Von Neumann and O. Morgenstern. Theory of games and economic behavior.
Princeton University Press, Princeton, 1953.
[244] X. Wan, L. Peng, and Y. Li. A Review and Comparison of Methods for Recreating
Individual Patient Data from Published Kaplan-Meier Survival Curves for Economic
Evaluations: A Simulation Study. PLoS One, 10(3):e0121353, mar 2015.
[245] R. G. Watkins III and R. G. Watkins IV. Surgical Approaches to the Spine. Springer
New York, New York, NY, 2015.
[246] H.-L. Wee, S.-C. Li, F. Xie, X.-H. Zhang, N. Luo, D. Feeny, Y.-B. Cheung, D. Machin,
K.-Y. Fong, and J. Thumboo. Validity, feasibility and acceptability of time trade-off
and standard gamble assessments in health valuation studies: a study in a multiethnic
Asian population in Singapore. Value Health, 11 Suppl 1:S3–10, 2008.
[247] M. G. M. Weernink, C. G. M. Groothuis-Oudshoorn, M. J. IJzerman, and J. A.

van Til. Valuing Treatments for Parkinson Disease Incorporating Process Utility:
Performance of Best-Worst Scaling, Time Trade-Off, and Visual Analogue Scales.
Value Health, 19(2):226–232, 2016.
[248] M. C. Weinstein, J. E. Siegel, M. R. Gold, M. S. Kamlet, and L. B. Russell. Rec-

ommendations of the Panel on Cost-effectiveness in Health and Medicine. JAMA,
276(15):1253–8, oct 1996.
[249] M. C. Weinstein, G. Torrance, and A. McGuire. QALYs: the basics. Value Health,
12 Suppl 1:S5–9, mar 2009.
[250] N. J. Welton and A. E. Ades. Estimation of markov chain transition probabilities

and rates from fully and partially observed data: uncertainty propagation, evidence
synthesis, and model calibration. Med. Decis. Making, 25(6):633–45, 2005.
[251] C. Williams, J. D. Lewsey, D. F. Mackay, and A. H. Briggs. Estimation of Sur-

vival Probabilities for Use in Cost-effectiveness Analyses: A Comparison of a Multi-
state Modeling Survival Analysis Approach with Partitioned Survival and Markov
Decision-Analytic Modeling. Med. Decis. Mak., 37(4):427–439, 2017.
[252] S. E. Wolowacz, A. Briggs, V. Belozeroff, P. Clarke, L. Doward, R. Goeree, A. Lloyd,

and R. Norman. Estimating Health-State Utility for Economic Models in Clini-
cal Studies: An ISPOR Good Research Practices Task Force Report. Value Heal.,
19(6):704–719, 2016.
187
[253] B. Woods, E. Sideris, S. Palmer, N. Latimer, and M. Soares. Partitioned survival
analysis for decision modelling in health care: a critical review. Technical Report
June, NICE Decision Support Unit, Sheffield, 2017.
[254] G. Wu and L. Peng. Pharmacokinetics of propofol in patients undergoing total hip

replacement : effect of acute hypervolemic hemodilution. Anaesthesist, 60(9):835–
840, 2011.
[255] R. F. Young, E. M. Post, and G. A. King. Treatment of spinal epidural metastases.

Randomized prospective comparison of laminectomy and radiotherapy. J. Neuro-
surg., 53(6):741–8, 1980.
[256] N. Zethraeus and M. Johannesson. A comparison of patient and social tariff values
derived from the time trade-off method. Health Econ., 8(6):541–5, sep 1999.
[257] L. Zhao, B. Claggett, L. Tian, H. Uno, M. A. Pfeffer, S. D. Solomon, L. Trippa, and

L. J. Wei. On the restricted mean survival time curve in survival analysis. Biometrics,
72(1):215–221, 2016.
188

Decision Analysis of Surgical Treatment Indications For Metastatic Epidural Spinal Cord Compression

Uploaded by

Copyright:

Available Formats

Decision Analysis of Surgical Treatment Indications For Metastatic Epidural Spinal Cord Compression

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Decision Analysis of Surgical Treatment Indications For Metastatic Epidural Spinal Cord Compression

Uploaded by

Copyright:

Available Formats

Decision Analysis of Surgical

Treatment Indications for Metastatic

Thesis submitted to the

School of Epidemiology and Public Health

This manuscript-based thesis consists of a series of articles prepared for publication in

2 SOAP-MESCC Psychometric Evaluation 9

3 General Population Utility Valuation Study 21

5 Approximating Transition Rates 47

7 Discussion and Conclusion 77

A Appendices for Chapter 2 84

C Appendices for Chapter 4 129

D Appendices for Chapter 5 132

E Appendices for Chapter 6 145

cS+RT laminectomy without stabilization followed by external beam radiotherapy

EORTC European Organisation for Research and Treatment of Cancer

HRQoL health-related quality-of-life

ICC intra-class correlation coefficient

MCID minimally clinically important difference

MESCC metastatic epidural spinal cord compression

mS+RT modern surgery followed by external beam radiotherapy

NSCLC non-small cell lung cancer

PFS progression-free survival

QALY quality-adjusted life year

RT-alone external beam radiotherapy alone

SBRT stereotactic body radiation therapy

SOAP Self-directed Online Assessment of Preferences

2.1 MESCC Health State Attribute Phrasing . . . . . . . . . . . . . . . . . . . 14

3.1 Participant Demographics . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

5.1 Data Abstracted from MESCC RCT PFS and OS Analyses . . . . . . . . . 54

6.1 Selection Criteria used in the MESCC RCT[171] . . . . . . . . . . . . . . 67

B.1 Quotas Based on the 2016 Canadian Census . . . . . . . . . . . . . . . . . 112

C.1 Parameters used for Interpolation of var [S (tm )] . . . . . . . . . . . . . . . 130

3.1 Flow of participants. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

5.1 State-transition diagram for illness-death model . . . . . . . . . . . . . . . 50

6.1 State-transition diagram for decision model . . . . . . . . . . . . . . . . . . 68

A.1 MESCC SOAP: page 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

B.1 Variance function check. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

1. For tissue diagnosis

2. Following deterioration during or relapse after RT-alone

1.2 Surgical Decision Making for MESCC

Barrier 1: Lack of Utility Data

Barrier 2: Skepticism Regarding Decision Analysis

1.4 Overview of Thesis

Barrier 1: Lack of Utility Data

Barrier 2: Skepticism Regarding Decision-Analysis

The Psychometric Properties of a

2 – Postgraduate Medical Education, The University of Ottawa, Ottawa, Ontario, Canada

3 – Undergraduate Medical Education, The University of Ottawa, Ottawa, Ontario,

4 – Division of Orthopaedic Surgery, The University of Ottawa, Ottawa, Ontario, Canada

6 – The Ottawa Hospital Research Institute, Ottawa, Ontario, Canada

7 – The Institute for Clinical Evaluative Sciences, Ottawa, Ontario, Canada

Attribute Symbol Item EORTC Code

Dependence D- You do not need help with eating, C05, BM49

Table 2.2: Participant Characteristics Stratified by Response Pattern

Valid Valid Invalid Invalid

Table 2.3: Agreement, Reliability, and Responsiveness Results

Number Smallest ICC (95% CI)† Guyatt’s

A General Population Utility

2 – Division of Orthopaedic Surgery, The University of Ottawa, Ottawa, Ontario, Canada