Author's Accepted Manuscript

Setting tolerance limits for statistical project

control using earned value management
Jeroen Colin, Mario Vanhoucke

www.elsevier.com/locate/omega

PII: S0305-0483(14)00074-7
DOI: http://dx.doi.org/10.1016/j.omega.2014.06.001
Reference: OME1408

To appear in: Omega

Received date: 14 May 2013

Revised date: 12 June 2014
Accepted date: 12 June 2014

Cite this article as: Jeroen Colin, Mario Vanhoucke, Setting tolerance limits for
statistical project control using earned value management, Omega, http://dx.doi.
org/10.1016/j.omega.2014.06.001

This is a PDF file of an unedited manuscript that has been accepted for
publication. As a service to our customers we are providing this early version of
the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting galley proof before it is published in its final citable form.
Please note that during the production process errors may be discovered which
could affect the content, and all legal disclaimers that apply to the journal
pertain.
 Setting tolerance limits for statistical project control using earned
value management

Jeroen Colina , Mario Vanhouckea,b,c,∗

a Faculty of Economics and Business Administration, Ghent University, Tweekerkenstraat 2, 9000 Gent
(Belgium)
b Technology and Operations Management, Vlerick Business School, Reep 1, 9000 Gent (Belgium)
c Department of Management Science and Innovation, University College London, Gower Street, London WC1E
6BT (United Kingdom)

Abstract

Project control has been a research topic since decades that attracts both academics and prac-
titioners. Project control systems indicate the direction of change in preliminary planning vari-
ables compared with actual performance. In case their current project performance deviates from
the planned performance, a warning is indicated by the system in order to take corrective ac-
tions.
Earned value management/earned schedule (EVM/ES) systems have played a central role in
project control, and provide straightforward key performance metrics that measure the devia-
tions between planned and actual performance in terms of time and cost. In this paper, a new
statistical project control procedure sets tolerance limits to improve the discriminative power be-
tween progress situations that are either statistically likely or less likely to occur under the project
baseline schedule. In this research, the tolerance limits are derived from subjective estimates for
the activity durations of the project. Using the existing and commonly known EVM/ES metrics,
the resulting project control charts will have an improved ability to trigger actions when variation
in a project’s progress exceeds certain predefined thresholds
A computational experiment has been set up to test the ability of these statistical project control
charts to discriminate between variation that is either acceptable or unacceptable in the duration of
the individual activities. The computational experiments compare the use of statistical tolerance
limits with traditional earned value management thresholds and validate their power to report
warning signals when projects tend to deviate significantly from the baseline schedule.
Keywords: Project Management, Scheduling, Risk, Simulation

1. Introduction

Project management and control have been research topics since the development of project plan-
ning approaches such as the critical path method (CPM, [1, 2]) and the Program Evaluation and
Review Technique (PERT, [3]). The majority of the research endeavors published in the academic
literature focuses on the construction of a project baseline schedule within the presence of limited
resource constraints (see e.g. a recent survey written by Hartmann and Briskorn [4]). Ever since,
the use of a project baseline schedule has been put into the right perspective, as it only acts as a

∗ Corresponding author
Email addresses: [email protected] (Jeroen Colin), [email protected] (Mario Vanhoucke)

Preprint submitted to Elsevier June 18, 2014

roadmap for the future project performance. If the project runs into trouble, the baseline sched-
ule can be used to detect such problems and to trigger corrective actions. The construction of a
project baseline schedule should go hand in hand with the development of a risk analysis [5] to
detect the weak parts in the schedule. These two sources of information should therefore be used
concurrently, both before and during the execution of the project. The need for the integration
of the project baseline scheduling construction, the risk analysis and management and the project
control step has been recently referred to as dynamic scheduling [6, 7].
In this paper, the construction of a baseline schedule is assumed to be a given, and the focus lies
on the presentation of a new project control system [8]. Rozenes et al. [9] have characterized a
project control system by stating that it should indicate the direction of change in preliminary
planning variables when compared with actual performance. If there is an unacceptable gap
between planning and performance, a warning should be generated by the system in order to take
corrective action. In the remainder of this text, we will further elaborate on this unacceptable gap
and quantify how it can be specified.
This study mainly focuses on the generation and interpretation of warning signals during project
control and hence it does not include results in the quality and input of corrective actions, nor on
the use of these new methods for contingency planning.
A well-known and widely used project control system is Earned Value Management (EVM) origi-
nally developed in the ’60s at the U.S. Department of Defense as a standard method to measure
a project’s performance. This system relies on a set of straightforward metrics to measure and
evaluate the general health of a project. These metrics serve as early warning signals to timely
detect project problems or to exploit project opportunities. EVM has been investigated and used
widely in project management practice and has been brought into the attention of researchers by
Fleming and Koppelman [10]. Recently, the research on EVM has produced the earned schedule
method (ES, [11]) and a large study on the integration of risk management and EVM/ES was
conducted [12].
EVM/ES calculations are based on the costs and durations for the activities in a project. Both in
the planning phase and during the execution of the project, some detail of the individual activities
is lost as the individual performance is summed up into the performance metrics observed and
controlled at the project level. Earned value management is known to have some shortcomings
when it comes to protecting the critical path in the schedule [10]. The summation of activity
earned value into the aggregate level for the project does not incorporate schedule risk analysis
(SRA, [13]) data, such as criticality. In spite of the complexity of this problem, EVM/ES use
in practice is characterized by decision making from practical experience, rules-of-thumb and
anecdotal evidence. In the academic literature, it is reasoned that the application of a statistical
process control approach could benefit the EVM/ES project control process. These previous
attempts depend on the availability of historical records that may or may not exist and may or
may not be valid. We therefore also allow the use of subjective estimates.
In this paper, a new statistical project control approach is presented. It is inspired by the practical
format of statistical process control charts and based on tolerance intervals produced by a simulated
sample. A controlled simulation experiment will be used to specifically produce two types of
project progress situations. First, project progress situations where the variation from plan is
limited to some acceptable margin are used to produce project control charts. Secondly, we test
the discriminative power of these control charts when the variation from plan exceed the pre-
defined acceptable margin.
The outline of this paper is as follows. In section 2, we present a literature review on the use
of statistical control principles in project management and control. We will present two project
control charts, the X chart and R chart. Section 3 illustrates these control charts on a fictitious
project example. In section 4, the design of the computational test experiments is outlined based
on various sources from literature. In section 5, test results obtained by computational experiments

2
using Monte Carlo simulation to imitate a fictitious project progress environment are discussed.
Section 6 draws general conclusions and highlights future research avenues.

2. Statistical Project Control

2.1. Literature overview

This literature review introduces the main concepts of statistical process control (SPC) as they
are applied to monitor and control processes in services and manufacturing. The application of
SPC techniques to earned value management is not new in literature. We introduce a statistical
project control procedure that aims to overcome some of the issues with the previously published
approaches, but which is not an implementation of SPC in the strict sense, although it repro-
duces some of its concepts and nomenclature. The last paragraph of this literature review is
dedicated to a summation of the differences between our statistical project control procedure and
the implications of applying SPC to project control in a strict sense.
In standard SPC applications a state of control is identified with a process generating samples
for which the subgroup averages are approximately normal under the central limit theory. Con-
trol charts such as the Shewhart, cumulative sum (CUSUM) and exponentially weighted mov-
ing avarage (EWMA) charts serve as on-line procedures to monitor process stability, to detect
assignable variation or to forecast process movements in industrial processes [14]. A process
is said to be in-control when only common cause variation is present. This type of variation
is characterized as coming from phenomena constantly active in a process, which can be pre-
dicted probabilistically. In his original work on process control, Shewhart [15] introduced the
term chance cause. A process is said to be out-of-control if a second type of variation is present
known as assignable cause variation. Assignable cause variation arises when a new, previously
unanticipated phenomenon is present in the system and should cause a signal.
Figure 1 illustrates how a control chart applied to project control might be interpreted. Periodic
observations of a performance measure for a fictitious project execution are outlined, along with
an illustrative upper control limit (UCL) and lower control limit (LCL). For illustrative purposes,
we will assume that the performance measure is of the type where a high value is good and a
low value acts as a trigger for actions due to potential problems. The frequently encountered
schedule performance index (SPI, section 2.3.1) and schedule performance index using earned
schedule (SPI(t), section 2.3.1) measures in EVM/ES are of this type, although control charts of
the opposite type have also been proposed where the natural logarithm of the reciprocal of SPI or
SPI(t) is taken [16]. In case the observations fall outside the project control limits, the charts report
a signal which could trigger actions to bring the project back to plan or to exploit opportunities.
More precisely, when an observation falls below the lower control limit, the performance has
likely dropped below the pre-defined acceptable margin on the baseline estimates. In this case,
this signal should be interpreted as a warning signal to the project manager to consider taking
corrective actions. When the performance indicator exceeds the upper control limit, it can be
seen as a signal to exploit opportunities, in which case the current project schedule might be
re-baselined to incorporate this opportunity. Different SPC control charts have been implemented
for project control using earned value management by [16–21]. A short overview of their respective
approaches and findings is listed below.
Lipke and Vaughn [17] apply a XmR chart (individuals and moving range chart) and calculate the
control limits based on a two-observation sample of earned value performance metrics. They use
the reciprocal measures of CPI (cost performance index) and SPI (schedule performance index),
as they have shown that these are normalized, such that many of the issues of applying SPC,
such as variability and homogeneity of the data, are avoided. Bauch and Chung [18] modify the
statistical process control charts proposed by Shewhart to be used with earned value management
considering three major aspects. Firstly, they asses single observations relative to historical project

3
 OutControl chart
of control signal
indication
→opportunity
→Re-baselining?
UCL

Central line % Completed

LCL

Out of control
Control chart indication
signal
→danger
→Take corrective actions

Figure 1: Illustrative control chart and out-of-control indications

data from 20 similar projects. Secondly, as projects span a finite time horizon, which is typically
different for each project, they normalize the different time lengths of each project into a consistent
number of time periods, chosen as 20 discrete measurements for the purpose of the paper. As a
third modification, they adopted the CUSUM approach [22] to incorporate the increasing nature
of period-to-period measurements of project performance. The typical y-axis of control charts
is transformed to plot cumulative project performance values. As a result, the central line and
upper and lower limits are non-stationary. Wang et al. [19] illustrate the use of SPC charts
using EVM/ES on a set of more than 30 software projects where abnormal progress situations are
effectively detected. [20] combine XmR charts with EVM/ES performance metrics, but implement
the logarithm of the reciprocal of SPI and the cost performance index (CPI). This was earlier
proposed by Lipke [16]. Control limits are calculated from historical data of 120 projects performed
by a management consulting company. Leu and Lin [20] introduce the division of different cases
based on historical factors influencing project performance and quantify the XmR chart procedure
for different trends observed in their data. Moreover, they accurately describe how “the quick
proliferation and complexity of project performance data indicate the need for a well-organized
project performance analysis process.” Aliverdi et al. [21] suggest to transform CPI data when
quality control charts are applied for cost control. SPC charts for both individual and moving
range measurements were successfully implemented. They conclude that the integrated approach
of EVM/ES and statistical quality control charts can contribute to a better and more reliable
project control process.
Overall, previous research on the integration of EVM/ES with SPC techniques for projects has
established that it improves project control by providing an objectively based and easily imple-
mented real-time monitoring system. The evidence provided in the literature is based on empirical
research of either post-hoc statistical analyses on projects [19] or the implementation of control
limits derived from historical data [16, 20]. The dependency on historical data does not pose an
immediate problem for statistical project control approaches, ideally the use of historical records
is preferred over subjective estimates for project control [23]. However, the concept of “similarity”
between projects, is often vaguely defined. When historical EVM/ES data is directly applied, the
project-specific EVM/ES dynamics should not be ignored. [24] illustrated these project-specific
dynamics with EVM/ES forecasting becoming less reliable for projects with a network close to
a parallel network structure. When the concepts of SPC are directly translated to incorporate
EVM/ES measurement, an assumption with respect to their distribution has to be made. In
previous literature this prerequisite was fulfilled by transforming the data to appear normally
distributed [17, 19] or lognormally distributed [16, 20] or appeared overlooked [21].
In section 2.3 we propose a different statistical project control procedure that will also include

4
the appealing aspects of using control charts. They are easy to set up, implement and interpret.
However, the imposed tolerance limits on the earned value metrics should be interpreted in the
more technical sense [25]. The tolerance limits are acquired from the empirical distribution function
(edf) of the EVM/ES metric under study at different points along the life-cycle of a project. This
edf is obtained from a Monte Carlo simulation model that allows variation to be added at the
activity level of the project. This variation can be introduced as subjective estimates of what
deviations are statistically plausible during the execution of the project or from calibration with
historical in-company data.

Table 1: Difference between SPC for projects and our statistical project control procedure

statistical process control statistical project control

for project control setting tolerance limits
Data type Process data Simulated data
Data measurements Data collected over time Data collected using simulations
Number of data points Small sample size per time unit Large sample size per simulation point
Process Infinite process Finite process
Analysis Ex post Ex ante
Control limits Deviations from normal progress Deviations from desired progress
Normal project progress Baseline schedule
Reference value
(with normal variation) (with acceptable tolerances)

Table 1 summarizes the main differences between our statistical project control procedure that
uses tolerance limits and a classical SPC implementation for project control. The main aspects
are given along the following lines.
• Data input: A first fundamental difference lies in the data input necessary to construct the
control limits. While process data needs to be collected at regular time intervals to monitor
the behavior of a process in progress when classical SPC is implemented, our control charts
rely on inputs of estimates that express the desired tolerance on the activity durations or
allows calibration to historical data at the activity level. Consequently, these inputs are
used to express an ex ante desired state of outcome on activity durations rather than an
observed ex post outcome in the measured variable. For this reason, the activity distribu-
tion inputs can and probably will be different than the real activity distributions, and the
tolerance limits are set up to detect these differences and to receive timely warnings when
real activity durations exceed the desired thresholds. Unlike classical applications of SPC
to project control, which aim at detecting deviations from the average on-going state of
the measurable variable in progress, the project control charts using tolerance limits aim at
detecting deviations from a predefined desired state in the variable defined before the start
of the progress.
• Dynamic project performance measurements: Given the ex ante approach of setting tolerance
for project control, the procedure requires a large sample of data points to predict the impact
of the acceptable activity tolerances on the tolerance limits set for EVM/ES. Therefore, the
data is collected using Monte-Carlo simulations using the desired tolerances on the activity
distributions and measuring their impact on the EVM/ES output metrics such as SV, SPI,
SV(t) and SPI(t). Consequently, these output metrics are used to construct tolerance limits
given the desired input parameters. This is in high contrast to the on-going data collection
method of observed process data used to construct control limits to detect deviations from
normal or average process behavior in a classical application of SPC.
• Control limits: Due to the fundamental difference between the control limits of SPC for
project control and the statistical tolerance limits based on the desired activity duration
tolerances of our project control approach, we have decided to use the words tolerance limit

5
 rather than control limit during the construction of the control charts. Consequently, the
control limits set on the EVM/ES output metrics SV, SPI, SV(t) and SPI(t) are used to define
the thresholds on the tolerances in project performance behavior given the desired tolerances
on the activity durations and hence act as warning signals to measure unacceptable project
progress behavior compared to the baseline schedule. This is exactly what an EVM/ES
system is trying to accomplish and this view is consistent with the definition of a project
control system [9]. Project based performance metrics along the project progress are used to
act as an action threshold to detect underlying problems that are responsible for the deviation
between planned outcome and project progress outcome. The tolerance limits used in this
paper define these action thresholds based on a statistical analysis of the simulated data
resulting from a pre-specified allowable variation.

2.2. A new statistical project control approach

Our statistical project control approach differs from those described in the literature overview by
setting tolerance limits. These tolerance limits are calculated by means of simulation and this
allows us to redefine and adapt the process control principles regularly found in manufacturing
and services to our requirements. Process control is generally an ongoing process where little
in-depth information is known for the process, and a state of statistical control is inferred from
the characteristics of the recorded process variables. Our approach, shares the use of control
charts, but is fundamentally and functionally different, as outlined in the last paragraph of section
2.1.
Practically, these differences brings about two main aspects:
• The statistical control procedure for a project will be finite in time. This follows directly
from the definition of a project as a one-time endeavour with a clearly defined start and
finish. The control chart will therefore only allow a predefined number of measurements at
distinct time-instances.
• Simulation allows us to apply a state of control, such as the user deems appropriate, and
characterize the resulting EVM/ES project progress variables. Historical data can play a
role in the input modelling of the simulation, but is no longer directly required to calculate
the control limits. Historical records of EVM/ES may be used to evaluate the congruence
of the simulation output with real-life executions.
The main research interest of this article is to develop and test a new project control approach
for schedule control. Since schedule control can be expressed in the deviations between the actual
durations and the planned durations of the activities of the project, these will be the primary
focus of our simulation model. The simulation model will add uncertainty to the durations of
the activities, which will result in fictitious project executions that deviate from the baseline. An
effective statistical project control chart is required to differentiate between the variation that is
acceptable for an activity and variation that is unacceptable. If an activity has a deviation between
its real duration and its planned durations that is unacceptably large, we wish to detect this on a
project control chart using EVM/ES. If all the activity durations lie within pre-defined acceptable
limits, we will consider the project to be executed as planned, and the control charts should not
generate any signal.
Our approach allows the user to set tolerances on the activity level, which will result in tolerance
limits for project progress variables to be used at the project level in control charts. The tolerance
limits for the project progress variables will be calculated in what is referred to as the static phase
of our procedure.
When actual project progress is recorded during the execution of a project, the observations can
be plotted against the tolerance limits on the project control charts. This phase is referred to as
the dynamic phase. In section 5, we will quantify the performance of the tolerance limits in the
control charts using a second large Monte Carlo simulation.

6
2.3. Project control charts

2.3.1. The static phase

Determine the as planned situation. If the project in progress exhibits an abnormal behavior
compared to a predefined situation, the project manager must be able to timely detect this.
Consequently, the normal project progress situation needs to be defined by the project manager in
order to construct control charts that distinguish between as planned and not as planned situations.
Therefore, it is assumed that the project baseline schedule is a forecast of what might happen in the
future during project progress and determines the as planned situation, subject to some acceptable
variation on the activity durations and possible project-wide effects of uncertain events. This is
exactly the starting point of dynamic scheduling, which proposes that the project baseline schedule
needs to be used to measure the project progress. Both positive and negative deviations within
a certain range are assumed to be inherent to any project and are considered as normal while
abnormal deviations exceeding certain thresholds should trigger further investigations.
This variation on the activity durations, and additional uncertain events, obviously have an impact
on the project performance which might differ from the initially expected project performance, as
stipulated in the project baseline schedule. This as planned project performance deviation will be
measured by Monte Carlo simulation runs.

Monte Carlo simulations. Monte Carlo simulations in project management and control have been
used widely in literature. Since its introduction [26] to analyze project networks, the methodology
has been used for a range of project management research projects and applications. In [27],
a classified bibliography of project risk management research is given where simulation plays a
central role. Ever since, Monte Carlo simulation studies have been a respected methodology in
project management and scheduling. The simulation studies in this paper are based on the project
control simulation studies described in [28, 29]. The methodology used has been initially developed
by Vanhoucke and Vandevoorde [24] and is formalized in a dynamic scheduling simulation tool,
P2 Engine [30].
In the current paper, the simulation runs will be used to simulate fictitious project progress in order
to obtain periodic project performance data using the well-known project performance indicators
obtained by the EVM/ES key metrics. Consequently, fictitious project progress is simulated for
a number nrs of Monte Carlo runs. Variation in the activity durations will have an impact on
the project performance over time and therefore, the variable Xsp will be used to denote the
performance indicator value from simulation run s (s = 1, . . . , nrs) measured at a certain moment
in time between the start and the finish of the simulated project progress when it is p percent
completed (p = (1, . . . , P )×∆PC). A total of P observations are taken with increment ∆PC.
The following performance indicators are periodically measured during the simulation runs
SV Schedule variance (SV = EV - PV)
SPI Schedule performance index (SPI = EV / PV)
SV(t) Schedule variance using earned schedule (SV(t) = ES - AD)
SPI(t) Schedule performance index using earned schedule (SPI(t) = ES / AD)

with
PV Planned value
EV Earned value
AC Actual cost
ES Earned schedule
AD Actual duration
BAC Budget at completion
PC Percentage project completion (PC = EV / BAC) (0% ≤ PC ≤ 100%)

7
 P Number of observations taken, with an increment ∆PC
For the sake of completeness, we even include SV and SV(t) in our research. Previous EVM/ES
studies using SPC often neglected these performance metrics and thus very few empirical evidence
for their distributions is known. Since we do not assume any distribution for the control measures
a priori, we can incorporate them into our research. Note that, we do not incorporate the cost
control metrics cost variance (CV = EV - AC) and cost performance index (CPI = EV / AC). In
our simulation model, cost is always strongly correlated with the time-performance of the activity,
which might not necessarily be the case in real-life. Since our focus is on controlling the time
performance of a project, we restrict the analysis to those metrics specifically designed to do
so.

Add tolerance limits. The simulated data is used to construct control limits for the project control
variables Xp =(SV, SV(t), SPI, SPI(t)) at a review period p for two distinct types of control
charts. The proposed X chart and the R chart resemble the Shewhart individuals (indX) and
moving range (mR) control charts in their ability to inspect individual observations [31]. Control
limits for the Shewhart charts are constructed based on the moving average over the range (i.e.
the difference between an observation xi and its predecessor xi−1 ), calculated using the previous
m − 1 observations.:
m

= 1 
MR |xi − xi−1 | (1)
m − 1 i=2

This concept of moving range is not adopted here to construct tolerance limits for project control,
since it implies that observations are taken for a consistently operating process. On the contrary, we
assume that our nrs simulation runs provide sufficient observations to reconstruct the distribution
function for the variable Xp at a review period p ∈ (1, . . . , P ) × ∆PC. Note that we do not assume
two random variables Xj and Xk (where j, k ∈ (1, . . . , P ) × ∆PC) to be independent, but we will
construct tolerance limits for each of them independently. For any two adjacent review periods,
we will look at the range Rj = |Xj − Xj−1 | (j ∈ (1, . . . , P ) × ∆PC) and treat this variable as a
new project performance variable.
• The X chart is used to monitor individual observations for the project performance variables
Xp at a review period p. Tolerance limits can be calculated from the empirical distribution
function F̂Xp (t).
nrs
1 
F̂Xp (t) = 1{Xsp ≤ t} (2)
nrs s=1

derived from a sample of independent identically distributed (i.i.d.) variables (X1p , . . . ,

Xnrs p ) taken at p. 1{A} represents the indicator function of an event A:

1 if A is TRUE
1{A} = (3)
0 if A is FALSE

By the strong law of large numbers, we know that the emperical distribution function (2)
asymptotically converges to the real cumulative distribution function F(t) for all possible
values of t.
The lower and upper statistical tolerance limits at a review period p for a level α (LT Lα Xp
and U T LαXp ) are then calculated as the α
th
and (1 − α)th quantile of the distribution F(t).
The quantiles can be calculated from the inverse of the empirical distribution function or
by estimation of the sample quantiles. Let Q̂(α)p for 0 < α < 1 denote the sample quan-
tile based on a a set of independent observations {X1p , X2p , . . . , Xnrs p } from the distribu-
tion F at review period p. If the order statistics of {X1p , X2p , . . . , Xnrs p } are denoted as

8
 {X(1)p , X(2)p , . . . , X(nrs)p }, then Q̂(α)p can be estimated as:

Q̂(α)p =(1 − γ)X(j)p + γX(j+1)p

j−α+1 j−2+α
where ≤α< (4)
nrs nrs
with γ = αnrs − α − j + 1 and j = αnrs − α + 1. u is used to represent the largest
integer smaller than u. We refer to [32] for a detailed comparison of different functions to
calculate continuous sample quantiles.
The lower and upper statistical tolerance limits at a review period p for a level α for the X
chart are:
LT Lα
Xp = Q̂(α)p
∀p ∈ (1, . . . , P ) × ∆PC (5)
U T Lα
Xp = Q̂(1 − α)p

The value for the level α can be chosen subsequently with respect to a total level of control
over al P review periods of the project. We will illustrate this in section 5.1.
• The R chart monitors the range R (or difference) between two adjacent observations. This
results in tolerance limits for a vector of observations which is in length one less than those
for the X chart. The empirical distribution function F̂Rp (t) for R at a period p can be found
using nrs simulation runs.
nrs
1 
F̂Rp (t) = 1{Rsp ≤ t} (6)
nrs s=1

with:

Rsp = |Xsp − Xsp−1 | ∀p ∈ (2, . . . , P ) × ∆PC

and the tolerance limits for a level α are:

LT Lα
Rp = Q̂(α)p
∀p ∈ (2, . . . , P ) × ∆PC (7)
U T Lα
Rp = Q̂(1 − α)p

where Q̂(α)p is calculated with the order statistics {R(1)p , R(2)p , . . . , R(nrs)p }. The intuitive
explanation for the use of the range chart is that the difference between two measurements
of project performance relates to the instantaneous change of the performance. Since a
project consists of activities that begin and end at discrete time instances, the transformed
variable Rp should better reflect the performance of activities that are being executed at
that time instance and the project control process is less cluttered by the performance of
past activities.

2.3.2. The dynamic phase

A project needs to be monitored by periodically measuring its performance. Each periodic project
performance observation can be plotted on the control chart. Yp is used to refer to the performance
indicator value during project progress at an instance p. It is different from Xsp in the sense that
it is not a priori assumed to be drawn from the distribution of Xsp . Rather, the calculated
tolerance limits for the control charts are used to asses whether a value Yp is likely to have been
produced under the null hypothesis (characterized by the empirical distribution function of Xp in
the simulation of the static phase).
Classical SPC applications often reject the notion that control charts are a form of hypothesis
testing. However, using tolerance limits for our statistical project control approach, we have taken

9
the liberty to present our control charts in such a way. Control charts and hypothesis tests can
be regarded as alternatives and the analogy is as presented below:
The null and alternative hypotheses of the project schedule control procedure proposed in this
paper are
H0 : The project is executed as planned
Ha : The project is executed not as planned

This hypothesis is tested using periodic control charts, by checking whether the current project
performance Yp lies within the control limits or not, as follows

Accept H0 : LT Lα α
Xp ≤ Yp ≤ U T LXp

Reject H0 : (Yp < LT Lα α

Xp ) ∨ (Yp > U T LXp )

for the X chart and

Accept H0 : LT Lα α
Rp ≤ |Yp − Yp−1 | ≤ U T LRp

Reject H0 : (|Yp − Yp−1 | < LT Lα α

Rp ) ∨ (|Yp − Yp−1 | > U T LRp )

for the R chart.

3. Illustrative example

This section illustraties the calculations to construct statistical project control charts based on
a fictitious project network example. Figure 2 displays the fictitious illustrative project network
from [12] with 10 non-dummy activities. The numbers above each node are used to display the
predefined baseline duration (in days) while the number below each node denotes the baseline
total activity cost.

4 4 8
2 5 9

80
80
8
0 9 3 0
1 3 10 12

36
6
7

0
0
1 5 1 8

21
4 6 7 3

25
20
120
11

60

Figure 2: An illustrative project network (Source: [12])

Table 2 displays five simulation runs (nrs = 5) where the activity durations have been randomly
chosen from a uniform distribution with a maximum relative deviation from the baseline duration
equal to 30%. The first row of table 2 shows the limits between which the fictitious activity
durations can be chosen.
Table 3 displays the earned value metric EV along the project duration (in days). It is assumed
that all activities follow a linear earned value accrue and, consequently, variation in an activity
duration has a linear impact on the actual activity cost.

10
 Table 2: Normal activity variation (in days) from five simulation runs

Activity 2 3 4 5 6 7 8 9 10 11 RD
cv
c±δ 3-5 6-12 1-1 3-5 4-6 1-1 5-9 6-10 2-4 2-4
1 3 8 1 5 4 1 9 6 2 4 15
2 5 6 1 5 6 1 7 6 3 3 16
3 4 7 1 4 5 1 6 7 2 3 15
4 4 12 1 3 5 1 9 8 4 2 16
5 3 11 1 3 4 1 8 10 2 4 16

Table 3: The Earned Value (EV) and Planned Value (PV) along the project duration

Scenario 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
1 56 92 129 154 180 320 358 396 417 439 443 446 450 454 456
2 47 72 98 123 148 174 195 333 374 413 437 442 446 450 455 456
3 50 79 107 136 164 193 221 364 392 416 441 446 450 455 456
4 48 75 102 129 163 196 346 382 419 425 431 438 442 447 452 456
5 55 90 105 160 195 225 349 385 422 429 436 442 448 452 455 456
PV 49 77 105 133 161 189 333 380 408 434 440 446 450 454 455 456

In practice, EV measurements will be collected as imposed by contract specifications or at distinct

time instances. It is easily understood that such time-instances will not always coincide with the
discrete percentage complete points for which tolerance limits for control charts are calculated.
This is also not the case for the project runs in our simulation software. If data are collected
at fixed time-intervals, the measurements along the time-axis need to be transformed towards
measurements equally spaced along the percentage complete axis. The last row of table 3 displays
the planned value along the project duration. The transformation towards measurements equally
spaced along the percentage complete axis is done by means of interpolation.
Interpolation might introduce additional variance in the observations made in both the static phase
and the dynamic phase. However, this error-term can be kept within reasonable value if both the
increment in percentage complete ∆P C and the periodic intervals at which data is collected are
chosen sufficiently small. We illustrate this procedure with the SPI data obtained for percentage
completion value of p = {10, 20, 30, 40, . . . , 90} for the first row of table 2. The SPI measurements
for this fictitious project example are obtained from the data recorded for each day of table 3 by
dividing the EV of the first row by the PV, as is illustrated in table 4. In order to find the SPI
value for this fictitious project execution at e.g. 50% of the project complete, we need to look for
the data we collected that have PC values lying the closest to 50%: in this case our measured
SPI after tA = 5 and tB = 6 days. To approximate SPI a simple calculation through linear
interpolation shows that:
SP ItB − SP ItA
SP Ip=50% = SP ItA + (50 − P CtA ) (8)
P C tB − P C tA
1.69 − 1.12
= 1.12 + (50 − 39.5) (9)
70.2 − 39.5
= 1.31 (10)

These calculations result in the SPI measurements found in table 5.

11
 Table 4: The SPI along the first fictitious project execution’s duration
Time 1 2 3 4 5 6 7 8
EV 56 92 129 154 180 320 358 396
PC 12.3 20.2 28.3 33.8 39.5 70.2 78.5 86.8
PV 49 77 105 133 161 189 333 380
SPI 1.14 1.19 1.23 1.16 1.12 1.69 1.08 1.04

Time 9 10 11 12 13 14 15 16
EV 417 439 443 446 450 454 456 456
PC 91.4 96.3 97.1 97.8 98.7 99.6 100 100
PV 408 434 440 446 450 454 455 456
SPI 1.02 1.01 1.01 1 1 1 1 1

Table 5: The SPI for PC intervals after interpolation

PC 10% 20% 30% 40% 50% 60% 70% 80% 90%
1 1.11 1.19 1.21 1.13 1.31 1.5 1.69 1.07 1.03
2 0.96 0.93 0.92 0.79 0.66 0.76 0.85 0.91 0.95
3 1.02 1.03 1.02 1.02 0.67 0.77 0.87 0.96 0.96
4 0.98 0.97 0.98 1.03 1.04 1.04 1.04 1.02 1.03
5 1.1 1.16 1.12 1.21 1.19 1.14 1.08 1.03 1.02
Xp 1.03 1.05 1.05 1.03 0.98 1.04 1.11 1.00 1.00
LT L10%
Xp 0.97 0.95 0.94 0.88 0.66 0.76 0.86 0.93 0.95
U T L10%
Xp 1.11 1.18 1.17 1.18 1.26 1.36 1.45 1.05 1.03

12
3.1. X chart

To calculate the tolerance limits for the X chart, we need to find the first and last α quantiles for
Xp at all P = 9 instances. This illustrative example only includes the SPI as a control variable.
For more details regarding the calculations of the other earned value metric (SV) and the earned
schedule metrics (SPI(t) and SV(t)), we refer the reader to the detailed work by Vanhoucke [12]
and the original work on earned schedule by Lipke [33].
Figure 3 illustrates the tolerance limits obtained for an X chart for SPI. The central line (CL)
is added for illustrative purposes and is the mean SPI along the project (X̄p ), the outer lines
represent the upper and lower tolerance limits (LT Lα α
Xp , U T LXp , for α = 10%) and the dots
illustrate how measurements made on a new run (with durations that exceed the predefined 30%
margins) are categorized as not as planned as soon as its value exceeds a tolerance limit.

Illustrative X chart Illustrative R chart

for project control for project control

0.4
1.4

U T L10%
Xp
0.3
1.2

∆SPI
SPI

0.2

U T L10%
1.0

X̄p Rp

LT L10%
Xp R̄p
0.1
0.8

LT L10%
Rp
0.0

10% 30% 50% 70% 90% 20% 40% 60% 80%

PC PC

Figure 3: X chart control limits for the illustrative Figure 4: R chart control limits for the illustrative
example example

3.2. R chart

The R chart investigates the difference between two consecutive measurements. To illustrate
the procedure, the SPI values from table 5 are transformed into the ∆SPI values of table 6.
Afterwards, the sample mean (Rp ) and the quantiles are calculated from table 6. Figure 4 depicts
the R control chart with the same run outlined as figure 3. This illustrates how both charts can
show completely different dynamic behaviour of the calculated control limits, but can be read and
processed in entirely the same way.

4. Research methodology and experimental test design

In this section, the data generation process to create a project benchmark set, the methodology
used, as well as the main research questions set during the computational experiment in order to
validate the project control approach are explained along the following subsections.

13
 Table 6: The ∆SPI for PC intervals after interpolation
PC 20% 30% 40% 50% 60% 70% 80% 90%
1 0.08 0.02 0.08 0.18 0.19 0.19 0.62 0.04
2 0.03 0.01 0.13 0.13 0.1 0.09 0.06 0.04
3 0.01 0.01 0 0.35 0.1 0.1 0.09 0
4 0.01 0.01 0.05 0.01 0 0 0.02 0.01
5 0.06 0.04 0.09 0.02 0.05 0.06 0.05 0.01
Rp 0.04 0.02 0.07 0.14 0.09 0.09 0.17 0.02
LT L10%
RP 0.01 0.01 0.02 0.01 0.02 0.02 0.03 0
U T L10%
RP 0.07 0.03 0.11 0.28 0.15 0.15 0.41 0.04

4.1. Project data

The X and R project control charts are tested on a set of 900 fictitious projects generated by a
project generator RanGen [34, 35] that has been used previously in project scheduling and control
studies. Vanhoucke [28] uses this dataset to connect network topology information to the optimal
control procedure found to monitor the time performance of a project.
The 900 projects are chosen in accordance to the serial/parallel (SP) indicator originally proposed
by Vanhoucke et al. [36] and used in previous project control simulation studies [24, 28, 29]. The
SP indicator measures how close a project matches a complete serial or parallel project and is based
on the maximal progressive level concept of Elmaghraby [37]. If m is the maximal progressive
level, it is defined as the maximum number of activities lying on a single path in the project.
If n is the number of activities in the project, it is clear that for a serial project m = n. The
SP = (m − 1)/(n − 1) then equals 1 for a serial project. If the largest path in a project consists
of only 1 activity, the project network is said to be parallel and SP = 0. The project dataset
consists of 100 projects for 9 intermediate SP values between 0 and 1 (SP = {0.1, 0.2, . . . , 0.9}).
If conclusions are formulated regarding a mean value over all 900 projects, it is conjectured that
they are general enough to extrapolate over all possible project networks. In the results section 5,
we did not find a significant influence of such serial/parallel network structure on the performance
of our control chart to monitor the schedule performance of a project. This result shows that our
statistical project control approach provides a more robust solution than regular use of EVM/ES,
which was found to improve notably for more serial project networks [28].

4.2. Methodology

Monte Carlo simulations are used in the static phase of our statistical project control procedure
to construct the tolerance limits used in the X and R charts, as well as to emulate real project
executions in the dynamic phase where fictitious project progress will be classified using these
charts. A distinction is made between both phases in the form of the variation present at the
activity level.
In order to generate fictitious project executions in both the static and dynamic phase of the Monte
Carlo simulation, we produce activity durations under interactive and compound uncertainties
[38]. Therefore, two sources of uncertainty are applied. In section 4.2.1, variation is modelled as
probability distributions that are applied to the duration of the project’s activities. This variation
will result in fictitious project executions that are either as planned or not as planned. In section
4.2.2, we discuss the potential dependencies between a project’s activities and how project progress
can be affected by uncertain events.
For the construction of the tolerance limits we will allow only variation that is acceptable at the
activity level. The projects are executed as planned, which represents the desired state of schedule
control in the presence of uncertain project-wide events. In the dynamic phase, fictitious project

14
progress is generated for which it is possible that the durations of the activities exhibit unacceptable
variation. These fictitious project executions are then used to quantify the performance of the
proposed control charts in their ability to accurately categorize whether the plan is still adequate
or not. Consequently, the project control charts are used to test whether the fictitious project
executions are being executed as planned or not as planned.

4.2.1. Project control input modelling

Since the occurrence of a project management simulation model in [39], the question has been asked
how to accurately represent the uncertainty present in the duration of activities in a real-life project
[40]. Since the PERT-beta distributions have been found to fall short, different authors have
suggested alternatives. Kuhl et al. [41] proposed the generalized beta distribution [28, 29]. Trietsch
et al. [42] theoretically support a claim to use the Parkinson distribution, while the lognormal
distribution [43], a mixture of beta and uniform distributions [44] and the doubly truncated normal
distribution [45] have also been applied to study stochastic activity networks.
For the purpose of our research we have chosen to use a uniform distribution. However, the
presented methodology can be reproduced with any distributional input that is able to accurately
represent the stochastic nature of project activity durations. Moreover, we strongly advise a
calibration to historical records for the input modelling, if our methodology would be applied
in a practical setting. The uniform distribution reflects an absence of interest in the precise
distributional characteristics of the project activities by the project manager. In this view on
schedule control, the only interest is in the relative deviation of a duration from its baseline
estimate. Most of the simulation experiments will incorporate a linear association variable (see
section 4.2.2) or have not as planned variation added to them. Therefore, the resulting distribution
for di /dˆi , where dˆi represents the baseline estimate for the duration of activity i, will most likely
not be uniform. Section 5 has a paragraph allocated to the validation of our simulation model
using the recently published lognormal core theory [42]. Moreover, a reasonable fit to the proposed
generalized beta distribution [41] was also found during the data analysis process. We give more
details concerning the parameters used to simulate activity durations in the remainder of this
section 4.

As planned project progress. Variation is added to the baseline duration dˆi of an activity i in
the project by a uniform distribution. We define the as planned project schedule executions
in the static phase using the lightgrey probability distribution functions presented in figure 5,
characterised by a choice for the maximum allowed deviation from the baseline estimate duration
δap . This distribution is presented relative to the baseline duration. The Monte Carlo samples
that are drawn from this distribution therefore need to be multiplied with the baseline duration
dˆi , in order to model the duration di of an activity i in a fictitious execution of the project.
The maximal relative deviation δap from the baseline estimate duration represents the subjective
estimate of the variation on the baseline schedule that is regarded as planned.

Detect both opportunities and dangers. In the dynamic phase, project progress situations are
simulated for which the baseline schedule might no longer be adequate. In order to simulate this,
variation is added using a not as planned uniform probability distribution that has a standard
deviation larger than that of the as planned uniform distribution. Figure 5 shows two illustrative
examples of this not as planned uniform distribution, hatched in black, for a low and a high value
of %Anp . The variable %Anp expresses the percentage of the total number of activities N that
has a duration that is not confined within the as planned margin δap . %Anp is therefore proposed
to provide an intuitive basis for creating not as planned project progress situations. The larger
%Anp , the more the project progress can be regarded as going not as planned and warning signals
generated by a project control procedure becomes more desirable.

15
 Uniform probability
distribution functions

as planned
not as planned
1 − δnp0-
1 − δap
1
1 + δap

1 + δnp+

0 = 1 − δnp-
1 − δap
1
1 + δap

1 + δnp+
low %Anp high %Anp

Figure 5: Defining the uniform input model for activity durations

The as planned parameter δap and the not as planned parameter %Anp will be set in our experiment
according to the values specified in table 7. From the δap value, the extremes of the (lightgrey) as
planned uniform probability distribution are calculated directly (1 − δap and 1 + δap ). Indirectly,
the extremes of the (hatched in black) not as planned uniform distribution function (1 − δnp−
and 1 + δnp+ ) can be found using equations 11 and 12. We will discuss how these extremes are
calculated along the following lines.
Let us assume that the N activities in the project have identical probability distributions and that
they are sampled independently. The reader should consider the fact that we add dependencies
through linear association in section 4.2.2. Let pnp denote the probability of drawing a value
outside of the as planned margin. Consider this a success/failure or coin toss experiment where
a draw from outside the as planned margin represents a success. The probability of drawing
a number of successes Nnp larger than %Anp N  from a total of N is given by the binomial
distribution.
P[Nnp > %Anp N ] = 1 − P[Nnp ≤ %Anp N ]
%Anp N 
 N!
=1− pi (1 − pnp )N −i (11)
i=0
i! (N − i)! np

In order to calculate pnp from equation 11, we have to assign a value to its left-hand side. This
value represents the chance that a minimum of %Anp N  successes is drawn. We chose this value
as 95% in our experiments, in accordance with a probability of 5% that a fictitious project with
less then %Anp N  successes is simulated. For a simulation of 1, 000 runs, an average of 50 of
such fictitious projects can then be expected to be included in the output. From equation 11, we
proceeded with a numerical approximation to calculate pnp .
pnp = P[|D − 1| > δap ]
2δap
=1− (12)
δnp+ + δnp−
From pnp and δap , the extremes of the not as planned uniform distribution (represented using
δnp− and δnp+ ) can then be found from equation 12, where D denotes the random variable that
is drawn.

16
In order to derive a value for both δnp− and δnp+ , an assumption has to be made with respect
to one or both of these unknowns. We chose to implement a linear relation (δnp+ = 1.5 δnp− )
which expresses the tendency of project activities to be late rather than early. However, in the
cases where this relation resulted in a negative value for the lower extreme of the distribution
(1 − δnp− < 0), δnp− was fixed at 1 and the only unknown δnp+ was found from equation 12. The
factor 1.5 shown for the linear relation is arbitrary and different choices were also implemented
during our tests.

4.2.2. Project progress simulation modelling

The Monte Carlo simulation generates activity durations for all activities according to the input
models described in section 4.2.1. In accordance to recent literature we implemented two structures
based on the linear association concept [23]. Linear association allows us to model the dependencies
between activities and test how project progress is affected by this.
• Independence: The first structure can be regarded as a special case, since complete inde-
pendence is assumed between activities. Activity durations are sampled independently from
each other, but precedence constraints within the project assure that activities influence
one another, as expressed in real start times, and the project in its whole. Although this
assumption of independence has been shown to fall short in accurately modelling project
executions in real life [42, 46, 47], it has been used in both analytical [48, 49] and simulation
[28, 29, 37, 50, 51] studies to produce insights into stochastic activity networks.
• Linear Association: The second structure models dependencies between activity durations.
If users are required to specify correlations between activities, they face a challenging task
since there is no established methodology for estimating correlation coefficients. In order
to reject the independence assumption in our simulations, we employ the linear association
technique [23, 42], where a systemic error B provides objective information on correlations.
We say that N positive random variables Zj are linearly associated if Zj = BXj where {Xj }
is a set of n independent positive random variables and B is a positive random variable,
independent from {Xj } that can be regressed from historical data. We follow [23] and model
B as a lognormal random variable, since the lognormal distribution is found appropriate to
reflect both additive and multiplicative factors, ubiquitous in real-life project activities. The
lognormal random variable is characterized by its mean µB and its coefficient of variation
cvB .
Our model assumes that the earned value (EV) for a single activity follows a linear accrue, starting
from its actual start up to its budget at completion (BAC) when it is finished [12]. Planned value
(PV) follows this same linear accrue from the planned start up to the planned finish time of the
activity. EV and PV are calculated in P2 Engine at the project level and are compared to produce
SV, SPI, SV(t) and SPI(t). The calculations for the X and R control charts and the recorded
performances in section 5 are implemented in the statistical programming language R [52].

4.3. Overview of experiments

Table 7 gives an overview of the simulation experiments conducted for this research. A total of
400 different simulation experiments were run for each project network according to the inputs for
the parameters µB , cvB , δap and %Anp provided in table 7. These parameters define respectively
to what extent dependencies between activities are incorporated or not and the as planned and
not as planned variation . It should be noted that although the independence assumption is
retained for some scenarios (denoted as cvB = 0), a constant B is still used in these to alter
the mean (µB ∈ {0.8, 1, 1.5, 2}). The 400 experiments produce 1,000 runs for all 900 projects
(section 4.1), for which the baseline estimates and costs are assigned according to the first column
of table 7. Periodic EVM/ES reports are generated at 19 distinct moments in the project (p =

17
 Table 7: Overview of the computational experiments
Baseline Simulation model Settings
as planned not as planned
Estimates
nrs = 1, 000
Uniform
µB ∈ {0.8, 1, 1.5, 2}
∀i ∈ N : dˆi ∈[8;56]
Independence cvB = 0
δap ∈ {20%, .., 80%} δap ∈ {20%, .., 80%}
Fixed Cost
%Anp = 0 %Anp ∈ {20%, .., 80%}
Uniform
[0; 500]
nrs = 1, 000
µB ∈ {0.8, 1, 1.5, 2}
Variable Cost
Linear association cvB ∈ {0.3, 0.8, 2}
Uniform
δap ∈ {20%, .., 80%} δap ∈ {20%, .., 80%}
[700; 1500]
%Anp = 0 %Anp ∈ {20%, .., 80%}

{5%, 10%, 15%, 20%, . . . , 90%, 95%}) and interpolation is done as shown in section 3 with a time-
increment chosen sufficiently small. The durations of the activities were expressed in days, whereas
the time-increment was in the order of hours.
We will use these experiments to demonstrate the value of statistical project control for project
management in section 5 using the control efficiency measures defined in section 4.4. We have
tried to incorporate the best practices for project control modelling as provided by the current
literature, to produce results that are as generalisable to project management practice as possible.
However, we will expand on the limitations of our research, and its results, in the concluding
section 6.

4.4. Project control efficiency measures

In order to measure and compare the efficiency of X and R statistical project control charts we
propose three simulation-based performance assessment measures [53].

The detection performance. The detection performance equals the frequentist approximation that
a certain not as planned situation will result in a signal during the dynamic phase, given the
as planned data from the static phase. A signal is generated as in section 2.3.2 if the project
performance Yp or |Yp − Yp−1 | is not confined to the tolerance limits of respectively the X or R
chart at any review period p.
The detection performance is calculated as:
nrs
1 
1s {∃p|(Yp < LT LαXp ) ∨ (Yp > U T LαXp )} (13)
nrs s=1

for the X chart, where 1s {A} is the indicator function of an event A for a simulation run s.
Otherwise stated, we represent the occurrence of a signal for the X charts at review period p as
the event PX :
nrs
1 
1s {∃p|PX } (14)
nrs s=1

For the R chart, the detection performance is calculated analogously as :

nrs
1 
1s {∃p|PR } (15)
nrs s=1

18
where:

1s {∃p|PR } = 1s {∃p|(|Yp − Yp−1 | < LT LαRp ) ∨ (|Yp − Yp−1 | > U T LαRp )} (16)

The probability of overreactions. The probability of overreactions approximates the probability

that a signal arises even though all activities are executed as planned. This is calculated by 3-fold
cross-validation. The runs of the static phase (going as planned ) are divided into 3 equal groups
and the probability of overreactions is calculated 3 times. Once for each group where the two other
groups are used to calculate the tolerance limits. We confirm that deviations between groups never
exceeded 1%, and in the remainder we will present the average over all 3 groups. Consequently,
the probability of overreactions represents situations where project managers get a false impression
that the project is being executed not as planned, which acts as a trigger to detect the reasons for
this behavior, while there is no unacceptable variation present at the activity level. These false
signals should be avoided since they lead to time-consuming and unnecessary overreactions made
by the project manager and his/her team in charge of controlling the project performance. Each
time an not as planned indication is reported, the project control chart is said to be overreacting
since it generates a signal, while at activity level only acceptable variation is present.
The probability of overreactions for one group can be calculated using nrs/3 runs from the static
phase as (14) and (15) for respectively the X and the R chart.

Area under the curve. The probability of overreactions and the detection performance can both be
used as approximations of the effectiveness of the proposed project control procedures. However,
in the results section 5, the need for a more holistic representation of the effectiveness will be
expressed. We find this in the concept of the calculated area under the curve (AUC) for a receiver
operating characteristic curve (ROC). The ROC representation is widely used in classification
testing and machine learning [54] and can be plotted directly from the probability of overreactions
and the detection performance.
Riemann integration of the ROC curve provides us with a single unitless measure for the area
under the curve that captures the characteristics of both the probability of overreactions and the
detection performance at the same time. The AUC should be more than 0.5, which is the area
under the no-discrimination line. This is the line where the true positive rate equals the false
positive rate, and which can be regarded as the characteristic of a classification action that is
purely based on a randomized process such as a coin toss.

5. Computational tests and results

In this section, we present our findings from our large computational experiment. First, we
introduce how the concept of the tolerance level α can be used to set a desired degree of probability
of overreactions and how the detection performance is affected, in section 5.1. Combined, these
two output measures allow us to draw a ROC curve and to calculate the AUC.
Section 5.2 provides the validation for our project progress simulation model according to the
recently postulated lognormal core theory [23, 42]. The efficiency of the described X and R
statistical project control charts is explored in section 5.3 and is compared to a more pragmatic
model of traditional EVM/ES use and how it can be implemented for different control points
throughout the project in section 5.4

5.1. Close control versus overreactions - Setting a tolerance level

When the statistical project control charts are constructed, we use a parameter α as the tolerance
level (see section 2.3). This represents the choice for the first and last αth quantile of the empirical

19
 Setting tolerance level Setting tolerance level
for a X chart for a R chart
1.0

1.0
0.8

0.8
0.6

0.6
0.4

0.4
0.2

0.2
Probability of overreactions Probability of overreactions
Detection performance Detection performance
0.0

0.0
0% 2% 4% 6% 8% 10% 40% 0% 2% 4% 6% 8% 10% 40%

α α

Figure 6: Setting the tolerance level α for a X chart, Figure 7: Setting the tolerance level α for a R chart,
for a (δap ,%Aap )=(20%,20%) scenario under the in- for a (δap ,%Aap )=(20%,20%) scenario under the in-
dependence assumption (µB = 1, cvB = 0). dependence assumption (µB = 1, cvB = 0).

distribution function constructed in the static phase to form the tolerance limits of the X and R
chart.
Is is clear that α will affect both the probability of overreactions and the detection performance of
the X and R charts. Therefore, an appropriate choice for α should reflect both a project manager’s
aversion to risk and his/her willingness to invest effort in false alarms. Figures 6 and 7 illustrate
the effect α has on the probability of overreactions and the detection performance for one specific
scenario.
It is easily perceived that the real strength of a chart is in the relative improvement of the detection
performance over a certain increase in the probability of overreactions. Optimally, we prefer the
probability of overreactions as close to 0 as possible. This would mean that the project management
team does not need to invest time and effort unnecessarily in drilling down the WBS, to find the
variation at the activity level to be confined within the acceptable margins. Analogously, we prefer
the detection performance as close to 1 as possible. For a certain scenario, this would imply that
the not as planned situation gets detected somewhere throughout the lifetime of the project. This
preference is accurately translated into a calculated AUC that needs to be as close to 1 as possible.
The ROC curves and the corresponding AUC values for the X chart and R chart from figures 6
and 7 are depicted in figure 8.

5.2. Validation of the project simulation model

In order to validate the project progress simulation model (presented in section 4.2.2), we anal-
ysed the simulation output according to the lognormal core theory [23, 42]. Figures 9 and 10
depict boxplots for respectively the calculated mean and standard deviations of our simulation
output.
Figure 9 presents the mean of the simulated activity durations. This figure clearly shows how the
average activity duration in a project is affected by µB . If B is a random variable (cvB = 0), rather
than a constant (cvB = 0), both the spread and the average of the mean duration is increased.
Figure 10 shows the effect that the coefficient of variation cvB has on the standard deviation of the
simulated activity durations. Both the average and the spread of the standard deviation increase
with increasing cvB . This effect is noticeably larger for the not as planned situations.

20
 Area under the curve

1.0
0.8
Detection performance

0.6
0.4
0.2

AUC(X)=0.84
AUC(R)=0.93
0.0

0.0 0.2 0.4 0.6 0.8 1.0

Probability of overreactions

Figure 8: Area under the curve for the X and R chart, for a (δap ,%Aap )=(20%,20%) scenario under the independence
assumption (µB = 1, cvB = 0).

In conclusion, we can safely state that in order to produce project progress simulations with
variation in the orders of magnitude that is discussed by Trietsch et al. [42], with high variation
instances having a standard deviation larger than 2, and large deviations in the mean, linear
association via a lognormal variable B is preferred.

5.3. Project control efficiency

This part of the results section explores the efficiency of the proposed control charts using tolerance
limits on EVM/ES by means of the area under the curve measure introduced in section 4.4.

5.3.1. Area under the curve

The area under the curve representation in figure 11 provides a nuanced and holistic view of the
project control efficiency for the X and the R chart. In general, the figure shows that the X and
R chart perform better than that a randomized classification process would. All AUC values are
significantly larger than 0.5, the AUC value of the no-discrimination line. Overall the X chart
outperforms the R chart. This can be seen more clearly from an aggregated AUC calculated over
all simulation experiments. This will be presented in section 5.4
For all cells along the diagonal of the static/dynamic phase grid, the choice for cvB is equal in both
phases and we can conclude that the R chart slightly outperforms the X chart. In other words,
if an historical calibration precedes the simulations of the static phase and new projects can be
expected to comply with the historical data, then the tolerance levels of the R chart prove to be
the most efficient project control option. The cells below and above the diagonal represent the
performance of the X and the R chart under parameter uncertainty [55]. When cvB is expected
to underestimate the real coefficient of variation of the linear association variable B during the
dynamic phase (above the diagonal in the static/dynamic grid), then the X chart outperforms the
R chart, although both produce very high AUC values. If however, the coefficient of variation of
B is chosen too high (below the diagonal), the R chart will be more favourable for efficient project
control during the dynamic phase.

21
 Mean Standard deviation
activity duration activity duration

=2 =2
µB cv B
.5 .8
=1 =0
µB c vB
=1 .3
µB =0
c vB

0.8 cvB ≠ 0 =0 Not as planned

µB
= cvB = 0 cv B As planned

0 2 4 6 8 10 0 5 10 15

m s

Figure 9: Distribution of the resulting mean in the project Figure 10: Distribution of the resulting standard devia-
simulation model. tions in the project simulation model.

5.3.2. Influencing factors for the X and R chart AUC

We proceed with a discussion in more detail for the AUC shown in figure 11. First, we provide
an intuitive basis for the observed variation within each cell of the static/dynamic phase grid.
Secondly, we investigate the effect that a change in the mean or coefficient of variation of the
observed distribution of di /dˆi has on the AUC.

Intuitive scenario representation.

Without loss of generality, we will illustrate the intuitive basis on a single cell of the static/dynamic
grid. We assume that the project management team has made the right choice for cvB = 0.3 and
µB = 1 in the static phase from calibration to historical data recorded within the organisation.
Moreover, we also assume that the real-life execution, for which the statistical project control
charts are calculated, can be correctly modelled using the same random variable B. Figure 12
shows the expected AUC over a subset of not as planned scenarios.
In order to improve the readability of figure 12, we applied a color coding where the lightest
shades of grey correspond to the lowest values and the darker to the highest values. The depicted
scenarios correspond to as planned situations for δap ∈ {20%, 40%, 60%, 80%} and not as planned
scenarios with %Anp ∈ {20%,40%, 60%, 80%}.
The color-coded variation in figure 12 was found to be identical for both the X and the R chart,
so only that of the X chart is shown. Figure 12 largely confirms the intuition behind the scenario
representation. The higher the percentage of activities %Anp that is likely to have a duration that
is not confined to the as planned margin δap , the more likely this scenario will produce a clear
signal that can be interpreted using the statistical project control charts. For any %Anp ≥ 20%,
an increase in AUC can also be observed for increasing values of the as planned margin δap .
From equation 12, it can be seen that if pnp remains fixed (resulting from a fixed %Anp ), then
(δnp− + δnp+ ) needs to increase if δap increases. The extremes of the not as planned uniform
thereby diverge, which results in a larger standard deviation and mean of the distribution. In the
next section, it will be shown that increasing values for the mean and standard deviation lead to
increasing AUC values.

22
 Area under the curve

=2
0.99 0.98 0.95 0.91 0.85 0.81 0.78 0.81

cv B
Dynamic Phase

.8 0.99 0.98 0.93 0.92 0.83 0.85 0.80 0.87

=0
cv B
.3 0.99 0.98 0.93 0.94 0.85 0.90 0.85 0.91
=0
cv B
=0
0.99 0.99 0.92 0.95 0.86 0.91 0.87 0.92

c vB
=0 =0
.3
=0
.8 =2
cv B cv B cv B cv B
X R
Static Phase

Figure 11: Area under the curve for different static/dynamic phase combinations

Area under the curve

=2
cv B
Dynamic Phase

.8
=0
cv B
.3 0.93
=0
cv B
=0
cv B Area under the curve
=0 =0
.3
=0
.8 =2
cv B cv B cv B cv B
Legend
Static
tic Phase
1
0.8
0.9

0.6 0.8
%Anp

0.4 0.7

0.6
0.2
0.5

0.2 0.4 0.6 0.8

δap

Figure 12: Area under the curve for different not as planned situations

23
 Influence of a change in Influence of a change in
mean standard deviation
1.0

1.0
Area under the curve

Area under the curve

0.8

0.8
0.6

0.6
0.4

0.4
0.2

0.2
X X
R R
0.0

0.0
) ) ) ) ) ) ) ) ) ) ) ) ) ) ) )
0.2 0.4 0.6 0.8 8,1 1.2 1.4 1.6 .4,
-1 -0.6 -0.2 ,0.2 ,0.6 .6,1 ,1.4 ,1.8
[0, [0.2, [0.4, [0.6, [0. [1, [1.2, [1.4, [-1
, ,
[-1 [-0.6 [-0.
2 0.2
[ [0 [1 [1.4

∆m ∆s
∆s ∈ [-0.1,0.1] ∆m ∈ [-0.1,0.1]

Figure 13: The effect a change in the mean has on Figure 14: The effect a change in the standard de-
the AUC. viation has on the AUC.

Change in mean or standard deviation.

This paragraph explores the effect the observed underlying distribution of di /dˆi has on the area
under the curve. We investigate the impact of a change in mean and standard deviation of the
observed distribution of di /dˆi for the dynamic phase relative to the static phase (which is used
to calculate the statistical tolerance limits) for the X and R control charts. The relative change
in mean ∆m = mdynamic − mstatic is the difference between the calculated mean of the activity
durations in the dynamic phase (mdynamic ) and the mean of the activity durations in the static
phase (mstatic ). Similarly, ∆s = sdynamic − sstatic is defined as the difference between the standard
deviation of these two phases.
Figure 13 presents the effect a relative change in the mean of the underlying distribution has on
the efficiency of the project control process. The x-axis represents the different ranges of values
that are found for ∆m, while the change in the standard deviation is kept close to zero between
the dynamic phase and the static phase (∆s ≤ ±0.1). In figure 13, we observe an increasing trend
for the AUC for increasing values of ∆m up to 1.2. If the mean of the static phase is e.g. 1, then
increasing values for the AUC can be expected for a mean in the dynamic phase up to 2.2. Both
the X and the R chart exhibit this behaviour, while the R chart significantly outperforms the X
chart.
The effect of a change in the standard deviation ∆s of the underlying distribution is shown in
figure 14 for approximately equal mean values (∆m ≤ ±0.1). The x-axis represents the ranges of
values found for ∆s. If the standard deviation of the underlying distribution of di /dˆi decreases
between the static and the dynamic phase (∆s < 0), we observe decreasing values for the AUC
of the X control chart and increasing values for the AUC of the R control chart. Moreover, the R
chart is preferred in order to accurately detect not as planned situations if ∆s < 0. For positive
∆s values, the AUC of both the X chart and the R chart increases. In these situations, the X
chart performs better than the R chart.

5.4. Comparing statistical project control to traditional EVM/ES

In this section we revisit the discussion on the comparison of the use of project control charts
using statistical tolerance limits (STL) to the more traditional use of EVM/ES metrics and some
recent advances from literature for project schedule control. Although practical use of EVM/ES

24
is typically characterized by on the spot decision-making from practical experience, we implement
four techniques to compare our statistical control charts with.
• Rules-of-thumb (RoT): This project schedule control approach combines the use of
project control charts with rules-of-thumb. Instead of our statical tolerance limits for the X
chart, these charts employ either static, widening or narrowing tolerance limits for schedule
control purposes. The tolerance limits, represented by a + b p, are implemented as symmet-
rical lines around a percentage complete axis. Therefore, the median of each performance
metric Q̂(0.5)p calculated in the static phase, is subtracted for all percentage project com-
plete p.
– Static tolerance limits mimic the “Target Performance Chart” [56]. The slope a equals
0, for these tolerance limits
– Widening tolerance limits should be used in combination with SV and SV(t). These
variance-based performance metrics grow larger in absolute figures during project progress
and therefore, an increasing upper tolerance limit should be used. Since the lower tol-
erance limit is symmetrical, overall the tolerance limits can be said to be widening
towards the end of the project.
– Narrowing tolerance limits should be used in combination with SPI and SPI(t). Simple
simulations indicate that the variation for these index-based performance metrics de-
creases along the percentage complete axis. Consequently, narrowing tolerance limits
are preferred during project progress.
Without loss of generality, we assume that these control strategies represent a best-case
decision making process from practical experience, since they are calibrated to the static
phase simulation data, much as the X and R control charts. This calibration is done for
the slope a and the intercept b, which are optimized using a coarse search algorithm over a
large solution space, in order to result in probabilities of overreactions that are competitive
with those chosen for the X and R chart. In practical decision-making from experience, this
calibration is impossible and the real-life probabilities of overreactions can be expected to
be much higher. We therefore consider this RoT control strategy to be best-case EVM/ES
decision making without the use of statistical tolerance limits.
• EVM/ES on multiple control points: The use of EVM/ES control metrics applied at
different (and multiple) control points in the project schedule has recently been discussed by
Colin and Vanhoucke [57]. Inspired by the concepts of the critical chain/buffer management
(CC/BM) methodology, different control points throughout the project are suggested at
which EVM/ES control charts should be monitored. The two alternative approaches are
discussed in the original paper and can be briefly summarized as follows
– Feeding paths (FP): The concept of feeding paths is adopted from the CC/BM litera-
ture. EVM/ES schedule control metrics are calculated for those paths in the project
network where a buffer should be place according to CC/BM in order to mitigate risks
and to protect the critical path. The schedule control metrics are then monitored on a
control chart against rules-of-thumb tolerance levels.
– Sub-networks (SN): The concept of sub-network control points expands on the concept
of feeding paths. Schedule control metrics are calculated for the collection of paths that
enter the critical path at a given point. Again, these EVM/ES metrics are referenced
against tolerance limits on control charts. The tolerance limits are calculated using
sample quantiles, in a manner similar to that presented in section 2.3.
It should be noted that the differnece between the two techniques mainly lies in the number
of control points, and figure 15 only gives a summary overview of the main results to compare
the alternative project control methods.

25
 Project schedule control using EVM/ES
1.0
Area under the curve

0.9
0.8
0.7
0.6
0.5

STL RoT FP SN LP

X R c g g 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5
ati in in
Stidenrrow
W Na

Figure 15: Area under the curve comparison of the X and R charts (STL) with other schedule control approaches
using EVM/ES

• Longest path monitoring (LP): Inspired by the fact that the EVM/ES schedule control
metrics are fully reliable for serial networks but have a decreasing performance for more
parallel networks [28], the longest path control approach was proposed by Lipke [58]. In this
method, SPI(t) is referenced against a static tolerance limit on a single project control chart.
The SPI(t) is however not calculated for all activities of the project, but only for those that
lie on the longest path in the project network. This longest path is updated dynamically
during project progress with actual durations for those activities that have already been
finished and the baseline estimate durations for the activities that are not yet started.
Figure 15 depicts the calculated area under the curve as boxplots for the statistical project control
charts proposed in this paper and the alternative schedule control procedures outlined in this sec-
tion. This comparison is based on all the data obtained from the simulations in the static/dynamic
phase grid of figure 11. A detailed discussion on the merits and the distinct characteristics of the
procedures presented in this section falls outside of the scope of this paper. In short, we conclude
that the X chart outperforms the R chart, based on the aggregated AUC over all performed sim-
ulation experiments. Both these charts, using statistical tolerance limits should be preferred over
all other tested schedule control procedures for the following four reasons
• The RoT strategies display a significantly lower AUC. The narrowing tolerance limits and
the widening tolerance limits can however be considered as useful alternatives.
• The FP method is shown for a combination of maximally 10 feeding paths monitored during
project progress. The maximal number of feeding paths can however be much larger, even for
small projects. The x-axis in figure 15 under FP depicts the number of control points that are
used. The first control point is always the critical path, while the others are used to monitor
the performance of the feeding paths. When not all feeding paths are monitored, information
about certain activities is lost. This results in an increased probability of overreactions, which
affects the detection performance and consequently the calculated AUC. Figure 15 shows that
this produces a decreasing schedule control performance for an increasing number of included
feeding paths.
• The SN method shows the opposite behaviour. AUC increases to the optimal unity area
under the curve for 5 sub-network control points included in the schedule control procedure.

26
 When its performance is compared to the X chart, it shows to be a valuable alternative.
However, while the X control chart relies on only one control point, the SN approach needs
up to 5 control points to only slightly outperform the X chart. Moreover, the SN approach
also relies on the STL procedure which is the topic of this paper.
• The LP method [58] is found to be outperformed by both the X and the R chart. Using
our definition and measures for schedule control it is not able to produce similar results.
However, this method was developed foremost to produce project duration forecasts, rather
than to detect not as planned project progress.

6. Discussion and conclusions

In this paper, a project control system is presented which is inspired by the well-known statisti-
cal process control charts widely used to monitor manufacturing processes but is fundamentally
different in its use of the progress data to construct tolerance limits. The use of earned value
management performance metrics in a control chart, based on simulated random variation, allows
to discriminate between as planned and not as planned project performance variation. This pa-
per measures the performance of the two proposed charts, using a wide range of fundamentally
different project networks in a large computational experiment. The first chart (X chart) mimics
a Shewhart chart for process control with adaptations to fit the finite nature of a project. The
X chart uses the commonly known schedule performance metrics provided in an earned value
management system. The second chart (R chart) proposes an adjustment on these metrics to
focus on the instantaneous changes in project performance. The assumption is that by performing
differential calculations, the accumulated performance of past activities is filtered and only active
tasks are being monitored.
Overall the X chart shows to be the most promising to be used in practice to monitor deviations
from the baseline schedule. Altogether, both show very comparable performance. The results
of various computational experiments show the relevance and usefulness of this approach in a
project control setting, although all the results are subject to the specific definition of as planned
and not as planned project progress and the characteristics of the simulation model as applied
in this research. In order to assess the performance of the statistical tolerance limit approach
discussed in the paper, the X and R chart’s performances are compared to add-hoc rules-of-
thumb strategies that would typically be used in project control using EVM/ES and some recently
developed EVM/ES extensions. Overall, the X and R chart can defensibly be preferred over all
tested procedures
Obviously, the results obtained in this study should be put in the right perspective. The use
of simulated data often shows some weaknesses and the results should therefore be interpreted
with care. A first assumption in our simulation study lies in the dependence structure of activity
durations during sampling. While the assumption of independence was invalidated in previous re-
search [42, 59], it has nevertheless been used in earlier reports on project control [28, 29]. We also
expanded our project progress simulation model with the concept of linear association [23]. How-
ever, the input for our simulation experiments depends largely on subjective estimates for activity
durations, where in practice calibration to historical data should always be preferred.
A second restriction of our simulated project control study is the strict focus on time performance
of projects. Although EVM/ES takes both a time and cost focus during project performance
management, this paper puts a strong and restricted focus on the time dimension of a project in
progress in order to validate the use of schedule performance metrics in a statistical project control
approach. Activity costs have been assumed to be linearly dependent on the activity duration
time, which is restricted to the use of periodic costs. In the future, our model could be adjusted to
incorporate other situations. Furthermore, other relevant project performance dimensions, such
as quality control or project scope control could not be taken into account in this study. Future

27
research will build on this paper, stretching its scope to incorporate the restrictions and expanding
on the idea of using process control procedures for project control.
Finally, it should be mentioned that a detailed comparison between the statistical project control
approach of this paper and the alternative control approaches available in literature is not within
the scope of this paper. A comparison between the quality of methods such as Critical Path
Method (CPM, [1]), the Programme Evaluation and Review Technique (PERT, [60]), its novel
extension to PERT21 [23], the bottom-up approach using Schedule Risk Analysis [29] and the
Critical Chain Project Management (CCPM, [61, 62]) is therefore considered as an interesting
future research avenue to further improve and optimize the knowledge on project control.
This study should be relevant for both practitioners and academics for the following reasons.
Performance indicators from earned value management/earned schedule have been used widely
by practitioners in controlling projects. However, the lack of guidelines towards reasonable toler-
ance limits to discriminate between acceptable and unacceptable performance variation has often
classified this technique as secondary to the necessary skills of intuition, existing experience and
knowledge a project manager must have. The methodology and control approach proposed in
this paper quantifies the use of tolerance limits. This paper shows that the discriminative power
between as planned and not as planned variation in a project is much better when a customized
control chart is used rather than when manual and intuitive thresholds (represented through the
rules-of-thumb) are used. The on the spot decision-making during project control is thereby as-
sisted. Next to the practical relevance, we also believe that this topic will contribute to new
research challenges where extensions of this approach might lead to an increased discriminative
power and better project control.

References

[1] J. Kelley, M. Walker, Critical path planning and scheduling: An introduction, Mauchly As-
sociates, Ambler, PA, 1959.

[2] J. Kelley, Critical path planning and scheduling: Mathematical basis, Operations Research 9
(1961) 296–320.

[3] D. Malcolm, J. Roseboom, C. Clark, W. Fazar, Application of a technique for a research and
development program evaluation, Operations Research (1959) 646–669.
[4] S. Hartmann, D. Briskorn, A survey of variants and extensions of the resource-constrained
project scheduling problem, European Journal of Operational Research 207 (2010) 1–15.

[5] M. Madadi, H. Iranmanesh, A management oriented approach to reduce a project duration

and its risk (variability), European Journal of Operational Research 219 (3) (2012) 751–761.

[6] E. Uyttewaal, Dynamic Scheduling With Microsoft Office Project 2003: The book by and for
professionals, Co-published with International Institute for Learning, Inc., 2005.

[7] M. Vanhoucke, Project Management with Dynamic Scheduling: Baseline Scheduling, Risk
Analysis and Project Control, Vol. XVIII, Springer, 2012.

[8] S. Rozenes, G. Vitner, S. Spraggett, Project control: literature review, Project Management
Journal 37 (2006) 5–14.

[9] S. Rozenes, G. Vitner, S. Spraggett, MPCS: Multidimensional project control system, Inter-
national Journal of Project Management 22 (2004) 109–118.

[10] Q. Fleming, J. Koppelman, Earned value project management. 3rd Edition, Newtown Square,
PA: Project Management Institute, 2005.

28
[11] W. Lipke, O. Zwikael, K. Henderson, F. Anbari, Prediction of project outcome: The appli-
cation of statistical methods to earned value management and earned schedule performance
indexes, International Journal of Project Management 27 (2009) 400–407.

[12] M. Vanhoucke, Measuring Time - Improving Project Performance using Earned Value Man-
agement, Vol. 136 of International Series in Operations Research and Management Science,
Springer, 2010.

[13] D. Hulett, Schedule risk analysis simplified, Project Management Network 10 (1996) 23–30.

[14] Y. Fang, J. Zhang, Performance of control charts for autoregressive conditional heteroscedastic
processes, Journal of Applied Statistics 26 (6) (1999) 701–714.

[15] W. A. Shewhart, Economic control of quality of manufactured product, Vol. 509, ASQ Quality
Press, 1931.

[16] W. Lipke, A study of the normality of earned value management indicators, The Measurable
News 4 (2002) 1,6,7,12–14,16.

[17] W. Lipke, J. Vaughn, Statistical process control meets earned value, CrossTalk: The Journal
of Defense Software Engineering June (2000) 16–20,28–29.

[18] G. T. Bauch, C. A. Chung, A statistical project control tool for engineering managers, Project
Management Journal 32 (2001) 37–44.

[19] Q. Wang, N. Jiang, L. Gou, M. Che, R. Zhang, Practical experiences of cost/schedule measure
through earned value management and statistical process control, Lecture Notes in Computer
Science 3966 (2006) 348–354.
[20] S. S. Leu, Y. C. Lin, Project performance evaluation based on statistical process control
techniques, Journal of Construction Engineering and Management 134 (2008) 813–819.

[21] R. Aliverdi, L. Moslemi Naeni, A. Salehipour, Monitoring project duration and cost in a
construction project by applying statistical quality control charts, International Journal of
Project Management.

[22] E. Page, Continuous inspection schemes, Biometrika (1954) 100–115.

[23] D. Trietsch, K. R. Baker, Pert 21: Fitting PERT/CPM for use in the 21st century, Interna-
tional Journal of Project Management 30 (4) (2012) 490–502.

[24] M. Vanhoucke, S. Vandevoorde, A simulation and evaluation of earned value metrics to fore-
cast the project duration, Journal of the Operational Research Society 58 (2007) 1361–1374.

[25] G. Burrows, Statistical tolerance limits–what are they, Tech. rep., Knolls Atomic Power Lab.,
Schenectady, NY (1962).
[26] R. Van Slyke, Monte Carlo methods and the PERT problem, Operations Research 11 (1963)
839–860.

[27] T. Williams, A classified bibliography of recent research relating to project risk management,
European Journal of Operational Research 85 (1995) 18–38.

[28] M. Vanhoucke, Using activity sensitivity and network topology information to monitor project
time performance, Omega The International Journal of Management Science 38 (2010) 359–
370.

[29] M. Vanhoucke, On the dynamic use of project performance and schedule risk information
during project tracking, Omega The International Journal of Management Science 39 (2011)
416–426.

29
[30] M. Vanhoucke, Integrated Project Management and Control: First come the theory, then the
practice, Management for Professionals, Springer, 2014.
[31] D. C. Montgomery, W. Woodall, Research issues and and ideas in statistical process control,
Journal of Quality Technology 31 (4) (1999) 376–387.
[32] R. J. Hyndman, Y. Fan, Sample quantiles in statistical packages, The American Statistician
50 (4) (1996) 361–365.
[33] W. Lipke, Statistical methods applied to EVM ...the next frontier, The Measurable News
Winter.
[34] L. Tavares, J. Ferreira, J. Coelho, The risk of delay of a project in terms of the morphology
of its network, European Journal of Operational Research 119 (1999) 510–537.
[35] E. Demeulemeester, M. Vanhoucke, W. Herroelen, Rangen: A random network generator for
activity-on-the-node networks, Journal of Scheduling 6 (2003) 17–38.
[36] M. Vanhoucke, J. Coelho, D. Debels, B. Maenhout, L. Tavares, An evaluation of the adequacy
of project network generators with systematically sampled networks, European Journal of
Operational Research 187 (2008) 511–524.
[37] S. Elmaghraby, On criticality and sensitivity in activity networks, European Journal of Op-
erational Research 127 (2000) 220–238.
[38] S. Wang, G. Huang, An integrated approach for water resources decision making under inter-
active and compound uncertainties, Omega The International Journal of Management Science
44 (2014) 32–40.
[39] T. Williams, What are PERT estimates?, Journal of the Operational Research Society 46
(1995) 1498–1504.
[40] M. E. Kreye, Y. M. Goh, L. B. Newnes, P. Goodwin, Approaches to displaying information to
assist decisions under uncertainty, Omega The International Journal of Management Science
40 (6) (2012) 682–692.
[41] M. E. Kuhl, E. K. Lada, N. M. Steiger, M. A. Wagner, J. R. Wilson, Introduction to mod-
eling and generating probabilistic input processes for simulation, in: S. Henderson, B. Biller,
M. Hsieh, J. Shortle, J. Tew, R. Barton (Eds.), Proceedings of the 2007 Winter Simulation
Conference, New Jersey: Institute of Electrical and Electronics Engineers, 2007, pp. 63–76.
[42] D. Trietsch, L. Mazmanyan, L. Govergyan, K. R. Baker, Modeling activity times by the
Parkinson distribution with a lognormal core: Theory and validation, European Journal of
Operational Research (2012) 386–396.
[43] S. Mohan, M. Gopalakrishnan, H. Balasubramanian, A. Chandrashekar, A lognormal ap-
proximation of activity duration in pert using two time estimates, Journal of the Operational
Research Society 58 (6) (2007) 827–831.
[44] E. Hahn, Mixture densities for project management activity times: A robust approach to
PERT, European Journal of Operational Research 188 (2008) 450–459.
[45] T. Kotiah, N. D. Wallace, Another look at the pert assumptions, Management Science 20 (1)
(1973) 44–49.
[46] C. Ragsdale, The current state of network simulation in project management theory and
practice, Omega The International Journal of Management Science 17 (1) (1989) 21–25.
[47] R. Schonberger, Why projects are “always” late: A rationale based on manual simulation of
a PERT/CPM network, Interfaces 11 (1981) 65–70.

30
[48] D. L. Fisher, D. Saisi, W. M. Goldstein, Stochastic pert networks: Op diagrams, critical paths
and the project completion time, Computers & operations research 12 (5) (1985) 471–482.

[49] B. Dodin, M. Sirvanci, Stochastic networks and the extreme value distribution, Computers
& Operations Research 17 (4) (1990) 397–409.

[50] T. Williams, Practical use of distributions in network analysis, Journal of the Operational
Research Society (1992) 265–270.

[51] F. Acebes, J. Pajares, J. M. Galán, A. López-Paredes, A new approach for project control
under uncertainty. going back to the basics, International Journal of Project Management.

[52] R Core Team, R: A Language and Environment for Statistical Computing, R Foundation for
Statistical Computing, Vienna, Austria (2013).
URL http://www.R-project.org

[53] T. Ponsignon, L. Mönch, Simulation-based performance assessment of master planning ap-

proaches in semiconductor manufacturing, Omega The International Journal of Management
Science 46 (2014) 21–35.

[54] T. Hastie, R. Tibshirani, J. Friedman, T. Hastie, J. Friedman, R. Tibshirani, The elements

of statistical learning, Vol. 2, Springer, 2009.

[55] S. Lim, A joint optimal pricing and order quantity model under parameter uncertainty and
its practical implementation, Omega The International Journal of Management Science 41 (6)
(2013) 998–1007.

[56] F. Anbari, Earned value project management method and extensions., Project Management
Journal 34(4) (2003) 12–23.

[57] J. Colin, M. Vanhoucke, A comparison of the performance of various project control methods
using earned value management systems, Submitted to an international journal.

[58] W. Lipke, Speculations on project duration forecasting, The Measurable News 3 (2012) 3–7.
[59] T. Williams, The contribution of mathematical modelling to the practice of project manage-
ment, IMA Journal of Management Mathematics 14 (1) (2003) 3–30.
[60] W. Fazar, Program evaluation and review technique, The American Statistician 13 (1959) 10.

[61] E. Goldratt, Critical Chain, North River Press, Great Barrington, MA., 1997.

[62] R. C. Ash, P. H. Pittman, Towards holistic project scheduling using critical chain methodology
enhanced with pert buffering, International Journal of Project Organisation and Management
1 (2) (2008) 185–203.

31