Omega 49 (2014) 107–122

Contents lists available at ScienceDirect

Omega
journal homepage: www.elsevier.com/locate/omega

Setting tolerance limits for statistical project control using earned

value management$
Jeroen Colin a, Mario Vanhoucke a,b,c,n
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

art ic l e i nf o a b s t r a c t

Article history: Project control has been a research topic since decades that attracts both academics and practitioners.
Received 14 May 2013 Project control systems indicate the direction of change in preliminary planning variables compared
Accepted 12 June 2014 with actual performance. In case their current project performance deviates from the planned
This manuscript was processed by Associate
performance, a warning is indicated by the system in order to take corrective actions.
Editor Kolisch
Earned value management/earned schedule (EVM/ES) systems have played a central role in project
Available online 23 June 2014
control, and provide straightforward key performance metrics that measure the deviations between
Keywords: planned and actual performance in terms of time and cost. In this paper, a new statistical project control
Project management procedure sets tolerance limits to improve the discriminative power between progress situations that
Scheduling
are either statistically likely or less likely to occur under the project baseline schedule. In this research,
Risk
the tolerance limits are derived from subjective estimates for the activity durations of the project. Using
Simulation
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 variations that are 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.
& 2014 Elsevier Ltd. All rights reserved.

1. Introduction 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
Project management and control have been research topics since information should therefore be used concurrently, both before and
the development of project planning approaches such as the critical during the execution of the project. The need for the integration of
path method (CPM, [1,2]) and the Program Evaluation and Review the project baseline scheduling construction, the risk analysis and
Technique (PERT, [3]). The majority of the research endeavors management and the project control step has been recently referred
published in the academic literature focus on the construction of a to as dynamic scheduling [6,7].
project baseline schedule within the presence of limited resource In this paper, the construction of a baseline schedule is
constraints (see e.g. a recent survey written by Hartmann and assumed to be given, and the focus lies on the presentation of a
Briskorn [4]). Ever since, the use of a project baseline schedule has new project control system [8]. Rozenes et al. [9] have character-
been put into the right perspective, as it only acts as a roadmap for ized a project control system by stating that it should indicate the
the future project performance. If the project runs into trouble, the direction of change in preliminary planning variables when
baseline schedule can be used to detect such problems and to trigger compared with actual performance. If there is an unacceptable
corrective actions. The construction of a project baseline schedule 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 unaccep-
$
This manuscript was processed by Associate Editor Kolisch. table gap and quantify how it can be specified.
n
Corresponding author at: Faculty of Economics and Business Administration,
Ghent University, Tweekerkenstraat 2, 9000 Gent, Belgium.
This study mainly focuses on the generation and the inter-
E-mail addresses: [email protected] (J. Colin), pretation of warning signals during project control and hence it
[email protected] (M. Vanhoucke). does not include results in the quality and input of corrective

http://dx.doi.org/10.1016/j.omega.2014.06.001
0305-0483/& 2014 Elsevier Ltd. All rights reserved.
108 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122

actions, or on the use of these new methods for contingency simulation to imitate a fictitious project progress environment
planning. are discussed. Section 6 draws general conclusions and highlights
A well-known and widely used project control system is future research avenues.
Earned Value Management (EVM) originally 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 2. Statistical project control
straightforward metrics to measure and evaluate the general
health of a project. These metrics serve as early warning signals 2.1. Literature overview
to timely detect project problems or to exploit project opportu-
nities. EVM has been investigated and used widely in project This literature review introduces the main concepts of statis-
management practice and has been brought into the attention of tical process control (SPC) as they are applied to monitor and
researchers by Fleming and Koppelman [10]. Recently, the research control processes in services and manufacturing. The application
on EVM has produced the earned schedule method (ES, [11]) and a of SPC techniques to earned value management is not new in the
large study on the integration of risk management and EVM/ES literature. We introduce a statistical project control procedure that
was conducted [12]. aims to overcome some of the issues with the previously pub-
EVM/ES calculations are based on the costs and durations for lished approaches, but which is not an implementation of SPC in
the activities in a project. Both in the planning phase and during the strict sense, although it reproduces some of its concepts and
the execution of the project, some detail of the individual activities nomenclature. The last paragraph of this literature review is
is lost as the individual performance is summed up into the dedicated to a summation of the differences between our statis-
performance metrics observed and controlled at the project level. tical project control procedure and the implications of applying
Earned value management is known to have some shortcomings SPC to project control in a strict sense.
when it comes to protecting the critical path in the schedule [10]. In standard SPC applications a state of control is identified with
The summation of activity earned value into the aggregate level for a process generating samples for which the subgroup averages are
the project does not incorporate schedule risk analysis (SRA, [13]) approximately normal under the central limit theory. Control
data, such as criticality. In spite of the complexity of this problem, charts such as the Shewhart, cumulative sum (CUSUM) and
EVM/ES use in practice is characterized by decision making from exponentially weighted moving average (EWMA) charts serve as
practical experience, rules-of-thumb and anecdotal evidence. In online procedures to monitor process stability, to detect assignable
the academic literature, it is reasoned that the application of a variation or to forecast process movements in industrial processes
statistical process control approach could benefit the EVM/ES [14]. A process is said to be in-control when only common cause
project control process. These previous attempts depend on the variation is present. This type of variation is characterized as
availability of historical records that may or may not exist and may coming from phenomena constantly active in a process, which
or may not be valid. We therefore also allow the use of subjective can be predicted probabilistically. In his original work on process
estimates. control, Shewhart [15] introduced the term chance cause.
In this paper, a new statistical project control approach is A process is said to be out-of-control if a second type of variation
presented. It is inspired by the practical format of statistical process is present known as assignable cause variation. Assignable cause
control charts and based on tolerance intervals produced by a variation arises when a new previously unanticipated phenom-
simulated sample. A controlled simulation experiment will be enon is present in the system and should cause a signal.
used to specifically produce two types of project progress situa- Fig. 1 illustrates how a control chart applied to project control
tions. First, project progress situations where the variation from might be interpreted. Periodic observations of a performance
plan is limited to some acceptable margin are used to produce measure for a fictitious project execution are outlined, along with
project control charts. Second, we test the discriminative power of an illustrative upper control limit (UCL) and lower control limit
these control charts when the variation from plan exceeds the pre- (LCL). For illustrative purposes, we will assume that the perfor-
defined acceptable margin. mance measure is of the type where a high value is good and a low
The outline of this paper is as follows. In Section 2, we present a value acts as a trigger for actions due to potential problems. The
literature review on the use of statistical control principles in frequently encountered schedule performance index (SPI, Section
project management and control. We will present two project 2.3.1) and schedule performance index using earned schedule (SPI
control charts, the X chart and the R chart. Section 3 illustrates (t), Section 2.3.1) measures in EVM/ES are of this type, although
these control charts on a fictitious project example. In Section 4, control charts of the opposite type have also been proposed where
the design of the computational test experiments is outlined based the natural logarithm of the reciprocal of SPI or SPI(t) is taken [16].
on various sources from the literature. In Section 5, test results In case the observations fall outside the project control limits, the
obtained by computational experiments using Monte Carlo charts report a signal which could trigger actions to bring the

Control chart signal

opportunity
Re-baselining?
UCL

Central line % Completed

LCL

Control chart signal

danger
Take corrective actions

Fig. 1. Illustrative control chart and out-of-control indications.

 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122 109

project back to plan or to exploit opportunities. More precisely, control by providing an objectively based and easily implemented
when an observation falls below the lower control limit, the real-time monitoring system. The evidence provided in the literature
performance has likely dropped below the pre-defined acceptable is based on empirical research of either post-hoc statistical analyses
margin on the baseline estimates. In this case, this signal should be on projects [19] or the implementation of control limits derived from
interpreted as a warning signal to the project manager to consider historical data [16,20]. The dependency on historical data does not
taking corrective actions. When the performance indicator exceeds pose an immediate problem for statistical project control approaches,
the upper control limit, it can be seen as a signal to exploit ideally the use of historical records is preferred over subjective
opportunities, in which case the current project schedule might be estimates for project control [23]. However, the concept of “similar-
re-baselined to incorporate this opportunity. Different SPC control ity” between projects is often vaguely defined. When historical EVM/
charts have been implemented for project control using earned ES data is directly applied, the project-specific EVM/ES dynamics
value management by [17,16,18–21]. A short overview of their should not be ignored. Vanhoucke and Vandevoorde [24] illustrated
respective approaches and findings is listed below. these project-specific dynamics with EVM/ES forecasting becoming
Lipke and Vaughn [17] apply a XmR chart (individuals and less reliable for projects with a network close to a parallel network
moving range chart) and calculate the control limits based on a structure. When the concepts of SPC are directly translated to
two-observation sample of earned value performance metrics. incorporate EVM/ES measurement, an assumption with respect to
They use the reciprocal measures of CPI (cost performance index) their distribution has to be made. In the previous literature this
and SPI (schedule performance index), as they have shown that prerequisite was fulfilled by transforming the data to appear nor-
these are normalized, such that many of the issues of applying SPC, mally distributed [17,19] or lognormally distributed [16,20] or
such as variability and homogeneity of the data, are avoided. appeared overlooked [21].
Bauch and Chung [18] modify the statistical process control charts In Section 2.3 we propose a different statistical project control
proposed by Shewhart to be used with earned value management procedure that will also include the appealing aspects of using
considering three major aspects. Firstly, they asses single observa- control charts. They are easy to set up, implement and interpret.
tions relative to historical project data from 20 similar projects. However, the imposed tolerance limits on the earned value
Secondly, as projects span a finite time horizon, which is typically metrics should be interpreted in the more technical sense [25].
different for each project, they normalize the different time The tolerance limits are acquired from the empirical distribution
lengths of each project into a consistent number of time periods, function (edf) of the EVM/ES metric under study at different points
chosen as 20 discrete measurements for the purpose of the paper. along the life-cycle of a project. This edf is obtained from a Monte
As a third modification, they adopted the CUSUM approach [22] to Carlo simulation model that allows variation to be added at the
incorporate the increasing nature of period-to-period measure- activity level of the project. This variation can be introduced as
ments of project performance. The typical y-axis of control charts subjective estimates of what deviations are statistically plausible
is transformed to plot cumulative project performance values. As a during the execution of the project or from calibration with
result, the central line and the upper and lower limits are non- historical in-company data.
stationary. Wang et al. [19] illustrate the use of SPC charts using Table 1 summarizes the main differences between our statis-
EVM/ES on a set of more than 30 software projects where tical project control procedure that uses tolerance limits and a
abnormal progress situations are effectively detected. Leu and classical SPC implementation for project control. The main aspects
Lin [20] combine XmR charts with EVM/ES performance metrics, are given along the following lines:
but implement the logarithm of the reciprocal of SPI and the cost
performance index (CPI). This was earlier proposed by Lipke [16].
Data input: A first fundamental difference lies on the data input
Control limits are calculated from historical data of 120 projects necessary to construct the control limits. While process data
performed by a management consulting company. Leu and Lin [20] needs to be collected at regular time intervals to monitor the
introduce the division of different cases based on historical factors behaviour of a process in progress when classical SPC is
influencing project performance and quantify the XmR chart implemented, our control charts rely on inputs of estimates
procedure for different trends observed in their data. Moreover, that express the desired tolerance on the activity durations or
they accurately describe how “the quick proliferation and com- allows calibration to historical data at the activity level. Con-
plexity of project performance data indicate the need for a well- sequently, these inputs are used to express an ex ante desired
organized project performance analysis process.” Aliverdi et al. state of outcome on activity durations rather than an observed
[21] suggest to transform CPI data when quality control charts are ex post outcome in the measured variable. For this reason, the
applied for cost control. SPC charts for both individual and moving activity distribution inputs can and probably will be different
range measurements were successfully implemented. They con- than the real activity distributions, and the tolerance limits are
clude that the integrated approach of EVM/ES and statistical set up to detect these differences and to receive timely
quality control charts can contribute to a better and more reliable warnings when real activity durations exceed the desired
project control process. thresholds. Unlike classical applications of SPC to project
Overall, previous research on the integration of EVM/ES with SPC control, which aim at detecting deviations from the average
techniques for projects has established that it improves project on-going state of the measurable variable in progress, the

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

Control procedure aspects Statistical process control for project control Statistical 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
Reference value Normal project progress (with normal variation) Baseline schedule (with acceptable tolerances)
110 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122

project control charts using tolerance limits aim at detecting The main research interest of this paper is to develop and test a
deviations from a predefined desired state in the variable new project control approach for schedule control. Since schedule
defined before the start of the progress. control can be expressed in the deviations between the actual

Dynamic project performance measurements: Given the ex ante durations and the planned durations of the activities of the
approach of setting tolerance for project control, the procedure project, these will be the primary focus of our simulation model.
requires a large sample of data points to predict the impact of The simulation model will add uncertainty to the durations of the
the acceptable activity tolerances on the tolerance limits set for activities, which will result in fictitious project executions that
EVM/ES. Therefore, the data is collected using Monte-Carlo deviate from the baseline. An effective statistical project control
simulations using the desired tolerances on the activity dis- chart is required to differentiate between the variation that is
tributions and measuring their impact on the EVM/ES output acceptable for an activity and the variation that is unacceptable.
metrics such as SV, SPI, SV(t) and SPI(t). Consequently, these If an activity has a deviation between its real duration and its
output metrics are used to construct tolerance limits given the planned durations that is unacceptably large, we wish to detect
desired input parameters. This is in high contrast to the on- this on a project control chart using EVM/ES. If all the activity
going data collection method of observed process data used to durations lie within pre-defined acceptable limits, we will consider
construct control limits to detect deviations from normal or the project to be executed as planned, and the control charts
average process behaviour in a classical application of SPC. should not generate any signal.

Control limits: Due to the fundamental difference between the Our approach allows the user to set tolerances on the activity
control limits of SPC for project control and the statistical level, which will result in tolerance limits for project progress
tolerance limits based on the desired activity duration toler- variables to be used at the project level in control charts. The
ances of our project control approach, we have decided to use tolerance limits for the project progress variables will be calcu-
the words tolerance limit rather than control limit during the lated in what is referred to as the static phase of our procedure.
construction of the control charts. Consequently, the control When actual project progress is recorded during the execution
limits that set on the EVM/ES output metrics SV, SPI, SV(t) and of a project, the observations can be plotted against the tolerance
SPI(t) are used to define the thresholds on the tolerances in limits on the project control charts. This phase is referred to as the
project performance behaviour given the desired tolerances on dynamic phase. In Section 5, we will quantify the performance of
the activity durations and hence act as warning signals to the tolerance limits in the control charts using a second large
measure unacceptable project progress behaviour compared to Monte Carlo simulation.
the baseline schedule. This is exactly what an EVM/ES system is
trying to accomplish and this view is consistent with the 2.3. Project control charts
definition of a project control system [9]. Project based perfor-
mance metrics along the project progress are used to act as an 2.3.1. The static phase
action threshold to detect underlying problems that are Determine the as planned situation: If the project in progress
responsible for the deviation between planned outcome and exhibits an abnormal behaviour compared to a predefined situa-
project progress outcome. The tolerance limits used in this tion, the project manager must be able to timely detect this.
paper define these action thresholds based on a statistical Consequently, the normal project progress situation needs to be
analysis of the simulated data resulting from a pre-specified defined by the project manager in order to construct control charts
allowable variation. 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
2.2. A new statistical project control approach progress and determines the as planned situation, subject to some
acceptable variation on the activity durations and possible project-
Our statistical project control approach differs from those wide effects of uncertain events. This is exactly the starting point
described in the literature overview by setting tolerance limits. of dynamic scheduling, which proposes that the project baseline
These tolerance limits are calculated by means of simulation and schedule needs to be used to measure the project progress. Both
this allows us to redefine and adapt the process control principles positive and negative deviations within a certain range are
regularly found in manufacturing and services to our require- assumed to be inherent to any project and are considered as
ments. Process control is generally an ongoing process where little normal while abnormal deviations exceeding certain thresholds
in-depth information is known for the process, and a state of should trigger further investigations.
statistical control is inferred from the characteristics of the This variation on the activity durations, and additional uncer-
recorded process variables. Our approach shares the use of control tain events, obviously has an impact on the project performance
charts, but is fundamentally and functionally different, as outlined which might differ from the initially expected project perfor-
in the last paragraph of Section 2.1. mance, as stipulated in the project baseline schedule. Thisas
Practically, these differences bring about two main aspects: planned project performance deviation will be measured by Monte
Carlo simulation runs.

The statistical control procedure for a project will be finite in Monte Carlo simulations: Monte Carlo simulations in project
time. This follows directly from the definition of a project as a management and control have been used widely in the literature.
one-time endeavour with a clearly defined start and finish. The Since its introduction [26] to analyze project networks, the metho-
control chart will therefore only allow a predefined number of dology has been used for a range of project management research
measurements at distinct time-instances. projects and applications. In [27], a classified bibliography of project

Simulation allows us to apply a state of control, such as the user risk management research is given where simulation plays a central
deems appropriate, and characterize the resulting EVM/ES role. Ever since, Monte Carlo simulation studies have been a
project progress variables. Historical data can play a role in respected methodology in project management and scheduling.
the input modelling of the simulation, but is no longer directly The simulation studies in this paper are based on the project control
required to calculate the control limits. Historical records of simulation studies described in [28,29]. The methodology used has
EVM/ES may be used to evaluate the congruence of the been initially developed by Vanhoucke and Vandevoorde [24] and is
simulation output with real-life executions. formalized in a dynamic scheduling simulation tool, P2 Engine [30].
 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122 111

In the current paper, the simulation runs will be used to Rj ¼ jX j X j  1 j (j A ð1; …; PÞ  ΔPCÞ and treat this variable as a new
simulate fictitious project progress in order to obtain periodic project performance variable:
project performance data using the well-known project perfor-
mance indicators obtained by the EVM/ES key metrics. Conse-
The X chart is used to monitor individual observations for the
quently, fictitious project progress is simulated for a number nrs of project performance variables Xp at a review period p. Toler-
Monte Carlo runs. The variation in the activity durations will have ance limits can be calculated from the empirical distribution
an impact on the project performance over time and therefore, the function F^ X p ðtÞ.
variable Xsp will be used to denote the performance indicator value 1 nrs  
from simulation run s (s ¼ 1; …; nrs) measured at a certain moment F^ X p ðtÞ ¼ ∑ 1 X sp rt ð2Þ
nrs s ¼ 1
in time between the start and the finish of the simulated project
progress when it is p percent completed (p ¼ ð1; …; PÞ  ΔPC). derived from a sample of independent identically distributed (i.
A total of P observations are taken with increment ΔPC. i.d.) variables ðX 1p , ..., X nrs p Þ taken at p. 1fAg represents the
The following performance indicators are periodically mea- indicator function of an event A:

sured during the simulation runs: 1 if A is TRUE
1fAg ¼ ð3Þ
0 if A is FALSE
SV schedule variance (SV¼EV  PV)
By the strong law of large numbers, we know that the empirical
SPI schedule performance index (SPI ¼EV/PV)
distribution function (2) asymptotically converges to the real
SV(t) schedule variance using earned schedule (SV(t)¼
cumulative distribution function FðtÞ for all possible values of t.
ES AD)
The lower and upper statistical tolerance limits at a review
SPI(t) schedule performance index using earned schedule (SPI
period p for a level α(LTLαX p and UTLαX p ) are then calculated as the
(t) ¼ES/AD)
αth and ð1  αÞth quantile of the distribution FðtÞ. The quantiles
with
can be calculated from the inverse of the empirical distribution
function or by the estimation of the sample quantiles. Let Q^ ðαÞp
PV planned value for 0 o α o1 denote the sample quantile based on a set of
EV earned value independent observations fX 1p ; X 2p ; …; X nrs p g from the distribu-
AC actual cost tion F at review period p. If the order statistics of
ES earned schedule fX 1p ; X 2p ; …; X nrs p g are denoted as fX ð1Þp ; X ð2Þp ; …; X ðnrsÞp g, then
AD actual duration Q^ ðαÞp can be estimated as
BAC budget at completion
PC percentage project completion (PC¼EV/BAC) (0% r Q^ ðαÞp ¼ ð1  γ ÞX ðjÞp þ γ X ðj þ 1Þp
PC r 100%) j α þ 1 j2þα
where rαo ð4Þ
P number of observations taken, with an increment ΔPC nrs nrs
For the sake of completeness, we even include SV and SV(t) in our with γ ¼ αnrs  α  j þ1 and j ¼ ⌊αnrs  α þ 1c. ⌊uc is used to
research. The previous EVM/ES studies using SPC often neglected represent the largest integer smaller than u. We refer to [32] for
these performance metrics and thus very few empirical evidence a detailed comparison of different functions to calculate con-
for their distributions is known. Since we do not assume any tinuous sample quantiles.
distribution for the control measures a priori, we can incorporate The lower and upper statistical tolerance limits at a review
them into our research. Note that, we do not incorporate the cost period p for a level α for the X chart are the following:
control metrics cost variance (CV ¼ EV AC) and the cost perfor-
mance index (CPI ¼EV/AC). In our simulation model, cost is always LTLαX p ¼ Q^ ðαÞp
8 p A ð1; …; PÞ  ΔPC ð5Þ
strongly correlated with the time-performance of the activity, UTLαX p ¼ Q^ ð1  αÞp
which might not necessarily be the case in real-life. Since our
focus is on controlling the time performance of a project, we The value for the level α can be chosen subsequently with
restrict the analysis to those metrics specifically designed to do so. respect to a total level of control over al P review periods of the
Add tolerance limits: The simulated data is used to construct control project. We will illustrate this in Section 5.1.
limits for the project control variables Xp ¼(SV, SV(t), SPI, SPI(t)) at a
The R chart monitors the range R (or difference) between two
review period p for two distinct types of control charts. The proposed adjacent observations. This results in tolerance limits for a
X chart and the R chart resemble the Shewhart individuals (indX) and vector of observations which is in length one less than those for
moving range (mR) control charts in their ability to inspect individual the X chart. The empirical distribution function F^ Rp ðtÞ for R at a
observations [31]. Control limits for the Shewhart charts are con- period p can be found using nrs simulation runs:
structed based on the moving average over the range (i.e. the
1 nrs  
difference between an observation xi and its predecessor xi  1 ), F^ Rp ðtÞ ¼ ∑ 1 Rsp r t ð6Þ
nrs s ¼ 1
calculated using the previous m  1 observations:
with
1 m
d¼
MR ∑ jx  xi  1 j ð1Þ Rsp ¼ jX sp  X sp  1 j; 8 p A ð2; …; PÞ  ΔPC
m1 i ¼ 2 i
and the tolerance limits for a level α are
This concept of moving range is not adopted here to construct
tolerance limits for project control, since it implies that observations
α
LTLRp ¼ Q^ ðαÞp
; 8 p A ð2; …; PÞ  ΔPC ð7Þ
are taken for a consistently operating process. On the contrary, we UTLαRp ¼ Q^ ð1  αÞp
assume that our nrs simulation runs provide sufficient observations to
reconstruct the distribution function for the variable Xp at a review where Q^ ðαÞp is calculated with the order statistics fRð1Þp ; Rð2Þp
period p A ð1; …; PÞ  ΔPC. Note that we do not assume two random ; :::; RðnrsÞp g. The intuitive explanation for the use of the range
variables Xj and Xk (where j; k A ð1; …; PÞ  ΔPC) to be independent, chart is that the difference between two measurements of
but we will construct tolerance limits for each of them independently. project performance relates to the instantaneous change of the
For any two adjacent review periods, we will look at the range performance. Since a project consists of activities that begin
112 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122

and end at discrete time instances, the transformed variable Rp This hypothesis is tested using periodic control charts, by checking
should better reflect the performance of activities that are whether the current project performance Yp lies within the control
being executed at that time instance and the project control limits or not, as follows:
process is less cluttered by the performance of past activities.
Accept H0 : LTLαX p r Y p r UTLαX p
Reject H0 : ðY p o LTLαX p Þ 3 ðY p 4UTLαX p Þ
2.3.2. The dynamic phase
A project needs to be monitored by periodically measuring its
for the X chart and
performance. Each periodic project performance observation can
be plotted on the control chart. Yp is used to refer to the Accept H0 : LTLαRp r jY p  Y p  1 j r UTLαRp
performance indicator value during project progress at an instance Reject H0 : ðjY p  Y p  1 j o LTLαRp Þ 3 ðjY p  Y p  1 j 4UTLαRp Þ
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 for the R chart.
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). 3. Illustrative example
Classical SPC applications often reject the notion that control
charts are a form of hypothesis testing. However, using tolerance This section illustrates the calculations to construct statistical
limits for our statistical project control approach, we have taken project control charts based on a fictitious project network
the liberty to present our control charts in such a way. Control example. Fig. 2 displays the fictitious illustrative project network
charts and hypothesis tests can be regarded as alternatives and the from [12] with 10 non-dummy activities. The numbers above each
analogy is as presented below: node are used to display the predefined baseline duration (in days)
The null and alternative hypotheses of the project schedule while the number below each node denotes the baseline total
control procedure proposed in this paper are activity cost.
H0 : The project is executed as planned Table 2 displays five simulation runs (nrs ¼5) where the activity
durations have been randomly chosen from a uniform distribution
Ha : The project is executed not as planned
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.
4 4 8 Table 3 displays the earned value metric EV along the project
2 5 9 duration (in days). It is assumed that all activities follow a linear
80 80 8 earned value accrue and, consequently, variation in an activity
0 9 3 0 duration has a linear impact on the actual activity cost. In practice,
1 3 10 12 EV measurements will be collected as imposed by contract
36 6
7 specifications or at distinct time instances. It is easily understood
0 0
1 5 1 8 that such time-instances will not always coincide with the discrete
21 percentage complete points for which tolerance limits for control
4 6 7 3 charts are calculated. This is also not the case for the project runs
25 20 120 in our simulation software. If data are collected at fixed time-
11
intervals, the measurements along the time-axis need to be
60 transformed towards measurements equally spaced along the
Fig. 2. An illustrative project network (Source: [12]). 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
Table 2 complete axis is done by means of interpolation.
Normal activity variation (in days) from five simulation runs. Interpolation might introduce additional variance in the obser-
vations made in both the static phase and the dynamic phase.
Activity 2 3 4 5 6 7 8 9 10 11 RD
c 7 δcv 3–5 6–12 1–1 3–5 4–6 1–1 5–9 6–10 2–4 2–4 However, this error-term can be kept within reasonable value if
both the increment in percentage complete ΔPC and the periodic
1 3 8 1 5 4 1 9 6 2 4 15 intervals at which data is collected are chosen sufficiently small.
2 5 6 1 5 6 1 7 6 3 3 16 We illustrate this procedure with the SPI data obtained for
3 4 7 1 4 5 1 6 7 2 3 15
percentage completion value of p ¼ f10; 20; 30; 40; …; 90g for the
4 4 12 1 3 5 1 9 8 4 2 16
5 3 11 1 3 4 1 8 10 2 4 16 first row of Table 2. The SPI measurements for this fictitious
project example are obtained from the data recorded for each

Table 3
The earned value (EV) and the 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
 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122 113

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

Fig. 3. X chart control limits for the illustrative example.

Table 5 Table 6
The SPI for PC intervals after interpolation. The ΔSPI for PC intervals after interpolation.

PC 10% 20% 30% 40% 50% 60% 70% 80% 90% PC 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 1 0.08 0.02 0.08 0.18 0.19 0.19 0.62 0.04
2 0.96 0.93 0.92 0.79 0.66 0.76 0.85 0.91 0.95 2 0.03 0.01 0.13 0.13 0.1 0.09 0.06 0.04
3 1.02 1.03 1.02 1.02 0.67 0.77 0.87 0.96 0.96 3 0.01 0.01 0 0.35 0.1 0.1 0.09 0
4 0.98 0.97 0.98 1.03 1.04 1.04 1.04 1.02 1.03 4 0.01 0.01 0.05 0.01 0 0 0.02 0.01
5 1.1 1.16 1.12 1.21 1.19 1.14 1.08 1.03 1.02 5 0.06 0.04 0.09 0.02 0.05 0.06 0.05 0.01

Xp 1.03 1.05 1.05 1.03 0.98 1.04 1.11 1.00 1.00 Rp 0.04 0.02 0.07 0.14 0.09 0.09 0.17 0.02
LTL10%
Xp
0.97 0.95 0.94 0.88 0.66 0.76 0.86 0.93 0.95 LTL10%
RP
0.01 0.01 0.02 0.01 0.02 0.02 0.03 0
UTL10% 1.11 1.18 1.17 1.18 1.26 1.36 1.45 1.05 1.03 UTL10% 0.07 0.03 0.11 0.28 0.15 0.15 0.41 0.04
Xp RP

durations that exceed the predefined 30% margins) are categorized

as not as planned as soon as its value exceeds a tolerance limit.
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 3.2. R chart
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 The R chart investigates the difference between two consecu-
to 50%: in this case our measured SPI after tA ¼5 and tB ¼ 6 days. To tive measurements. To illustrate the procedure, the SPI values from
approximate SPI a simple calculation through linear interpolation Table 5 are transformed into the ΔSPI values in Table 6. After-
shows that wards, the sample mean (R p ) and the quantiles are calculated from
Table 6. Fig. 4 depicts the R control chart with the same run
SPI t B SPI tA
SPI p ¼ 50% ¼ SPI tA þ ð50 PC tA Þ ð8Þ outlined as Fig. 3. This illustrates how both charts can show
PC t B PC tA
completely different dynamic behaviour of the calculated control
limits, but can be read and processed in entirely the same way.
1:69 1:12
SPI p ¼ 50% ¼ 1:12 þ ð50 39:5Þ ð9Þ
70:2 39:5
4. Research methodology and experimental test design
SPI p ¼ 50% ¼ 1:31 ð10Þ

These calculations result in the SPI measurements found in In this section, the data generation process to create a project
Table 5. benchmark set, the methodology used and the main research
questions set during the computational experiment in order to
validate the project control approach are explained along the
3.1. X chart following subsections.

To calculate the tolerance limits for the X chart, we need to find 4.1. Project data
the first and last α quantiles for Xp at all P ¼9 instances. This
illustrative example only includes the SPI as a control variable. For The X and R project control charts are tested on a set of 900
more details regarding the calculations of the other earned value fictitious projects generated by a project generator RanGen [34,35]
metric (SV) and the earned schedule metrics (SPI(t) and SV(t)), we that has been used previously in project scheduling and control
refer the reader to the detailed work by Vanhoucke [12] and the studies. Vanhoucke [28] uses this dataset to connect network
original work on earned schedule by Lipke [33]. topology information to the optimal control procedure found to
Fig. 3 illustrates the tolerance limits obtained for an X chart for monitor the time performance of a project.
SPI. The central line (CL) is added for illustrative purposes and is The 900 projects are chosen in accordance to the serial/parallel
the mean SPI along the project (X p ), the outer lines represent the (SP) indicator originally proposed by Vanhoucke et al. [36] and
upper and lower tolerance limits (LTLαX p , UTLαX p , for α ¼ 10%) and used in the previous project control simulation studies [24,28,29].
the dots illustrate how measurements made on a new run (with The SP indicator measures how close a project matches a complete
114 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122

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
Fig. 4. R chart control limits for the illustrative example. stochastic activity networks.
For the purpose of our research we have chosen to use a
serial or parallel project and is based on the maximal progressive uniform distribution. However, the presented methodology can be
level concept of Elmaghraby [37]. If m is the maximal progressive reproduced with any distributional input that is able to accurately
level, it is defined as the maximum number of activities lying represent the stochastic nature of project activity durations. More-
on a single path in the project. If n is the number of activities over, we strongly advise a calibration to historical records for the
in the project, it is clear that for a serial project m ¼n. The input modelling, if our methodology would be applied in a
SP ¼ ðm  1Þ=ðn  1Þ then equals 1 for a serial project. If the largest practical setting. The uniform distribution reflects an absence of
path in a project consists of only 1 activity, the project network is interest in the precise distributional characteristics of the project
said to be parallel and SP¼0. The project dataset consists of activities by the project manager. In this view on schedule control,
100 projects for 9 intermediate SP values between 0 and 1 the only interest is in the relative deviation of a duration from its
(SP ¼ f0:1; 0:2; :::; 0:9g). If conclusions are formulated regarding a baseline estimate. Most of the simulation experiments will incor-
mean value over all 900 projects, it is conjectured that they are porate a linear association variable (see Section 4.2.2) or have not
general enough to extrapolate over all possible project networks. as planned variation added to them. Therefore, the resulting
In Section 5, we did not find a significant influence of such serial/ distribution for di =d^ i , where d^ i represents the baseline estimate
parallel network structure on the performance of our control chart for the duration of activity i, will most likely not be uniform.
to monitor the schedule performance of a project. This result Section 5 has a paragraph allocated to the validation of our
shows that our statistical project control approach provides a simulation model using the recently published lognormal core
more robust solution than regular use of EVM/ES, which was theory [42]. Moreover, a reasonable fit to the proposed generalized
found to improve notably for more serial project networks [28]. 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.
4.2. Methodology As planned project progress: The variation is added to the
baseline duration d^ i of an activity i in the project by a uniform
Monte Carlo simulations are used in the static phase of our distribution. We define the as planned project schedule executions
statistical project control procedure to construct the tolerance in the static phase using the lightgrey probability distribution
limits used in the X and R charts, as well as to emulate real project functions presented in Fig. 5, characterized by a choice for the
executions in the dynamic phase where fictitious project progress maximum allowed deviation from the baseline estimate duration
will be classified using these charts. A distinction is made between δap. This distribution is presented relative to the baseline duration.
both phases in the form of the variation present at the The Monte Carlo samples that are drawn from this distribution
activity level. therefore need to be multiplied with the baseline duration d^ i , in
In order to generate fictitious project executions in both the order to model the duration di of an activity i in a fictitious
static and the dynamic phase of the Monte Carlo simulation, we execution of the project. The maximal relative deviation δap from
produce activity durations under interactive and compound uncer- the baseline estimate duration represents the subjective estimate
tainties [38]. Therefore, two sources of uncertainty are applied. of the variation on the baseline schedule that is regarded as
In Section 4.2.1, variation is modelled as probability distributions planned.
that are applied to the duration of the project's activities. This Detect both opportunities and dangers: In the dynamic phase,
variation will result in fictitious project executions that are either project progress situations are simulated for which the baseline
as planned or not as planned. In Section 4.2.2, we discuss the schedule might no longer be adequate. In order to simulate this,
potential dependencies between a project's activities and how variation is added using a not as planned uniform probability
project progress can be affected by uncertain events. distribution that has a standard deviation larger than that of the as
For the construction of the tolerance limits we will allow only planned uniform distribution. Fig. 5 shows two illustrative exam-
variation that is acceptable at the activity level. The projects are ples of this not as planned uniform distribution, hatched in black,
executed as planned, which represents the desired state of sche- for a low and a high value of %Anp . The variable %Anp expresses the
dule control in the presence of uncertain project-wide events. percentage of the total number of activities N that has a duration
In the dynamic phase, fictitious project progress is generated for that is not confined within the as planned margin δap. %Anp is
which it is possible that the durations of the activities exhibit therefore proposed to provide an intuitive basis for creating not as
unacceptable variation. These fictitious project executions are then planned project progress situations. The larger the %Anp , the more
used to quantify the performance of the proposed control charts in the project progress can be regarded as going not as planned and
their ability to accurately categorize whether the plan is still warning signals generated by a project control procedure becomes
adequate or not. Consequently, the project control charts are used more desirable.
 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122 115

The as planned parameter δap and the not as planned parameter calculate
%Anp will be set in our experiment according to the values
pnp ¼ P½jD  1j4 δap 
specified in Table 7. From the δap value, the extremes of the
(lightgrey) as planned uniform probability distribution are calcu- 2δap
¼ 1 ð12Þ
lated directly (1  δap and 1 þ δap ). Indirectly, the extremes of the δnp þ þ δnp 
(hatched in black) not as planned uniform distribution function
From pnp and δap, the extremes of the not as planned uniform
(1  δnp  and 1 þ δnp þ ) can be found using Eqs. (11) and (12). We
distribution (represented using δnp  and δnp þ ) can then be found
will discuss how these extremes are calculated along the
from Eq. (12), where D denotes the random variable that is drawn.
following lines.
In order to derive a value for both δnp  and δnp þ , an assumption
Let us assume that the N activities in the project have identical
has to be made with respect to one or both of these unknowns. We
probability distributions and that they are sampled independently.
chose to implement a linear relation (δnp þ ¼ 1:5δnp  ) which
The reader should consider the fact that we add dependencies
expresses the tendency of project activities to be late rather than
through linear association in Section 4.2.2. Let pnp denote the
early. However, in the cases where this relation resulted in a
probability of drawing a value outside of the as planned margin.
negative value for the lower extreme of the distribution
Consider this success/failure or coin toss experiment where a draw
(1  δnp  o 0), δnp  was fixed at 1 and the only unknown δnp þ
from outside the as planned margin represents a success. The
was found from Eq. (12). The factor 1.5 shown for the linear
probability of drawing a number of successes Nnp larger than
relation is arbitrary and different choices were also implemented
⌊%Anp Nc from a total of N is given by the binomial distribution:
during our tests.
P½N np 4 ⌊%Anp Nc ¼ 1  P½Nnp r ⌊%Anp Nc
⌊%Anp Nc
N!
¼ 1 ∑ pi ð1  pnp ÞN  i ð11Þ 4.2.2. Project progress simulation modelling
i¼0 i!ðN  iÞ! np The Monte Carlo simulation generates activity durations for all
In order to calculate pnp from Eq. (11), we have to assign a value to activities according to the input models described in Section 4.2.1.
its left-hand side. This value represents the chance that a mini- In accordance to recent literature we implemented two structures
mum of ⌊%Anp Nc successes is drawn. We chose this value as 95% in based on the linear association concept [23]. Linear association
our experiments, in accordance with a probability of 5% that a allows us to model the dependencies between activities and test
fictitious project with less then ⌊%Anp Nc successes is simulated. how project progress is affected by this.
For a simulation of 1000 runs, an average of 50 of such fictitious

Independence: The first structure can be regarded as a special
projects can then be expected to be included in the output. From
Eq. (11), we proceeded with a numerical approximation to case, since complete independence is assumed between activ-
ities. 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 [46,47,42], it has been
used in both analytical [48,49] and simulation [50,37,28,29,51]
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 associa-
tion technique [42,23], where a systemic error B provides
objective information on correlations.
We say that N positive random variables Zj are linearly
associated if Zj ¼ BXj where fX j g is a set of n independent
positive random variables and B is a positive random variable,
Fig. 5. Defining the uniform input model for activity durations. independent of fX j g that can be regressed from historical data.

Table 7
Overview of the computational experiments.

Baseline Simulation model Settings

Estimates As planned Not as planned

Uniform INDEPENDENCE nrs ¼1,000
8 iA N : d^ A ½8; 56
i
μB A f0:8; 1; 1:5; 2g
cvB ¼ 0
Fixed Cost δap A f20%; ‥; 80%g δap A f20%; ‥; 80%g
Uniform %Anp ¼ 0 %Anp A f20%; ‥; 80%g
½0; 500
LINEAR ASSOCIATION nrs ¼1,000
Variable Cost μB A f0:8; 1; 1:5; 2g
Uniform cvB A f0:3; 0:8; 2g
½700; 1500 δap A f20%; ‥; 80%g δap A f20%; ‥; 80%g
%Anp ¼ 0 %Anp A f20%; ‥; 80%g
116 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122

We follow [23] and model B as a lognormal random variable, event P X :

since the lognormal distribution is found appropriate to reflect
1 nrs 
both additive and multiplicative factors, ubiquitous in real-life ∑ 1s ( pjP X g ð14Þ
nrs s ¼ 1
project activities. The lognormal random variable is character-
ized by its mean μB and its coefficient of variation cvB. For the R chart, the detection performance is calculated analogously
as
Our model assumes that the earned value (EV) for a single 1 nrs 
activity follows a linear accrue, starting from its actual start up to ∑ 1s ( pjP R g ð15Þ
nrs s ¼ 1
its budget at completion (BAC) when it is finished [12]. Planned
value (PV) follows this same linear accrue from the planned start where
up to the planned finish time of the activity. EV and PV are 1s f (pjP R g ¼ 1s f ( pjðjY p  Y p  1 jo LTLαRp Þ 3 ðjY p  Y p  1 j 4 UTLαRp Þg
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 ð16Þ
control charts and the recorded performances in Section 5 are The probability of overreactions: The probability of overreactions
implemented in the statistical programming language R [52]. 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
4.3. Overview of experiments divided into 3 equal groups and the probability of overreactions is
calculated 3 times. Once for each group where the two other
Table 7 gives an overview of the simulation experiments groups are used to calculate the tolerance limits. We confirm that
conducted for this research. A total of 400 different simulation deviations between groups never exceeded 1%, and in the remain-
experiments were run for each project network according to the der we will present the average over all 3 groups. Consequently,
inputs for the parameters μB , cvB, δap and %Anp provided in Table 7. the probability of overreactions represents situations where project
These parameters define to what extent dependencies between managers get a false impression that the project is being executed
activities are incorporated or not and the as planned and not as not as planned, which acts as a trigger to detect the reasons for this
planned variation . It should be noted that although the indepen- behaviour, while there is no unacceptable variation present at the
dence assumption is retained for some scenarios (denoted as activity level. These false signals should be avoided since they lead
cvB ¼0), a constant B is still used in these to alter the mean to time-consuming and unnecessary overreactions made by the
(μB A f0:8; 1; 1:5; 2g). The 400 experiments produce 1000 runs for project manager and his/her team in charge of controlling the
all 900 projects (Section 4.1), for which the baseline estimates and project performance. Each time an not as planned indication is
costs are assigned according to the first column in Table 7. Periodic reported, the project control chart is said to be overreacting since
EVM/ES reports are generated at 19 distinct moments in the it generates a signal, while at activity level only acceptable
project (p ¼ f5%; 10%; 15%; 20%; :::; 90%; 95%g) and interpolation variation is present.
is done as shown in Section 3 with a time-increment chosen The probability of overreactions for one group can be calculated
sufficiently small. The durations of the activities were expressed in using nrs/3 runs from the static phase as (14) and (15) for
days, whereas the time-increment was in the order of hours. respectively the X and the R chart.
We will use these experiments to demonstrate the value of Area under the curve: The probability of overreactions and the
statistical project control for project management in Section 5 detection performance can both be used as approximations of the
using the control efficiency measures defined in Section 4.4. We effectiveness of the proposed project control procedures. However,
have tried to incorporate the best practices for project control in Section 5, the need for a more holistic representation of the
modelling as provided by the current literature, to produce results effectiveness will be expressed. We find this in the concept of the
that are as generalizable to project management practice as calculated area under the curve (AUC) for a receiver operating
possible. However, we will expand on the limitations of our characteristic curve (ROC). The ROC representation is widely used
research, and its results, in Section 6. in classification testing and machine learning [54] and can be
plotted directly from the probability of overreactions and the
detection performance.
4.4. Project control efficiency measures Riemann integration of the ROC curve provides us with a single
unitless measure for the area under the curve that captures the
In order to measure and compare the efficiency of X and R characteristics of both the probability of overreactions and the
statistical project control charts we propose three simulation- detection performance at the same time. The AUC should be more
based performance assessment measures [53]. than 0.5, which is the area under the no-discrimination line. This is
The detection performance: The detection performance equals the the line where the true positive rate equals the false positive rate,
frequentist approximation that a certain not as planned situation and which can be regarded as the characteristic of a classification
will result in a signal during the dynamic phase, given the as action that is purely based on a randomized process such as a
planned data from the static phase. A signal is generated as in coin toss.
Section 2.3.2 if the project performance Yp or jY p  Y p  1 j is not
confined to the tolerance limits of respectively the X or the R chart
at any review period p.
5. Computational tests and results
The detection performance is calculated as

1 nrs n o In this section, we present our findings from our large compu-
∑ 1s ( pjðY p o LTLαX p Þ 3 ðY p 4 UTLαX p Þ ð13Þ tational experiment. First, we introduce how the concept of the
nrs s ¼ 1
tolerance level α can be used to set a desired degree of probability
for the X chart, where 1s fAg is the indicator function of an event A of overreactions and how the detection performance is affected, in
for a simulation run s. Otherwise stated, we represent the Section 5.1. Combined, these two output measures allow us to
occurrence of a signal for the X charts at review period p as the draw a ROC curve and to calculate the AUC.
 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122 117

Section 5.2 provides the validation for our project progress increase in the probability of overreactions. Optimally, we prefer the
simulation model according to the recently postulated lognormal probability of overreactions as close to 0 as possible. This would
core theory [23,42]. The efficiency of the described X and R mean that the project management team does not need to invest
statistical project control charts is explored in Section 5.3 and is time and effort unnecessarily in drilling down the WBS, to find the
compared to a more pragmatic model of traditional EVM/ES use variation at the activity level to be confined within the acceptable
and how it can be implemented for different control points margins. Analogously, we prefer the detection performance as close
throughout the project in Section 5.4 to 1 as possible. For a certain scenario, this would imply that the
not as planned situation gets detected somewhere throughout the
5.1. Close control versus overreactions – setting a tolerance level 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
When the statistical project control charts are constructed, we curves and the corresponding AUC values for the X chart and the R
use a parameter α as the tolerance level (see Section 2.3). This chart respectively from Figs. 6 and 7 are depicted in Fig. 8.
represents the choice for the first and the last αth quantile of the
empirical distribution function constructed in the static phase to
form the tolerance limits of the X and the R chart. 5.2. Validation of the project simulation model
Is is clear that α will affect both the probability of overreactions
and the detection performance of the X and R charts. Therefore, an In order to validate the project progress simulation model
appropriate choice for α should reflect both a project manager's (presented in Section 4.2.2), we analysed the simulation output
aversion to risk and his/her willingness to invest effort in false according to the lognormal core theory [23,42]. Figs. 9 and 10
alarms. Figs. 6 and 7 illustrate the effect α has on the probability of depict boxplots for respectively the calculated mean and standard
overreactions and the detection performance for one specific deviations of our simulation output.
scenario. Fig. 9 presents the mean of the simulated activity durations.
It is easily perceived that the real strength of a chart is in the This figure clearly shows how the average activity duration in a
relative improvement of the detection performance over a certain project is affected by μB . If B is a random variable (cvB a 0), rather
than a constant (cvB ¼ 0), both the spread and the average of
the mean duration are increased. Fig. 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

Fig. 6. Setting the tolerance level α for a X chart, for a (δap, %Aap ) ¼(20%, 20%)
scenario under the independence assumption (μB ¼ 1; cvB ¼ 0).
Fig. 8. Area under the curve for the X and R charts, for a (δap, %Aap ) ¼(20%, 20%)
scenario under the independence assumption (μB ¼ 1; cvB ¼ 0).

Fig. 7. Setting the tolerance level α for a R chart, for a (δap, %Aap )¼ (20%, 20%)
scenario under the independence assumption (μB ¼ 1; cvB ¼ 0). Fig. 9. Distribution of the resulting mean in the project simulation model.
118 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122

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] respectively. 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.
Fig. 10. Distribution of the resulting standard deviations in the project
simulation model.
5.3.2. Influencing factors for the X and the R chart AUC
We proceed with a discussion in more detail for the AUC shown
in Fig. 11. First, we provide an intuitive basis for the observed
variation within each cell of the static/dynamic phase grid. Second,
we investigate the effect that a change in the mean or coefficient
of variation of the observed distribution of di =d^ i 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 ¼0.3 and μB ¼ 1 in the static phase
from calibration to historical data recorded within the organiza-
tion. 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. Fig. 12
shows the expected AUC over a subset of not as planned scenarios.
In order to improve the readability of Fig. 12, we applied a
colour coding where the lightest shades of grey correspond to the
Fig. 11. Area under the curve for different static/dynamic phase combinations. lowest values and the darker to the highest values. The depicted
scenarios correspond to as planned situations for δap A f20%; 40%;
60%; 80%g and not as planned scenarios with %Anp A f20%;
the standard deviation increase with increasing cvB. This effect is 40%; 60%; 80%g.
noticeably larger for the not as planned situations. The colour-coded variation in Fig. 12 was found to be identical
In conclusion, we can safely state that in order to produce for both the X and the R chart, so only that of the X chart is shown.
project progress simulations with variation in the orders of Fig. 12 largely confirms the intuition behind the scenario repre-
magnitude that is discussed by Trietsch et al. [42], with high sentation. The higher the percentage of activities %Anp that is
variation instances having a standard deviation larger than 2, and likely to have a duration that is not confined to the as planned
large deviations in the mean, linear association via a lognormal margin δap, the more likely this scenario will produce a clear signal
variable B is preferred. that can be interpreted using the statistical project control charts.
For any %Anp Z 20%, an increase in AUC can also be observed for
increasing values of the as planned margin δap. From Eq. (12), it can
5.3. Project control efficiency be seen that if pnp remains fixed (resulting from a fixed %Anp ), then
(δnp  þ δnp þ ) needs to increase if δap increases. The extremes of the
This part of the results section explores the efficiency of the not as planned uniform thereby diverge, which results in a larger
proposed control charts using tolerance limits on EVM/ES by standard deviation and mean of the distribution. In the next
means of the area under the curve measure introduced in Section section, it will be shown that increasing values for the mean and
4.4. standard deviation lead to increasing AUC values.
Change in mean or standard deviation: This paragraph explores
the effect the observed underlying distribution of di =d^ i has on the
5.3.1. Area under the curve area under the curve. We investigate the impact of a change in
The area under the curve representation in Fig. 11 provides a mean and standard deviation of the observed distribution of di =d^ i
nuanced and holistic view of the project control efficiency for the for the dynamic phase relative to the static phase (which is used to
X and the R chart. In general, the figure shows that the X and the R calculate the statistical tolerance limits) for the X and R control
chart perform better than that a randomized classification process charts. The relative change in mean Δm ¼ mdynamic  mstatic is the
would. All AUC values are significantly larger than 0.5, the AUC difference between the calculated mean of the activity durations
value of the no-discrimination line. Overall the X chart outper- in the dynamic phase (mdynamic ) and the mean of the activity
forms the R chart. This can be seen more clearly from an durations in the static phase (mstatic ). Similarly, Δs ¼ sdynamic sstatic
aggregated AUC calculated over all simulation experiments. This is defined as the difference between the standard deviation of
will be presented in Section 5.4 these two phases.
 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122 119

Legent

Fig. 12. Area under the curve for different not as planned situations.

Fig. 13. The effect a change in the mean has on the AUC.

Fig. 14. The effect a change in the standard deviation has on the AUC.

Fig. 13 presents the effect a relative change in the mean of the

underlying distribution has on the efficiency of the project control between the static and the dynamic phase (Δs o 0), we observe
process. The x-axis represents the different ranges of values that decreasing values for the AUC of the X control chart and increasing
are found for Δm, while the change in the standard deviation is values for the AUC of the R control chart. Moreover, the R chart is
kept close to zero between the dynamic phase and the static phase preferred in order to accurately detect not as planned situations if
(Δs r 7 0:1). In Fig. 13, we observe an increasing trend for the AUC Δs o 0. For positive Δs values, the AUC of both the X chart and the
for increasing values of Δm up to 1.2. If the mean of the static R chart increases. In these situations, the X chart performs better
phase is e.g. 1, then increasing values for the AUC can be expected than the R chart.
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 out- 5.4. Comparing statistical project control to traditional EVM/ES
performs the X chart.The effect of a change in the standard
deviation Δs of the underlying distribution is shown in Fig. 14 In this section we revisit the discussion on the comparison of
for approximately equal mean values (Δm r 70:1). The x-axis the use of project control charts using statistical tolerance limits
represents the ranges of values found for Δs. If the standard (STL) to the more traditional use of EVM/ES metrics and some
deviation of the underlying distribution of di =d^ i decreases recent advances from the literature for project schedule control.
120 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122

Although practical use of EVM/ES is typically characterized by on points throughout the project are suggested at which EVM/ES
the spot decision-making from practical experience, we imple- control charts should be monitored. The two alternative
ment four techniques to compare our statistical control charts approaches are discussed in the original paper and can be
with briefly summarized as follows:

Feeding paths (FP): The concept of feeding paths is adopted

Rules-of-thumb (RoT): This project schedule control approach from the CC/BM literature. EVM/ES schedule control metrics
combines the use of project control charts with rules-of-thumb. are calculated for those paths in the project network where
Instead of our statical tolerance limits for the X chart, these a buffer should be placed according to CC/BM in order to
charts employ either static, widening or narrowing tolerance mitigate risks and to protect the critical path. The schedule
limits for schedule control purposes. The tolerance limits, control metrics are then monitored on a control chart
represented by a þbp, are implemented as symmetrical lines against rules-of-thumb tolerance levels.
around a percentage complete axis. Therefore, the median of
Sub-networks (SN): The concept of sub-network control points
each performance metric Q^ ð0:5Þp calculated in the static phase expands on the concept of feeding paths. Schedule control
is subtracted for all percentage project complete p: metrics are calculated for the collection of paths that enter the
○ Static tolerance limits mimic the “Target Performance Chart” critical path at a given point. Again, these EVM/ES metrics are
[56]. The slope a equals 0, for these tolerance limits referenced against tolerance limits on control charts. The
○ Widening tolerance limits should be used in combination tolerance limits are calculated using sample quantiles, in a
with SV and SV(t). These variance-based performance manner similar to that presented in Section 2.3.
metrics grow larger in absolute figures during project
progress and therefore, an increasing upper tolerance limit It should be noted that the difference between the two
should be used. Since the lower tolerance limit is symme- techniques mainly lies on the number of control points, and
trical, overall the tolerance limits can be said to be widening Fig. 15 only gives a summary overview of the main results to
towards the end of the project. compare the alternative project control methods:
○ Narrowing tolerance limits should be used in combination with
Longest path monitoring (LP): Inspired by the fact that the EVM/
SPI and SPI(t). Simple simulations indicate that the variation ES schedule control metrics are fully reliable for serial networks
for these index-based performance metrics decreases along but have a decreasing performance for more parallel networks
the percentage complete axis. Consequently, narrowing toler- [28], the longest path control approach was proposed by Lipke
ance limits are preferred during project progress. [58]. In this method, SPI(t) is referenced against a static
tolerance limit on a single project control chart. The SPI(t) is
Without loss of generality, we assume that these control however not calculated for all activities of the project, but only
strategies represent a best-case decision making process from for those that lie on the longest path in the project network.
practical experience, since they are calibrated to the static phase This longest path is updated dynamically during project pro-
simulation data, much as the X and R control charts. This gress with actual durations for those activities that have
calibration is done for the slope a and the intercept b, which are already been finished and the baseline estimate durations for
optimized using a coarse search algorithm over a large solution the activities that are not yet started.
space, in order to result in probabilities of overreactions that are
competitive with those chosen for the X and the R chart. In
practical decision-making from experience, this calibration is Fig. 15 depicts the calculated area under the curve as boxplots
impossible and the real-life probabilities of overreactions can be for the statistical project control charts proposed in this paper and
expected to be much higher. We therefore consider this RoT the alternative schedule control procedures outlined in this sec-
control strategy to be best-case EVM/ES decision making with- tion. This comparison is based on all the data obtained from the
out the use of statistical tolerance limits. simulations in the static/dynamic phase grid of Fig. 11. A detailed

EVM/ES on multiple control points: The use of EVM/ES control discussion on the merits and the distinct characteristics of the
metrics applied at different (and multiple) control points in the procedures presented in this section falls outside of the scope of
project schedule has recently been discussed by Colin and this paper. In short, we conclude that the X chart outperforms the
Vanhoucke [57]. Inspired by the concepts of the critical chain/ R chart, based on the aggregated AUC over all performed simula-
buffer management (CC/BM) methodology, different control tion experiments. Both these charts using statistical tolerance

Fig. 15. Area under the curve comparison of the X and R charts (STL) with other schedule control approaches using EVM/ES.
 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122 121

limits should be preferred over all other tested schedule control hoc rules-of-thumb strategies that would typically be used in
procedures for the following four reasons: project control using EVM/ES and some recently developed EVM/
ES extensions. Overall, the X and the R chart can defensibly be

The RoT strategies display a significantly lower AUC. The preferred over all tested procedures
narrowing tolerance limits and the widening tolerance limits Obviously, the results obtained in this study should be put in
can however be considered as useful alternatives. the right perspective. The use of simulated data often shows some

The FP method is shown for a combination of maximally 10 weaknesses and the results should therefore be interpreted with
feeding paths monitored during project progress. The maximal care. A first assumption in our simulation study lies on the
number of feeding paths can however be much larger, even for dependence structure of activity durations during sampling. While
small projects. The x-axis in Fig. 15 under FP depicts the the assumption of independence was invalidated in the previous
number of control points that are used. The first control point research [59,42], it has nevertheless been used in earlier reports
is always the critical path, while the others are used to monitor on project control [28,29]. We also expanded our project progress
the performance of the feeding paths. When not all feeding simulation model with the concept of linear association [23].
paths are monitored, information about certain activities is lost. However, the input for our simulation experiments depends
This results in an increased probability of overreactions, which largely on subjective estimates for activity durations, where in
affects the detection performance and consequently the calcu- practice calibration to historical data should always be preferred.
lated AUC. Fig. 15 shows that this produces a decreasing A second restriction of our simulated project control study is
schedule control performance for an increasing number of the strict focus on time performance of projects. Although EVM/ES
included feeding paths. takes both a time and a cost focus during project performance

The SN method shows the opposite behaviour. AUC increases to management, this paper puts a strong and restricted focus on the
the optimal unity area under the curve for 5 sub-network time dimension of a project in progress in order to validate the use
control points included in the schedule control procedure. of schedule performance metrics in a statistical project control
When its performance is compared to the X chart, it shows to approach. Activity costs have been assumed to be linearly depen-
be a valuable alternative. However, while the X control chart dent on the activity duration time, which is restricted to the use of
relies on only one control point, the SN approach needs up to periodic costs. In the future, our model could be adjusted to
5 control points to only slightly outperform the X chart. incorporate other situations. Furthermore, other relevant project
Moreover, the SN approach also relies on the STL procedure performance dimensions, such as quality control or project scope
which is the topic of this paper. control could not be taken into account in this study. Future

The LP method [58] is found to be outperformed by both the X research will be built on this paper, stretching its scope to
and the R chart. Using our definition and measures for schedule incorporate the restrictions and expanding on the idea of using
control it is not able to produce similar results. However, this process control procedures for project control.
method was developed foremost to produce project duration Finally, it should be mentioned that a detailed comparison
forecasts, rather than to detect not as planned project progress. between the statistical project control approach of this paper and
the alternative control approaches available in the literature is not
within the scope of this paper. A comparison between the quality
6. Discussion and conclusions of methods such as Critical Path Method (CPM, [1]), the Pro-
gramme Evaluation and Review Technique (PERT, [60]), its novel
In this paper, a project control system is presented which is extension to PERT21 [23], the bottom-up approach using Schedule
inspired by the well-known statistical process control charts Risk Analysis [29] and the Critical Chain Project Management
widely used to monitor manufacturing processes but is fundamen- (CCPM, [61,62]) is therefore considered as an interesting future
tally different in its use of the progress data to construct tolerance research avenue to further improve and optimize the knowledge
limits. The use of earned value management performance metrics on project control.
in a control chart, based on simulated random variation, allows us This study should be relevant for both practitioners and
to discriminate between as planned and not as planned project academics for the following reasons. Performance indicators from
performance variations. This paper measures the performance of earned value management/earned schedule have been used
the two proposed charts, using a wide range of fundamentally widely by practitioners in controlling projects. However, the lack
different project networks in a large computational experiment. of guidelines towards reasonable tolerance limits to discriminate
The first chart (X chart) mimics a Shewhart chart for process between acceptable and unacceptable performance variations has
control with adaptations to fit the finite nature of a project. The X often classified this technique as secondary to the necessary skills
chart uses the commonly known schedule performance metrics of intuition, existing experience and knowledge a project manager
provided in an earned value management system. The second must have. The methodology and control approach proposed in
chart (R chart) proposes an adjustment on these metrics to focus this paper quantifies the use of tolerance limits. This paper shows
on the instantaneous changes in project performance. The that the discriminative power between as planned and not as
assumption is that by performing differential calculations, the planned variations in a project is much better when a customized
accumulated performance of past activities is filtered and only control chart is used rather than when manual and intuitive
active tasks are being monitored. thresholds (represented through the rules-of-thumb) are used.
Overall the X chart is shown to be the most promising to be The on the spot decision-making during project control is thereby
used in practice to monitor deviations from the baseline schedule. assisted. Next to the practical relevance, we also believe that this
Altogether, both show very comparable performance. The results topic will contribute to new research challenges where extensions
of various computational experiments show the relevance and of this approach might lead to an increased discriminative power
usefulness of this approach in a project control setting, although and better project control.
all the results are subject to the specific definition of as planned
and not as planned project progresses and the characteristics of the References
simulation model as applied in this research. In order to assess the
performance of the statistical tolerance limit approach discussed [1] Kelley J, Walker M. Critical path planning and scheduling: an introduction.
in the paper, the X and R chart's performances are compared to ad Ambler, PA: Mauchly Associates; 1959.
122 J. Colin, M. Vanhoucke / Omega 49 (2014) 107–122

[2] Kelley J. Critical path planning and scheduling: mathematical basis. Operations [33] Lipke W. Statistical methods applied to EVM ...the next frontier. The Measur-
Research 1961;9:296–320. able News (Winter) 2006, PMI College of Performance Management (1/1/
[3] Malcolm D, Roseboom J, Clark C, Fazar W. Application of a technique for a 2006).
research and development program evaluation. Operations Research [34] Tavares L, Ferreira J, Coelho J. The risk of delay of a project in terms of the
1959:646–69. morphology of its network. European Journal of Operational Research
[4] Hartmann S, Briskorn D. A survey of variants and extensions of the resource- 1999;119:510–37.
constrained project scheduling problem. European Journal of Operational [35] Demeulemeester E, Vanhoucke M, Herroelen W. Rangen: a random network
Research 2010;207:1–15. generator for activity-on-the-node networks. Journal of Scheduling
[5] Madadi M, Iranmanesh H. A management oriented approach to reduce a 2003;6:17–38.
project duration and its risk (variability). European Journal of Operational [36] Vanhoucke M, Coelho J, Debels D, Maenhout B, Tavares L. An evaluation of the
Research 2012;219(3):751–61. adequacy of project network generators with systematically sampled net-
[6] Uyttewaal E. Dynamic scheduling with microsoft office project 2003: works. European Journal of Operational Research 2008;187:511–24.
the book by and for professionals. Plantation, FL, USA: J. Ross Publishing, [37] Elmaghraby S. On criticality and sensitivity in activity networks. European
Inc.; New York, NY, USA: International Institute for Learning World Journal of Operational Research 2000;127:220–38.
Headquarters /www.iil.comS; 2005. [38] Wang S, Huang G. An integrated approach for water resources decision
[7] Vanhoucke M. Project management with dynamic scheduling: baseline making under interactive and compound uncertainties. Omega The Interna-
scheduling, risk analysis and project control, vol. XVIII. Springer; 2012. tional Journal of Management Science 2014;44:32–40.
(Springer Heidelberg Dordrecht London New York, DOI: 10.1007/978-3-642- [39] Williams T. What are PERT estimates? Journal of the Operational Research
25175-7, Springer-Verlag GmbH, Heidelberger Platz 3, 14197 Berlin, Germany). Society 1995;46:1498–504.
[8] Rozenes S, Vitner G, Spraggett S. Project control: literature review. Project [40] Kreye ME, Goh YM, Newnes LB, Goodwin P. Approaches to displaying
Management Journal 2006;37:5–14. information to assist decisions under uncertainty. Omega The International
[9] Rozenes S, Vitner G, Spraggett S. MPCS: multidimensional project control Journal of Management Science 2012;40(6):682–92.
system. International Journal of Project Management 2004;22:109–18. [41] Kuhl ME, Lada EK, Steiger NM, Wagner MA, Wilson JR. Introduction to
[10] Fleming Q, Koppelman J. Earned value project management. 3rd ed. Newtown modeling and generating probabilistic input processes for simulation. In:
Square, PA: Project Management Institute; 2005. Henderson S, Biller B, Hsieh M, Shortle J, Tew J, Barton R, editors. Proceedings
[11] Lipke W, Zwikael O, Henderson K, Anbari F. Prediction of project outcome: the of the 2007 winter simulation conference. New Jersey: Institute of Electrical
application of statistical methods to earned value management and earned and Electronics Engineers; 2007. p. 63–76.
schedule performance indexes. International Journal of Project Management [42] Trietsch D, Mazmanyan L, Govergyan L, Baker KR. Modeling activity times by
2009;27:400–7. the Parkinson distribution with a lognormal core: theory and validation.
[12] Vanhoucke M. Measuring time—improving project performance using earned
European Journal of Operational Research 2012:386–96.
value management international series in operations research and manage- [43] Mohan S, Gopalakrishnan M, Balasubramanian H, Chandrashekar A. A log-
ment science, vol. 136. Springer (Springer Heidelberg Dordrecht London New
normal approximation of activity duration in pert using two time estimates.
York, DOI: 10.1007/978-1-4419-1014-1, Springe Science þ Business Media, LLC,
Journal of the Operational Research Society 2007;58(6):827–31.
223 Spring Street, NY, 10013, USA, 14197 Berlin, Germany); 2010.
[44] Hahn E. Mixture densities for project management activity times: a robust
[13] Hulett D. Schedule risk analysis simplified. Project Management Network
approach to PERT. European Journal of Operational Research 2008;188:450–9.
1996;10:23–30.
[45] Kotiah T, Wallace ND. Another look at the pert assumptions. Management
[14] Fang Y, Zhang J. Performance of control charts for autoregressive conditional
Science 1973;20(1):44–9.
heteroscedastic processes. Journal of Applied Statistics 1999;26(6):701–14.
[46] Ragsdale C. The current state of network simulation in project management
[15] Shewhart WA. Economic control of quality of manufactured product, vol. 509.
theory and practice. Omega The International Journal of Management Science
ASQ Quality Press (American Society for Quality, P.O. Box 3005, Milwaukee, WI
1989;17(1):21–5.
53201-3005, USA); 1931.
[47] Schonberger R. Why projects are “always” late: a rationale based on manual
[16] Lipke W. A study of the normality of earned value management indicators. The
simulation of a PERT/CPM network. Interfaces 1981;11:65–70.
Measurable News 2002;4:1,6,7,12–4,16.
[48] Fisher DL, Saisi D, Goldstein WM. Stochastic pert networks: OP diagrams,
[17] Lipke W, Vaughn J. Statistical process control meets earned value. CrossTalk:
critical paths and the project completion time. Computers & Operations
The Journal of Defense Software Engineering 2000;(June):16–20,28–29.
Research 1985;12(5):471–82.
[18] Bauch GT, Chung CA. A statistical project control tool for engineering
[49] Dodin B, Sirvanci M. Stochastic networks and the extreme value distribution.
managers. Project Management Journal 2001;32:37–44.
[19] Wang Q, Jiang N, Gou L, Che M, Zhang R. Practical experiences of cost/schedule Computers & Operations Research 1990;17(4):397–409.
[50] Williams T. Practical use of distributions in network analysis. Journal of the
measure through earned value management and statistical process control. In:
Lecture notes in computer science, vol. 3966; 2006. p. 348–54. Operational Research Society 1992:265–70.
[20] Leu SS, Lin YC. Project performance evaluation based on statistical process [51] Acebes F, Pajares J, Galán JM, López-Paredes A. A new approach for project
control techniques. Journal of Construction Engineering and Management control under uncertainty. Going back to the basics, International Journal of
2008;134:813–9. Project Management 2014;32(3):423–34.
[21] Aliverdi R, Moslemi Naeni L, Salehipour A. Monitoring project duration and [52] R Core Team, R: a language and environment for statistical computing, R
cost in a construction project by applying statistical quality control charts, foundation for statistical computing, Vienna, Austria; 2013. Url: 〈http://www.
International Journal of Project Management 2013;31(3):411-23. R-project.org〉.
[22] Page E. Continuous inspection schemes. Biometrika 1954:100–15. [53] Ponsignon T, Mönch L. Simulation-based performance assessment of master
[23] Trietsch D, Baker KR. Pert 21: Fitting PERT/CPM for use in the 21st century. planning approaches in semiconductor manufacturing. Omega The Interna-
International Journal of Project Management 2012;30(4):490–502. tional Journal of Management Science 2014;46:21–35.
[24] Vanhoucke M, Vandevoorde S. A simulation and evaluation of earned value [54] Hastie T, Tibshirani R, Friedman J, Hastie T, Friedman J, Tibshirani R. The
metrics to forecast the project duration. Journal of the Operational Research elements of statistical learning, vol. 2. Springer (Springer-Verlag GmbH,
Society 2007;58:1361–74. Heidelberger Platz 3, 14197 Berlin, Germany); 2009.
[25] Burrows G. Statistical tolerance limits–what are they. Technical report. [55] Lim S. A joint optimal pricing and order quantity model under parameter
Schenectady, NY: Knolls Atomic Power Lab.; 1962. uncertainty and its practical implementation. Omega The International
[26] Van Slyke R. Monte Carlo methods and the PERT problem. Operations Journal of Management Science 2013;41(6):998–1007.
Research 1963;11:839–60. [56] Anbari F. Earned value project management method and extensions. Project
[27] Williams T. A classified bibliography of recent research relating to project risk Management Journal 2003;34(4):12–23.
management. European Journal of Operational Research 1995;85:18–38. [57] J. Colin, M. Vanhoucke, A comparison of the performance of various project
[28] Vanhoucke M. Using activity sensitivity and network topology information to control methods using earned value management systems, Unpublished
monitor project time performance. Omega The International Journal of results.
Management Science 2010;38:359–70. [58] Lipke W. Speculations on project duration forecasting. The Measurable News
[29] Vanhoucke M. On the dynamic use of project performance and schedule risk 2012;3:3–7.
information during project tracking. Omega The International Journal of [59] Williams T. The contribution of mathematical modelling to the practice of
Management Science 2011;39:416–26. project management. IMA Journal of Management Mathematics 2003;14
[30] Vanhoucke M. Integrated project management and control: first come the (1):3–30.
theory, then the practice, management for professionals. Springer (Springer- [60] Fazar W. Program evaluation and review technique. The American Statistician
Verlag GmbH, Heidelberger Platz 3, 14197 Berlin, Germany); 2014. 1959;13:10.
[31] Montgomery DC, Woodall W. Research issues and ideas in statistical process [61] Goldratt E. Critical chain. Great Barrington, MA: North River Press; 1997.
control. Journal of Quality Technology 1999;31(4):376–87. [62] Ash RC, Pittman PH. Towards holistic project scheduling using critical chain
[32] Hyndman RJ, Fan Y. Sample quantiles in statistical packages. The American methodology enhanced with pert buffering. International Journal of Project
Statistician 1996;50(4):361–5. Organisation and Management 2008;1(2):185–203.