SIMPLIFIED PROGRAM EVALUATION AND REVIEW

TECHNIQUE (PERT)
By Wayne D. Cottrell,1 P.E.

ABSTRACT: A simplified version of the program evaluation and review technique (PERT) for project planning
is developed and tested. The simplification is to reduce the number of estimates required for activity durations
from three, as in conventional PERT, to two. This is accomplished by applying the normal distribution, rather
than the beta, to an activity duration. The two required duration estimates are the ‘‘most likely’’ and the
‘‘pessimistic.’’ These modifications reduce the level of effort needed to apply PERT. Simplified PERT durations
are subject to errors of greater than 10% when the skewness of the actual distribution is greater than 0.28 or
less than ⫺0.48. In analyzing 12 project networks, though, the simplified PERT produced values similar to those
Downloaded from ascelibrary.org by University of Alberta on 06/18/13. Copyright ASCE. For personal use only; all rights reserved.

of conventional PERT for activity and project durations and variances and project duration probabilities. Hence,
when activity duration distributions are not highly skewed, results similar to those of conventional PERT can
be obtained using the simpler technique. Two suggestions for future research are to survey practitioners on the
usefulness of the simplified PERT and to find a fixed, skewed distribution that can approximate activity durations
having long tails.

INTRODUCTION probability of a shorter duration is zero. Similarly, b is the

maximum duration; the probability that the duration will be
The current paper presents and evaluates a simplified pro- less than or equal to b is 100%. No assumption is made about
gram evaluation and review technique (PERT), a planning tool the position of m relative to a and b. In statistical terms, a and
for application to construction and other projects. The outline b are the extreme ends of a hypothetical distribution of dura-
of the paper is as follows: First, PERT is described; second, tion times. The mode of the distribution is m. To accommodate
criticisms of PERT in the literature are reviewed; third, the flexibility in the positions of these parameters, the beta distri-
simplified PERT is presented; fourth, the simplified technique bution is used, as shown in Fig. 2 (Malcolm et al. 1959; Clark
is tested and evaluated, both on activity duration means and 1962). The beta distribution is useful for describing empirical
on a set of project networks. The results are compared with data and can be either symmetric or skew (Benjamin and Cor-
those of conventional PERT. Finally, the conclusions are nell 1970).
stated. The objective of the present research is to present a The third step is to compute the expected value and variance
simplified technique that might be more easily and readily ap- of the duration of each activity in the project network. The
plied in industry, particularly in construction project planning, mean of a beta distribution is a cubic equation. The PERT
than conventional PERT. The simplified PERT algorithm could equation for the mean [(1)] is a linear approximation to this
be embedded into construction scheduling software to facili-
tate its application. ␶e = (a ⫹ 4m ⫹ b)/6 (1)
where ␶e = expected duration of an activity. Badiru (1991)
Description of PERT shows that (1) is exact when m is equal to the mode, which
occurs when a and b are symmetrical about m.
PERT was originated by the U.S. Navy in 1958 as a tool
In unimodal probability distributions, the standard deviation
for scheduling the development of a complete weapons system
of the distribution is equal to approximately one-sixth of the
(Malcolm et al. 1959). The technique considers a project to be
range (Watson et al. 1993). With 100% of the possible dura-
an acyclic network of events and activities. The duration of a
tions bound by a and b, the estimated variance of the duration
project is determined by a system flow plan in which the du-
is as follows:
ration of each task has an expected value and a variance. The
critical path includes a sequence of activities that cannot be ␴2100(␶e) = [(b ⫺ a)/6]2 (2)
delayed without jeopardy to the entire project. PERT can be
used to estimate the probability of completing either a project where ␴ = variance of the activity duration. Moder and Rod-
2

or individual activities by any specified time. It is also possible gers (1968) argue that the exact endpoints of the range of the
to determine the time duration corresponding to a given prob- duration are impossible to define. Their alternative is to define
ability (Callahan et al. 1992). a and b as the 5% and 95% thresholds of the range, respec-
The first step in applying PERT is to diagram the project tively. Then, the variance is as follows:
network, where each arc represents an activity and each node ␴290(␶e) = [(b ⫺ a)/3.2]2 (3)
symbolizes an event (such as the beginning or completion of
a task), as in Fig. 1. Alternatively, each node can symbolize Perry and Greig (1975), alternatively, use 3.25 in the de-
an activity. The second step is to designate three time estimates nominator of (3), rather than 3.2. They argue that subjective
for each task: optimistic (a), pessimistic (b), and most likely probability distributions tend to be rounded (platykurtic) rather
(m). Small probabilities are associated with a and b. In the than peaked. The denominator of 3.25 is more appropriate for
original PERT, a is the minimum duration of an activity; the platykurtic, bell-shaped curves (Perry and Greig 1975). Moder
and Rodgers’ result seems to be cited more frequently in the
1
Independent Res., 3009 West Hyde Ave., Visalia, CA 93291. literature, though.
Note. Discussion open until July 1, 1999. To extend the closing date The fourth step is to order the activities sequentially, from
one month, a written request must be filed with the ASCE Manager of the beginning to the end of the project, in a tabular format,
Journals. The manuscript for this paper was submitted for review and
possible publication on October 21, 1997. This paper is part of the Jour-
listing the optimistic, pessimistic, most likely, and expected
nal of Construction Engineering and Management, Vol. 125, No. 1, durations and the variances. Fifth, forward and backward
January/February, 1999. 䉷ASCE, ISSN 0733-9634/99/0001-0016 – 0022/ passes through the network are performed to identify the crit-
$8.00 ⫹ $.50 per page. Paper No. 16835. ical path, just as in the widely used critical path method. The
16 / JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / JANUARY/FEBRUARY 1999

J. Constr. Eng. Manage. 1999.125:16-22.

 (1996), indicate that ‘‘optimistic’’ and ‘‘pessimistic’’ are am-
biguous and are subject to interpretation. For example, b has
been described as having a ‘‘small chance’’ or a ‘‘one in 100’’
chance of being exceeded (Swanson and Pazer 1971). Little-
field and Randolph (1987) state that, based on past research,
‘‘people are not very good estimators of the extreme values.’’
Moder and Rodgers (1968) relax this requirement, as shown
in (3). Lau et al. (1996) further state that little is known about
modal estimation. The simplification of PERT proposed in the
present paper reduces the dependence on subjective estimation
of activity durations by decreasing the number of time esti-
FIG. 1. Project Network for Gas Pipeline Operation: a-m-b Ac- mates from three to two.
tivity Durations Second, the mean and variance of an activity duration, as
calculated using (1) and (2) or (1) and (3), are estimates of
the actual mean and variance of a beta distribution. Badiru
Downloaded from ascelibrary.org by University of Alberta on 06/18/13. Copyright ASCE. For personal use only; all rights reserved.

central limit theorem is then applied as follows: The distri-

bution of the sum of the expected durations of the activities (1991) summarizes and discusses both the estimated and actual
along the critical path is approximately normal, particularly as equations. MacCrimmon and Ryavec (1964) compute the max-
the number of activities increases. The expected duration of imum possible errors between the estimated and actual means
each sum is equal to the sum of the expected durations. Sim- and variances. If (2) is assumed to be accurate, then the max-
ilarly, the variance of each sum is the sum of the variances. imum possible error in the mean is about 33%; if (1) is pre-
These applications of the central limit theorem enable the sumed accurate, then the maximum possible error in the stan-
computation of project duration probabilities using the devia- dard deviation is about 17%. McBride and McClelland (1967),
tions from a zero mean of the standard normal variable (Z). alternatively, calculate a maximum possible error of 18.8% in
These probabilities can be critical in making financial deci- the mean when (2) is held as accurate. (The writer verified
sions about the viability of a project (Callahan et al. 1992). MacCrimmon and Ryavec’s result.) Sasieni (1986) finds no
basis for the values of the beta distribution parameters as-
EVALUATION OF PERT IN LITERATURE sumed in the derivation of (1). Littlefield and Randolph (1987)
refute Sasieni, restating the rationale given by the developers
Problems with PERT and early reviewers of PERT.
Third, the beta distribution is presumed to be applicable to
Various criticisms and proposed modifications to PERT have all project activities. Grubbs (1962) and MacCrimmon and Ry-
appeared in the literature since the early 1960s. There are five avec (1964) criticize this aspect of PERT. At the time of
recognized problems with PERT. First, it is difficult for project PERT’s development, no empirical study had been done to
engineers and planners to accurately estimate the optimistic, determine the typical distribution of activity times of repre-
most likely, and pessimistic durations of an activity. Grubbs sentative projects. MacCrimmon and Ryavec (1964) estimate
(1962) and Moder et al. (1983) note that subjective estimates the level of error introduced by an incorrect activity distribu-
of a, m, and b are based on judgment and may not be closely tion assumption. AbouRizk and Halpin (1992) show, through
related to statistical sampling of the actual times. The latter their analysis of empirical construction activity duration data,
authors note that the subjectivity is compounded by the fact that the beta distribution is appropriate. AbouRizk and Halpin
that the activity duration distribution is purely hypothetical, as (1994) develop a procedure to fit beta distributions to construc-
well, as discussed later. MacCrimmon and Ryavec (1964) cal- tion operations.
culate the sensitivity of (1) and (2) to incorrect estimates of a, Fourth, PERT considers only the critical path in computing
m, and b. Swanson and Pazer (1971), along with Lau et al. project completion time probabilities. The method ignores

FIG. 2. Beta Distributed Activity Duration Example

JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / JANUARY/FEBRUARY 1999 / 17

J. Constr. Eng. Manage. 1999.125:16-22.

 near-critical paths that possess a not in significant probability (1990) eliminates m and uses only a and b. He argues that m
of becoming critical (Callahan et al. 1992). Numerous authors is ‘‘practically useless,’’ but offers no evaluation of his new
have described this problem and have proposed solutions, in- procedure. Contrarily, other research has indicated that a and
cluding time-probability computations (Fulkerson 1962), crit- b are more difficult to estimate than is m (Littlefield and Ran-
icality indexes (Van Slyke 1963; Dodin 1984; Bandopadhyay dolph 1987; Lau et al. 1996). Finally, Lau et al. (1996) state
and Sundararajan 1987), activity time adjustments (Lindsey that the estimation of all three of PERT’s time estimates is
1972), and total project duration probability distributions subject to ambiguity. They replace a, m, and b with sets of
(Hartley and Wortham 1966; Robillard and Trahan 1977; either five or seven quantiles (the 0.25 quantile, for example,
Kamburowski 1985; Anklesaria and Drezner 1986; Sculli and is greater than 25% of the numbers in the distribution). The
Shum 1991). One of the results of missing near-critical paths estimation of quantiles is argued as being more straightforward
in PERT is a ‘‘merge event bias.’’ The magnitude of this prob- than the estimation of modes and extreme values. Their
lem, which worsens as the number of parallel paths in a net- method has merit, but the number of time estimates required
work increases, has been evaluated by a number of authors is increased from three to five or seven.
(Crandall 1976; Moder et al. 1983; Sculli 1983). The simpli-
fied PERT proposed in the present paper does not attempt to SIMPLIFIED PERT PROCEDURE
Downloaded from ascelibrary.org by University of Alberta on 06/18/13. Copyright ASCE. For personal use only; all rights reserved.

tackle the near-critical paths or the merge event bias problem.

This would be a subject for further research. The proposed simplification of PERT is to reduce the num-
Finally, Callahan et al. (1992) state that ‘‘one of the major ber of time estimates required for each task from three to two.
drawbacks of PERT in construction applications is that it re- This reduction decreases both the level of effort needed to
quires multiple time estimates, which can be time-consuming apply PERT and the required knowledge of activity durations.
to develop.’’ The authors indicate that PERT is rarely used on To retain a probabilistic procedure, the time estimates must be
construction projects. Moder et al. (1983) propose that PERT inputs to determining the expected value and variance of an
is not used either because ‘‘top project managers do not un- activity duration. The only choice is to assume that the distri-
derstand the basic principles of probability and statistics’’ or, bution of a duration is symmetric, i.e., normal, as in Fig. 3,
regardless of their understanding of the subject, ‘‘they have rather than beta. A unique normal distribution is defined by
not learned how to use PERT.’’ One of the objectives of the any given pair of mean and standard deviation values (Watson
present paper is to enhance PERT’s applicability by reducing et al. 1993). Thus, a unique normal distribution can be defined
the number of time estimates needed for its use. Project man- by any two points on one side of the curve. The other elements
agers would still need to know a few basics about probability of the PERT technique remain the same.
and statistics; however, the availability of scheduling software Given that Izuchukwu (1990) uses a and b in his procedure,
would make knowledge of the mathematics and of the intri- the question remains as to which of a, m, and b to use in the
cacies of the algorithm less essential. new, simplified procedure. Moder et al. (1983) report that most
time estimates are on the optimistic side, resulting in actual
project durations being longer than those forecast. This finding
Modifications to PERT
is congruous with Izuchukwu’s on the closeness of m to a.
Numerous authors have developed modifications to PERT, Hence, relying on a and m only may result in optimistic time
including the adjustments to (2) (Moder and Rodgers 1968; estimates. The more conservative approach is to use m and b.
Perry and Greig 1975) and criticality indexes (see prior text). Here, m, the mode, is equal to the mean, since the distribution
Some authors have, as is done in the present paper, modified is symmetric.
PERT’s time estimates. Troutt (1989) replaces the mode with The expected duration of an activity in simplified PERT can
the median in (1). He states that this produces a good estimate be determined as follows:
of the mean regardless of the probability distribution assumed. Te = m (4)
Three time estimates are still required, however. Izuchukwu

FIG. 3. Normally Distributed Activity Duration Example

18 / JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / JANUARY/FEBRUARY 1999

J. Constr. Eng. Manage. 1999.125:16-22.

 where Te = expected duration. The two variances can be com- m␤ = ␣/(␣ ⫹ ␤) (7)
puted as follows: The standard normal variable, Z, is equal to
3.44 when b is the upper bound on 100% of all durations, and ␮␤ = (␣ ⫹ 1)/(␣ ⫹ ␤ ⫹ 2) (8)
is 1.645 when b is greater than 95% of the durations. Z = (b ␴ = [(␣ ⫹ 1)(␤ ⫹ 1)]/[(␣ ⫹ ␤ ⫹ 2) (␣ ⫹ ␤ ⫹ 3)]
2
␤
2
(9)
⫺ m)/␴, so
In the preceding equations, ␣ and ␤ = parameters of the
␴2100(Te) = [(b ⫺ m)/3.44]2 (5) beta distribution. The conventional PERT equations for the
mean and variance of an activity duration are approximations.
and
For example, (1) is exact only when ␣ ⫹ ␤ = 4 (Gallagher
␴290(Te) = [(b ⫺ m)/1.645]2 (6) 1987). This restriction was established to simplify the equation
for the mean. Badiru (1991) indicates that, regardless of the
In choosing between (5) and (6) for the variance, it is rec- values of ␣ and ␤, (1) is exact when the PERT estimates are
ognized that the normal distribution is not bound on either end. symmetric about m.
Hence, b as the upper limit on 100% of the durations would Another approach, based on stronger assumptions, is to let
be, by definition, infinite. Eq. (6), therefore, is preferable. The the standard deviation of the beta distribution be equal to one-
Downloaded from ascelibrary.org by University of Alberta on 06/18/13. Copyright ASCE. For personal use only; all rights reserved.

remainder of the simplified PERT procedure is the same as sixth of the range. Then, the variance, shown in (9), is equal
that of conventional PERT. Hence, the merge event bias prob- to 1/36. Solving this equation, and using (7) and (8) to sub-
lem is not explicitly corrected. stitute terms, produces a cubic equation in terms of ␣ and m␤,
as shown in (10) (Littlefield and Randolph 1987)
TESTING AND EVALUATION
␣3 ⫹ (36m␤3 ⫺ 36m␤2 ⫹ 7m␤)␣2 ⫺ 20m␤2 ␣ ⫺ 24m␤3 = 0 (10)
Normal Distribution Assumption for
Varying m␤ between 0 and 1 produces (␣, ␤) pairs that fa-
Activity Durations
cilitate the computation of the actual mean ␮␤. Then, ␮␤ can
The first step in evaluating the simplified PERT procedure be compared to the estimated mean ␶e. To evaluate simplified
is to review the normal distribution assumption for individual PERT’s time estimate, ␮␤ can also be compared to Te from
activity durations. AbouRizk and Halpin (1992) perform a sta- (4).
tistical analysis of construction activity duration data. The au- The comparison between the actual mean of the beta distri-
thors developed a database of 71 construction activities, in- bution, the estimated mean of conventional PERT, and the
cluding various forms of loading, dumping, bulldozing, pipe mean of simplified PERT, over a 0 – 1 range [i.e., in (1), a =
jacking, and hauling. The data are not described in detail, but 0 and b = 1], is shown in Table 1. Note that this approach
the authors state that at least 20 observations of each activity requires that the original formula for the standard deviation be
were available, with over 100 observations in some samples. used — the variance is shown in (2) — rather than Moder and
Most of the sample duration distributions are associated with Rodgers’ (1968) modification. Since a different definition for
some degree of skewness (⫺1.207 ⱕ g1 ⱕ 4.869) and kurtosis the variance is used in simplified PERT, the evaluation focused
(1.008 ⱕ g2 ⱕ 31.346). The beta distribution is found to cover on comparing the means. For values of m␤ (=Te) ranging from
most of the density shapes. 0.01 to 0.99 (column 1), both ␶e and ␮␤ were computed, as
The normal distribution is associated with a skewness of 0 shown in columns 4 and 5, along with the percent differences
and a kurtosis of 3 (AbouRizk and Halpin 1992). Demenais between Te and ␮␤ and between ␶e and ␮␤, as shown in col-
et al. (1986), in discussing how major genes are analyzed in umns 6 and 7. The difference between ␶e and ␮␤ ranges from
family studies, identify three ranges of skewness for sample 16.8% at m␤ = 0.01, to 0.0% at m␤ = 0.50, to ⫺4.4% at m␤ =
sizes of 100 nuclear families: low, intermediate, and high. 0.99. By comparison, the difference between Te and ␮␤ ranges
When there is low skewness (⫺0.2 ⱕ g1 ⱕ 0.2), the sampled from 95.2% at m␤ = 0.01, to 0.0% at m␤ = 0.50, to ⫺25.1%
distribution can be considered normal. For intermediate skew- at m␤ = 0.99. The skewness of the beta distribution, computed
ness (⫺0.4 ⱕ g1 ⱕ ⫺0.2 and 0.2 ⱕ g1 ⱕ 0.4), the distribution from (14) in Appendix I, ranges from 1.01 at m␤ = 0.01, to
is nearly normal. In cases of high skewness (g1 < ⫺0.4 or g1 0.00 at m␤ = 0.50, to ⫺1.01 at m␤ = 0.99. The indication is
> 0.4), the normal distribution assumption is questionable. Of that the simplified PERT’s Te is subject to greater error than
the 71 construction activities sampled by AbouRizk and Hal- the conventional PERT’s ␶e. The error is particularly large at
pin (1992), 13 have low skewness, 10 have intermediate skew- small m␤ values, which occur when the beta distribution has
ness, and 48 have high skewness. There does not appear to be a large positive skewness.
any pattern in the degree of skewness by type of activity.
Hence, in about one-third of the cases, a normal distribution TABLE 1. Means of Simplified and Conventional PERT Activ-
assumption is reasonable. ity Durations
The indication is that the normal distribution is viable only m␤ = % Difference % Difference
part of the time. The questions, then, are: How much error Te ␣ ␤ ␶e ␮␤ ␮␤ ⫺ Te ␮␤ ⫺ ␶e g1␤
might be introduced into the mean of an activity duration by (1) (2) (3) (4) (5) (6) (7) (8)
assuming that its distribution is normal, when, in fact, it is 0.01 0.03 2.91 0.173 0.208 95.2 16.8 1.01
skewed? What cumulative effect might such errors have on 0.10 0.36 3.25 0.233 0.243 58.8 3.8 0.81
project duration probabilities? To answer these questions, a 0.20 0.89 3.57 0.300 0.293 31.7 ⫺2.4 0.59
theoretical evaluation of activity durations and a practical eval- 0.30 1.57 3.67 0.367 0.355 15.5 ⫺3.2 0.38
uation of some project durations were conducted. 0.40 2.32 3.48 0.433 0.426 6.0 ⫺1.8 0.18
0.50 3.00 3.00 0.500 0.500 0.0 0.0 0.00
0.60 3.48 2.32 0.567 0.574 ⫺4.5 1.3 ⫺0.18
Evaluation of Activity Duration Means 0.70 3.67 1.57 0.633 0.645 ⫺8.6 1.8 ⫺0.38
0.80 3.57 0.89 0.700 0.707 ⫺13.1 1.0 ⫺0.59
The beta distribution has four parameters that enable its 0.90 3.25 0.36 0.767 0.758 ⫺18.8 ⫺1.2 ⫺0.81
flexibility. For a standardized (0, 1) beta distribution, the mode 0.99 2.91 0.03 0.827 0.792 ⫺25.1 ⫺4.4 ⫺1.01
(m␤), mean (␮␤), and variance (␴2␤) are as shown in (7), (8), Note: m␤ and ␣ were determined from (10); ␤ was deduced using (7);
and (9), respectively (Benjamin and Cornell 1970; Swanson ␶e was calculated using (1) with a = 0 and b = 1; ␮␤ was computed from
and Pazer 1971) (8). Skewness g1␤ is that of beta distribution, as computed using (14).

JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / JANUARY/FEBRUARY 1999 / 19

J. Constr. Eng. Manage. 1999.125:16-22.

 It is apparent that the potential for an error in estimating the TABLE 2. Critical Path PERT Duration Data
mean of an activity duration is greater with simplified PERT Conventional Simplified
than with conventional PERT. An error of ⫾10% in the mean Merge
time of a simplified PERT activity occurs at standardized Te Project Activities events ␶e ␴290(␶e) Te ␴290(Te)
values of 0.35 and 0.75, respectively. Standardized durations (1) (2) (3) (4) (5) (6) (7)
of between 0.35 and 0.75 occur at degrees of skewness be- A 5 1 89.0 36.7 89.0 48.0
tween 0.28 and ⫺0.48. Thus, when the skewness of the dis- B 7 4 55.0 16.2 55.0 15.2
C 5 2 41.7 13.3 41.0 19.2
tribution is greater than 0.28 or less than ⫺0.48, the error is
D 10 3 58.3 29.3 58.0 29.6
greater than 10%. In AbouRizk and Halpin (1992), 51 of the E 16 4 286.2 396.6 285.0 385.1
71 construction activities studied have such skewed distribu- F 3 2 12.2 4.4 12.0 5.2
tions. Hence, applying the normal distribution to the mean of G 6 4 64.0 200.4 57.0 405.8
an activity duration distribution would be adequate in 30%, H 3 1 13.0 3.1 13.0 3.0
but inaccurate in about 70%, of all construction activities. This I 4 1 12.0 2.7 12.0 2.6
J 10 4 47.7 38.3 45.1 92.2
conclusion assumes that the activities surveyed by AbouRizk
K 3 1 7.2 2.8 7.0 3.3
and Halpin (1992) are common and would be critical in a
Downloaded from ascelibrary.org by University of Alberta on 06/18/13. Copyright ASCE. For personal use only; all rights reserved.

L 4 3 66.0 211.9 63.0 290.1

typical project.

Practical Evaluation of Conventional and Simplified alence of the two means was conducted using the Satterthwaite
PERT Project Durations test (Watson et al. 1993). This test is appropriate for two sam-
ples with unequal variances. The test is described in Appendix
The second phase in evaluating the simplified PERT is to II. The null hypothesis that the mean values of m ⫺ a and
compare its results for entire projects with those obtained us- b ⫺ m are equal was rejected. Hence, even though the con-
ing the conventional PERT. Comparisons of multiple networks dition shown in (13) occurs with the greatest frequency, b ⫺
have been used to assess the computing speed of probabilistic m > m ⫺ a in enough cases to offset this.
scheduling methods (Diaz and Hadipriono 1993). To apply a The most extreme example of the difference between Te and
similar approach, expected durations and variances were com- ␶e occurs in the network shown in Fig. 1; the simplified PERT
puted for a set of 12 project networks, each featuring a, m, expected duration is 10.9% less than that of conventional
and b estimates for the activities. One of these networks is PERT. An examination reveals that, for four of the six activ-
shown in Fig. 1. The duration data and network layouts were ities on the critical path, (b ⫺ m) = 3*(m ⫺ a). In the other
obtained from a host of sources (Van Slyke 1963; Mac- 11 project networks, these ratios are not nearly as great.
Crimmon and Ryavec 1964; Moder et al. 1983; Dodin 1984; Column 5 of Table 2 lists the project variances computed
Callahan et al. 1992; Diaz and Hadipriono 1993). using (3). These can be compared with the simplified PERT
In each of the 12 project networks, the critical path is the variances listed in column 7. The value of ␴290(␶e) ranges from
same for both PERT procedures; this does not always have to 2.7 to 396.6 while 2.6 ⱕ ␴290(Te) ⱕ 405.8. The simplified
be the case. The results are shown in Table 2. Column 1 dis- PERT variances are greater than the conventional PERT vari-
tinguishes the projects; project ‘‘G’’ is the network shown in ances in eight of the 12 networks. This indicates that, in gen-
Fig. 1. Column 2 lists the number of activities on the critical eral, the simplified PERT variances are greater than those com-
paths; the range is from 3 to 16. Column 3 lists the number puted using conventional PERT.
of merge events on the critical paths. The probability that an- Project duration probabilities are shown in Table 3. Column
other path is critical increases as the number of merge events 3 lists the project durations expected to occur with 80% prob-
increases (Moder et al. 1983). ability using conventional PERT (␶80). Column 5 lists T80 val-
Column 4 lists the expected project durations using conven- ues for simplified PERT. Columns 4 and 6 list the probabilities
tional PERT. These values can be compared with those in col- that the expected project durations will be exceeded by 10%
umn 6, which are the expected durations using simplified (P10). Column 4 lists P10 values for conventional PERT, while
PERT. In all 12 networks, Te ⱕ ␶e. The differences between column 6 lists P10 values for simplified PERT. The probabili-
the expected durations range from ⫺10.9% to zero. There are ties were calculated by applying the central limit theorem to
three possibilities for the expected durations computed by the the sums of the individual project activities. Then, the standard
two procedures, each of which can be demonstrated mathe- normal variable (Z) along with cumulative normal distribution
matically. By definition, b ⱖ m ⱖ a
(a ⫹ 4m ⫹ b)/6 = m, so a ⫹ b = 2m = m ⫹ m; TABLE 3. PERT Project Duration Probabilities

thus b ⫺ m = m ⫺ a (11) Conventional, b > 95% Simplified, b > 95%

Project Activities ␶80 P10 T80 P10
(a ⫹ 4m ⫹ b)/6 > m, so a ⫹ b > 2m or m ⫹ m;
(1) (2) (3) (4) (5) (6)
thus b ⫺ m > m ⫺ a (12) A 5 94.1 7.1% 94.8 10.0%
B 7 58.4 8.6% 58.3 7.9%
(a ⫹ 4m ⫹ b)/6 < m, so a ⫹ b < 2m or m ⫹ m; C 5 44.7 12.7% 44.7 17.5%
D 10 62.9 14.1% 62.6 14.3%
thus b ⫺ m < m ⫺ a (13) E 16 302.9 7.5% 301.5 7.3%
F 3 13.9 28.1% 13.9 29.9%
When (11) holds, a and b are symmetric about m. When G 6 75.9 32.6% 74.0 38.9%
(12) holds, the duration distribution is skewed left. When (13) H 3 14.5 23.1% 14.4 22.4%
holds, the distribution is skewed right. The a, m, and b values I 4 13.4 23.4% 13.4 22.8%
of the 172 activities included in the 12 project networks were J 10 52.8 22.1% 53.2 31.9%
K 3 8.6 33.5% 8.5 35.1%
studied. Izuchukwu (1990) states that a and m are typically
L 4 78.3 32.5% 77.3 35.6%
closer together than are m and b. This actually occurs in only
67 of the 172 activities. However, the overall average values Note: ␶80, T80 = project duration estimated to occur with 80% proba-
bility. P10 = probability that project’s duration is 10% longer than ex-
of m ⫺ a and b ⫺ m are 3.05 and 4.16, respectively, with pected duration.
variances of 10.06 and 21.67. A hypothesis test on the equiv-
20 / JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / JANUARY/FEBRUARY 1999

J. Constr. Eng. Manage. 1999.125:16-22.

 TABLE 4. Hypothesis Tests on Activity Durations and Variances
Statistic X1 X2 s21 s22 sX1⫺X2 tcalc ␯ t␣/2,␯
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Expected duration 9.10 8.91 86.93 86.21 1.10 0.185 171 1.98
Variance, b > 95% 14.42 10.19 1,379.24 651.87 3.44 1.229 171 1.98
Note: Symbols and equations are detailed in Appendices II and IV.

tables were used to compute probabilities. The T80 values range hypothesis tests on 172 project activities. These results are in
from 2.5% less than to 0.8% greater than ␶80. Seven of the ␶80 contrast to those obtained in the first phase of the evaluation.
values are greater than, three are equal to, and two are less The implication is that the 12 networks tested did not include
than the T80 values. Eight of the simplified PERT P10 values activities with highly skewed duration distributions. If the 12
are greater than those of conventional PERT, while the re- project networks are truly representative, then the simplified
maining four are less. PERT produces results similar to those obtained with conven-
Downloaded from ascelibrary.org by University of Alberta on 06/18/13. Copyright ASCE. For personal use only; all rights reserved.

It is difficult to draw firm conclusions from this analysis of tional PERT, but with less effort.
the variances and probabilities of simplified and conventional Further analysis of the simplified PERT procedure is rec-
PERT durations in the 12 project networks. The indication is ommended, possibly on other networks, and via simulation of
that simplified PERT produces shorter expected project dura- activity durations on a single network. The new procedure
tions, but greater project duration variances, than does con- does not explicitly improve upon PERT’s merge event bias
ventional PERT. The effects of these cancel each other, such problem. Further research is suggested toward developing a
that the project duration probabilities of simplified PERT are probabilistic procedure that is easy to apply and that eliminates
not much different from those of conventional PERT. Thus, in merge event bias and the ignorance of near-critical paths. Fur-
these and similar project networks, results at least as reliable ther analysis of construction activity duration distributions is
as those of conventional PERT can be obtained using the sim- also suggested. If the critical activities in a given project fea-
pler technique. ture highly skewed duration distributions, then simplified
To extend the analysis, the durations of the 172 activities PERT may be inappropriate. An alternative might be to use a
comprising the 12 networks were examined. Two null hypoth- fixed, positively skewed distribution, defined by two param-
eses were tested: that the means of the 172 expected durations eters, for all activities. A survey of practitioners is suggested
computed using conventional and simplified PERT are equal, to ascertain the usefulness of the new procedure. To facilitate
and that the means of the 172 variances with b > 95% of the its usage, the simplified PERT algorithm could be embedded
durations are equal. The Satterthwaite test was used (see Ap- into construction scheduling software.
pendix II).
The results are shown in Table 4. The null hypotheses that APPENDIX I. DEGREE OF SKEWNESS OF
the means of both the expected durations and the variances BETA DISTRIBUTION
are equal are not rejected in any case. The overall conclusion
is that the conventional and simplified PERTs produce similar The degree of skewness of a standardized (0, 1) beta dis-
values for activity durations and variances. tribution can be computed as follows:

CONCLUSIONS
A simplified version of PERT has been developed. The new
g1␤ =
1 (␣ ⫹ 1)
␴ (␣ ⫹ ␤ ⫹ 2)
3 冋 (␣ ⫹ 3)(␣ ⫹ 2)
(␣ ⫹ ␤ ⫹ 4)(␣ ⫹ ␤ ⫹ 3)
technique reduces the level of effort required by conventional
PERT because only two time estimates, rather than three, are
required for each activity. The remainder of the procedure is
⫺
3(␣ ⫹ 1)(␣ ⫹ 2)
(␣ ⫹ ␤ ⫹ 2)(␣ ⫹ ␤ ⫹ 3)
⫹
2(␣ ⫹ 1)2
(␣ ⫹ ␤ ⫹ 2)2
册 (14)
identical to the conventional method. The reduced effort may The parameters of the beta distribution are ␣ and ␤, while ␴
result in a significant time savings for large projects in which = standard deviation.
there are many tasks. Two evaluations of the new method were
conducted. In the first phase, the estimated means of a range APPENDIX II. SATTERTHWAITE TEST
of modal duration values were computed for both simplified
and conventional PERT. The estimated means were compared The Satterthwaite test (Aspin test; Welch test) is applied to
with actual means obtained from the standardized beta distri- a hypothesis test on two means when the samples have un-
bution. The activity duration means computed using simplified equal or separate variances (Watson et al. 1993). The test is
PERT are subject to greater error than are those computed as follows:
using conventional PERT, especially when the distribution is
H0: ␮1 = ␮2 and Ha: ␮1 ≠ ␮2 (15)
highly skewed. At degrees of skewness between 0.28 and
⫺0.48, the error in the mean would be less than or equal to Reject H0 if tcalc < ⫺t␣/2,␯ or tcalc > t␣/2,␯, where (16)
10% of the actual value. Based on the AbouRizk and Halpin
(1992) study, about 30% of all construction activities have ¯1 ⫺ X
tcalc = (X ¯ 2)/sX1⫺X2
¯ ¯ (17)
skewness levels within this range.
¯ = [(s1 ⫹ s2)/n] , and
2 2 0.5
This second phase of the evaluation featured the computa- sX1⫺X2
¯ (18)
tion of project durations for 12 networks using both simplified ␯=n⫺1 (19)
and conventional PERT. The expected durations and variances
of individual project activities, computed using simplified In (15) – (19), H0 = null hypothesis; Ha = alternative hy-
PERT, are essentially equal to those computed using the con- pothesis; ␮ = population mean; t = value of the statistic from
ventional procedure. Simplified PERT produces shorter project Student’s t distribution; ␣ indicates the confidence level; ␯ =
durations, but greater project duration variances, than does degrees of freedom; X̄ = sample mean; s2 = sample variance;
conventional PERT. The combination of these two effects re- and s X1⫺X2 = pooled sample standard deviation. Eqs. (18) and
sults in similar project duration probabilities. These conclu- (19) are valid when the number of observations in each sample
sions are drawn from the analysis of the networks and via is the same.
JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / JANUARY/FEBRUARY 1999 / 21

J. Constr. Eng. Manage. 1999.125:16-22.

 ACKNOWLEDGMENTS Moder, J. J., and Rodgers, E. G. (1968). ‘‘Judgement estimates of the
moments of PERT type distributions.’’ Mgmt. Sci., 15(2), B76 – B83.
The present paper is an expanded and significantly enhanced version Perry, C., and Greig, I. D. (1975). ‘‘Estimating the mean and variance of
of a term project that the writer completed for a planning and scheduling subjective distributions in PERT and decision analysis.’’ Mgmt. Sci.,
course, instructed by Dr. Hosin Lee, in the Department of Civil and En- 21(12), 1477 – 1480.
vironmental Engineering at the University of Utah in Salt Lake City. Fig. Robillard, P., and Trahan, M. (1977). ‘‘The completion time of PERT
1 was extracted from his course materials. networks.’’ Operations Res., 25(1), 15 – 29.
Sasieni, M. W. (1986). ‘‘A note on PERT times.’’ Mgmt. Sci., 32(12),
1652 – 1653.
APPENDIX III. REFERENCES Sculli, D. (1983). ‘‘The completion time of PERT networks.’’ J. Opera-
tional Res. Soc., 34(2), 155 – 158.
AbouRizk, S. M., and Halpin, D. W. (1992). ‘‘Statistical properties of Sculli, D., and Shum, Y. W. (1991). ‘‘An approximate solution to the
construction duration data.’’ J. Constr. Engrg. and Mgmt., ASCE, PERT problem.’’ Comp. and Mathematics Applications, 21(8), 1 – 7.
118(3), 525 – 544. Swanson, L. A., and Pazer, H. L. (1971). ‘‘Implications of the underlying
AbouRizk, S. M., and Halpin, D. W. (1994). ‘‘Fitting beta distributions assumptions of PERT.’’ Decision Sci., 2(October), 461 – 480.
based on sample data.’’ J. Constr. Engrg. and Mgmt., ASCE, 120(2), Troutt, M. D. (1989). ‘‘On the generality of the PERT average time for-
288 – 305. mula.’’ Decision Sci., 20, 410 – 412.
Downloaded from ascelibrary.org by University of Alberta on 06/18/13. Copyright ASCE. For personal use only; all rights reserved.

Anklesaria, K. P., and Drezner, Z. (1986). ‘‘A multivariate approach to Van Slyke, R. M. (1963). ‘‘Monte Carlo methods and the PERT prob-
estimating the completion time for PERT networks.’’ J. Operational lem.’’ Operations Res., 11(5), 839 – 860.
Res. Soc., 37(8), 811 – 815. Watson, C. J., Billingsley, P., Croft, D. J., and Huntsberger, D. V. (1993).
Badiru, A. B. (1991). ‘‘A simulation approach to PERT network analy- Statistics for management and economics, 5th Ed., Allyn and Bacon,
sis.’’ Simulation, 57(4), 245 – 255. Boston.
Bandopadhyay, S., and Sundararajan, A. (1987). ‘‘Simulation of a long-
wall development — Extraction network.’’ CIM Bull., 80(903), 62 – 70.
Benjamin, J. R., and Cornell, C. A. (1970). Probability, statistics, and APPENDIX IV. NOTATION
decision for civil engineers. McGraw-Hill, New York.
Callahan, M. T., Quackenbush, D. G., and Rowings, J. E. (1992). Con- The following symbols are used in this paper:
struction project scheduling. McGraw-Hill, New York.
Clark, C. E. (1962). ‘‘The PERT model for the distribution of an activity a = optimistic activity duration; first shape parameter of beta
time.’’ Operations Res., 10, 405 – 406. distribution;
Crandall, K. C. (1976). ‘‘Probabilistic time scheduling.’’ J. Constr. Div., b = pessimistic activity duration; third shape parameter of
ASCE, 102(3), 415 – 423. beta distribution;
Demenais, F., Lathrop, M., and Lalouel, J. M. (1986). ‘‘Robustness and g1 = sample skewness coefficient;
power of the unified model in the analysis of quantitative measure-
ments.’’ Am. J. Human Genetics, 38(2), 228 – 234.
g1␤ = skewness coefficient of standardized beta distribution;
Diaz, C. F., and Hadipriono, F. C. (1993). ‘‘Nondeterministic networking g2 = sample kurtosis coefficient;
methods.’’ J. Constr. Engrg. and Mgmt., ASCE, 119(1), 40 – 57. m = most likely or modal activity duration;
Dodin, B. (1984). ‘‘Determining the K most critical paths in PERT net- m␤ = mode of beta distribution;
works.’’ Operations Res., 32(4), 859 – 877. n = number of observations in sample;
Fulkerson, D. R. (1962). ‘‘Expected critical path lengths in PERT net- P10 = probability that project’s duration is 10% longer than
works.’’ Operations Res., 10(6), 808 – 817. expected duration;
Gallagher, C. (1987). ‘‘A note on PERT assumptions.’’ Mgmt. Sci., s2 = variance of sample of values;
33(10), 1360. s X1⫺X2 = pooled sample standard deviation;
Grubbs, F. E. (1962). ‘‘Attempts to validate certain PERT statistics or
‘picking on PERT.’ ’’ Operations Res., 10(6), 912 – 915.
Te = expected duration time of activity using simplified PERT
Hartley, H. O., and Wortham, A. W. (1966). ‘‘A statistical theory for PERT procedure;
critical path analysis.’’ Mgmt. Sci., 12(10), B469 – B481. T80 = simplified PERT project duration estimated to occur
Izuchukwu, J. I. (1990). ‘‘Project management: Shortening the critical with 80% probability;
path.’’ Mech. Engrg., 112(2), 59 – 60. t, tcalc = value of statistic in Student’s t distribution;
Kamburowski, J. (1985). ‘‘Normally distributed activity durations in X̄ = mean of sample of values;
PERT networks.’’ J. Operational Res. Soc., 36(11), 1051 – 1057. Z = standard normal variable;
Lau, A. H.-L., Lau, H.-S., and Zhang, Y. (1996). ‘‘A simple and logical ␣ = parameter of beta distribution; 1 ⫺ (desired statistical
alternative for making PERT time estimates.’’ IIE Trans., 28, 183 – confidence level/100);
192.
Lindsey, J. H. II. (1972). ‘‘An estimate of expected critical-path length
␤ = parameter of beta distribution;
in PERT networks.’’ Operations Res., 20(4), 800 – 812. ␮ = population mean;
Littlefield, T. K., and Randolph, P. H. (1987). ‘‘An answer to Sasieni’s ␮␤ = mean of beta distribution;
question on PERT times.’’ Mgmt. Sci., 33(10), 1357 – 1359. ␯ = degrees of freedom;
MacCrimmon, K. R., and Ryavec, C. A. (1964). ‘‘An analytical study of ␴2␤ = variance of beta distribution;
the PERT assumptions.’’ Operations Res., 12(1), 16 – 37. ␴290 = variance of activity’s duration when a and b are greater
Malcolm, D. G., Roseboom, J. H., Clark, C. E., and Fazar, W. (1959). than 5% and 95%, respectively, of activity’s duration;
‘‘Application of a technique for research and development program ␴2100 = variance of activity’s duration when a and b are lower
evaluation.’’ Operations Res., 11(5), 646 – 669. and upper bounds, respectively, on activity’s possible
McBride, W. J. Jr., and McClelland, C. W. (1967). ‘‘Pert and the beta
distribution.’’ IEEE Trans. on Engrg. Mgmt., Piscataway, N.J., 14(4),
durations;
166 – 169. ␶e = expected duration time of activity using conventional
Moder, J. J., Phillips, C. R., and Davis, E. W. (1983). Project management PERT procedure; and
with CPM, PERT and precedence diagramming, 3rd Ed., Van Nostrand ␶80 = conventional PERT duration estimated to occur with
Reinhold, New York. 80% probability.

22 / JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / JANUARY/FEBRUARY 1999

J. Constr. Eng. Manage. 1999.125:16-22.