Skip to content

Lesson 20: PERT and Probabilistic Project Scheduling

20.1 Learning Outcomes

By the end of this hour a student will be able to:

  1. Explain the three time estimates and the Beta distribution assumption.
  2. Compute the expected time \(t_e\) and variance \(\sigma^2\) of an activity.
  3. Compute the project variance and apply the Central Limit Theorem.
  4. Compute the probability of completion by a target date, and the date for a given confidence.
  5. State PERT's limitations, especially the merge-event bias.

20.2 Why PERT?

CPM assumes durations are known. For novel work — R&D, construction, software — they are not. PERT (Program Evaluation and Review Technique, US Navy, 1958) models them as random variables.

\[ \boxed{\ \textbf{CPM asks "how long?" PERT asks "how likely?"}\ } \]

20.3 The Three Time Estimates

For each activity, obtain from the person who will do the work:

Estimate Symbol Meaning
Optimistic \(a\) everything goes right; ~1 % chance of being this fast
Most likely \(m\) the modal duration under normal conditions
Pessimistic \(b\) everything that can reasonably go wrong does; ~1 % chance of being this slow

Always \(a \le m \le b\).

The Beta distribution assumption

PERT assumes activity duration follows a Beta distribution on \([a,b]\) with mode \(m\). The Beta is chosen because it is:

  • bounded (unlike the normal, a task cannot take negative time or infinite time),
  • flexible in shape — it can be symmetric or skewed either way,
  • unimodal, matching how people estimate.

Real activity durations are typically right-skewed: things can go badly wrong (long tail) more easily than they can go spectacularly right.

The formulas

\[ \boxed{\ t_e \ =\ \frac{a + 4m + b}{6} \qquad \textbf{(expected time)}\ } \]
\[ \boxed{\ \sigma^2 \ =\ \left(\frac{b-a}{6}\right)^{2} \qquad \textbf{(variance)}\ } \]

Where they come from.

  • \(t_e\) is a weighted average giving the mode four times the weight of either extreme — a standard approximation to the Beta mean.
  • \(\sigma\) uses the six-sigma range heuristic: for many distributions, essentially the whole range spans about \(6\sigma\), so \(\sigma \approx (b-a)/6\).

Note that \(\sigma\) depends ONLY on the spread \(b-a\), not on \(m\). Two activities with the same range have the same variance regardless of skew.

20.4 Project Analysis

The procedure

1. For each activity compute t_e and sigma^2
2. Run CPM (L19) using t_e as the duration
3. Identify the critical path; T_e = its length
4. Project variance = SUM of the variances of the CRITICAL activities only
5. Project standard deviation = sqrt(project variance)
6. Apply the normal approximation for probability questions

Why the variances add

The critical-path length is a sum of activity durations. For independent random variables,

\[ \mathrm{Var}\left(\sum_i X_i\right) = \sum_i \mathrm{Var}(X_i) \]

Variances add; standard deviations do NOT.

\[ \boxed{\ \sigma_{\text{project}} = \sqrt{\sum_{i\in\text{critical path}} \sigma_i^{2}}\ } \]

The normal approximation

By the Central Limit Theorem, the sum of several independent random variables is approximately normal, whatever their individual distributions. So

\[ T \ \sim\ N\left(T_e,\ \sigma^2_{\text{project}}\right) \]

Probability of completion by a target date \(D\):

\[ \boxed{\ z = \frac{D - T_e}{\sigma_{\text{project}}}, \qquad \Pr[T \le D] = \Phi(z)\ } \]

Date for a given confidence \(p\):

\[ \boxed{\ D = T_e + z_p\,\sigma_{\text{project}}\ } \]
Confidence \(z_p\)
50 % 0.000
75 % 0.674
90 % 1.282
95 % 1.645
99 % 2.326

20.5 Worked Example

Activity \(a\) \(m\) \(b\) Predecessors
A 2 4 6 —
B 3 5 13 —
C 1 3 5 A
D 4 6 14 A
E 2 3 10 B, C
F 3 5 7 D
G 1 2 3 E, F

Expected times and variances

Activity \(a\) \(m\) \(b\) \(t_e = \dfrac{a+4m+b}{6}\) \(\sigma^2 = \left(\dfrac{b-a}{6}\right)^2\)
A 2 4 6 \(\dfrac{2+16+6}{6} = \mathbf{4}\) \(\left(\dfrac{4}{6}\right)^2 = \mathbf{0.444}\)
B 3 5 13 \(\dfrac{3+20+13}{6} = \mathbf{6}\) \(\left(\dfrac{10}{6}\right)^2 = \mathbf{2.778}\)
C 1 3 5 \(\dfrac{1+12+5}{6} = \mathbf{3}\) \(\left(\dfrac{4}{6}\right)^2 = \mathbf{0.444}\)
D 4 6 14 \(\dfrac{4+24+14}{6} = \mathbf{7}\) \(\left(\dfrac{10}{6}\right)^2 = \mathbf{2.778}\)
E 2 3 10 \(\dfrac{2+12+10}{6} = \mathbf{4}\) \(\left(\dfrac{8}{6}\right)^2 = \mathbf{1.778}\)
F 3 5 7 \(\dfrac{3+20+7}{6} = \mathbf{5}\) \(\left(\dfrac{4}{6}\right)^2 = \mathbf{0.444}\)
G 1 2 3 \(\dfrac{1+8+3}{6} = \mathbf{2}\) \(\left(\dfrac{2}{6}\right)^2 = \mathbf{0.111}\)

CPM on the expected times

Activity ES EF LS LF TF Critical
A 0 4 0 4 0 ✓
B 0 6 5 11 5
C 4 7 8 11 4
D 4 11 4 11 0 ✓
E 7 11 11 15 4
F 11 16 11 16 0 ✓
G 16 18 16 18 0 ✓
\[ \boxed{\ \textbf{Critical path: } A\to D\to F\to G, \qquad T_e = 4+7+5+2 = \mathbf{18\ \text{weeks}}\ } \]

Project variance

Sum the variances of CRITICAL activities only.

\[ \sigma^2_{\text{project}} = \sigma^2_A + \sigma^2_D + \sigma^2_F + \sigma^2_G \]
\[ = 0.444 + 2.778 + 0.444 + 0.111 = \mathbf{3.778} \]
\[ \boxed{\ \sigma_{\text{project}} = \sqrt{3.778} = \mathbf{1.944\ \text{weeks}}\ } \]

Note what dominates. Activity D contributes 2.778 of the 3.778 — 73.5 % of the total variance — because its range is \(14-4 = 10\) weeks. D is the single largest source of schedule risk, and that is where risk-reduction effort belongs.

Probability questions

(a) Probability of finishing within 20 weeks.

\[ z = \frac{20 - 18}{1.944} = \mathbf{1.029} \]
\[ \Pr[T\le20] = \Phi(1.029) = \mathbf{0.848} = \mathbf{84.8\%} \]

(b) Probability of finishing within 22 weeks.

\[ z = \frac{22-18}{1.944} = 2.058 \qquad\Longrightarrow\qquad \Pr = \Phi(2.058) = \mathbf{98.0\%} \]

(c) Date for 95 % confidence.

\[ D = T_e + 1.645\sigma = 18 + 1.645(1.944) = 18 + 3.198 = \mathbf{21.2\ \text{weeks}} \]
\[ \boxed{\ \textbf{Promise 22 weeks to be 95 \% confident; the expected duration is only 18.}\ } \]

(d) Probability of finishing within the expected duration of 18 weeks.

\[ z = \frac{18-18}{1.944} = 0 \qquad\Longrightarrow\qquad \Pr = \Phi(0) = \mathbf{50\%} \]
\[ \boxed{\ \textbf{A project promised at its EXPECTED duration has only a 50 \% chance of success.}\ } \]

This is the single most important practical lesson in PERT, and the most widely ignored in practice.

20.6 Limitations of PERT

The merge-event bias

PERT computes the variance of the critical path only. But a near-critical path may overtake it.

Concrete illustration. Suppose two paths have \(T_e = 18\) and \(T_e = 17\), each with \(\sigma = 2\). PERT reports \(T_e = 18\). But the project finishes when both are done:

\[ T_{\text{project}} = \max(T_1,\ T_2) \]

and \(E[\max(T_1,T_2)] > \max(E[T_1], E[T_2]) = 18\) — by Jensen's inequality, since \(\max\) is convex.

\[ \boxed{\ \textbf{PERT is OPTIMISTICALLY BIASED. The true expected duration exceeds $T_e$.}\ } \]

The bias grows with the number of near-critical paths. For the illustration above the bias is roughly \(0.5\)–\(1\) week; with five near-critical paths it can be several weeks.

The remedy: Monte Carlo simulation. Sample each activity's duration from its distribution, compute the network's completion time, repeat 10 000 times, and read the empirical distribution. This handles merge bias, correlated durations and non-Beta shapes automatically, and is what modern project software does.

Other limitations

Limitation Consequence
Beta assumption is arbitrary the \(\frac{a+4m+b}{6}\) weights have no deep justification
Estimates are subjective optimism bias inflates confidence; \(b\) is usually understated
Independence assumed in reality a shared resource or bad weather correlates delays
Only the critical path's variance counted merge bias (§5.1)
Durations assumed fixed once started no learning, no mid-course correction
\[ \boxed{\ \textbf{PERT's greatest value is forcing explicit discussion of uncertainty, not the third decimal of a probability.}\ } \]

20.7 Solved Problems

Problem 1 — Computing t_e and σ

For activities with \((a,m,b)\) as below, compute \(t_e\), \(\sigma^2\) and \(\sigma\), and comment on the skew.

Activity \(a\) \(m\) \(b\)
P 5 8 11
Q 2 4 12
R 6 6 6
S 3 10 11

Solution.

Activity P (symmetric: \(m\) is the midpoint of \([5,11]\)).

\[ t_e = \frac{5+32+11}{6} = \frac{48}{6} = \mathbf{8}, \qquad \sigma^2 = \left(\frac{6}{6}\right)^2 = \mathbf{1}, \qquad \sigma = \mathbf{1} \]

\(t_e = m = 8\) — symmetric, so mean = mode.

Activity Q (right-skewed: long pessimistic tail).

\[ t_e = \frac{2+16+12}{6} = \frac{30}{6} = \mathbf{5}, \qquad \sigma^2 = \left(\frac{10}{6}\right)^2 = \mathbf{2.778}, \qquad \sigma = \mathbf{1.667} \]

\(t_e = 5 > m = 4\) — the long tail pulls the mean above the mode, exactly as it should.

Activity R (certain).

\[ t_e = \frac{6+24+6}{6} = \mathbf{6}, \qquad \sigma^2 = 0, \qquad \sigma = \mathbf{0} \]

No uncertainty — a deterministic activity, which PERT handles as a special case ✓.

Activity S (left-skewed: long optimistic tail).

\[ t_e = \frac{3+40+11}{6} = \frac{54}{6} = \mathbf{9}, \qquad \sigma^2 = \left(\frac{8}{6}\right)^2 = \mathbf{1.778}, \qquad \sigma = \mathbf{1.333} \]

\(t_e = 9 < m = 10\) — the mean is pulled below the mode.

Summary and interpretation.

Activity \(t_e\) \(m\) Skew \(\sigma\) Risk
P 8 8 none 1.000 moderate
Q 5 4 right 1.667 highest
R 6 6 none 0.000 none
S 9 10 left 1.333 moderate
\[ \boxed{\ \textbf{Q is the riskiest: largest } \sigma \textbf{ AND right-skewed (bad surprises more likely than good).}\ } \]

A caution about relying on \(\sigma\) alone. \(\sigma\) depends only on \(b - a\), so it treats right-skew and left-skew identically. Activity Q (right-skewed) is operationally far more dangerous than a left-skewed activity with the same \(\sigma\), because overruns are the outcome that hurts. Read the skew as well as the variance.

Problem 2 — Full probabilistic analysis

A project has critical path with activities and estimates below. Compute \(T_e\), \(\sigma\), and answer the probability questions.

Activity \(a\) \(m\) \(b\)
1 4 6 8
2 6 9 18
3 2 5 8
4 5 7 15

Solution.

Step 1 — expected times and variances.

Act \(t_e\) \(\sigma^2\)
1 \(\dfrac{4+24+8}{6} = \mathbf{6}\) \(\left(\dfrac{4}{6}\right)^2 = 0.444\)
2 \(\dfrac{6+36+18}{6} = \mathbf{10}\) \(\left(\dfrac{12}{6}\right)^2 = 4.000\)
3 \(\dfrac{2+20+8}{6} = \mathbf{5}\) \(\left(\dfrac{6}{6}\right)^2 = 1.000\)
4 \(\dfrac{5+28+15}{6} = \mathbf{8}\) \(\left(\dfrac{10}{6}\right)^2 = 2.778\)
\[ T_e = 6+10+5+8 = \mathbf{29\ \text{weeks}} \]
\[ \sigma^2_{\text{proj}} = 0.444+4.000+1.000+2.778 = \mathbf{8.222} \]
\[ \sigma_{\text{proj}} = \sqrt{8.222} = \mathbf{2.867\ \text{weeks}} \]

Step 2 — probability questions.

(a) \(\Pr[T\le 32]\).

\[ z = \frac{32-29}{2.867} = 1.046 \qquad\Longrightarrow\qquad \Phi(1.046) = \mathbf{85.2\%} \]

(b) \(\Pr[T\le 27]\).

\[ z = \frac{27-29}{2.867} = -0.698 \qquad\Longrightarrow\qquad \Phi(-0.698) = \mathbf{24.3\%} \]

(c) \(\Pr[29 \le T \le 33]\).

\[ z_1 = 0, \quad z_2 = \frac{33-29}{2.867} = 1.395 \]
\[ \Pr = \Phi(1.395) - \Phi(0) = 0.9185 - 0.5 = \mathbf{41.9\%} \]

(d) The 90 %-confidence date.

\[ D = 29 + 1.282(2.867) = 29 + 3.676 = \mathbf{32.7\ \text{weeks}} \]

(e) The 99 %-confidence date.

\[ D = 29 + 2.326(2.867) = 29 + 6.669 = \mathbf{35.7\ \text{weeks}} \]

Step 3 — the risk profile.

Promise Probability of meeting it
27 weeks 24 %
29 weeks (= \(T_e\)) 50 %
32 weeks 85 %
32.7 weeks 90 %
34 weeks 96 %
35.7 weeks 99 %
\[ \boxed{\ \textbf{Buying confidence from 50 \% to 90 \% costs 3.7 weeks — a 13 \% schedule buffer.}\ } \]

Where the risk lies. Activity 2 contributes \(4.000/8.222 = \mathbf{48.6\%}\) of the total variance, activity 4 another 33.8 %. Together they account for 82 % of the schedule risk while being only 62 % of the duration.

\[ \boxed{\ \textbf{Reducing the uncertainty on activities 2 and 4 is worth far more than shortening activities 1 and 3.}\ } \]

Concretely: if better planning narrowed activity 2's range from \([6,18]\) to \([7,13]\), its variance would fall from 4.000 to \(\left(\frac{6}{6}\right)^2 = 1.000\), giving

\[ \sigma_{\text{new}} = \sqrt{8.222 - 3.000} = \sqrt{5.222} = 2.285 \]

and the 90 % date would drop to \(29 + 1.282(2.285) = \mathbf{31.9}\) weeks — 0.8 weeks of buffer saved by reducing uncertainty, not duration.

Problem 3 — The merge-event bias, quantified

A project has two parallel paths, each ending at the final event: path 1 with \(T_e = 20\), \(\sigma = 3\); path 2 with \(T_e = 19\), \(\sigma = 3\). What does PERT report, and what is the truth?

Solution.

What PERT reports. The critical path is path 1 (\(T_e = 20\), the longer).

\[ T_e = 20, \qquad \sigma = 3 \]
\[ \Pr[T\le20] = \Phi(0) = \mathbf{50\%} \]

The truth. The project finishes when both paths are done:

\[ T = \max(T_1,\ T_2) \]
\[ \Pr[T \le 20] = \Pr[T_1\le20]\cdot\Pr[T_2\le20] \quad \text{(assuming independence)} \]
\[ \Pr[T_1\le20] = \Phi\left(\frac{20-20}{3}\right) = \Phi(0) = 0.500 \]
\[ \Pr[T_2\le20] = \Phi\left(\frac{20-19}{3}\right) = \Phi(0.333) = 0.630 \]
\[ \boxed{\ \Pr[T\le20] = 0.500\times0.630 = \mathbf{0.315} = \mathbf{31.5\%}\ } \]
\[ \boxed{\ \textbf{PERT says 50 \%; the truth is 31.5 \%. PERT is optimistic by 18.5 percentage points.}\ } \]

The magnitude of the bias in time terms. Solve for the date giving a true 50 % probability:

\[ \Phi\left(\frac{D-20}{3}\right)\cdot\Phi\left(\frac{D-19}{3}\right) = 0.5 \]

Trying \(D = 21.5\): \(\Phi(0.5)\Phi(0.833) = 0.691\times0.798 = 0.551\). Trying \(D = 21\): \(\Phi(0.333)\Phi(0.667) = 0.630\times0.748 = 0.471\).

\[ \boxed{\ \textbf{The true median completion is about 21.3 weeks, not 20 — a bias of ~1.3 weeks.}\ } \]

How the bias grows. With \(k\) independent near-critical paths each having 50 % chance of meeting \(T_e\):

\[ \Pr[\text{all done by } T_e] = 0.5^{k} \]
Number of paths at \(T_e\) \(\Pr[\text{on time}]\)
1 50 %
2 25 %
3 12.5 %
5 3.1 %
10 0.1 %
\[ \boxed{\ \textbf{A project with 5 equally critical paths has a 3 \% chance of meeting its PERT date.}\ } \]

The practical remedies.

  1. Monte Carlo simulation — the correct answer, and standard in project software.
  2. Add an explicit buffer — Goldratt's critical chain method removes safety from individual activities and pools it into a single project buffer, sized at roughly \(\frac12\sqrt{\sum\sigma^2}\) across all paths.
  3. Count near-critical paths as a risk indicator. Report "this project has 4 paths within 2 weeks of critical" alongside the PERT date.
\[ \boxed{\ \textbf{Never quote a PERT probability without checking how many paths are near-critical.}\ } \]

Problem 4 — Setting a defensible deadline

Using the §4 example (\(T_e = 18\), \(\sigma = 1.944\)), a client demands a fixed delivery date with a penalty of ₹5 lakh per week late. Internal cost of finishing early is ₹1 lakh per week of idle resources. What date should be promised?

Solution.

The trade-off. Promising date \(D\):

\[ E[\text{cost}] = 5\,E[\max(T-D,\ 0)] + 1\cdot E[\max(D-T,\ 0)] \]

This is the newsvendor structure (which reappears in inventory theory, L32). The optimal \(D\) satisfies

\[ \Pr[T \le D] = \frac{c_u}{c_u+c_o} \]

where \(c_u = 5\) (underage: penalty per week late) and \(c_o = 1\) (overage: idle cost per week early).

\[ \boxed{\ \Pr[T\le D^{*}] = \frac{5}{5+1} = \frac56 = 0.8333\ } \]

Find \(D^{*}\). From the normal table, \(\Phi(z) = 0.8333\) gives \(z = 0.967\).

\[ D^{*} = T_e + 0.967\sigma = 18 + 0.967(1.944) = 18 + 1.880 = \mathbf{19.88 \approx 20\ \text{weeks}} \]
\[ \boxed{\ \textbf{Promise 20 weeks.}\ } \]

Verify the economics at \(D = 20\).

\[ \Pr[\text{late}] = 1 - \Phi\left(\frac{20-18}{1.944}\right) = 1 - \Phi(1.029) = 1-0.848 = \mathbf{15.2\%} \]

Expected weeks late (for a normal, \(E[\max(T-D,0)] = \sigma\left[\phi(z) - z(1-\Phi(z))\right]\)):

\[ \phi(1.029) = 0.2349, \qquad E[\text{late}] = 1.944\left[0.2349 - 1.029(0.152)\right] \]
\[ = 1.944(0.2349 - 0.1564) = 1.944(0.0785) = \mathbf{0.153\ \text{weeks}} \]
\[ \text{Expected penalty} = 5\times0.153 = ₹0.76\ \text{lakh} \]

Compare with alternative promises.

Promise \(D\) \(\Pr[\text{late}]\) \(E[\text{penalty}]\) \(E[\text{idle cost}]\) Total
18 50.0 % ₹3.88 L ₹0.78 L ₹4.65 L
19 30.3 % ₹1.88 L ₹1.38 L ₹3.26 L
20 15.2 % ₹0.77 L ₹2.15 L ₹2.92 L ←
21 6.1 % ₹0.26 L ₹3.05 L ₹3.31 L
22 2.0 % ₹0.07 L ₹4.01 L ₹4.08 L
\[ \boxed{\ \textbf{20 weeks minimises expected total cost at ₹2.92 lakh.}\ } \]

Advice to the negotiator.

"Promise 20 weeks. Our expected duration is 18, so this gives a 2-week buffer and an 85 % chance of delivering on time. The expected penalty exposure is ₹0.77 lakh. Promising 18 weeks would look attractive but carries a 50 % chance of penalty and an expected cost 63 % higher."

A crucial caveat. This uses PERT's \(\sigma = 1.944\), which ignores merge bias (§5.1). If the project has near-critical paths, the true \(\sigma\) is larger and the true \(T_e\) higher.

\[ \boxed{\ \textbf{Recommend running a Monte Carlo simulation before signing a penalty clause.}\ } \]

Problem 5 — A complete PERT study with risk management

A pharmaceutical firm is planning a clinical trial. Analyse and produce a risk-management plan.

Activity Description \(a\) \(m\) \(b\) Predecessors
A Protocol design 4 6 14 —
B Ethics approval 8 12 28 A
C Site selection 3 5 7 A
D Patient recruitment 12 20 40 B, C
E Treatment phase 24 26 34 D
F Data analysis 4 6 8 E
G Report writing 3 4 11 F

Solution.

Step 1 — expected times and variances.

Act \(a\) \(m\) \(b\) \(t_e\) \(\sigma^2\) \(\sigma\)
A 4 6 14 \(\dfrac{4+24+14}{6} = \mathbf{7}\) \(\left(\dfrac{10}{6}\right)^2 = 2.778\) 1.667
B 8 12 28 \(\dfrac{8+48+28}{6} = \mathbf{14}\) \(\left(\dfrac{20}{6}\right)^2 = 11.111\) 3.333
C 3 5 7 \(\dfrac{3+20+7}{6} = \mathbf{5}\) \(\left(\dfrac{4}{6}\right)^2 = 0.444\) 0.667
D 12 20 40 \(\dfrac{12+80+40}{6} = \mathbf{22}\) \(\left(\dfrac{28}{6}\right)^2 = 21.778\) 4.667
E 24 26 34 \(\dfrac{24+104+34}{6} = \mathbf{27}\) \(\left(\dfrac{10}{6}\right)^2 = 2.778\) 1.667
F 4 6 8 \(\dfrac{4+24+8}{6} = \mathbf{6}\) \(\left(\dfrac{4}{6}\right)^2 = 0.444\) 0.667
G 3 4 11 \(\dfrac{3+16+11}{6} = \mathbf{5}\) \(\left(\dfrac{8}{6}\right)^2 = 1.778\) 1.333

Step 2 — CPM on expected times.

\[ \text{ES}_A = 0,\ \text{EF}_A = 7 \]
\[ \text{ES}_B = 7,\ \text{EF}_B = 21; \qquad \text{ES}_C = 7,\ \text{EF}_C = 12 \]
\[ \text{ES}_D = \max(21, 12) = \mathbf{21},\ \text{EF}_D = 43 \]
\[ \text{ES}_E = 43,\ \text{EF}_E = 70; \qquad \text{ES}_F = 70,\ \text{EF}_F = 76 \]
\[ \text{ES}_G = 76,\ \text{EF}_G = \mathbf{81} \]
\[ \boxed{\ \textbf{Critical path: } A\to B\to D\to E\to F\to G, \quad T_e = 81\ \text{weeks}\ } \]

(C has float: \(\text{LF}_C = 21\), \(\text{LS}_C = 16\), so \(\text{TF}_C = 16-7 = \mathbf{9}\) weeks.)

Step 3 — project variance.

\[ \sigma^2 = 2.778 + 11.111 + 21.778 + 2.778 + 0.444 + 1.778 = \mathbf{40.667} \]
\[ \sigma = \sqrt{40.667} = \mathbf{6.377\ \text{weeks}} \]

Step 4 — the risk decomposition — the key analytical output.

Activity \(\sigma^2\) % of total risk \(t_e\) % of duration
A 2.778 6.8 % 7 8.6 %
B 11.111 27.3 % 14 17.3 %
D 21.778 53.6 % 22 27.2 %
E 2.778 6.8 % 27 33.3 %
F 0.444 1.1 % 6 7.4 %
G 1.778 4.4 % 5 6.2 %
\[ \boxed{\ \textbf{D (recruitment) and B (ethics) contribute 81 \% of the risk but only 44 \% of the duration.}\ } \]

Note that E (treatment) is the longest activity at 27 weeks but contributes only 6.8 % of risk — because its duration is protocol-fixed and predictable.

Step 5 — completion probabilities.

Target \(z\) \(\Pr\)
81 weeks 0.00 50.0 %
85 weeks 0.627 73.5 %
90 weeks 1.412 92.1 %
91.5 weeks 1.645 95.0 %
96 weeks 2.353 99.1 %

Step 6 — the risk-management plan.

Priority 1 — patient recruitment (D): 53.6 % of all risk.

  • Range \([12, 40]\) weeks — the pessimistic case is more than three times the optimistic.
  • Actions: open more sites than strictly needed; pre-screen patient databases before D begins; build recruitment incentives into site contracts; establish weekly enrolment tracking with a trigger to add sites if behind plan.
  • Target: narrow the range to \([15, 30]\), reducing \(\sigma^2\) from 21.778 to \(\left(\frac{15}{6}\right)^2 = 6.25\).

Priority 2 — ethics approval (B): 27.3 % of risk.

  • Range \([8, 28]\) weeks; delays are usually caused by protocol queries.
  • Actions: pre-submission consultation with the ethics committee; engage an experienced regulatory writer for the protocol (activity A); prepare responses to common queries in advance.
  • Target: narrow to \([10, 20]\), reducing \(\sigma^2\) from 11.111 to \(\left(\frac{10}{6}\right)^2 = 2.778\).

Impact of both improvements.

\[ \sigma^2_{\text{new}} = 2.778 + 2.778 + 6.25 + 2.778 + 0.444 + 1.778 = \mathbf{16.806} \]
\[ \sigma_{\text{new}} = \mathbf{4.100\ \text{weeks}} \]
\[ \text{95 \% date: } 81 + 1.645(4.100) = 81 + 6.75 = \mathbf{87.7\ \text{weeks}} \]
\[ \boxed{\ \textbf{Risk mitigation cuts the 95 \% date from 91.5 to 87.7 weeks — a 3.8-week saving with NO reduction in expected duration.}\ } \]

Priority 3 — exploit C's float. Site selection has 9 weeks of float. Start it early and in parallel with ethics approval, so that recruitment (D) can begin the moment ethics clears.

Step 7 — the recommendation to the board.

"Expected completion is 81 weeks, but we recommend committing externally to 92 weeks (95 % confidence). Patient recruitment and ethics approval carry 81 % of the schedule risk despite being under half the duration. With the mitigation plan above, the 95 % date falls to 88 weeks. We also recommend a Monte Carlo simulation before any regulatory filing commitment, since PERT systematically understates duration when parallel paths exist."

20.8 Practice Problems

  1. Compute \(t_e\) and \(\sigma\) for \((a,m,b) = (3, 7, 11)\). [Ans: \(t_e=7\), \(\sigma=1.333\)]
  2. A critical path has variances 1, 4, 2, 2. Find \(\sigma_{\text{project}}\). [Ans: 3]
  3. \(T_e = 40\), \(\sigma = 4\). Find \(\Pr[T\le45]\). [Ans: \(z=1.25\), 89.4 %]
  4. \(T_e = 30\), \(\sigma = 5\). What date gives 90 % confidence? [Ans: 36.4]
  5. Why do variances add but standard deviations not?
  6. Explain the merge-event bias in three sentences.
  7. A project has 4 near-critical paths. What is the approximate probability of meeting \(T_e\)? [Ans: \(\approx 0.5^4 = 6\%\)]

20.9 Summary

  • PERT replaces CPM's single duration with three estimates \(a, m, b\), assuming a Beta distribution: \(t_e = \frac{a+4m+b}{6}\) and \(\sigma^2 = \left(\frac{b-a}{6}\right)^2\). Note \(\sigma\) depends only on the range, not the skew.
  • Run CPM on the \(t_e\) values; the project variance is the sum of the CRITICAL activities' variances — variances add, standard deviations do not.
  • By the Central Limit Theorem, \(T\sim N(T_e, \sigma^2)\), so \(\Pr[T\le D] = \Phi\!\left(\frac{D-T_e}{\sigma}\right)\) and the \(p\)-confidence date is \(T_e + z_p\sigma\).
  • A project promised at its expected duration has only a 50 % chance of success — the most important and most ignored fact in project management.
  • Risk decomposition is PERT's most valuable output: identify which activities contribute most variance, not most duration, and target mitigation there. Reducing uncertainty saves buffer without shortening any activity.
  • Merge-event bias: PERT counts only the critical path's variance, so it is optimistically biased whenever parallel near-critical paths exist. With \(k\) equally critical paths, \(\Pr[\text{on time}]\approx0.5^{k}\).
  • The remedy is Monte Carlo simulation; PERT's enduring value is forcing an explicit conversation about uncertainty.

Unit II is now complete. You can solve games, sequence jobs, formulate and solve integer programmes, and schedule projects deterministically and probabilistically.

Next lesson: L21 — Characteristics of queuing systems (Unit III begins).

Practise now: unit2-games-networks-lab.html → CPM/PERT Lab: switch to PERT mode, enter three estimates per activity, and see the probability curve, risk decomposition and Monte Carlo comparison.