PERT Analysis Revisited

I received an email notification this week of a new article on TechRepublic through the Blogger Tom Mochal titled "Use PERT technique for more accurate estimates." (To get at this article may require a login process).
    ow I know every project management tool provides a PERT analysis function and most of the Project Management literature targeted at "education" speaks to PERT but PERT has serious problems, significant flaws and must be used with care.

Quick Overview of the Issues
    The PERT process was developed in the mid-1950's in response to the political pressure on the Polaris Special Projects Office (SPO) under Vice Adm William F. Raborn [1]. PERT was a public relations success, but later studies raised doubts about whether PERT contributed much to the management success of Polaris. The classic study of PERT and Polaris is [2].
    PERT makes many assumptions about its formula of :

Expected Duration = (A+4M+B) / 6

that simply are NOT true in practice. While this formula can be found in every project management tool, it has many flaws, some of which will cause serious problems for a project manager. Let's start with some assumptions that are false. [4,5]

1. The Critical Path does not change
2. The task completion times are normally distributed with a symmetric probability distribution curve
3. The task completion times are statistically independent
4. The critical path usually comprises fewer activities than the Central Limit Theorem requires for statistical validity of the 6 in the denominator as the standard deviation. [3]
5. The PERT approximation formula (A+4M+B) of activity mean and variances deviate significantly from actual completion durations. [3]

The result of these assumptions not being met yields:

1. The near critical paths often turn critical
2. Failure to properly model the probability distribution in a non-symmetrical manner using Beta or Triangle distributions (PERT assumes a Normal - Gaussian - symmetrical distribution)

The result of these assumptions and outcomes produce significant errors in the Estimated Duration of the project when compared with a Monte Carlo Simulation of identical networks of activities.

Monte Carlo Simulation
    The Monte Carlo simulation process defines a probability distribution for each activities duration. This distribution can be Beta, Triangle or similar. But it should always be non-symmetrical to represent "real world" durations and completion times. The Monte Carlo tool picks a random duration from each distribution and uses that for the actual  duration for the activity. It does this for every task in the network. The it calculates (using the native tool) the completion date of observed activities. It does this many 1,000's of time, until a histogram of completion dates is accumulated.

The Result
When compared to Monte Carlo, PERT can exhibit 20% to 50% variances depending on the duration of the activities and the structure of the network. [3] provides a very detailed account of this assessment.
    Once this effect is realized the use of PERT becomes very unlikely for anything but the most trivial project.

References
[1] "Quantitative Risk Analysis for Project Management,: A Critical Review," Lionel Galway, Rand Working Paper, WR-112-RC, February 2004.
[2] "The Polaris System Development: Bureaucratic and Programmatic Success in Government," Harvey M. Sapolsky, Harvard University Press, 1972.
[3] "PERT Completion Times Revisited," Ted Williams, University of Michigan-Flint, School of Management, Working Paper 2005-02, September 2005.
[4] "Hidden Assumptions in Project Management Tools," Dr. Eva Reginier, DRMI Newsletter, January 10, 2006, Naval Postgraduate School, Monterey CA.
[5] "Activity Completion Times in PERT and Scheduling Network Simulation," Dr. Eva Reginier, DRMI Newsletter, April 8, 2005, Naval Postgraduate School, Monterey CA.