The robustness of mean and variance approximations in risk analysis

Abstract

This paper examines further the problem of estimating the mean and variance of a continuous random variable from estimates of three points within the distribution, typically the median or mode and two extreme fractiles. The problem arises most commonly in PERT and risk analysis where it can usually be assumed that the distribution in question is bell-shaped and positively skewed, often typified by a Beta distribution. Over the years, a number of alternative approximations have been proposed, usually as modifications to the original PERT formulae. The accuracy of a number of these approximations is investigated based not only on a Beta distribution, but also for three other commonly used bell-shaped, positively skewed distributions, namely the Gamma, Lognormal and F distributions. It is shown that a balanced weighted average of the median and the 4% fractiles provides a consistent estimator of the distribution mean across all four distributions. Furthermore, reasonably accurate estimates of the variance can also be obtained by treating the three fractiles as defining an equivalent discrete distribution with the same probability weight as in the formula for the mean.