Importance of Distributional Form in Characterizing Inputs to Monte Carlo Risk Assessments

The selection among distributional forms for inputs into uncertainty and variability (e.g., Monte Carlo) analyses is an important task. This paper considers the importance of distributional selection by examining the overall and tail behavior of the lognormal, Weibull, gamma, and inverse gaussian distributions. It is concluded that at low relative standard deviation (below 1), there is less of a difference between upper tail behavior among the distributions than at higher RSD values. Sample sizes in excess of 200 are required to reliably distinguish between distributional forms at the higher RSD values. The likelihood statistic appears to offer a reasonable approach to distributional discrimination, and it, or a similar approach, should be incorporated into distributional fitting procedures used in risk analysis.