Does Distributional Shape Matter in Monte Carlo Analysis?* G. Harris, B. Binkowitz, and D. Wartenberg; Merck Co. and Robert Wood Johnson Medical School
Monte Carlo analyses are used frequently in quantitative risk analysis. For most such analyses, investigators must specify both a statistical distribution or shape and associated parameter values for each input variable to be treated stochastically in the model. Many investigators have emphasized the importance of selecting the correct distributional shape for accurate results. Collecting enough data to differentiate between candidate distributions often requires a substantial sampling burden. To test the importance of correctly specifying the distributional shape we replicated three published Monte Carlo analyses, substituting reasonable alternative statistical distributions for those specified by the original investigator while holding the mean and standard deviation of the distribution constant, where possible. Except when the coefficient of variation of the input variable is exceptionally large, we find that most results vary by less than a factor of 2. If these results are valid, selection of distributional shapes generally do not have substantial influence on the estimated cumulative risk. Sampling efforts should be limited or focused on other aspects of the problem under study, such as specification of the appropriate risk model or estimation of the appropriate parameter values.
This paper is an extension of work funded by the NJ Department of Environmental Protection.
*Best paper winner.
Go to . . .
2001 SRA Annual Meeting Table of Contents
2001 SRA Annual Meeting Author Index
Main Abstracts Menu Page
RiskWorld Home Page