Choosing Distributional Forms for Use in Monte Carlo Exposure Assessments. M. A. Youngren, MITRE Corporation, Reston, VA; and S. H. Youngren, Novigen Sciences, Inc., Washington, DC
Use of distributional analyses has become very common for exposure and risk assessments. However, many exposure assessors are unsure how to choose the appropriate distributional families (e.g., normal, log-normal, exponential, etc.) to be used in the assessment. For example, multiple software packages will "fit" data to different distributions, but not all "fits" are equal. Standard distributional fits to data will provide the best fit for the bulk of the data that will be found at the mode of the distribution. As a result, the fit in the tails may be poor. However, the data in the tails (i.e., the 95th through 99.9th percentile) is often the area of most interest. The packages that rank-order distributional fits by Anderson-Darling, Chi-square, etc. will also provide the best fit in the mode of the distribution and fail to accurately reflect the behavior of the data in the tail. However, they do suggest the distribution with the best "fit". But is this the "fit" needed by the assessor? Additionally, many data sets, particularly small data sets, will "fit" multiple distributional families. This presentation will illustrate methods to help the practitioner choose the "best" distributional family (as there is no absolute correct choice).
Go to . . .