A MORE COMPLETE CHARACTERIZATION OF UNCERTAINTY : CAN IT BE DONE ?

Discrete event simulation model output analysis provides characterizations of the uncertainty in the simulation results. This allows decision makers to judge risk in decisions made based on simulation model output. These characterizations, whether they are confidence intervals, variance estimates or quantile estimates, ignore important sources of uncertainty. In particular, results obtained during the model validation phase are usually not incorporated in the uncertainties presented in the usual output analysis. Further, many validation approaches assume that the data used to fit input distributions are complete, and the output analysis does not include uncertainties caused by the use of finite samples to determine input distributions. Bootstrap methods have been used successfully in these settings, yet the formal requirement that the bootstrapped statistic be a continuous function of the data does not hold. Knowing this, it is easy to create scenarios where bootstrapping fails. This paper examines whether this shortcoming can be fixed, and presents one possible strategy.