UC Berkeley

Sequential stopping rules are often used to determine the run length of a simulation experiment. They can be used within confidence interval procedures for simulation output analysis. We analyze the effect of sequential stopping rules on the coverage of confidence interval procedures. Our research develops methods for evaluating coverage of rules that are not operating in the limit. We assess the quality of different types of rules and introduce an analytical method for evaluating the loss in coverage for a certain class of stopping rules. This method includes a calculation of the distribution of the stopping time of the procedure. We find optimal parameters for absolute precision stopping rules applied to normally distributed data and use these results to suggest methods for obtaining nominal coverage with minimal replications. We relate confidence interval procedures to decision making to develop improved stopping techniques, and extend the discussion on stopping rules to other simulation areas.