Two-phase screening procedure for simulation experiments

Analysts examining complex simulation models often conduct screening experiments to identify important factors. The controlled sequential bifurcation screening procedures CSB and CSB-X use a sequence of tests to classify factors as important or unimportant, while controlling Type I error and power. These procedures require analysts to identify the directions of the effects prior to experimentation, which can be problematic. We propose hybrid two-phase approaches, FFCSB and FFCSBX, as alternatives. Phase 1 uses an efficient fractional factorial to estimate factor effect directions; phase 2 uses CSB or CSB-X. Empirical investigations show these outperform CSB(X) in efficiency and effectiveness for many situations of practical interest.