Two-Stage Controlled Fractional Factorial Screening for Simulation Experiments

Factor screening with statistical control makes sense in the context of simulation experiments that have random error, but can be run automatically on a computer and thus can accommodate a large number of replications. The discrete-event simulations common in the operations research field are well suited to controlled screening. In this paper, two methods of factor screening with control of Type I error and power are compared. The two screening methods are both robust with respect to two-factor interactions and nonconstant variance. The first method is an established sequential method called controlled sequential bifurcation for interactions (CSB-X). The second method uses a fractional factorial design in combination with a two-stage procedure for controlling power. The two-stage controlled fractional factorial (TCFF) method requires less prior information and is more efficient when the percentage of important factors is 5% or higher.