Robust sequential bifurcation for simulation factor screening under data contamination

Abstract

In the past decades, sequential bifurcation (SB) has been extensively employed as a popular factor screening method to identify active factors in simulation experiments due to its high efficiency. However, previous work on SB assumes that the response follows a normal distribution without contamination, which is actually inevitable in practice. In this paper, we propose robust sequential bifurcation (RSB), a procedure that can ensure insensitive screening results even if response data are contaminated. In order to achieve this goal, a robust statistic, which is based on sample median and median absolute deviation, is employed and incorporated with a fixed-width interval method to determine a suitable sample size and test the importance of factors under specified Type I and Type II errors. Simulation experiments are conducted to compare RSB with classic SB methods in terms of effectiveness and efficiency, verifying the robustness of the proposed method under different contamination scenarios.