Abstract
AbstractNonnormally distributed response values, such as count data for instance, create challenges for factor screening. One problem is that variances may vary from run to run. Another is the choice of screening design for such responses. In this paper, we assess some screening performances for three popular screening designs: a definite screening design, a minimum resolution IV design, and a Plackett‐Burman design. Four distributions, two binomials, one gamma, and one Poisson are chosen for the response values. For each distribution, we test out if it is best to use the raw data, a variance‐stabilizing transformation of the data, or perform a generalized linear modeling assuming three factors are active. From our investigations, two‐level nonregular designs gave the highest success rate in identifying the subset of active factors and a variance‐stabilizing transformation turned out to perform equally good or better than generalized linear modeling in most cases.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have