Application And Simulation Study Of Shrinkage Methods In Analysis For The Large Supersaturated Designs

Supersaturated designs have provided a tool for screening out the active factors from a set of potentially active factors during the initial phase of an experiment. However, data analysis of such designs is still in the preliminary stage, particularly for large supersaturated designs. In this thesis, we discuss the issue of the performance of shrinkage methods for the analysis of large supersaturated design by applying one kind of supersaturated designs with the same number of rows and an increasing the number of columns in linear regression model. Simulation results show that the ability of S.EB to analysis supersaturated design is relatively higher than other shrinkage methods, and that the ability of it to analysis supersaturated design is affected to a certain extent by the number of active factors and the number of columns of design matrix in the model. When the number of active factors is 1 or 2, both the number of active factors and the number of columns of the design matrix have little impact on the performance of S.EB. When the number of active factor is 1, S.EB can identify the true model; when 2, it can identify a model close to the true model. But when the number of active factors is larger than 2, the performance of S.EB is affected significantly by the number of active factors and the number of columns of the design matrix, with the fact that the ability of it to identify the true model decreases significantly when the number of active factors and/or the number of columns increase(s).