Model Averaging Method for Supersaturated Experimental Design

Abstract

In this paper, a new modified model averaging method was proposed. The candidate model construction was performed by distinguishing the covariates into focus variables and auxiliary variables whereas the weights selection was implemented using Mallows criterion. In addition, the illustration result shows that the applied model averaging method could be considered as a new alternative method for supersaturated experimental design as a typical form of high dimensional data. A supersaturated factorial design is an experimental series in which the number of factors exceeds the number of runs, so its size is not enough to estimate all the main effect. By using the model averaging method, the estimation or prediction power is significantly enhanced. In our illustration, the main factors are regarded as focus variables in order to give more attention to them whereas the lesser factors are regarded as auxiliary variables, which is along with the hierarchical ordering principle in experimental research. The limited empirical study shows that this method produces good prediction.