Variable elimination in nested DEA models: A statistical approach

Abstract

In this study, a new definition of a relevant variable in a DEA model is proposed for variable selection. The selection procedure is the conventional iterative backward elimination procedure with multiple statistical comparisons. The multiple tests of null hypothesis are reduced to a simple hypothesis test using either the binomial probability or the McNemar test with Bonferroni correction of significant level. From the results based on two simulation populations of moderately and lowly correlated input variables, the proposed procedure using either one of the suggested statistical tests can identify the relevant variables with high accuracy and eliminate the irrelevant variables effectively. In the dataset from a large scale experiment in the US public school education, the reduced model selected by the proposed procedure is shown to be the better approximation of the full model than the ones selected by the Pastor et al. method.