Implementing Variable Selection Techniques in Regression

Abstract

The most commons methods of variable selection (forward, backward, all possible subsets) were considered. Criticisms and common misuses of stepwise methods were presented. Suggestions were made for each method concerning appropriate procedures to follow in running computer programs and the information that should be reported with the results. An example was presented which showed how proper selection should be done. When variables are selected for a regression model, the stepwise method can be helpful if the initial choice of variables is chosen as much as possible using theory, the defaults of the computer program used are not used automatically, more than one computer run is done using different variable selection methods, and the final model is chosen through an intelligent process, not automatically using the final model generated by the computer program. When the model is described, all subjective decisions made in the model selection process should be reported. 

An earlier version of this paper was presented at the annual meeting of the American Educational Research Association in Boston, Massachusetts, April, 1990, as part of presidential address.