Standardized versus unstandardized regression weights

Abstract

Behavioral scientists appear widely divided about the merits of standardized versus unstandardized regression weights. The present paper has therefore attempted to clarify the issue by illustrating how the two kinds of weights respond to the following circumstances: selection, or sampling variation, on the independent variable; presence of measurement error; changes in units of measurement; scales that are relative to norms rather than absolute; changes in test length; and selection on dependent and mediating variables. Standardized weights are invariant over changes in the units of measurement; unstandardized weights are invariant over selection on the independent variable (but only on the independent variable) and in the presence of dependent variable error. In general, however, the illustrated circumstances affected both kinds of weights, sometimes quite strongly. Therefore, neither kind of weight appears very robust, neither can be recommended unequivocally for comparing groups, and neither provides a panacea for overcoming problems of analyzing behavioral data.