Simpler Isn't Better: Ordinary Least Squares Regression Weights Make Better Predictions than Simple Alternative Weights in Realistic Simulations

Abstract

Some researchers advocate using alternative regression weights, as opposed to ordinary least squares (OLS) regression weights, when model predictability is low or when there are few observations per predictor. Perhaps the most comprehensive simulation study to date on the topic, "The Superiority of Simple Alternatives to Regression for Social Science Predictions" by Dana and Dawes (2004), concluded that, in almost every circumstance likely to be encountered by a social scientist, researchers should weight predictors equally or by their correlations with the criterion. Though this recommendation has not been adopted widely, their study remains unreplicated and their conclusion unchallenged in the literature. The present set of studies partially replicate Dana and Dawes (2004) and show that, when parameters are restricted to a very plausible range for the social sciences, the performance of equal and correlation weights can be quite different from that described by Dana and Dawes. The resulting recommendation for researchers is that OLS regression ought not be abandoned except in the very rare case of strong a priori knowledge of the relationships among model parameters.