Differentially Weighting Linear Models Of Behavior An Empirical Comparison Of Six Methods

Six methods of estimating regression weights for a linear model of behavior were compared in 51 samples of National Guardsmen. Ordinary least squares, Bayesian m-group regression, ridge regression, equal weighting, and two related methods were used. Weights were estimated in one-half of each sample and then applied to data in the other half. Ratios of observations to predictors ranged from 4:1 to 19:1. Cross validation R sq was used as the index of model or equation stability. Results support earlier findings that least squares weights are relatively unstable in small samples, but do not indicate the superiority of any one other method. Future research and implications for using these regression techniques in testing behavioral models are discussed. (Author).