Determining relative importance of predictors with the observational design.

Abstract

A recent comparison of typical extreme-groups designs and observational designs (G. H. McClelland & C. M. Judd, 1993) showed that extreme-groups designs have greater power to detect interactions than do observational designs and that extreme-groups designs provide estimates of unstandardized parameter values that have smaller standard errors than do estimates provided by observational designs. In this study, this discussion is taken a step further by investigation of the advantages and disadvantages associated with inferences drawn from extreme-groups and observational designs for the estimation of standardized effects. Observational designs, through accurate estimation of predictor variances, are concluded to be superior for the purposes of standardized parameter estimation. Finally, various ways of adapting extreme-groups designs to better justify inferences concerning population distributions are suggested.