The analysis of interaction effects in multiple regression has received considerable attention in recent years, but problems with the valid identification of moderating variables have been noted by researchers. G. McClelland and C. Judd (1993), in their discussion of the statistical difficulties of detecting interactions and moderating effects, warned against the use of a four-corners subsample approach to moderated multiple regression, but they did not present empirical evidence that such an approach provides less power than the use of the full random sample. This study was conducted to produce evidence of the extent of power loss that is associated with the subsample strategy. The effectiveness of the four-corners subsample procedure was investigated through a Monte Carlo study that used regression models to generate data from populations with linear, nonlinear, and nonadditive relationships. In all, 2,304 conditions were examined, for 3 models, 4 levels of population "R" squared, 4 levels of regressor correlation, 4 levels of regressor reliability, 3 levels of sample size, and 4 levels of effect size for the nonlinear or nonadditive component. Results suggest that the use of the four-corners strategy rather than full sample analysis shows better specificity at the expense of reduced statistical power, or sensitivity, relative to full sample analysis. Despite the improved specificity of the four-corners approach, model misidentification rates were high in many of the conditions examined. The utility of either the four-corners approach or the full sample approach for testing theory is limited. (Contains 3 tables and 26 references.) (SLD)