| Abstract: | ABSTRACT The article by Bissonnette, Ickes, Bernstein, and Knowles (1990) presented in this issue contains a series of simulations which vary widely in the degree to which they resemble the parameters that have been found in actual data. Their conclusions feature warnings based on simulations containing large confounds between trait extremity and item variance. In real data, however, the corresponding confounds typically have been small. When Bissonnette et al. used appropriately small confounds in their simulations, their results showed that median split techniques failed to inflate Type I error rates. The constructive implications of Bissonnette et al.'s simulations indicate that moderator-variable researchers should always perform a simple check for range restriction and for variance-extremity confounds, and that there is no danger of spurious conclusions as long as these checks conform to what has typically been found with real data. |
