Cautions in assessing spurious "moderator effects."

Lubinski and Humphreys (1990) illustrated how moderator effects uncovered using hierarchical multiple regression procedures may be spurious and that the predictor-criterion functional relationship may be better characterized by a nonlinear function. They recommended concurrently inspecting nonlinear functions when examining moderator effects. This article reviews the evidence presented by Lubinski and Humphreys to substantiate their argument and the generality of their findings beyond their data. Although an examination of nonlinear functions may be justified in their specific illustration, the blanket recommendation of inspecting polynomial terms when examining moderator effects is troublesome and may itself lead to spurious conclusions. In a recent article in Psychological Bulletin, Lubinski and Humphreys (1990) presented a thoughtful discussion of how moderator effects uncovered using hierarchical multiple regression analyses (HMRA) may be spurious. Using data from the Project Talent Data Bank (Flanagan et al, 1962), Lubinski and Humphreys illustrated empirically that a nonlinear term (e.g., X1 or X1'1) may characterize the functional relationship between predictors (in their case, quantitative and spatial abilities) and the criterion (exceptional levels of mathematical sophistication) better than the interaction term. As a consequence, Lubinski and Humphreys concluded that researchers examining moderator effects should concurrently inspect nonlinear functions. In this article I review the evidence presented by Lubinski and Humphreys to substantiate their argument and the generality of their findings beyond their data set. Although an examination of nonlinear functions may be justified in their specific illustration, the blanket recommendation of inspecting polynomial terms when using HMRA to examine moderator effects is troublesome and may itself lead to spurious conclusions.