Moderator Variable

A moderator variable is an independent or predictor variable (e.g., Z) that interacts with another independent or predictor variable (e.g., X) in predicting scores on and accounting for variance in a dependent or predicted variable (e.g., Y). Note that the terms independent variable and dependent variable are appropriate only in the case of experimental research. In nonexperimental research, the appropriate, analogous terms are predictor variable and predicted variable.

Moderator variables (referred to hereinafter as moderators) imply conditional relations. That is, the strength and/or form (e.g., linear, quadratic) of the relation between X and Y varies as a function of the value of the moderator, Z. There are many examples of moderators in both theory and research in such disciplines as psychology, sociology, management, education, political science, biology, epidemiology, and medicine. For instance, in industrial and organizational psychology, extant theory specifies that the relation between ability and performance is moderated by motivation. The greater the level of motivation, the stronger the relation between ability and performance.

By studying moderators, researchers learn about how relations between variables of interest vary across levels of the moderator. Interestingly, it has been argued that the amount of progress in any discipline can be indexed by the degree to which its theory and research have considered the role of moderators.

An Important Distinction

Figure 1 Illustration of a Mediating Effect

A review of the literature reveals that moderators are often confused with mediator variables (referred to below as mediators). In contrast to the just noted role played by moderators, mediators transmit the effects of independent variables to dependent variables. This is illustrated in the causal chain shown in Figure 1, which depicts M as a mediator of the relation between X and Y. There are many examples of mediation. In psychology, for instance, stress (M) has been shown to mediate the relation between stressors (X) and strain (Y).

Strategies for Detecting Moderators

Over the years, a number of strategies have been proposed for the detection and description of moderators. Unfortunately, many strategies are unsound (e.g., single group validity). In addition, others (e.g., comparing correlation coefficients for artificially created subgroups) have lower levels of statistical power than others (e.g., moderated multiple regression). In the interest of brevity, only three strategies are described here: (a) the analysis of variance based test for an interaction, (b) the multiple regression based test for an interaction, and (c) the test for the homogeneity of correlation coefficients.

The Analysis of Variance Strategy

A commonly used technique for testing interaction effects in experimental research is analysis of variance (ANOVA). Assuming a 2 × 2 experimental design involving independent variables A (a1, a2,… aj) and B (b1, b2,… bk), ANOVA tests for an interaction by determining if the A × B effect explains variance over and above the additive effects of A and B. The respective population and sample effect models for a two-way ANOVA are as follows:

[Page 625]

In terms of the population model, a main effect of A is implied by the rejection of the null hypothesis of αj = 0, a main effect of B is indicated by the rejection of the null hypothesis of βj = 0, and an interaction effect is signaled by the rejection of the null hypothesis of (αβ)jk = 0.

...