Although interaction effects are frequently found in experimental studies, field researchers report considerable difficulty in finding theorized moderator effects. Previous discussions of this discrepancy have considered responsible factors including differences in measurement error and use of nonlinear scales. In this article we demonstrate that the differential efficiency of experimental and field tests of interactions is also attributable to the differential residual variances of such interactions once the component main effects have been partialed out. We derive an expression for this residual variance in terms of the joint distribution of the component variables and explore how properties of the distribution affect the efficiency of tests of moderator effects. We show that tests of interactions in field studies will often have less than 20% of the efficiency of optimal experimental tests, and we discuss implications for the design of field studies.
It is common for researchers discovering a significant interaction of a measured variable X with a manipulated variable Z to examine simple effects of Z at different levels of X. These “spotlight” tests are often misunderstood even in the simplest cases, and it appears that consumer researchers are unsure how to extend them to more complex designs. The authors explain the general principles of spotlight tests, show that they rely on familiar regression techniques, and provide a tutorial demonstrating how to apply these tests across an array of experimental designs. Rather than following the common practice of reporting spotlight tests at one standard deviation above and below the mean of X, it is recommended that when X has focal values, researchers should report spotlight tests at those focal values. When X does not have focal values, it is recommended that researchers report ranges of significance using a version of Johnson and Neyman's test the authors term a “floodlight.”
Moderated multiple regression models allow the simple relationship between the dependent variable and an independent variable to depend on the level of another independent variable. The moderated relationship, often referred to as the interaction, is modeled by including a product term as an additional independent variable. Moderated relationships are central to marketing (e.g., Does the effect of promotion on sales depend on the market segment?). Multiple regression models not including a product term are widely used and well understood. The authors argue that researchers have derived from this simpler type of multiple regression several data analysis heuristics that, when inappropriately generalized to moderated multiple regression, can result in faulty interpretations of model coefficients and incorrect statistical analyses. Using theoretical arguments and constructed data sets, the authors describe these heuristics, discuss how they may easily be misapplied, and suggest some good practices for estimating, testing, and interpreting regression models that include moderated relationships.
Marketing researchers frequently split (dichotomize) continuous predictor variables into two groups, as with a median split, before performing data analysis. The practice is prevalent, but its effects are not well understood. In this article, the authors present historical results on the effects of dichotomization of normal predictor variables rederived in a regression context that may be more relevant to marketing researchers. The authors then present new results on the effect of dichotomizing continuous predictor variables with various nonnormal distributions and examine the effects of dichotomization on model specification and fit in multiple regression. The authors conclude that dichotomization has only negative consequences and should be avoided.
This article is a primer on issues in designing, testing, and interpreting interaction or moderator effects in research on family psychology. The first section focuses on procedures for testing and interpreting simple effects and interactions, as well as common errors in testing moderators (e.g., testing differences among subgroup correlations, omitting components of products, and using median splits). The second section, devoted to difficulties in detecting interactions, covers such topics as statistical power, measurement error, distribution of variables, and mathematical constraints of ordinal interactions. The third section, devoted to design issues, focuses on recommendations such as including reliable measures, enhancing statistical power, and oversampling extreme scores. The topics covered should aid understanding of existing moderator research as well as improve future research on interaction effects.