The moderator effect that wasn't there: Statistical problems in ambivalence research.

Abstract

Ambivalence researchers often collapse separate measures of positivity and negativity into a single numerical index of ambivalence and refer to it as objective, operative, or potential ambivalence. The authors argue that this univariate approach to ambivalence models undermines the validity of subsequent statistical analyses because it confounds the effects of the index and its components. To remedy this situation, they demonstrate how the assumptions underlying the indices derived from the conflicting reactions model and similarity-intensity model can be tested using a multivariate approach to ambivalence models. On the basis of computer simulations and reanalyses of published moderator effects, the authors show that the frequently reported moderating influence of ambivalence on attitude effects may be a statistical artifact resulting from unmodeled correlations of positivity and negativity with attitude and the dependent variable. On the basis of extensive power analyses, they conclude that it may be extremely difficult to detect moderator effects of ambivalence in observational data. Therefore, they encourage ambivalence researchers to take an experimental approach to study design and a multivariate approach to data analysis.