One-mediator model of exposure effects is still viable

Michael H. Birnbaum, Barbara A. Mellers
1979 Journal of Personality and Social Psychology  
Birnbaum and Mellers criticized the use of partial correlation and multiple regression by Moreland and Zajonc to argue for two independent effects of stimulus exposure on liking. The null hypothesis that one variable mediates the effect of the independent variable on the dependent variables was not tested by their analyses. In response, Moreland and Zajonc reanalyzed their data, using structural equations analysis, and replied that there is evidence to support their previous conclusions.
more » ... , the present article shows that the small residuals from the one-mediator model may be due to shared nonlinearity (correlated errors) in three of the dependent variables. This simpler interpretation achieves a better fit to the Moreland and Zajonc data than the two-mediator model they advocated. Since the null hypothesis of one mediator is still viable, the burden of proof rests on those who contend that there is more to the exposure effect than stimulus recognition. Moreland and Zajonc (1977) replicated the exposure effect, that is, the finding that stimuli that are presented with greater frequency will be rated more favorably. They also asked their subjects to rate their familiarity with the stimuli, and they found a significant partial correlation between liking and exposure frequency when rated familiarity was partialed out. This partial correlation (and related regression analyses) led them to conclude that there is an additional, "independent" effect of exposure frequency on liking that is not mediated by stimulus recognition. Birnbaum and Mellers (1979) questioned the use of partial correlation and regression in this argument. They pointed out that the null hypothesis that a single variable (e.g., recognition) mediates the effect of the independent variable on both dependent variables predicts that both partials should have the same sign as the original correlations. Only if the dependent variable measuring recognition is assumed to be perfectly correlated with the mediator does the regression analy-Requests for reprints should be sent to Michael H. Birnbaum, Department of Psychology, University of Illinois, Champaign, Illinois 61820. sis test the null hypothesis. Otherwise, lack of perfect validity of the dependent variable vitiates the analysis. Birnbaum and Mellers (1979) presented path models both for the null hypothesis of one mediator and for the alternative hypothesis that a second mediator is required. Implications of the models and methods to distinguish them were described. These methods were applied to the three major variables in Moreland and Zajonc (1977) with the result that no evidence was found to require the hypothesis of two mediators over the simpler hypothesis of one. In response to these arguments, Moreland and Zajonc (1979) reanalyzed their data, using structural equations analysis. They argued that when all five variables are considered, a case can be made for the existence of an extra path from stimulus frequency to liking, in addition to the path via subjective recognition. However, the present article will show that the residuals from the one-mediator model, fit to all five variables, are small in magnitude and are not of the form anticipated by the Moreland and Zajonc model. Instead, they can be explained by the simpler hypothesis that some of the dependent variables are nonlinearly related to the single mediator.
doi:10.1037//0022-3514.37.6.1090 fatcat:vw7t3ylibjc7tepdho33dixypq