Moderating Effects: The Myth of Mean Centering

Mean centering of variables is often advocated for estimating moderated regressions to reduce the multicollinearity that results from introducing the product term of two variables (x1x2) as an independent variable in the regression equation. The authors establish that, contrary to conventional wisdom, mean centering does not reduce multicollinearity at all, even if the bivariate correlation between x1 and x1x2 is decreased. Using Monte Carlo simulation, they also address the current admonition to systematically include all the product-term components in the moderated regression model; they conclude that such an automatic full-model specification is ill advised due to the structural multicollinearity it introduces. Finally, the authors propose a varying parameter model (VPM) to test more effectively moderating effects and show under what conditions VPM outperforms OLS estimations.