In my work I am analyzing moderation effects and following recommendations of Carte and Russell (2003) and Frazier et al. (2004). One specific aspect is examing multicolinearity between X, Z, and the XZ product term ( X= predictor, Y= Dependent variable, Z= moderator, XZ= interaction term) since high multicolinearity distorts beta (path coefficient) of XZ and consequently interpretability of results.
Although centering/standardizing reduces problems associated with multicollinearity (i.e., high correlations) among the variables in the regression equation (Carte and Russell 2003) , I still have high multicollinearity - See screenshot from SmartPLS output (Y = UO, X=EG, Z=KIT , XZ= EG*KIT) coloured yellow. The Interaction term EG*KIT is highly correlated to EG (i.e. 0.90).
However, when I use the Software "Interactions" which is used to designed to draw and analyze statistical interactions (downloadable at http://www.danielsoper.com/interaction/default.aspx) its output file reveals that the correlation is practically absent (0,048) - see screenshot coloured purple.
While values of other correlations are almost the same between SmartPLS & Interactions Software (e.g. UO with EG) coloured blue, there seems to be this huge difference only in the case of the Interaction terms (XZ).
So I was just wondering why “SmartPLS “suggests high multicollinearity while “Interactions” implies no multicollinearity problem. Maybe I am just interpreting the Output files wrong, but some clarification on this issue would be very helpful.
SmartPLS Output
Interactions Software Output using SMartPLS Latent variable Scores
Carte, T.A. and Russell, C.J. (2003). In Pursuit of Moderation: Nine Common Errors and Their Solutions. MIS Quarterly, 27 (3), 479-501.
Frazier, P., Tix, A. and Barron, K. (2004). Testing Moderator and Mediator Effects in Counseling Psychology Research. Journal of Counseling Psychology, 51(1), p.p. 134, 115.