0

I am trying to assess the impact of multicollinearity in a regression because I have two separately measured variables which have the reversed signs problem (one predictor is +b regression weight, the other is -b). The two variables are fairly highly correlated (Spearman rank correlation = .78) although VIFs are around 1.5 to 1.7.

I remember mean centering in the past for dealing with multicollinearity whilst using interaction terms in moderation analyses (e.g. Smith & Sasaki, 1979). Would it have any impact here despite not being a moderation analyses? I'm just doing a standard linear regression with no predictor problems.

I know there is some debate on mean centering anyway (for instance... acknowledged in Shieh, 2011). However, I'm just wondering about people's views on whether there is any benefit.

If not, any ideas? I'm trying to avoid using a Log transformation on one of the variables as I end up losing data.

Simon.

Justification for why I do not consider this question to be a duplicate question: This question deals with mean centring outside of moderation/interaction terms for the specific purpose of multicollinearity. The posts I have found so far either 1) deal centring for interaction terms, 2) don't deal with centring for interaction terms but make little reference to multicollinearity in that scenario.

SimonsSchus
  • 719
  • 5
  • 13
  • 1
    It is strange to see the Spearman rank correlation coefficient being used to assess collinearity! Moreover, that coefficient of 0.78 and--especially--the low VIFs--indicate there is no collinearity problem whatsoever. As explained in multiple answers to the duplicate question, mean centering will have no effect on any regression model that already includes a constant term. – whuber Oct 15 '14 at 16:13
  • Hi there. Thank you for taking the time to help me out. For clarity, I was using Spearman as an indicator of potential collinearity, given that my covariables are not normally distributed. It was one of the various indicators I used to assess whether multicollinearity is a problem (the others being VIF, Condition number, correlation matrix although the latter two). Is this a poor idea? – SimonsSchus Oct 15 '14 at 20:12
  • Also, thank you so much for explaining that mean centring does not have any effect on a regression model not already having a constant term. I did not find this very clear in the write-up of earlier posts, and now I understand why this was deemed a duplicate post. I really appreciate your help. Simon. – SimonsSchus Oct 15 '14 at 20:13

0 Answers0