Hi Eins,
If for instance v1=1:100 and v2=1:100, v1*v2=v3, v1 has a very high multiple correlation with v1 and v2 (which may cause Multicollinearity Problem). A simple solution for that is to substract the means of v1 and v2 before multiplicating)
(compute v3=(v1-50)*(v2-50).)
Now the correlation of v1 and v2 with v3 is much lower. The correlation is just an artifact.
Frans
-------------------
dr F.H.G. (Frans) Marcelissen
[hidden email]
Van: SPSSX(r) Discussion [mailto:[hidden email]] Namens E. Bernardo
Verzonden: dinsdag 2 april 2013 5:34
Aan: [hidden email]
Onderwerp: Multicollinearity Problem in Moderation Model
Dear Experts,
Our moderation model has four continuous independent variables(IVs), only one dichotomous moderator variable(Mo) and continuous dependent. All variables are not latent, so we tried to use linear regression model with the product terms of IV and Mo serve as the moderating variable. We used centered IVs. We've examined the multicollinearity statistics and found out that some have tolerance of 0 to .03 and VIF of 5 and above. Our sample size is just n=218. Results of the regression analysis (the moderation model above) show that there is no significant moderation effect.
On the other hand, we tried to do two separate regression analysis: one model for each group of the moderating variable ( with n1= 100, n2=118). Results show that two of the IVs in one group are significant (p<.01), while all of the IVs in the other group are nonsignificant (p>.05). Do these results indicate significant moderation effect?
Please advise how can we address the multicollinearity problem?
Thank you.
Eins
| Free forum by Nabble | Edit this page |