Identifying causes of multicollinearity

[Agnes von Scheele]

Hi!

We are struggling with high VIF's in our model (in which we are using panel data with our main explanatory variable as dummies), but can't seem to find out why that is. When checking for the relationship between our outcome and independent variables, they seem to point in the expected direction. As we run our regression however, many variables flips sign and turn insignificant. So we figured that it must be due to the high VIF's. When trying to identify which variables are causing these issues in an attempt to find a solution, we have looked at the correlation matrix but can't seem to find any particularly high correlations. Rather the opposite, they are very low. We've tried standardizing the control variables, but that did not help. As far as we can tell when running the calculating VIF's for the dummies, they do not seem to be the problem either.

Is this a structural issue in the sense that our model isn't good enough perhaps? We would very much appreciate some advice to point us in the right direction. However, we are also very new to Stata, so please don't hesitate to ask for clarification!

Many thanks in advance!

[Oscar Ozfidan]

In my experience it is not unusual to have different signs or different significance of coefficients in a uni-variate specification vs in a specification with all dummies together. The dummies may not be highly correlated individually but there may be a correlation between a group of them vs one or another group of them. I would try a stepwise elimination to see where it takes me.

[Rich Goldstein]

your situation is not completely clear to me but I think that the user-written coldiag2 will help; this is at SSC

[Agnes von Scheele]

Pardon me for this very late response, thank you both very much for taking the time to reply to my question!