Signs of correlation flip in estat vce

[Marc Pelow]

Hi there,

I'm feeling quite puzzled after examining the results of the -estat vce, corr- command after the estimation of a GEE model. It appears that the correlation of the coefficients of certain control variables exhibits flip in it's sign compared to what I observe when using the -corr- command on the variables. For example, two of my control variables, which measure the size of a firm and the number of analysts following the firm, should logically be positively correlation (analysts tend to follow larger, more well known firms). This is precisely what I find when running the corr command on these two variables. However, after integrating these controls into my GEE model, -estat vce, corr- indicates a negative correlation between their estimated coefficients. It's clear that there's something I'm overlooking, but I'm struggling to pinpoint exactly what it might be. It's the first time I use the -estat vce, corr- command but is there a rational reasoning why this might happen? I considered the possibility of multicollinearity being a factor, and would greatly appreciate your opinion.

Thanks in advance

[Maarten Buis]

The correlation between variables and the correlation implied by the sampling distribution of coefficients are two completely different things, so it is not surprising that the two are different.