regdhfe and warning "adjustment from Cameron, Gelbach & Miller applied"

[Leon Schmidt]

Dear Statalist,

I am using the great user-written program reghdfe and sometimes I am getting the error "VCV matrix was non-positive semi-definite; adjustment from Cameron, Gelbach & Miller applied." Despite the warning, I get estimates for all statistics and coefficients, so nothing seems to be wrong. Does anybody know if I can disregard this warning as there seems to be an adjustment applied and everything is estimated?

Note 1: I also asked this question here but have not received an answer (yet), this is why I am posting here https://github.com/sergiocorreia/reghdfe/issues/221

Note 1: There is a similar discussion here but this discussion is about missing statistics, which is not the case with my regressions https://www.statalist.org/forums/for...-about-reghdfe

Thank you very much for your help!

All the best
Leon

[Hong Il Yoo]

Like my -vcemway- command (see [here]) for robust inference with multiway clustering, -reghdfe- must have applied the eigendecomposition adjustment described in the last paragraph of p.241 in:

A. Colin Cameron, Jonah B. Gelbach & Douglas L. Miller (2011) Robust Inference With Multiway Clustering, Journal of Business & Economic Statistics, 29:2, 238-249.

Their Monte Carlo results suggest that you may proceed to the analysis as usual. Having said that, this does not mean that you can simply disregard the caveat. As two of the authors point out in the last paragraph of p.337 in:

A. Colin Cameron & Douglas L. Miller (2015) A practitioner’s guide to cluster-robust inference, Journal of Human Resources, 50(2), 317-372

one of the reasons why the covariance matrix may not be p.s.d. is if you have too small a number of clusters in some dimension(s). If that is the case, you may want to reconsider the wisdom of adjusting your standard errors for clustering in those dimensions.

[Leon Schmidt]

Thank you very much for your answer and the links! They were very helpful. As you and the authors point out, few clusters might be the reason for the covariance matrix not being positive semi-definite. In the second article they point out that another reason for this behavior can be that clustering is done "over the same groups as the fixed effects". I suspect this is the case in my regression as I include firm-level fixed effects and industry-by-year FE (firms do switch industries) and also cluster at these levels. When removing these clusters the warning goes away and nothing major changes.

[Leon Schmidt]

Just for completeness: Sergio answered me similarly here https://github.com/sergiocorreia/reghdfe/issues/221

[Hong Il Yoo]

Great, thanks for sharing the post.