Driscoll-Kraay Standard Errors Give Different Point Estimates Than Cluster SEs with ivreghdfe

[Roberto Santos]

Hi, I am running a robustness check to my main regression specification using Driscoll-Kraay standard errors, which are implemented in the package ivreghdfe. My main specification uses clustered SEs. I was expecting that my SEs would change once I ran them with Driscoll-Kraay SEs, but besides them, my point estimates also changed! I have taken a look at the original Driscoll-Kraay paper and I am not super clear why that would be the case. What is the reason for that? I am kind of puzzled, since I don't recall other cases in which the kind of SEs one choose change their point estimates. To make it more clear, I write below my cluster and Driscoll-Kraay SEs specifications.

Main specification (clustered SEs):

ivreghdfe outcome control1 control2 (endogenous=instrument), absorb(fixed_effect1 fixed_effect2 fixed_effect3) cluster(cluster_level) savefirst, if month<=90 & year>2001

Robustness Check (Driscoll-Kraay SEs):

ivreghdfe outcome control1 control2 (endogenous=instrument), absorb(fixed_effect1 fixed_effect2 fixed_effect3) dkraay(4) savefirst, if month<=90 & year>2001

Thanks!

[Carlo Lazzaro]

Roberto.
couldn't it be that Driscoll-Kraay implies lags?

[Roberto Santos]

Carlo Lazzaro , thanks for your reply. I think the lags are on spatial dependence on the error terms, they are not lags of regressors, but I could be wrong. I was looking again at the original DK

https://www.jstor.org/stable/2646837

and They only show one set of coefficients and multiple sets of SEs. That just makes me more puzzled that I am getting distinct results with DK. There is even an example with 2SLS in the original paper (and I am running an IV regression).

I was running the code again and I noticed two things:
-With clustered SEs it is dropping 33 singleton observations, while it drops 37 singleton observation in the DK case. This could explain small differences in the coefficients, but not large ones since my dataset is large (350k observations) and I have done robustness checks dropping more observations than that and the estimates are still comparable in magnitude.
-When I run the DK SE, I get the following message before the output that I am not sure what it means:

Warning: time variable cluster_level has 2462 gap(s) in relevant range

. Maybe it has something to do with it. My panel is unbalanced and my cluster_level is month-year-municipality. When a municipality enters the dataset, it almost always shows up every month until it is removed from the survey form where my data comes from. So it is unbalanced in the sense that some municipalities are available for all month-years, while some are only available for part of my study period (though they are available for continuous periods).

Sergio Correia , sorry to bother, but I wonder if you have thoughts on that. I feel this is not an issue/bug, so I haven't open a request in your Github.