Determining the use of clustered standard errors

[Constantin Domizlaff]

Hi all,

I am running a regression model regarding the impact of board characteristics on corporate emissions. I know heteroscedasticity is present and I want to use robust standard errors to deal with that. I now want to know how to determine whether the normal option

 reg Y X, vce(r)

should be used or rather a cluster approach. Through the use of

loneway Y Z

where Z is my cluster variable, I have found an intraclass correlation of about .97 for my 503 different clusters which I assume to be quite high and thus I assume I should use the clustered standard errors rather than the normal white standard errors. Am I correct here or did I make a mistake somewhere?

Thanks a lot in advance!

[Maxence Morlet]

Do you have panel data? Within what unit is correlation between residuals most probable? Cluster by that unit. How many of these units do you have? If you have less than 40, I would recommend using Wild Cluster Bootstrap.

Also, using vce(r) will invoke HC1 standard errors, which are often too optimistic in small samples. I recommend HC3 standard errors:

, hc3

[Constantin Domizlaff]

Hi Maxence, thanks for the reply.

Yes, I do have panel data. There are 503 different units (companies) where the correlation between the residuals is most probable. The sample is ~2600 observations. So you would recommend the HC3 standard errors over the clustered standard errors as well?

[Maxence Morlet]

No in this case CV1 standard erros, so vce(cluster companyID) should cut it.

Otherwise you can consult this post for more info on the topic, for CV3 standard errors (more conservative):

https://www.statalist.org/forums/for...tandard-errors

[Constantin Domizlaff]

Alright, thank you very much!