difference vce (robust) vce (cluster ID)

[Maria Kohnen]

Hi forum members,

I am doing a fixed effect panel regression and wanted avoid heteroskedasticity and autocorrelation.

Does

vce (robust

already suffice? or do i need to apply

 vce (cluster ID)

?
or are they intercangeable?
Lot of confusing info on this topic...

thank you,
best

[Abdul Adam]

Maria: see @Clyde Schechter #2 answer here: https://www.statalist.org/forums/for...-id-vce-robust

[Clyde Schechter]

Yes, this topic can be confusing. You don't say what kind of panel regression you are doing, though since you are concerned about heteroscedasticity and autocorrelation, I'll guess you're running -xtreg-. If it is -xtreg, fe-, then the non-cluster robust VCE is not available, and if you specify -vce(robust)-, Stata automatically uses -vce(cluster ID)- instead (assuming ID is the panel variable in your -xtset- command). So you don't have to think about it.

If you search the Statalist archives, I believe that you can find a post there from somebody at StataCorp explaining why -vce(robust)- is not valid with -xtreg, fe- and justifying their decision to implement -vce(robust)- as -vce(cluster panelvar)- in this command.

Added: Looking into it a bit farther, if you go to the PDF Documentation section on -xtreg-, you will find an explanation about the use of -vce(cluster panelvar)- in this context, along with some references. Let me also just remind you that the correct use of any kind of clustered VCE requires that the number of clusters be sufficiently large. There is no uniform agreement about how many clusters constitutes "sufficiently large," and I have seen rules of thumb from as low as 10 to as high as 100.

[Maria Kohnen]

Dear Clyde,

I thank you for your answer, which is much appreciated.
Sorry, yes, I am running -xtreg, fe- . I did specify vce(robust). So from what I understood that was perfectely okay to do, since stata automatically just uses vce (cluster ID)? so i do not have to change anything? ID is the panel variable in my xtset command. The number of clusters is 122, which should be okay i guess?
So including vce(robust)/ which means vce(cluster ID) makes my SE robust aganst heteroskedasticity and autocorrelation?

best regards

[Clyde Schechter]

Yes, it sounds like -vce(robust)- is appropriate here, and does what you want. And Stata will interpret it to mean -vce(cluster ID)-. And I don't think anyone will worry that 122 clusters are too few.

[Maria Kohnen]

I thank your for the information.
best regards