xttobit with robust and clustered standard errors || pooled tobit with robust and clustered standard errors

[Martin Tao]

Hi all,

I hope this finds you well!

As I unfortunately could not find a thread to answer my question, I opened a new one.

(1) Is it possible to adjust standard errors using Huber/White standard errors and/or clustering standard errors when using a random-effects tobit panel regression model? I have tried the following code.

xttobit y x, vce(robust) vce(cluster firmid)

Error: "option vce() specified multiple times"

xttobit y x, vce(robust)

Error: "vcetype 'robust' not allowed"

xttobit y x, vce(cluster firmid)

Error: "vcetype 'cluster' not allowed"

After I concluded that it is likely not possible to use xttobit with robust and/or clustered standard errors, I tested the same code with the -tobit comand which worked.

tobit y x, vce(cluster firmid)
tobit y x, vce(robust)

(2) However, it still seems not possible to use both commands vce(robust) and vce(cluster firmid) together. Is it generally possible to use both commands together? Or do clustered standard errors are automatically robust? I am asking as I found in one paper that used robust Huber/White standard errors (White, 1980) AND clustered standard errors at the firm level.

Background: I have panel data on a firm-year level. I chose to use a tobit model as my y variable is censored at 0. I tried to use -xttobit instead of -tobit first as I assumed it to be generally more appropriate when having panel data.

Thanks.

Best,
Martin

[Carlo Lazzaro]

Martin:
welcome to this forum.
1) it's true that the options you mentioned are not availabe for -xttobit- (surely for sound statistical reasons that, unfortunately, I do not remember). An alternative option is -bootstrap-ping the standard errors.-
2) unlike -xt- commands, in cross-sectional commands the -robust- option is developed for dealing with heteroskedasticity only, whereas -cluster- is conceived to take autocorrelation into account.
While it is not possible to use both options at the same time, in case you detected both heteroskedastcity and autocorrelation, it's wise to go -cluster-.

[Martin Tao]

Thank you very much Carlo, that helps a lot!