Fix effect model with autocorrelation and heteroscedasticity

[aifar baig]

hi... my model is based on the data for 17 banks for 15 years.. my diagnostic tests shows presence of autocorrelation and heteroscedasticity. I am applying is fix effect model. I have been suggested i should use panel corrected and Driscoll-Kraay to resolve above said issues. As i am not familiar with both of these models can you please guide me on this and let me know if these are right models for me.

[Carlo Lazzaro]

Aifar:
welcome to this forum.
Assuming that you're referring to -xtreg,fe-, if you do not have across-panel correlatiion too, you can go:

xtreg <depvar> <indepvars> <controls>, vce(cluster idcode)

[aifar baig]

Aifar:
welcome to this forum.
Assuming that you're referring to -xtreg,fe-, if you do not have across-panel correlatiion too, you can go:

xtreg <depvar> <indepvars> <controls>, vce(cluster idcode)

Thanks Carlo for your reply. yes I am referring to xtreg,fe and no I do not have a cross correlation, but auto correlation and heteroscedasticity. The probability of wooldrige and modified wald test is "0.0000".

So just to confirm. As per your suggestion through vce(cluster idcode) these issues will be resolved!!.

Thanks again and much appreciate your help.

[Carlo Lazzaro]

Correct.
Clustered robust standard error in -xtreg- take both heteroskedasticity and autocorrelation into account.
You can invoke it via -vce(cluster idcode)- or -robust-.
Personally, I prefer the former to avoid possible mistakes when I switch from -xtreg- to -regress- (where the two options do not do the same job, as -robust- deals with heteroskedastcity only, whereas clustered standard error is imposed via -vce(cluster clusterid)-)

[aifar baig]

Thank you so much for your help Carlo👍👍.. i will apply it...

[Jeff Wooldridge]

Aifar: The method for computing standard errors is different from the model you're estimating, and even the estimation method. With N = 17 and T = 15 it's tough to know the right thing to do. As Carlo pointed out, you can cluster at the level of the bank; N = 17 is pretty small but you can see what happens. Driscoll-Kraay is a very different way of computing a standard error. It allows cross-sectional correlation but dampens the correlations by using a Newey-West estimator. T = 15 is pretty small to apply that, but it's worth a try. Doing both is a good idea, with the caveat that both are problematical with small dimensions. There's not guarantee they're similar, unfortunately.

Carlo showed you the clustering command. You probably want to include i.year so that you also have time fixed effects. And use the fe option for two-way fixed effects.

xtscc is a user written command that does Driscoll-Kraay. You need to choose a lag for the maximum amount of serial correlation. With T = 17, probably two or three at most.

xtscc y x1 ... xK i.year, fe lag(2)

[aifar baig]

Thank you Jeff.. I applied both.. the results of main IVs are significant in both.. but there is a difference in results of some control variables.. By applying lag(2) or (3) there is not much of a difference so i am sticking to lag(2).. So now I am applying xtreg,fe with vce as well as Driscoll Kraay...

I will grateful of you can suggest me any further reading on using vce and Driscoll Kraay that can help me in understanding it better and explaining it in my analysis..

[Carlo Lazzaro]

Aifar:
see:
1) Driscoll, J. C., and A. C. Kraay. 1998. Consistent covariance matrix estimation with spatially dependent panel data. Review of Economics and Statistics 80: 549-560;
2) Daniel Hoechle 's article at: https://www.stata-journal.com/articl...article=st0128

[aifar baig]

Thank again Carlo..