Does using Discoll-Kraay standard errors correct heteroskedasticity and autocorrelation problem?

[Nicole Peredo]

Hello guys,

I am running a fixed effect regression model. I have a problem of heteroskedasticity and was wondering if I can correct this by using Discroll-Kraay standard errors.
First I tried using just robust errors using the command " xtreg, y x, fe vice(robust) " and the model lost significance, meaning my p-values > .0.1.
When using the Discroll-Kraay standard errors my model didn't lose any significance, meaning my p-values equal 0.00. Am I solving my problem by using these Discroll-Kraay standard errors?
I am very new to stata and panel data analysis and would appreciate the help!

[Joro Kolev]

The appropriate method depends on what panel you have.

Driscoll-Kraay standard errors take care of heteroskedasticity and both cross sectional and time series autocorrelation, but require a large T.

[Nicole Peredo]

Thank you Joro! I have currently T=16 years. If I understood correctly, it requires a minimum of T=20 years for Discoll-Kraay standard errors to take care for the problem. I am right?

[Joro Kolev]

No, there are no such hard thresholds like 20 or any other number.

16 periods is a bit short, but it is also not crazy to apply the estimator with 16 time periods. You should be fine.

Thank you Joro! I have currently T=16 years. If I understood correctly, it requires a minimum of T=20 years for Discoll-Kraay standard errors to take care for the problem. I am right?

[Jeff Wooldridge]

Nicole: How large is your N? T = 16 is pushing it, as the D-K standard errors work entirely off of the time series. It is like applying Newey-West HAC to a time series with T = 16. It's done, but the statistical properties cannot be very good. It's difficult to know whether the reduction in standard errors is simply due to the downward bias caused by assuming the serial correlation drops off after one or two periods -- which, I assume, is what you have specified in the HAC lag length.

If N is large -- or at least quite a bit larger than T -- you can dry using GLS applied to first differences to improve efficiency. But you still have to compute robust standard errors. Still, you may get enough of an efficiency gain.

[Nicole Peredo]

Thanks for the answers! Jeff I am currently working with T=19 and N=64! So I suppose I should try using GLS then. And I used a lag(4) when using D-K standard errors is that wrong?

[Jeff Wooldridge]

I would use something like this, where d2, ..., d19 are the time period dummies.

xtset id year
xtgee D.(y x1 ... xK d2 ... d19), corr(uns) vce(robust)

Whether this improves on usual FE remains to be seen -- as we say, an "empirical question."

Four lags seems a bit high with T = 19. What happens if you let the routine choose the lags?

[Nicole Peredo]

Thanks Jeff! I changed the lag to 1. How could I let the routine choose the lags?