Panel regression model with N>T and serial autocorrelated error.

[Ivan Gufler]

Dear all,

I've been trying to estimate a panel regression model on a dataset with N>T, where N is the number of cross-sectional units and T is number of time observations. I want to include a fixed effect.

I ran the Wooldridge test for serial autocorrelation and I rejected the null: so the model has serial autocorrelated errors.

Due to the fact that N>T I understand that I cannot rely on -xtreg-. My question is: given these conditions is it good to estimate the model by -xtregar- including a fixed effect? Is this model consistent with the fact that N>T and that the error is autocorrelated of order 1?


Many thanks to those who can help me

[Jeff Wooldridge]

I'm not sure where you got the impression that you cannot use xtreg just because there's serial correlation. It's more of a curiosity unless you want to try to improve efficiency. You may use xtreg with the fe option as long as you also use vce(cluster id) to make your standard errors robust to serial correlation.

What are your N and T? N > T is preferred for using xtreg.

[Ivan Gufler]

I'm not sure where you got the impression that you cannot use xtreg just because there's serial correlation. It's more of a curiosity unless you want to try to improve efficiency. You may use xtreg with the fe option as long as you also use vce(cluster id) to make your standard errors robust to serial correlation.

What are your N and T? N > T is preferred for using xtreg.

I'm sorry, there is a typo: I meant that N (number of countries in my sample) is much less than T (number of time observations). In particular I have 7 countries and 193 time observations (quarterly). My doubt is on xtreg, which requires, for asymptotic properties, that N is much greater than T. Unfortunately, my sample does not satisfy such condition.

I wonder whether estimating the model by a GLS with fixed effects and AR(1) disturbance (command: -xtregar) is the best (and most consistent) option.

Thank you

[Eric de Souza]

With some 50 years I don't see the advantage of pooling the countries together in a panel and imposing identical coefficients across countries.

[Jeff Wooldridge]

Okay, while that's technically "panel data," it's not amenable to standard panel data methods. It's more like multiple time series analysis. Eric's point is well taken. Now the issue is whether you want allow allow unrestricted time effects; that's the main reason for pooling across countries. But I doubt you need that much flexibility. Maybe a time trend and seasonal effects are enough.

You can allow correlation across country by using the Driscoll-Kraay approach. The standard errors are also robust to general forms of serial correlation. The command is user written: xtscc.