testing for heteroskedasticity in long panel datasets

[Michael Lee]

Hello everyone,

I'm working with a long panel dataset (T>N) and so I've chosen to use -xtpcse- command.
1. How can I test for existence of heteroskedasticity in this case?
2. to check autocorrelation I've used xtserial, is it right to do this?

Thank you in advance,
King regards,
Michael

[Joro Kolev]

The user contributed -xtserial- is solving a problem that you do not have -- it is for fixed effects panel data where N is large and T is small.

I think that whatever autocorrelation and heteroskedasticity tests you find in Stata should work for your case.

[Michael Lee]

The user contributed -xtserial- is solving a problem that you do not have -- it is for fixed effects panel data where N is large and T is small.

I think that whatever autocorrelation and heteroskedasticity tests you find in Stata should work for your case.

Thank you Joro.
But my problem is this: commands that I knew like -hettest-, -imtest-, -rvfplot-, -xttest*-, just works after commands like -xtreg-, -xtgls-, and -reg-. and the results of tests after xtgls and xtreg for heteroskedasticity are different! (for first one p-value for Wald test (het) is 7.1% and for the first one 3.6% for example!).
What should I do in this case?

[Carlo Lazzaro]

Michael:
set aside the test issue for a while, please note that you're comparing two way different estimators. No wonder that p-values differ.
If you have a T>N panel dataset, you should focus on -xtgls- and -xtregar- only.

[Michael Lee]

Michael:
set aside the test issue for a while, please note that you're comparing two way different estimators. No wonder that p-values differ.
If you have a T>N panel dataset, you should focus on -xtgls- and -xtregar- only.

Dear Carlo, due to heteroskedasticity, autocorrelation and cross-sectional dependence and having an unbalanced panel data I preferred to use -xtpcse-, according a discussion in this forum....

[Maxence Morlet]

Dear Michael,

How many time periods and cross-sectional units do you have?

[Michael Lee]

Dear Maxence,
T=190 (time periods) and N=6 (number of panel)

[Maxence Morlet]

You may want to consider the community contributed command xtscc, and check out the paper by Hoechle (2007).

[Michael Lee]

You may want to consider the community contributed command xtscc, and check out the paper by Hoechle (2007).

I have unbalanced panel data...

...and having an unbalanced panel data

[Jeff Wooldridge]

xtscc should allow unbalanced panels. It’s fixed effects with standard errors that allow cross-sectional dependence, serial correlation, heteroskedasticity.

[Jeff Wooldridge]

xtscc should allow unbalanced panels. It’s fixed effects with standard errors that allow cross-sectional dependence, serial correlation, heteroskedasticity.

[Michael Lee]

What is the advantage of -xtscc- over -xtpcse-?
In long unbalanced panel data?

[Maxence Morlet]

Professor Wooldridge, please correct if I'm wrong, but this is my understanding of the situation here:

The paper I mentioned, Hoechle (2007), provides a good overview of this issue: you have T>N, but T/N does not tend to infinity in your case, relatively far from it. Through Monte Carlo simulations, Hoechle (2007) demonstrates that when large T asymptotic properties cannot be achieved due to a finite number of time periods (as is the case here), Driscoll-Kraay (1998) standard errors outperform Beck and Katz (1995) panel-corrected standard errors.

[Michael Lee]

What is the advantage of -xtscc- over -xtpcse-?
In long unbalanced panel data?

Can anyone help in this regard?

[Maxence Morlet]

Can anyone help in this regard?

Have you read #13??