[Linh Nguyen]

Lihn:
1) the issue is that, even if -xtreg- control for autocorrelation, you cannot chose the model of autocorrelation (whereas you can with -xtgls- option).
2) What above does not mean that -xtreg.fe- does not approapriately deal with both heteroskedsticity and autocorrelation: it simply does not allow you to model the autocorrealtion as -xtgls- does.
The underlying reason is that modelling autocorrelation is an issue for T>N (long) panel datasets (-xtgls-), not (usually at least) for N>T (short) panel datasets (-xtreg-).

Thank you so much!

[Daniel Bond]

Frans:
for your panel I would recommend -xtgls-, not -xtreg-.
The -hausman- test outcome points you to -fe- or -re- specification, without explicitly considering the N and T dimension.

Dear Carlo,

Apologize for my prompt question.

I just do no understand why you recommended Frans to use -xtgls- for his panel while his data set was short panel (N>T) datasets that you suggested not to use -xtgls- in #2.

Since I have a similar datasets and questions of Frans, I want to figure them out. thanks!

Best regards.
Daniel

[Eric de Souza]

Concerning -xtregar-, please see this thread:
https://www.statalist.org/forums/for...ge-t-estimator

[Carlo Lazzaro]

Daniel:
with an average number of observations that approaches 62, the T dimension is relevant and modelling autocorrelation is probably an issue that is worth considering.

[Daniel Bond]

Daniel:
with an average number of observations that approaches 62, the T dimension is relevant and modelling autocorrelation is probably an issue that is worth considering.

Well noted! thank you very much!

[Linh Nguyen]

Lihn:
1) the issue is that, even if -xtreg- control for autocorrelation, you cannot chose the model of autocorrelation (whereas you can with -xtgls- option).
2) What above does not mean that -xtreg.fe- does not approapriately deal with both heteroskedsticity and autocorrelation: it simply does not allow you to model the autocorrealtion as -xtgls- does.
The underlying reason is that modelling autocorrelation is an issue for T>N (long) panel datasets (-xtgls-), not (usually at least) for N>T (short) panel datasets (-xtreg-).

Hi Carlo,

I accidentally read this post again and have one question:

You said "you cannot chose the model of autocorrelation". I don't understand what you mean about the word "chose"

For example: xtreg y x1 x2, re cluster (id) vs xtgls y x1 x2, corr(ar1) panels (h)

Does "chose the model" mean that we can delete some independent variables in our model? or what?

Could you please clarify your idea in further detail?

Thanks

[Carlo Lazzaro]

Linh:
I meant that, under -xtreg., -robust- and -vce(cluster clusterid)- options consider the same autocorrelation pattern for all the panels.
As you surmised, this is not an issue with -xtgls- options.

[Jeff Wooldridge]

It's not true that xtreg with vce(robust) imposes the same autocorrelation pattern for all panels. It's not model the autocorrelations but the standard errors allow any kind of serial correlation pattern -- heterogeneous across i or not. Now, this only holds when T << N. When T and N are comparable, Chris Hansen (2007, Journal of Econometrics) showed that one does need to assume the autocorrelations die out sufficiently quickly over time. But, of course, that is implied by the very special AR(1) structures usually estimated using xtgls.

Unless you really know what you're doing, I would avoid xtgls. It doesn't allow robust inference when it should allow a Newey-West HAC estimator to account for how the simple model of serial correlation could be wrong -- it is wrong -- especially with large T. It's really more like multiple time series even though, technically, it's a panel data set.

Without seeing the example, I would tend to recommend using two-way fixed effects. When T is sufficiently large, you can obtain valid standard errors that allow cross-sectional correlation because it applies Newey-West in the time series dimension. The original user-written command is xtscc. Now it is allowed in reghdfe, too.

[Carlo Lazzaro]

Thanks Jeff for correcting me.
I should have been more detailed in my previous reply.

[mausumi mohanty]

Hii
I am doing a panel data analysis with Time periods greater than No.of cross sections (T>N). I did a hausman test and also used xtoverid and verified that Random effect model is appropriate. but my model has serial correlation and cross section dependence problem. Can i directly use the xtgls command without running any random effect estimation technique? Is my model and command correct?

xtgls depvar indepvar controls