-xtgls- vs -xtscc-?

[Mohsin Javed]

I am analysing a panel data with n=19 (ID/panel variable i.e. countries) and T=44 (time variable) to understand the drivers of equity flows to emerging markets. Before estimating the model, I run following diagnostic tests:

1. Using -xttest3-, I find presence of Heteroskedasticity.
2. Using -xtserial-, no autocorrelation is found.
3. Using -xttest2-, I find presence of cross sectional dependence.

Moving on, based on these, I compute Robust Hausman Test using -xtoverid- after running -xtreg, re(robust)-. Robust Hausman Test suggests using FE model.

Now, I have to decide between -xtgls- & -xtscc-.

By using help xtscc, I find this:

-xtscc, fe- performs fixed-effects (within) regression with Driscoll-Kraay standard errors. These standard errors are robust to very general forms of cross-sectional ("spatial") and temporal dependence (provided that T is sufficiently large). If the residuals are assumed to be heteroscedastic only: use xtreg, fe robust.

On the other hand, -xtgls- fits panel-data linear models by using feasible generalized least squares. This command allows estimation in the presence of AR(1) autocorrelation within panels and cross-sectional correlation and heteroskedasticity across panels.

So, I have to choose between the following two commands:

1. xtscc depvar indepvars, fe
2. xtgls depvar indepvars, panels(correlated) corr(independent)

Which one should I choose?

Thanks.

[Carlo Lazzaro]

Mohsin:
I would go -xtscc- as -xtgls-is a type of pooled OLS estimator.

[Mohsin Javed]

Carlo:

Thanks a lot.

[Ali Maleki]

Hello, I am new to the stata community. Can you help me whether xtreg, xtgls or xtscc should be used for the panel data with the following characteristics?
I have three panel data (N=12 countries T=40 years) with three different dependent variables (but same independent variables) . data is not sample but the populations.
all three have autocorrolation
2 of them have heteroskedasticity, but one does not have.

Husman test and Breusch and Pagan test for the three says
1- niether fixed effext nor random effect is not suitable for two of them
2-fixed effext is suitable for one else.

Can you please help me how can I choose between the three xtreg, xtgls or xtscc and also ( fixed effect or random effect?)?

I read some helps but more confuses how to go with the three commands.

[Carlo Lazzaro]

Ali:
the T>N structure of your panel dataset rules out -xtreg-.
If there is autocorrelation withing and across panels, you can go -xtscc-.
I would also take a look at -xtregar-.

[Ali Maleki]

Thanks Carlo
Using fixed effect is ok, given that I do not have random sample?
Is the any advantage using xtregar vs. xtscc?

[Carlo Lazzaro]

Ali:
most of the times, panel datasets include non random samples. Therefore, this one is not an issue.
Take a look at the helpfiles/entries of the two modules and compare their capabilities to choose which one is better for your research purposes.

[Ali Maleki]

Many thanks, I read the manuals, still could not distinguish the advantage of xtscc over xtgls, as they are both seems good when we have autocorrolation and T>N.
I would be very grateful if you could explain me.

[Carlo Lazzaro]

Ali:
see xtreg re vs xtgls - Statalist;
-xtscc- allows the -fe- estimator.

[Ali Maleki]

Thanks, I look at them. Husman test and Breusch and Pagan Lagrangian multiplier test suggest that none of the fixed effect models and random effect models are not suitable for me.
I am kind of confused what to do now. Is xtgls now which you mentioned is a type of pooled OLS works for me?

[Carlo Lazzaro]

Ali:
if there's no evidence of a panel-wise effect in your dataset, you can simply go pooled OLS with clustered stsndard errors.

[Jeff Wooldridge]

Ali: Most of these tests are of questionable value when N is so small. Moreover, the usual Hausman and B-P tests aren't robust to anything. You very likely have serial correlation. Unfortunately, xtgls has no options for robust standard standard errors and so you have to assume your model of serial correlation is AR(1) and there cannot be heteroskedasticity. In your setting, I would use time and country fixed effects with xtscc and the Newey-West standard errors. With T not very large, you should limit the lags in computing the standard errors to a fairly small number -- with T = 40, probably no more than three or four.