Panel data with N < T heteroscedasticity and autocorrelation. Should I include country dummies?

[Natasha Kara]

I have panel data of N=18 countries with T=72 months. Heteroskedasticity and autocorrelation are present in the dataset. I was working on Stata with xtreg fixed effects with clustered countries since hausman test indicated so (I was suspecting it as well). My first question since I read this (p.11) "However, the Hausman test is only valid under homoscedasticity":

1. I know that heteroscedasticity is present in my data from likelihood-ratio test. Should I not take hausman results into account? If so, do I have any alternative?

In addition, I was advised that clusters do not work if they are very few. Since my clusters of 18 countries are way fewer than the 42 indicated on "Mostly Harmless Econometrics" by Angrist & Pischke I tried to find another way to do so.
Then reading "Microeconometrics using Stata" by Cameron & Trivedi I found the xtpcse command which was ideal for cases like mine with smaller N and larger T. The command was correcting for heteroscedasticity and auto-correlation. I also used the psar1 correlation and everything seemed perfect until I started wondering whether I should include country dummies in the model. No clear answer was given to this question in fora I read. I was quite sure that there is a country effect (from hausman test and from intuition). So I started searching for the differences between xtreg and xtpcse by running regressions when I realised that xtpcse with country dummies is the same as xtreg, fe and xtpcse with no dummies is the same as xtreg, re. Then I couldn't understand why this xtpcse command is proposed for cases of N < T since it gives the same results as xtreg. Anyone could explain this?
So I'm coming to my second and third questions:

1. Should I or should I not include country dummies with my xtpcse command? My main problem is that country dummies change the sign of the variable I want to investigate (positive and significant without the dummies, negative and significant with) whereas the signs and significance of control variables remain (almost) unchanged. As you understand the question whether should I include them is crucial.
2. Is xtpcse the answer to my problem or in the end it doesn't really help?

I also posted the question here: http://stats.stackexchange.com/quest...ld-i-include-c. If I receive any useful comment I will let you know. Any suggestions would be appreciated. Thank you in advance.

[Kenulina Schneider]

Hi Natasha,

I have the same problem. I don t know which option to use. When I run xtreg depvar indepvar, fe robust, however the results are not consistent with the literature, but with xpcse depvar indepvar,correlation(ar1) rhotype(freg) hetonly option I get very similar results. So, my questions are:

1)if my data is suffering from both serial correlation and heteroskedasticity. So, is it better to apply xpcse option? ( my panel data consists of 9 countries and 12 years).
2)since the xpcse option does not work with fe and re, then how to apply hausman test? are the following commands in the right position?

xtreg depvar indepvar, fe robust
estimates store fixed
xtreg depvar indepvar, re robust
estimates store random

xtoverid
p<0.05 (so fixed is appropriate)
xtpcse depvar indepvars, correlation(ar1) rhotype(freg) hetonly

I know the post is old, but maybe you have already solved this problem.
I would appreciate any suggestion!

King regard,
Kenulina

[Carlo Lazzaro]

Kenulina may find at least one reply to her already posted query at: http://www.statalist.org/forums/foru...xpcse-vs-xtreg