Pooled OLS/ FE model/ RE model: sigma_u and rho = 0 ---> Breusch Pagan gives prob>chi2 = 1.000

[Yannick van Weeren]

Dear All,

I am working with a panel dataset with 12 countries (the groups) and 5 years, so in total 60 obs (I know that's not a lot). I have reasons to assume that there are fixed effects for each country and for each year. Over time all the 3 dependent variables (on each of which I will regress a set of 11 independent variables) tend to increase. I did a Breusch-Pagan test to test whether to use Pooled OLS or panel regresssion. For all 3 dependent variables I get sigma_u = 0 and rho = 0 when performing a Random Effect model: xtreg dv iv's, re. When using the xttest0 command after that, I get prob>chi2 = 1.000 for all 3 models.

When I ommit one independent variable for one of the models, it suddenly gives a value for sigma_u and rho and for prob>chi2 it gives a value < 0.05, which means that I should use a panel regression. When performing a hausman test afterwards, I end up at the Fixed Effect model. The other models with the other 2 dependent variables remain giving sigma_u = 0, rho = 0 and consequetively prob>chi2 = 1.000 when ommitting that same independent variable.

How should I interpret prob>chi2 = 1.000? Which model should I use?

Underneath is the regression with ND as dependent variable. After dropping GFCFgr, sigma_u and rho gave a value and prob>chi2 < 0.05 (Breusch Pagan): the Hausman test indicated I should use the FE model, which corresponds with my expectations considering the real life situation.

Would be great if someone could help me out here.

Thanks in advance.

Kind regards,

Yannick

[Carlo Lazzaro]

Yannick:
the main problem seems to rest on an having a too limited sample size for too many predictors: you're simply asking too much out of your data. That's why the test you performed gave you back contradictory results. I would switch to a more parsimoniuos model and see whether those drawbacks still persist.