What is "pooled OLS"? And is i.country a fixed effects estimator?

[Jonas Wereld]

1) I'm trying to understand what "pooled OLS" would be. Does the inclusion of "vce(cluster country)" (clustered standard errors) or the inclusion of e.g. i.country (country specific effects) turn a normal linear regression into a pooled one?

2) Is not i.country itself a fixed effect? Then why do we have xtreg / xtlogit etc.? If I just want to control for country fixed effects, is i.country sufficient?

Thanks in advance!

[Carlo Lazzaro]

Jonas:
-in POLS, vce(cluster clusterid) mitigates the existence of non independent observations (that is, the panel data structure of your dataset);
- in limited instances POLS outperforms -xtreg- when it comes to panel data regression models. One case in point is when the F-test reported at the foot of the outcome table of -xtreg,fe- fails to reach statistical significance (i.e., when there's no evidence that the individual effects, taken together, are different from zero);
- i.country is a fixed effect (but you should be more detailed about what you're after to elicit more positive replies);
- serial correlation is another reason why -xt- suite commands should be preferred to the ones developed for one-wave data regression models;
- as an aside, please note that -fe- specification for -xtreg- and -xtlogit- differs.

[Jesse Wursten]

2) At some point you may have encountered the term "LSDV"-estimator. This stands for "Least Squares Dummy Variable"-estimator and is one of the ways to filter out fixed effects. If I recall correctly, xtreg instead uses the within-differencing method, which subtracts the mean over time per panel unit from all variables. Both methods normally produce the same beta coefficients. The xtreg method is normally faster.

Note that the standard errors might differ if you use robust/cluster(), I think xtreg uses a better finite-sample adjustment than reg (the difference won't be large though). See "A Practitioner's guide to cluster-robust inference" by Cameron and Miller.

[Jonas Wereld]

Thank you both very much. I have an additional question. I ran the following code

logit depvar v1 v2 v3 v4 i.country, vce(cluster country)

which gave me a strange missing values result:
Wald chi2(3) = .
Prob > chi2 = .
( Log pseudolikelihood & Pseudo R² seemed fine though).

In contrast, running clogit or xtlogit, re with the "vce(cluster country)" option does not seem to produce similar problems.

What is going on here? Why could I not combine the i.country fixed effect with the clustered standard errors in the logit command, but in clogit or xtlogit, it computes fine?

[Richard Williams]

You should use xtlogit or clogit rather than adding i.country. A brief explanation is given in the appendix of

http://www3.nd.edu/~rwilliam/xsoc739...edVsRandom.pdf