Hi Alessandro,

There are two issues here. The first is the "appropriate" estimator which is really specific to the question that you are trying to address, independent of sample size. An OLS specification is only appropriate in the very unlikely scenario that there is no country-specific component to you error term. If this is not true, then your estimated error will exhibit serial correlation and your estimates may be bias.

In these country-panels the answer is almost always a fixed effects specification, but that really depends on the question being asked. If you think that there is a country specific component to your error structure which is correlated with one or more of the regressors, then a fixed effects specification is appropriate (otherwise estimates will be biased.) If there is a country specific component to your error structure which is uncorrelated with the regressors, then a random effects estimator is fine.

The potentially bigger problem you face is sample size. Note that the “known” sample properties of OLS with small samples is under the assumption that the errors are normally distributed (this assumption does not need to be made large sample OLS) and uncorrelated with your regressors, both strong assumptions.

Given what you have said about the problem, I would strongly recommend fixed effects regression (forget the more complicated non-linear stuff). However, you will face an issue with sample size, you will need some big variation in your regressor. I would strongly recommend, if at all possible, getting your hands on more data.

Thanks Jesse,

For the moment I employed a fixed effects specification (within groups estimator), but also FGLS controlling for heteroskedasticity, and non parametric estimators (to don't make strong assumptions about distributions). Unfortunately we are constructing the series of data and there is no possibility of going back further.

With 22 countries and 13 years, I think you can estimate a country fixed effects model with 12 year dummies and 7/8 regressors. Given that there is heteroscedasticity and autocorrelation, better to use robust standard errors. Given the small sample size, it is better to provide additional evidence in favor of your results (using non-parametric techniques/ alternate definition of dependent variable and controls) so that the reader is convinced regarding the estimated relation.