Hamza,

The right approach is to do a thorough methodological review of related studies and benchmark their procedures.

Thanks.

I agree with Ngozi ADELEYE (above). Start with his advice. Beyond this, considering the limitations of your data (78 observations) you are likely to obtain useful estimates only from least squares estimation. Why? Because the characteristics - unbiasedness, consistency, efficiency (i.e. precision) - of all other panel estimators (IV, GLS, MLE, GMM) are asymptotic, i.e. dependent on large samples (in practice, at least many hundred, but better many thousand or more observations). Only least squares estimation has known and good small-sample properties (and 78 is a small sample). This consideration restricts you to a static fixed effects model, which is the workhorse panel model and estimated by least squares. Note also that your small sample has implications not only for the validity of your point estimates but also for your statistical inference. Generally in panel analysis, you would want to estimate with cluster-robust standard errors. However, the desirable characteristics of cluster-robust standard errors are also dependent on asymptotic reasoning, but in this case with respect to the number of clusters (countries, in your case) (not the number of observations). 13 clusters (countries) is unlikely to be sufficient to yield valid cluster-robust standard errors; indeed they are likely to be smaller than the default standard errors calculated by your software package. With a small number of clusters, good practice is to report both cluster-robust standard errors and default standard errors. Then adopt a conservative approach to inference by using the largest SEs. Bottom line. If you want the advantages of large-sample estimation and inference, you need a much larger dataset.

Without knowing what resources you have available, recommendations are not so easy to make. However, in any case, I suggest that you focus on the finite-sample (which in practice means small-sample) properties of estimators as opposed to their asymptotic (i.e. large-sample) characteristics. To the best of my knowledge, it is only least squares estimation that has known and good small-sample properties (unbiasedness, consistency and relative efficiency) that can be demonstrated analytically. For other estimators, small-sample characteristics have to investigated by means of numerical studies, in which the assumptions rarely fit precisely the specific datasets available for actual research projects.

I am a bit sceptic about the possibility of being able to obtain something reasonable.

It is not a micropanel as N is small, but T is also small so panel time-series is not choice. As most, you can look at fixe-effects OLS without looking at serial correlation issues, but I do not think one can get something insightful there...

Dear precious researchers, then could not the Structural equation Modelling be a good technique for getting results with this data?

I am sceptical about Hamza' suggestion, because the desirable properties of Structural Equation Modelling are asymptotic (i.e. require a large sample - at least many hundred and preferably many thousand - or more - observations). As I have argued above, and by Jose-Ignacio above, a small panel - especially one with limited time-series depth - is almost certainly best approached by fixed-effects estimation, which has known and good small-sample properties.

Dear Geoffrey Thomas Pugh,

I further proceed, through using the SEM approach of Path Anlysis and PLS-SEM (which is best for small sample data).

With respect to bias, consistency and efficiency, how do the properties of PLS-SEM compare with the small-sample properties of least squares estimation? The point is not whether some variant of SEM is the best available SEM approach when trying to analyse a small dataset, but whether it is as good in this respect as least squares estimation.

Dear Respected Researchers,

I hope you all are doing well. Almost, you all have recommended me for using OLS estimation technique as it was small sample panel data with N=13, T=6. But now, i have made few modifications in my data by adding 7 more countries in my sample data so now it's N=20, T=6. hence, i would like to ask you whether should i go with same OLS estimation technique with fixed effect model as I was recommended before or now should i go with GMM estimation technique as my N=20 & T=6 (more n of individuals)? What estimation techniques suit this sample of data now? I am looking to all of you for your kind responses, thanks!

Note: I may be looking to employ PCA analysis as well, as i have many numbers of indicators.

Geoffrey Thomas Pugh José-Ignacio Antón Bosede Ngozi Adeleye Suhail Ahmad

Following this question

N=20 & T=6 leaves you with the same problem and solution. For more sophisticated estimation you need much more data. There must be more than 20 developing countries, so enlarge your dataset to the limit. That will also enable you to explore heterogeneities between distinct groups of developing countries. Greater time-series depth would also be useful, although at a certain point (say T=20) you would need to consider panel time-series approaches rather than conventional panel analysis.

Thanks! Geoffrey Thomas Pugh for your kind recommendations.

You are no longer require to use PCA. As your total number of variables becomes (2 x 5 = 10). Therefore, under total fixed effects (TFE) you may process the data smoothly.