I have data on GDP, GDP growth, FDI, and growth factors for N = 4 countries from 1990 to 2020 (so T = 30). As I am performing macro analysis, I was primarily interested in comparable cases, but this led to a problem of small sample size.
I'm attempting to determine whether differing levels of human capital (measured by education and health measures) was a determining factor in FDI attraction and growth in these four comparable countries over their development paths (with dependent variable being GDP growth).
OLS and FE work in small samples, but my sample is very small, and I am reasonably certain that the trends are related and dynamic - there is almost certainly cointegration and unit roots, and serial correlation both within and across regressors. However, most tests for these things are designed to work within one country, and are not comparative.
What method would work best to estimate this?
Hi Alexandria,
You can try to use multilevel (linear mixed effect) regression where you set the slopes of each factors to vary accross the 4 countries, then you also add dummies to account for the temporal variation (by setting the first year as the baseline). Also set these dummies to vary accross countries. That way you can compare the impact of these factors accross these countries in a more parsimonous way. The problem is you have a very small number of group (country) and quite short temporal variation. However you might want to look it up and try...
Look at the sides of the OECD, and Google Data Explorer. You can easily find more data there to increase your N. Then you can use a PVAR approach. Moreover, notice that it might be smart to divide the sample into 2-3 time periods, as you otherwise assume that the relationships between your variables are more or less identical over 30 years, which seems to be a strong assumptio.
Dynamic panel models like Arellano bond (xtabond,xtdpdsys) are used for large N and small T. For small N and large T, you can use FE, or Zellner's seemingly unrelated regressions (SUR)-with Stata command sureg. or FGLS (xtgls).
Considering the have a dynamic model with the large time-dimension and comparatively small number of units of observation, you may want to consider cross-sectional ARDL (Pesaran and Smith 1995; Chudik et al. 2016).
Chudik, A., K. Mohaddes, M. H. Pesaran, and M. Raissi. 2016. Long-Run Effects in Large Heterogeneous Panel Data Models with Cross-Sectionally Correlated Errors Essays in Honor of Aman Ullah. 85-135.
Pesaran MH, Smith R (1995) Estimating long-run relationships from dynamic heterogeneous panels. J Econ 68(1):79–113
- Badges
- Science topic
- Similar topics
- Correction