Panel Unit Root Test

[Prashant Bhandari]

Hello Everyone,
I am working on studying the relationship between military expenditure (independent variable) and FDI inflow (dependent variable) using the panel of 63 countries spanning the period 1990-2018.

I think that the level of FDI inflow in year t will be impacted by the level of FDI inflow in the year (t-1). Thus, I was thinking of adding lagged FDI as one of my control. So one question is:

I) Is it reasonable to add lagged FDI as a control?

More importantly, my control variables include macroeconomic series like Inflation, Openness to Trade, GDP per capita growth rate, etc. I came across some papers (Cho, H.C. and Ramirez, M.D., 2016 and Bashier, A.A. and Siam, A.J., 2014) where they test for panel unit roots and co-integration. As an undergraduate student, I have taken a couple of Econometrics courses. But I had never come across any examples where the stationarity of the dependent variable was checked. I explored a couple of textbooks (Woolridge and Stock and Watson) but did not find this approach described in these books. Do you think it is necessary that I follow this route for my project? Also, since my data is an unbalanced panel due to a missing data problem, I found that there are only two different tests that allow me to check for panel stationarity.


Any suggestions would be helpful.


Best,
Prashant Bhandari

[Fei Wang]

Prashant, I'll skip your questions and directly give a proper model specification in my mind.

xtset country year
xtscc lfdi lme inflation openness gdpgr others year, fe

-xtscc- is a user-written command being able to account for general forms of cross-sectional and serial correlations -- For a panel data at a geographic level, taking both correlations into account matters. "lfdi" and "lme" are log terms of FDI inflow and military expenditure. It's very likely that both original terms are non-stationary and even non-linear along the time dimension. Taking log of them not only simplifies interpretation (elasticity) but also makes both trends more linear (particularly when annual growth rates are nearly constant). Then including the linear form of "year" would on the one hand mostly detrend the two variables, and on the other hand won't affect the inclusion of macro series. Option -fe- means fixed effects estimation which is equivalent to controlling for country-specific effects and ruling out any country-level time-invariant unobservables. I would graph all variables along time before justifying the model specification, and do no more tests beyond that.