Dynamic panel data with small T & N (T>N) with endogenous independent variables.

[Michael McCaffrey]

Hi Stata Users,

I have a panel data set of 12 countries across 38 quarterly periods for a variety of financial indicators in the Islamic financial sector at a macro level.
Following the literature, I was going to use the two-step System GMM as my regressor, however, as both T and N are small and T>N, I've seen various threads advise against this regressor.

The generic version of my equation is:

Pi_t = a0 + a1Pi_t _1 + a2SUKUi_t + a3DIVi_t + a4KAR_t + a5CIR_t + a5X_t ei_t

where P is profitability, SUKU is an abbreviation for Sukuk holdings to Capital, KAR is the capital-to-asset ratio, CIR is the cost-to-income ratio, X are the macro control variables and e is the error term.

As many of the independent variables affect one another, is there a most-suited estimator for my regression?

(on a side note) A paper using the same database used the two-step system GMM and justified it by its ability to tackle endogeneity issues, does this mean their results are potentially inaccurate as their T was also greater than N?

Many thanks,

Michael

[Sebastian Kripfganz]

To tackle any endogeneity issue with suitable instruments, simply use the xtivreg command. You can also add country dummies explicitly to your regression model and estimate it with ivregress. With 38 time periods, the bias from including a lagged dependent variable might not be very large anymore (although this is generally hard to say). A (two-step) system GMM estimator is generally inappropriate with small N because it relies on estimating an optimal weighting matrix, which has a cluster structure at the country level. Even if you just use the inefficient one-step estimator, standard errors are still typically clustered at the group level, and therefore statistical inference will be inaccurate.

Also, most of the conventional model specification tests rely on large-N approximations of the test statistic's distribution.