Is there a solution coded in R to estimate models of the form

$$ y_{igt} = \alpha_i + P_{gt} + \beta_1y_{igt-1}+ \beta_2y_{igt-2} + X_{igt}'\gamma + \epsilon_{igt} $$ ?

plm offers the pgmm package, which implements the Arellano-Bond estimator, but it doesn't seem able to handle FWL-based demeaning of factors other than the cross-sectional unit, or the simple addition of time dummies. lfe on the other hand doesn't seem to be able to handle dynamic panel GMM estimators.

I've got N= 2000, n = 65K, G = 39, and t = 25, so including time:group effects as simple factor variables is not an option, particularly given that I'll need to fit multiple models to figure out what lag structure I need to wipe out the autocorrelation.

If nothing is coded yet, can anyone recommend any clever workarounds?

  • $\begingroup$ Have you considered the strategies in personal.ceu.hu/staff/matyas/docs/BMW_ER_OnLine_Paper.pdf? They work if you include only one lag of the dependent variable. See model (7) and (12), and section 5.2. I do not how to implement the strategies proposed in the paper in R (or Stata) though. $\endgroup$
    – Elias
    Oct 21, 2017 at 21:29