Test: Biases in Long run coefficients in a Dynamic Panel Data?

[Manuel Almodovar]

Dear stata users:

I am learning the dynamic panel data models to use them in my research (at the moment with the xtabond2 command). To enrich the analysis I would like to include the Long run coefficients (in principle using the nlcom command).

The problem is that I have found a paper (quote at the end) that warns about the bias of the long run coefficients. In this respect my questions are:

1. Are the long run coefficients obtained with an xtabond2 and nlcom reliable?
2. Is it advisable to use the long run coefficients in publications or are they usually not relevant?

I indicate the paper mentioned above: W.R. Reed & M. Zhu (2016) On Estimating Long-Run Effects In Models with Lagged Dependent Variables. WORKING PAPER No. 16/2016 https://repec.canterbury.ac.nz/cbt/econwp/1616.pdf

Thank you for your answers

[Sebastian Kripfganz]

From personal experience, I agree with the authors' assessment that small-ish biases in the coefficient of the lagged dependent variable can lead to sizable bias in the long-run coefficents, especially when the persistence of the process is high. Those estimates should therefore be interpreted with caution. Yet, often those long-run effects are actually the main effects of interest.

The best one can do is to aim for an estimator that is relatively robust to changes in the specification and has a relatively small variance. Unfortunately, this is often not the case with GMM estimators, especially when the cross-sectional sample size is relatively small. Stronger assumptions (stronger instruments) can help.

(The nlcom command is fine. If there is a problem, it originates in the initial estimation of the coefficients.)

[Manuel Almodovar]

Thank you very much for your reply. It has been helpful.