problems with using GMM Using

[Martin Eden]

Hi there,

I am writing an article on the impact of various factors on the share of intangible assets in the company. In a number of works on which I rely, the authors used the GMM estimator.

I used FE, RE for my panel data, evaluated a more suitable model using the Hausman test. I also tried using GMM Estimator.

The problem, it seems to me, is that there are a significant number of omissions. In particular, for some IDs (in my case, these are companies) there are observations from 2000 to 2010 and none from 2011 to 2021. Can someone suggest how to improve the reliability of the results obtained?

For clarity, I attach the results of four models:

[Sebastian Kripfganz]

First of all, you need to decide which model you want to estimate. Your FE and RE estimations are for a static model, your GMM estimations are for a dynamic model with a lagged dependent variable. If you believe your data-generating model is indeed dynamic, then you can just forget about the FE/RE estimators. If there is no good theoretical reason for a lagged dependent variable, then do not bother with estimating a dynamic model.

Second, the FE/RE estimators rely on the assumption that all regressors are strictly exogenous; i.e., there is no contemporaneous or dynamic (lagged) feedback from the dependent variable to the regressors. In many applications, this is a quite strong assumption. If that assumption is justified, but you would like to estimate a dynamic model instead, a bias-corrected FE/RE or a ML estimator might be an alternative to GMM estimators:

Third, your GMM estimators likely suffer from too many instruments, which biases the results and weakens specification tests. If you want to delve deeper into dynamic panel model GMM estimation, the following presentation slides might be a helpful start:

• Kripfganz, S. (2019). Generalized method of moments estimation of linear dynamic panel data models. Proceedings of the 2019 London Stata Conference.