In dynamic paneld data (DPD) setting, FE or RE estimation would be biased since the within transformation will drive correlation between transformed lagged variable and transformed error. The common way to deal with it is A-B estimator which uses GMM. The original estimation they came up with is called difference GMM where lagged level DVs is used as instruments. However, it is well known that lagged variable is not proper IV. Later, they provided system GMM where lagged level DVs and differenced DVs are used together as instruments.
The attached slides by Baum would provide you with a general idea and corresponding Stata code, hopefully.
|