Estimation Methods in Panel Data Model
Static linear panel data model: Yit=bXit+aZi+Eit, can be divided into two groups according to its unobservable heterogeneity Zi’s property: one is called fixed effect when Zi is correlated with Xit which is common to see, and the other is random effect when Zi is orthogonal with Xit, just a special case.
Basically, there are 4 estimation methods for panel model:
1. Pooled-OLS. It is feasible, effective and consistent only if Zi are several unknown constants, therefore pooled-OLS is always wrong, but it tells us that the choice of estimation methods doesn’t depend upon data structure (panel, cross-section or time series) but the property of error term Eit;
2. LSDV. It is the most common estimator in panel model, which is applied in fixed effect models and is always feasible and consistent, however, with great sacrifice of freedom degrees, so the effectiveness is reduced. Another problem with LSDV is it automatically excludes observable group effects, you cannot add dummies into LSDV model;
3. GLS (or Feasible GLS). It is feasible when good luck and random effect occur. Usually, we use Hauseman test to identify fixed effect or random effect;
4. Hauseman-Taylor IV-GLS. It tries to conquer the disadvantages of fixed and random effects. With complex computation it gives us consistent estimations, but you should make clear the principle of instrument variable method, and which Xit’s are correlated with Zi and which ones not before using HT-IV-GLS.