Panel data conditions for consistency and unbiasedness of estimators

[@MelodyZ-0222]

Thank you for the video! Could you please explain why the random effect estimator is biased?

[@dr.swapnilsoni]

an excellent & lucid narration of a difficult topic on panel models. thanks a lot

[@officialmintt]

Simple breakdown of the essence. Thank you!

[@MelodyZ-0222]

Could you please explain why the random effect estimator is biased? Thank you!

[@ahmadghaemi2192]

Are there additional conditions that would make random effects unbiased? And thanks again for all your videos. I've watched the whole undergraduate course pt 1 and 2 and it's been immensely helpful.

[@lastua8562]

I think RE is never unbiased as it is an fGLS estimation. By definition, we need large samples for it to be reliable. Correct me if I am wrong.

[@nosslar]

How can we have random sampling in cross section when panel data have dependent variables?

[@ahmadghaemi2192]

What does one do if there is some unobservable individual *and* time variant effect in uit that is still correlated to xi?

[@lastua8562]

by definition, uit should be uncorrelated with all x. If not, our model is incorrectly specified and we should add omitted variables at the very least. Otherwise I would try with interactions and higher orders after testing for functional misspecification. I cannot think of additional measures, and would love to have input from others too below this comment.

[@abongilevika4358]

That could mean that you should use the FE estimation (within transformation) which works to remove the unobservable individual specific effect or time effect. The whole point of using the FE estimation is so that we can remove this unobserved heterogeneity, however if you want to measure the effect of an explanatory variable that is a dummy on the dependent variable , for example you want to measure the disparity between men and women and the effect this has on earnings then FE would be inappropriate as the FE (within transformation) would remove the gender dummy variable along with the unobserved heterogeneity, this is on of the main criticisms of the FE estimation. Random effects ca be used if for example you assume that the unobserved heterogeneity is uncorrelated with the explanatory variables, i.e Cov( ai, Xit) = 0. However, it might be very difficult to justify such an assumption in real life. Perhaps then you should go for instrumental variable regressions such as IV 2SLS or IV-GMM