the parameters in _{g } Varv_{g}, where v_{g }is the T 1 error vector for each g, is difficult when group effects have been removed. Estimates based on the FE residuals, v̂_{gt}, disappear as T → , but can be substantial. In AR(1) case, ̂ comes from

v̂_{gt }on v̂_{g,t−1}, t 2, . . . , T, g 1, . . . , G.

∙ One way to account for bias in ̂: use fully robust inference. But, as Hansen (2007b) shows, this can be very inefficient relative to his suggestion to bias-adjust the estimator ̂ and then use the bias-adjusted estimator in feasible GLS. (Hansen covers the general ARp model.) ∙ Hansen shows that an iterative bias-adjusted procedure has the same asymptotic distribution as ̂ in the case ̂ should work well: G and T both tending to infinity. Most importantly for the

14

(9)