Ryan Godwin and I have a new paper - "Improved Analytic Bias Correction for Maximum Likelihood Estimators". You can download it from here. (This is a revised version, 19 July 2017.)
This paper proposes a modification of the Cox-Snell/Cordeiro-Klein bias correction technique that we've used in our earlier research (including work with Helen Feng and Jacob Schwartz). For some more information about that work, see this earlier post.
This issue is also extensively discussed in the panel data literature, where the standard fixed effects estimator is biased/inconsistent for a fixed value of time-series observations. The estimator that you are suggesting sometimes can be also seen as the estimator that solves bias-corrected (up to O(n^{-1})) First order conditions of the ML estimator. One interesting example is a bias-corrected estimator for stationary AR(1) model with a constant term. In that case ``standard'' and ``improved'' bias-corrected estimators have closed form solutions.
ReplyDelete