A0794
Title: The bias of the modified limited information maximum likelihood estimator in static simultaneous equation models
Authors: Gareth Liu-Evans - University of Liverpool (United Kingdom) [presenting]
Garry Phillips - University of Exeter (United Kingdom)
Abstract: The Modified LIML (MLIML) estimator has received a resurgence of interest recently in the context of weak instruments and many instruments. Like the original LIML estimator MLIML is consistent, and in the case of many instruments, it has been found asymptotically optimal. The MLIML estimator has all necessary moments and is unbiased to order $O(1/T)$, making it an important alternative to the 2SLS estimator. We find the bias of the MLIML estimator to order $O(1/T^2)$, and similarly, find the LIML estimator pseudo-bias to this order. The MLIML (and LIML) bias can be substantial, and different ways of correcting this are considered in Monte Carlo experiments. As an application of bias-corrected MLIML estimation, we re-estimate the effect of shifting the relative supply of young college workers on the US college graduate wage premium
