A0614
Title: Almost unbiased variance estimation in IV regression
Authors: Yongdeng Xu - Cardiff University (United Kingdom) [presenting]
Abstract: Using Nagar's approximation, bias approximations are derived for the asymptotic variance estimators of 2SLS and Fuller estimators in instrumental variable regression, up to the order of $T^{-2}$. These bias approximations are then employed to develop bias-corrected variance estimators. Findings indicate that the asymptotic variance estimator for 2SLS exhibits an upward bias, while the Fuller estimator shows an upward bias in cases of just identification or low overidentification. Simulation results reveal that the bias-corrected variance estimator substantially reduces the bias of the asymptotic variance estimator, demonstrating minimal bias or near-unbiasedness. Importantly, the properties of the $t$-test improve significantly with bias-corrected variance estimates, consistently demonstrating better size properties and higher statistical power compared to the asymptotic variance. The practical relevance of these findings is illustrated through several well-known applications.
