Almost unbiased variance estimation in simultaneous equation models

While a good deal of research in simultaneous equation models has been conducted to examine the small sample properties of coefficient estimators there has not been a corresponding interest in the properties of estimators for the associated variances. In this paper we build on Kiviet and Phillips (2000) and explore the biases in variance estimators. This is done for the 2SLS and the MLIML estimators.The approximations to the bias are then used to develop less biased estimators whose properties are examined and compared in a number of simulation experiments. In addition, a bootstrap estimator is included which is found to perform especially well. The experiments also consider coverage probabilities/test sizes and test powers of the t-tests where it is shown that tests based on 2SLS are generally oversized while test sizes based on MLIML are closer to nominal levels. In both cases test statistics based on the corrected variance estimates generally have a higher power than standard procedures.