A General Method for Third-Order Bias and Variance Corrections on a Nonlinear Estimator

Motivated by a recent study of Bao and Ullah (2007a) on finite sample properties of MLE in the pure SAR (spatial autoregressive) model, a general method for third-order bias and variance corrections on a nonlinear estimator is proposed based on stochastic expansion and bootstrap. Working with concentrated estimating equation simplifies greatly the high-order expansions for bias and variance; a simple bootstrap procedure overcomes a major difficulty in analytically evaluating expectations of various quantities in the expansions. The method is then studied in detail using a more general SAR model, with its effectiveness in correcting bias and improving inference fully demonstrated by extensive Monte Carlo experiments. Compared with the analytical approach, the proposed approach is much simpler and has a much wider applicability. The validity of the bootstrap procedure is formally established. The proposed method is then extended to the case of more than one nonlinear estimator.