Computer-assisted improvement of mean squared error in statistical estimation

A computer-assisted method for improving the mean squared error (MSE) in estimation for parametric models is presented. Assuming existence of nontrivial sufficient statistics, the method involves generation of Monte Carlo samples from the conditional distribution of the observables, given the sufficient statistic(s). The method is illustrated in connection with a simple back-propagation neural network model for estimating a logistic regression function, and a specific numerical example related to logistic regression is presented.