This paper discusses methods for reducing the bias of consistent estim
ators that are biased in finite samples. These methods are available w
henever the bias function, which relates the bias of the parameter est
imates to the values of the parameters, can be estimated by computer s
imulation or by some other method. If so, bias can be reduced by one f
ull order in the sample size and, in some cases that may not be unreal
istic, virtually eliminated. Unfortunately, reducing bias may increase
the variance, or even the mean squared error, of an estimator. Whethe
r it does so depends on the shape of the bias function. The results of
the paper are illustrated by applying them to two problems: estimatin
g the autoregressive parameter in an AR(1) model with a constant term,
and estimating a legit model. (C) 1998 Elsevier Science S.A. All righ
ts reserved.