BIAS AND VARIANCE REDUCTION IN ESTIMATION OF MODEL DIMENSION

The problem of estimating the number of regressors to include in a lin ear regression model is considered. Estimators based on the final pred iction error and Akaike's criterion frequently have large positive bia s. Shrinkage correction factors and bootstrapping are used to produce new estimators with reduced bias. The asymptotic bias and mean-squared errors of these estimators are derived analytically. Finite-sample es timates are obtained by simulation.