Weighted-Average Least Squares (WALS): Confidence and Prediction Intervals

Abstract

We consider inference for linear regression models estimated by weighted-average least squares (WALS), a frequentist model averaging approach with a Bayesian flavor. We propose a new simulation method that yields re-centered confidence and prediction intervals by exploiting the bias-corrected posterior mean as a frequentist estimator of a normal location parameter. We investigate the performance of WALS and several alternative estimators in an extensive set of Monte Carlo experiments that allow for increasing complexity of the model space and heteroskedastic, skewed, and thick-tailed regression errors. In addition to WALS, we include unrestricted and fully restricted least squares, two post-selection estimators based on classical information criteria, a penalization estimator, and Mallows and jackknife model averaging estimators. We show that, compared to the other approaches, WALS performs well in terms of the mean squared error of point estimates, and also in terms of coverage errors and lengths of confidence and prediction intervals.