An Information Theoretic Approach to Estimation in the Case of Multicollinearity

We propose a data-based extremum formulation that extends theempirical-likelihood and information-theoretic methods of estimation andinference. It is demonstrated how this method may be used in a general linearmodel context to mitigate the problem of an ill-conditioned design matrix. Adual loss criterion function, which can be biased in finite samples, producesan estimator that is consistent and asymptotically normal. Limiting chi-squaredistributions are obtained that may be used for hypothesis testing andconfidence intervals. Empirical-risk sampling experiments suggest theestimator has excellent finite-sample properties under a squared error lossmeasure. Copyright Kluwer Academic Publishers 2003