Prior Information: The Mixed Prediction Approach

This paper addresses the combination of incomplete prior and sample information. In difference to the mixed estimation approach developed by H. Theil and A.S. Goldberger, dealing with prior knowledge of regression coefficients, we consider prior information on future observations of the dependent variable(s). This prior information could be given in the form of experts' expectations, or from estimations and/or projections of additional models. A framework for the incorporation of this prior knowledge in least squares estimation and prediction is developed. The approach is particularly useful when only aggregated information on the endogenous variables is available, as is often the case with regional level data.