Choice between Disaggregate and Aggregate Specifications Estimated by Instrumental Variables Methods.

A choice criterion is proposed for discriminating between disaggregate and aggregate models estimated by the instrumental variables method. The criterion, based on prediction errors, represents a generalization of criteria developed in the context of classical regressions models. The paper also derives general tests for aggregation bias in the instrumental variables context. The criterion and tests are applied in an analysis of U.K. employment demand. It is shown that a model disaggregated by forty industries predicts aggregate employment better than an aggregate model, and that significant biases exist in estimates of the long-run wage and output elasticities obtained from the aggregate model.