Are policy variables exogenous? The econometric implications of learning while maximizing

This dissertation explores the econometric implications of learning by economic agents. A distinction between active and passive learning is made. On the basis of an argument on the curvature of the value function arising in a dynamic programming approach to the general formulation of the problem, active learning is shown to be the rule rather than the exception. To provide a framework for the analysis, a paradigmatic model is presented in which the government maximizes the discounted sum of tax revenues subject to the constraint imposed by a Laffer curve involving a parameter initially not precisely known but about which Bayesian learning occurs. The paradigmatic model is nontrivial and dynamic by virtue of the presence of learning about the unknown parameter. The government strikes an optimal balance between maximization of current payoff and generation of future information which enhances the efficiency of maximization in subsequent periods. The issue of exogeneity of policy variables is addressed. It is demonstrated that learning affects the exogeneity status of policy variables and has implications analogous to the phenomenon in the focus of the Lucas critique. An additional constraint to augment the exogeneity definitions in Engle, Hendry and Richard (1983) is proposed. A simulation exercise based on the model presented supplies additional insights and quantitative evidence on the structure of the problem. Active learning is proven to be a distinct cause of time inconsistency of optimal plans, the extent of which is quantified for some specific examples. It is argued that learning is not dismissable as a merely transitory source of time inconsistency and loss of parameter invariance. Finally, the data generated are used to perform empirical exogeneity tests in the manner of Granger (1980) and Sims (1972), utilizing results from the survey of Geweke (1984). The results support the claim that the effects of Bayesian learning can be empirically detectable