This paper proposes a procedure for the estimation of discrete Markov decision models and studies its statistical and computational properties. Our method is similar to Rust's Nested Fixed-Point algorithm (NFXP), but the order of the two nested algorithms is swapped. First, we prove that this method produces the maximum likelihood estimator under the same conditions as NFXP. However, our procedure requires significantly fewer policy iterations than NFXP. Second, based on this algorithm, we define a class of sequential consistent estimators, K -stage Policy Iteration (PI) estimators, that encompasses MLE and Holz-Miller, and we obtain a recursive expression for their asymptotic covariance matrices. This presents the researcher with a 'menu' of sequential estimators reflecting a trade-off between efficiency and computational cost. Using actual and simulated data we compare the relative performance of these estimators. In all our experiments, the benefits in efficiency of using a two-stage PI estimator instead of a one-stage estimator (i.e., Hotz-Miller) are very significant. More interestingly, the benefits of MLE relative to two-stage PI are small.
More papers in Computing in Economics and Finance 1999 from Society for Computational Economics Address: CEF99, Boston College, Department of Economics, Chestnut Hill MA 02467 USA Contact information at EDIRC. Series data maintained by Christopher F. Baum ().