Divide and Conquer: A New Approach to Dynamic Discrete Choice with Serial Correlation

In this paper, we develop a method to efficiently estimate dynamic discrete choice models with AR(n) type serial correlation of the errors. First, to approximate the expected value function of the underlying dynamic problem, we use Gaussian quadrature, interpolation over an adaptively refined grid, and solve a potentially large non-linear system of equations. Second, to evaluate the likelihood function, we decompose the integral over the unobserved state variables in the likelihood function into a series of lower dimensional integrals, and successively approximate them using Gaussian quadrature rules. Finally, we solve the maximum likelihood problem using a nested fixed point algorithm. We then apply this method to obtain point estimates of the parameters of the bus engine replacement model of Rust [Econometrica, 55 (5): 999–1033, (1987)]: First, we verify the algorithm’s ability to recover the parameters of an artificial data set, and second, we estimate the model using the original data, finding significant serial correlation for some subsamples.