Data and Code for: Estimation and Inference of Structural Models by Stochastic Optimization

<br>Repeated optimizations required to compute parameter estimates and bootstrap standard errors of complex models can be computationally burdensome. In Forneron and Ng (2020), we design a resampled Newton-Raphson algorithm (rnr) that provides consistent estimates and valid standard errors in one run of the optimizer. The key insight is that the algorithm serves as a resampling device to produce a Markov chain of iterates with desirable properties. In this paper, we illustrate that rnr can speed up BLP estimation from almost five hours using standard (n out of n) bootstrap to just over an hour and can be further reduced to fifteen minutes using a resampled quasi-Newton (rqn) algorithm that does not directly compute the Hessian. A Monte-Carlo exercise using Probit IV regressions shows that rnr and rqn provide accurate estimates and coverage. The appeal of the proposed approach goes beyond faster computation. A re-sampling based indirect inference estimator not only produces standard errors easily, but is also more efficient than one obtained by classical optimization. This is illustrated by a dynamic panel model example.<br>