Robustness Checks in Structural Analysis

This paper introduces a computationally efficient methodology for estimating variants of structural models. Our approach approximates the relationship between moments and parameters, offering a low-cost alternative to traditional estimation methods. We establish general convergence conditions, primarily requiring model-based moments to be continuous functions of parameters. While this continuity does not necessitate a continuous economic model, it does require the model to have only sparse discontinuities, a concept we define. We also provide convergence rate bounds for Kernel and Neural Net approximations, with the latter demonstrating superior performance in higher dimensions.
We apply this methodology to two standard structural models: (1) dynamic corporate finance and (2) life-cycle portfolio choice. We demonstrate the reliability of our approach through simulations and then use it to explore identification, robustness to sample splits and moment selection, and model misspecification. These explorations are computationally infeasible with standard techniques, but become trivial with our methodology.