Sensitivity and Counterfactual Analysis of Structural Econometric Models

I propose a framework to characterize the sensitivity of estimates and counterfactuals to the distributional assumptions about latent variables in structural econometric models. I characterize the lower and upper bounds on the counterfactual as the distributional assumption is perturbed infinitesimally and locally along the gradient flow curve of the counterfactual. The framework applies to structural models with general smooth dependence on the distributional assumption, allows for sensitivity perturbations over neighborhoods of a general metric, and promises to be computationally tractable and automatable. I relate infinitesimal sensitivity measures to von Mises calculus and influence functions of semiparametric efficiency theory, and I relate local sensitivity analysis to information geometry. I will illustrate the framework via an application to the Rust87 model of optimal replacement of bus engines and note that it promises to be useful beyond sensitivity analysis in structural models.