Minimizing sensitivity to model misspecification

We propose a framework for estimation and inference when the model may be misspecified. rely on local asymptotic approach where degree of misspecification is indexed by sample size. construct estimators whose mean squared error minimax in neighborhood reference model, based one?step adjustments. In addition, we provide confidence intervals that contain true parameter under misspecification. As tool to interpret misspecification, map it power specification test model. Our allows systematic sensitivity analysis interest partially or irregularly identified. illustrations, study three applications: an empirical impact conditional cash transfers Mexico stems from presence stigma effects program, cross?sectional binary choice distribution misspecified, dynamic panel data number time periods small individual