A1914
Title: Inference for constrained extremum estimators
Authors: Jessie Li - University of California, Santa Cruz (United States) [presenting]
Abstract: Inference for constrained M-estimators and constrained GMM estimators with possibly nonsmooth or nonconvex objectives and possibly nonconvex constraint sets is studied. Test-inversion is used to construct a uniformly valid simultaneous confidence set which covers parameters either in the interior or on the boundary of the constraint set as well as parameters drifting towards the boundary at arbitrary rates. The method works for both $\sqrt{n}$-consistent estimators and estimators with a slower-than-$\sqrt{n}$ rate of convergence to non-standard asymptotic distributions. Monte Carlo simulations illustrate the uniformly correct coverage of the method in a boundary constrained maximum likelihood model and a nonsmooth GMM model. An empirical application conducts uniformly valid inference in a nested logit model of cereal demand which imposes the constraint that the correlation parameter among choices within the same nest must lie between 0 and 1.
