Some Properties of Tests for Possibly Unidentified Parameters

It is well known that confidence intervals for weakly identified parameters are unbounded with positive probability (e.g. Dufour, Econometrica 65, pp. 1365-1387 and Staiger and Stock, Econometrica 65, pp. 557-586), and that the asymptotic risk of their estimators is unbounded (Pötscher, Econometrica 70, pp.1035-1065). In this note we extend these "impossibility results" and show that uniformly consistent tests for weakly identified parameters do not exist. We also show that all similar tests of size a < 1/2 concerning possibly unidentified parameters have type II error probability that can be as large as 1 - a.