Generalized Inverses and Asymptotic Properties of Wald Tests

Abstract

We consider Wald tests based on consistent estimators of g-inverses of the asymptotic covariance matrix ∑ of a statistic that is n^1/2-asymptotically normal distributed under the null hypothesis. Under the null hypothesis and under any sequence of local alternatives in the column space of ∑, these tests are asymptotically equivalent for any choice of g-inverses. For sequences of local alternatives not in the column space of ∑ and for a suitable choice of g- inverse, the asymptotic power of the corresponding Wald test can be made equal to zero or arbitrarily large. In particular, the test based on a consistent estimator of the Moore-Penrose inverse of ∑ has zero asymptotic power against sequences of local alternatives in the orthogonal complement to the column space of ∑.