Third-order efficiency of conditional tests in exponential models: The lattice case

As is well known, in full rank multivariate exponential families, tests of Neyman structure are uniformly most powerful unbiased for one-sided problems. For the case of lattice distributions, the power of these tests--evaluated at contiguous alternatives--is approximated by asymptotic expansions up to errors of order o(n-1). Surprisingly the tests with Neyman structure are not third-order efficient in the class of all asymptotically similar tests unless the problem is univariate.