A simulation approach to the problem of computing Cox's statistic for testing nonnested models

This paper proposes a new procedure for computing the Cox statistic for tests of nonnested hypotheses using the method of stochastic simulation. The procedure is applicable to a wide class of probability distributions, is relatively simple to implement, and does not require an analytic derivation of the pseudo-true estimators that enter the Cox test statistic. The paper also contains an application of the proposed method to the test of probit versus logistic formulations of the univariate binary choice models. The empirical results show that the simulation method works reasonably well even for a moderate number of replications.