FINITE SAMPLE PROPERTIES OF NONSTATIONARY BINARY RESPONSE MODELS: A MONTE CARLO AND RESPONSE SURFACE ANALYSIS

This paper investigates the finite sample distributions of maximum likelihood estimators for nonstationary probit models. We find that, analogous to standard OLS models, commonly used tests statistics almost always reject the null hypothesis of no relationship between xt and a latent yt, even when they are, in fact, generated by independent random walks. However, if cointegrating relationships are present in the model, parameter distributions are better behaved and standard z and Wald test statistics are consistent.