MULTIPLE ROOTS OF THE TOBIT LOG-LIKELIHOOD

We show that the occurence of multiple roots of the Tobit log-likelihood function in the untransformed parameter space is not nearly so likely as heretofore accepted. The conditions which must be met for them to occur require data configurations which could not be met in random sampling. Moreover, to find a 'second root' in practice would require computer programs far more accurate than those which currently exist. Similar results are obtained for a number of other limited dependent variable models. © 1990.