The estimation of censored probit models can result in an estimated correlation between the disturbances approaching [plus]1.0 or -1.0 when most of the observations are selected into the sample and the outcomes are unequally distributed. Outcomes of 0 can induce an estimated correlation of -1.0, and outcomes of 1 can induce an estimated correlation of [plus]1.0. This paper analyzes the population problem, derives corresponding sample conditions, proposes a solution to the problem, and offers a computer program. Copyright 1996 by MIT Press.