One estimation scheme, maximum likelihood, is known to be adversely affected by noisy data, so a second method, robust in the face of deviations from assumptions of the model, was also used. In a 1970 two-year follow-up data set from the United States Parole Commission, the robust latent trait model is shown to predict parole performance more accurately than is possible with the other estimation scheme. The differences lie in the tail of the probity distribution, where the robust estimates prove to be clearly superior. It is proposed that these techniques also be tested in other parts of the criminal justice system. (Author abstract)