The effect on criterion-related validity of nonfitting response vectors (NRVs) on a predictor test was investigated. Using simulated data, it was shown that there was a substantial decrease in validity when the type of misfit was severe (i.e., guessing the correct answers to all test items), when the correlation between the predictor and the criterion scores was p = .3 or p = .4, and when the percent of NRVs was 15% or higher. The effect on test validity of removing NRVs from a dataset using the person-fit statistic Iz was also investigated. Only a small increase in validity (at most A = .03) was found. However, it was possible using lz to distinguish groups of examinees with low correlations with the criterion test. Regression analysis showed that the regression lines for fitting and nonfitting simulees did not cross and that only when the correlation between predictor scores and criterion scores was p = .3 or p = .4 and the number of simulees was large (30% NRVs) did lz scores improve prediction.