Attrition in an experimental osteoarthritis intervention was studied among 364 (130 male) volunteers (ages 60-87). Subjects were randomly assigned to control, social support, education, or combined treatment groups. A series of discriminant function analyses showed that the final intervention groups were more homogeneous than the original samples. The most robust predictor of attrition was having either high or very low depression scores. Social support variables were predictors of poor attendance in the social support group, indicating possible attrition bias in evaluating this treatment. The results indicate that psychosocial variables may be the best predictors of attrition in health intervention research among older subjects, and that variables related to attrition can be related to the content of the intervention.