Information was collected on 95 people with mental retardation who had been identified seven years previously as showing severe self-injurious behavior. At follow up 71% of participants were still showing self-injurious behavior of a severity which presented a management problem for care staff. The occurrence of specific topographies of self-injury was extremely stable among the group showing persistent self-injury. Finally, self-injury status at follow-up was predicted with 76% accuracy by a logistic regression model containing three variables: site of injury (higher persistence being shown by people exhibiting head directed self-injury); reported (greater) stability of self-injury when first identified; and (younger) age.