The construction sample consisted of 3,761 male borstal trainees, 23 percent of whom absconded; the validation sample consisted of 2,387 male borstal trainees, 20 percent of whom absconded. A total of 80 variables covering previous criminal and institutional history, family background, occupational status, ethnic origins, and indicators of psychiatric and physical disorders were cross-tabulated against absconding. Of these, 31 proved to be significant predictors. Six of the most significant variables were found to be especially successful in predicting absconding: previous absconding from a custodial institution, previous absconding from a noncustodial institution, motor vehicle theft, race, age at sentencing, and employment status at arrest. A log linear model was used to generate expected absconding rates for each combination of predictor variables, and a simple weighting scheme was devised that corresponded to expected absconding rates. The prediction scale was operationalized as an algorithm and applied to the validation sample. Logistic analysis was used to test the significance of relationships between risk categories and absconding, and regression models were fitted with the computer program. The linear trend was significant; departures from linearity were nonsignificant. Thus, the risk categories appear to be a straightforward linear index of proneness to abscond. Shrinkage was found to be neglible. Data indicate that on the basis of simple background information, it is possible to assess proneness to abscond. Tabular data and 10 references are provided.