Five methods of selecting and combining variables were compared -- Burgess method, Glueck method, multiple regression, predictive attribute analysis, and logistic regression. The 411 mostly white boys of working class background were divided into 2 groups using a table of random numbers, producing a construction sample of 205 and a validation sample of 206. Twenty-five predictor variables were included in the analysis in addition to the juvenile official and self-reported delinquency predictors. It was difficult to identify a group with much more than a 50-percent chance of juvenile delinquency. Conversely, this meant that it was difficult to identify more than 50 percent of the juvenile delinquents. It was easier to predict official convictions and adult offending than juvenile delinquency. The more sophisticated multiple regression, predictive attribute analysis, and logistic regression techniques were worse than the simpler Burgess and Glueck methods, but in most instances the latter two methods were not markedly more efficient than the best single predictor. It seems more realistic and feasible to predict not delinquency in general but the most persistent or 'chronic' offenders, who account for a significant proportion of all crime. Tabular data and 34 references are included.