A comprehensive set of dental variables was investigated to find the "best" combination of predictors for high caries increment in 7/8-year-old and 10/11-year-old children. Four populations with widely different caries prevalence were studied. Logistic regression analysis supplied multiple-input models by stepwise selection of predictors. A "low number of sound primary molars" was the best and most consistent predictor of high caries increment. The second best predictors were "high numbers of pre-cavity lesions on permanent first molars" (discolored pits and fissures in the younger age group and white spots on the smooth parts of buccolingual surfaces in the older age group). Inclusion of radiological variables did not substantially increase the quality of prediction. For practical application, models with various multiple inputs selected by stepwise procedures were compared with "fixed" three-input models. These three-input models resulted in predictive quality nearly equal to those of the multiple models. Traditional one-input models, containing DMFT or dmft, were inferior to the three-input models, particularly in the older age class. The lower the caries prevalence of the source data, the better was the prediction. As a summary measure characterizing the predictive performance of a model, we used the index "area under the receiver operating characteristic curve" A. For the 1984 data and the three-input models, the area was approximately 80%, and for the 1972 data, the area was 65-70%.