Recently, the distribution of dental caries has been shown to be skewed, and precise prediction models cannot be obtained using all the data. We applied a balancing technique to obtain more appropriate and robust models, and compared their accuracy with that of the conventional model. The data were obtained from annual oral check-ups for schoolchildren conducted in Japan. Five hundred children were followed from ages 5 to 8, and the three-year follow-up data were used. The variables used were salivary levels of mutans streptococci and lactobacilli, 3-min stimulated saliva volume, salivary pH, fluoride usage, and frequency of consumption of sweet snacks and beverages. Initially, conventional models were constructed by logistic regression analysis, neural network (a kind of prediction method), and decision analysis. Next, the balancing technique was used. To construct new models, we randomly sampled the same number of subjects with and without new dental caries. By repeated sampling, 10 models were constructed for each method. Application of the balancing technique resulted in the most robust model, with 0.73 sensitivity and 0.77 specificity obtained by C 5.0 analysis. For data with a skewed distribution, the balancing method could be one of the important techniques for obtaining a suitable and robust prediction model for dental caries.