We present and evaluate a method for predicting individual treatment benefits based on random effects logistic regression models of binary outcomes that change over time. The method uses empirical Bayes predictors based on patients' characteristics and responses to treatment. It is applicable to both 1-dimentional and 2-dimentional personalized medicine models. Comparisons between predicted and true benefits for simulated new patients using correlations, relative biases and mean-squared errors were used to evaluate prediction performance. The predicted benefits had relatively small biases and relatively high correlations with the true benefits in the simulated new patients. The predictors also captured estimated overall population trends in the evolution of individual benefits. The proposed approach can be used to retrospectively evaluate patients' responses in a clinical trial, or to retrospectively or prospectively predict individual benefits of different treatments for new patients using patients' characteristics and previous responses. The method is used to examine changes in the disorganized dimension of antipsychotic-naïve patients from an antipsychotic randomized clinical trial. Retrospective prediction of individual benefits revealed that more cannabis users had slower and lower responses to antipsychotic treatment as compared to non-cannabis users, revealing cannabis use as a negative prognostic factor for psychotic disorders in the disorganized dimension.
Keywords: Individual benefits; cannabis; empirical Bayesian prediction; longitudinal binary outcomes; psychosis.