Optimal predictive sample size for case-control studies