We examined whether probability-based decisions for streptococcal pharyngitis, using probabilities derived from predictive models along with Tompkins' decision rules, could be more cost-effective than the actual decisions of ten physicians. We retrospectively calculated the probability of a positive throat culture ("disease") for each of 310 patients using four different models based on discriminant analysis (1), a branching algorithm (2), and logistic regression (3 and 4). "Projected decisions" were based on these probabilities and Tompkins' rules. We calculated direct medical and indirect costs per correct action taken (diseased patient-treated or nondiseased patient-not-treated). Two models' projected decisions were more cost-effective than the physicians'. Model 1 primarily would have reduced treatment costs (leaving no diseased patient untreated); model 4 primarily would have reduced throat culture costs (with 15% projected undertreatment). While using statistical decision rules may be cost-effective in this setting, their adoption should be consistent with physician and patient priorities.