A total of 149 students enrolled in an undergraduate nursing research methods course participated in a study comparing three strategies for using formative evaluation (test feedback throughout a course) to predict students at risk of failure at summative evaluation (the final examination). Students took 12 weekly multiple-choice quizzes, which were graded and returned for self-study, and a final 60-item multiple-choice exam. Three 4-week quiz subtotals were the discriminating variables used to predict membership in three final-exam score categories: Group 1 (poor); Group 2 (fair); Group 3 (good). Separate discriminant analyses tested three patterns of assigning prior probabilities of group membership: (1) equal (each .333); (2) proportional to actual numbers of students in each group; (3) weighted by setting cost of misclassifying poor students as three times more serious than cost of misclassifying fair or good students. A significant discriminant function emerged, and confirming previous results, effect size (a standardized measure of the discrepancy between performance and the overall mean) for poor students decreased over time, showing that they were "closing the gap." Assigning probabilities proportional to cases gave best overall classification accuracy (53.02%), but Bayesian weighted adjustment best predicted students at risk of failure (82.1% correctly classified) while sacrificing some overall predictive power (42.95% correct). (LPG)