A replicable process for improving instruction through the consistent use of student data collected before, during, and after instruction is proposed. A rational analysis of different types of error rates (theoretical, base, posttest, instructional) and discrimination indices (base, posttest) leads to a set of rules for identifying test items and sections of instruction that require revision. The application of these rules is illustrated through the analysis of student responses to a subset of test items in a computer assisted instruction program in micro-economics. Finally, a discussion is presented that relates the proposed decision process to some theoretical issues in criterion-referenced testing and the formative evaluation of instruction. (Author/LR)