A concrete-semiconcrete-abstract (CSA) instructional approach derived from discovery learning (DIS) was embedded in a direct instruction (DI) methodology to teach eight elementary students with math disabilities. One-minute abstract-level probes were the primary metric used to assess student performance on subtraction problems (minuends 0-9). A single-subject, multiple-baseline across-participants design was used to identify the differential effects of discontinuing instruction at crossover (i.e., the point when the correct response rate exceeded the incorrect response rate). Results indicate that the commonly accepted practice of teaching an entire CSA unit of lessons may not be the most efficacious approach for classroom teachers. Instead, through daily data collection and application of the "crossover decision rule" (discontinue rule), teachers can selectively target those students appropriate for additional concrete- and/or semiconcrete-level instruction and those students for whom continued practice at the abstract level is more appropriate. Implications for teaching computational skills are examined. (Contains 3 tables and 3 figures.)