Generalizability of Large-Scale Performance Assessments in Science: Promises and Problems

This study provides empirical evidence about the sampling variability and generalizability (reliability) of a statewide science performance assessment. Results at both individual and school levels indicate that task-sampling variability was the major source of measurement error in the performance assessment; rater-sampling variability was negligible. Adding more tasks improves the generalizability of the measurement. For the school-level assessment, the variation of performance among students within a school was larger than the variation among schools. Increasing the number of students taking a test within a school thus increases the generalizability of the assessment. Finally, the allocation of students in a matrix-sampling design is compared to a studentscrossed-with-tasks design. The former would require fewer tasks per student than the latter to build a generalizable measure of school performance.