Computerized Adaptive Testing With the Bifactor Model

An algorithm designed to implement CAT with the dichotomous bifactor model is described. Performance of the algorithm was evaluated with several datasets from a 615-item personality instrument that scored on a general scale and four content scales. Post-hoc simulation, including cross-validation, and live testing bifactor CAT data were analyzed in terms of reductions in test length for each scale, correlations with trait estimates from all items in each scale, bias, and accuracy. Results showed very substantial reductions in scale and overall test length while maintaining correlations with full-scale scores above .90. For the general scale, mean test length reductions of about 95% were observed in both post-hoc simulation and live testing; only about 25 to 30 items were required, on average, to recover scale scores with a correlation above .90. Mean reductions of 68% to 90% were observed for the content scales. Across all scales combined, the bifactor CAT algorithm reduced test length by an average of about 80% and resulted in an actual testing time mean decrease of approximately 93 minutes (82%).