Arithmetic mean, Harmonic mean, and Jensen equality were applied to marginalize observed standard errors (OSEs) to estimate CAT reliability. Based on different marginalization method, three empirical CAT reliabilities were compared with true reliabilities. Results showed that three empirical CAT reliabilities were underestimated compared to true reliability in short test length (<40), whereas the magnitude of CAT reliabilities was followed by Jensen equality, Harmonic mean, and Arithmetic mean when mean of ability population distribution is zero. Specifically, Jensen equality overestimated true reliability when the number of items is over 40 and mean ability population distribution is zero. However, Jensen equality was recommended for computing reliability estimates because it was closer to true reliability even if small numbers of items was administered regardless of the mean of ability population distribution, and it can be computed easily by using a single test information value at θ = 0. Although CAT is efficient and accurate compared to a fixed-form test, a small fixed number of items is not recommended as a CAT termination criterion for 2PLM, specifically for 3PLM, to maintain high reliability estimates.
Keywords: classical test theory; computerized adaptive testing; item response theory (IRT); measurement; reliability.