Content area
This dissertation is comprised of two studies that examine the specifications set in Bayesian multidimensional adaptive testing (MAT). The first study examines the benefits of incorporating item response choice as nominal−response data by comparing the performance of the Multidimensional Nominal Response Model (MNRM) against the Multidimensional Three−Parameter Logistic Model (M3PL) in developing item banks and item selection criteria. Specifically, the item selection criteria D−optimality and mutual information are extended for use with nominal−response data and nominal−response item parameters. Results indicate that incorporating item response choice via the MNRM yields slight advantages over the traditional dichotomous scoring approach (operationalized via the M3PL). However, the perceived benefit appears to be moderated by the choice of item selection criterion, with D−optimality performing better with nominal−response data and mutual information performing better with dichotomous−response data.
The second study addresses to what extent the informativity of a prior distribution, which can be used in MAT updating algorithms, affects examinee ability estimates. In this study, a prior distribution is placed on examinee ability; the degree of informativity is manipulated by setting the prior distribution’s variance to either 1,10,1000, or 1,000,000. The effect is tested across MATs that vary in item bank parameterization (MNRM versus M3PL) and item selection criterion (D−optimality versus mutual information). Overall results show a difference when the variance is set to 1 versus 10, 1000, and 1,000,000; negligible differences exist among the larger variance values. Similar to the first study, however, outcomes appear dependent on both item bank and item selection criterion choices.