Multidimensional CAT item selection methods for domain scores and composite scores: theory and applications
Conventional unidimensional CAT only considers one construct (simple structure). It was found that multiple constructs, meaning multidimensional CAT in the same testing can highly improve person estimates with less test length comparing with conventional one. For multidimensional CAT, items may involve in more than one construct (complex structure), in this study authors, however, only took simple structure into account. This paper compared five item selection methods to investigate domain scores and total scores. Among these methods, V2 was proposed to weight the domain scores to form total scores by optimizing the weighting. The method is not equal weight for domain scores but depend on different score points. Other selection methods including V1 of equal weight for domain scores and maximum (or minimum) the information (or the volume and error variance). A series of simulation were conducted for 3 contents, 5 methods, 3 test lengths, 4 populations and 3 item pools of manipulated variable, and ABSBIAS, correlation, time, test reliability, and item usage of dependent variables. Results showed that the proposed V2 method performed better for both domain scores and overall scores and large item usage when content is 1 or 2 but performed worse when content is 0.