Computerized Adaptive Test Generalizability Theory

In a CAT, different examinees receive different subsets of items chosen by an algorithm based on their ability estimates. This adaptive feature makes standard reliability coefficients such as Cronbach's alpha inappropriate, since they assume all examinees respond to the same items. G-theory addresses this by treating the adaptive item selection as one facet of measurement and partitioning score variance into contributions from person ability, item difficulty, and their interaction. A G-coefficient then summarises how consistently examinees would be ranked if the measurement were replicated under similar conditions.