Real-Time Generative Adaptive Digit Span Testing
This paper explores the subject of generative adaptive testing using the digit span test as an example. A large-sample study of computer-generated and administered digit-span items on Navy recruits showed an almost perfect correlation (.98-.99) between digit span length and IRT difficulty. Predicted IRT parameters can be used for adaptive testing using items generated in real-time. Our results suggest that the best research strategy for developing generative adaptive tests may be to start with the most elementary cognitive tasks, and then build toward more complete psychometric models of complex mental tasks. The results of this study are sufficiently encouraging so that the same research approach should be tried with other forms of memory span tests and more complex working memory tests, including tests for figures, colors, and words. The paper advances the conjecture that the test information function of a generative CAT system has a mathematical relationship to the model fit and the distribution of the model-specified item parameters, independent of the content domain of the test.