CREATE Open Seminar - Improving estimation of student proficiency with joint models for item responses, response times, and interaction counts

Computer-based tests allow for the collection of detailed data, such as a combination of item scores, response times, and interaction counts. We examine the utility of joint latent variable models for such data to estimate student proficiency when varying the number of items, size of latent correlations, and the magnitude of residual dependence. The results indicate that considerable improvement in estimation performance can be realized when the latent variables are highly correlated. Explicitly modeling residual dependence improves estimation performance with small numbers of items but has minor impact otherwise. We discuss possible improvements to the estimation of standard errors for proficiency estimators and outline how joint latent variable modeling can be used in adaptive testing.