Unamortized black-box variational inference for joint response and response time models

An important aspect of psychometric model application is the choice of a suitable algorithm for parameter estimation. Particularly in the case of large datasets and high dimensional models, conventional algorithms may be inapplicable or relatively slow. As a remedy, variational inference was suggested in machine learning. While amortized variational inference algorithms are already established for several psychometric models, there has been little research on unamortized variational inference algorithms. In this article, we investigate the accuracy of this estimation method for a multidimensional variation of the joint model for responses and response times proposed by Van der Linden (Psychometrika 72: 287-308, 2007). Our results indicate that the proposed algorithm can be used for accurate item parameter estimation of these models even in relatively small datasets. In addition it is computationally fast in large datasets. We applied the method to a multidimensional empirical dataset in clinical psychology.