Unveiling local dependencies in accuracy and speed: A mixture hierarchical modeling approach.

Abstract

Hierarchical models (HMs) are commonly used to jointly model response accuracies and response times (RTs). However, their relationship cannot be fully captured by the population-level correlation between ability and speed, as local dependencies may arise and threaten valid inferences about individuals and items. In this study, a mixture-based hierarchical model (Mix-HM) that allows respondents to switch between different pacing speeds and identifies positive and negative item-level dependencies is proposed. Two simulation studies were conducted: Simulation 1 examined parameter recovery effects, and Simulation 2 evaluated the effectiveness of Bayesian information-based criteria in model selection processes. The results showed that the Mix-HM achieved satisfactory parameter recovery effects, whereas the conventional HM yielded biased estimates when local dependencies were present. In addition, the model fit criteria were generally able to correctly identify the true model across most conditions. An empirical analysis conducted using large-scale assessment data further showed that the new model provided an improved degree of fit and more stable parameter estimates, while the conventional HM distorted the relationship between ability and speed. These findings highlight the importance of accounting for the heterogeneity of latent speed and local dependencies when incorporating RTs as collateral information.