Parameter estimation approaches for multinomial processing tree models: A comparison for models of memory and judgment

Abstract Multinomial processing tree (MPT) models are commonly used in cognitive psychology to disentangle and measure the psychological processes underlying behavior. Various estimation approaches have been developed to estimate the parameters of MPT models for a group of participants. These approaches are implemented in various programs (e.g., MPTinR, TreeBUGS) and differ with regard to how data are pooled across participants (no pooling, complete pooling, or partial pooling). The partial-pooling approaches differ with regard to whether correlations between individual-level parameters are explicitly modeled (latent-trait MPT) or not (beta-MPT). However, it is currently unclear whether the theoretical advantages of the partial-pooling approaches actually yield the best results in standard practice (i.e., with typical parameter values and amounts of data). We conducted parameter recovery analyses comparing the accuracy and precision of four estimation approaches for two MPT models: the source-monitoring model and the hindsight-bias model. For essential (“core”) parameters of the two models, the partial-pooling approaches yielded the best results overall. Importantly, there were also model-specific differences between the approaches. For the source-monitoring model, the latent-trait approach achieved the best results. For more complex hindsight-bias model, the latent-trait approach appeared to be overparameterized for typical amounts of data; here, the beta-MPT approach was better. We derive recommendations for applications of the two MPT models.