Hierarchical MPT Modeling - Methodological Comparisons and Application Guidelines

• Applicant: Dr. Julia Groß
• Subject Area: General, Cognitive and Mathematical Psychology
• Term: from 2017 to 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 389614336

Multinomial processing tree (MPT) models are stochastic measurement models for categorical data. They allow investigators to explain participants' behavioral responses in experimental tasks by estimating the contribution of different latent cognitive processes that underlie these responses. MPT models have been developed for a variety of psychological paradigms (e.g., memory, judgment, reasoning, attitude measurement), and have become very popular over the last decades. They are increasingly used to investigate diverse populations, for example within clinical or developmental research. In customary MPT modeling, parameters are estimated for groups of participants or items. This aggregate-MPT approach assumes that there are no substantial differences between participants or items (i.e. parameter homogeneity). However, this assumption is likely violated in most applications, with the potential of dire consequences (e.g. biased parameter estimates and model-fit statistics). Reliability and validity of published aggregate-MPT modeling results therefore needs to be questioned. Yet, the nature, degree, and conditions of these biases in real empirical data has so far not been subject to systematic scientific study. It is therefore one of the two major aims of the work in the proposed network to conduct a large-scale reanalysis of existing MPT data from various paradigms. Individual MPT model fitting, due to reliability issues, is often not an adequate alternative for researchers interested in the investigation of populations that inherently show large within-group variation, and/or in obtaining individual (instead of group-) parameter estimates (e.g. for correlational analyses). As a compromise, hierarchical versions of MPT (h-MPT) models have been recently developed (e.g., Klauer, 2010; Smith & Batchelder, 2010). These incorporate heterogeneity in parameters by specifying a population-level distribution of the parameters across participants and/or items. They are a promising and useful improvement of the aggregate-MPT modeling practice. However, application of h-MPT is not straightforward. The different approaches rely on different statistics and different distributional assumptions; and there are numerous evaluation criteria for the model estimation process. Further, substantial knowledge and programming skills are currently required to fit h-MPT models. These aspects preclude the stringent application of h-MPT models and interpretation of their results in research. The second major aim of the work in the proposed network is therefore to establish agreed-upon h-MPT application guidelines. The network will bring together researchers from h-MPT model development with those seeking to apply h-MPT modeling in their research. With this fruitful mixture of the necessary methodological expertise and breadth of empirical expertise in terms of paradigms and participant populations we aim at facilitating the use of h-MPT modeling in psychological research.

• DFG Programme: Scientific Networks
• Co-Investigator: Professorin Dr. Beatrice G. Kuhlmann