psychotree: Recursive Partitioning Based on Psychometric Models

Achim Zeileis, Carolin Strobl, Florian Wickelmaier, Basil Komboz, Julia Kopf, Lennart Schneider, David Dreifuss, Rudolf DebelakOpenalex Unknown 2009 · DOI · OpenAlex

Abstract

Recursive partitioning based on psychometric models, employing the general MOB algorithm (from package partykit) to obtain Bradley-Terry trees, Rasch trees, rating scale and partial credit trees, and MPT trees, trees for 1PL, 2PL, 3PL and 4PL models and generalized partial credit models.

Introduction

In psychometrics, models such as the Rasch model or the Bradley-Terry model are fundamental for measuring latent traits and preferences. A core assumption of these models is parameter invariance, which posits that model parameters—such as item difficulty or object "worth"—remain constant across the entire population. However, in practice, this assumption is often violated by Differential Item Functioning (DIF), where different subgroups (e.g., categorized by age, gender, or education) interact with items differently.

Traditionally, identifying DIF required researchers to manually test pre-defined groups. The psychotree package for R automates this process by integrating psychometric models with recursive partitioning, allowing for the discovery of population heterogeneity without prior specification of subgroups.

Methodology: Model-Based Recursive Partitioning (MOB)

The psychotree package leverages the Model-Based Recursive Partitioning (MOB) algorithm, provided by the partykit package. This approach transforms the search for population heterogeneity into an algorithmic process:

  1. Global Model Fitting: A psychometric model is initially fitted to the full dataset.
  2. Instability Testing: The algorithm evaluates whether model parameters are unstable across a set of available partitioning variables (covariates).
  3. Recursive Splitting: If significant instability is detected, the data is split at the optimal point of the most influential covariate.
  4. Iteration: This process repeats recursively until no further significant differences are found or a minimum node size is reached.

Unlike standard decision trees that predict a simple mean or class, each terminal node in a psychotree contains a complete psychometric model tailored to that specific subgroup.

Implementation and Supported Models

The package provides a unified interface for several classes of psychometric trees:

  • Bradley-Terry Trees (bttree): Used for paired comparison and preference data.
  • Rasch Trees (raschtree): Designed for basic Item Response Theory (IRT).
  • IRT Trees (npltree): Supports 1PL, 2PL, 3PL, and 4PL models.
  • Partial Credit & Rating Scale Trees (pctree, rstree): For polytomous (Likert-scale) data.
  • MPT Trees (mptree): For Multinomial Processing Tree models in cognitive psychology.

The flexibility of these functions is illustrated by their argument structures, which allow for complex model specifications and handling of data nuances.

Core arguments for the `bttree` function, illustrating the integration of formulas and partitioning variables.

Empirical Applications and Insights

The utility of psychotree is demonstrated through several benchmark datasets:

  • General Knowledge (SPISA): A raschtree analysis of general knowledge quiz data revealed significant variations in item difficulty based on students' fields of study and age.
  • Preference Modeling (Topmodel2007): Using bttree, researchers identified distinct preference segments for TV show contestants based on the judges' gender and age, providing a more nuanced "worth" parameter than a global ranking.
  • DIF Detection (DIFSim): In simulated datasets, the algorithm successfully recovered the exact split points and covariates used to generate bias, validating its efficacy as a tool for DIF detection.

A key advantage of this tree-based approach is its ability to naturally capture interactions. For instance, it can identify cases where an item is biased only for a specific intersection of demographics (e.g., young, low-education males), which would be difficult to specify a priori in standard regression models.

Conclusion

The psychotree package offers a powerful, automated framework for exploring heterogeneity in psychometric data. By moving beyond global averages to subgroup-specific models, it provides researchers with deeper insights into how latent traits and preferences vary across populations. While computationally intensive for large-scale IRT models, its ability to visualize latent shifts and handle complex interactions makes it an essential tool for modern psychometric analysis.

Summary updated

Abstract

Recursive partitioning based on psychometric models, employing the general MOB algorithm (from package partykit) to obtain Bradley-Terry trees, Rasch trees, rating scale and partial credit trees, and MPT trees, trees for 1PL, 2PL, 3PL and 4PL models and generalized partial credit models.
BibTeX
@inproceedings{zeileis2009psychotree,
  author = {Achim Zeileis and Carolin Strobl and Florian Wickelmaier and Basil Komboz and Julia Kopf and Lennart Schneider and David Dreifuss and Rudolf Debelak},
  title = {{psychotree: Recursive Partitioning Based on Psychometric Models}},
  booktitle = {Openalex Unknown},
  year = {2009},
  doi = {10.32614/cran.package.psychotree},
}
Figure