EUVIMEDEUVIMED entdecken
Crossref

When DIF Goes Unmodeled: Assessing the Viability of Random Forest for Diagnostic Classification

Catherine Bain, Patrick D. Manapat, Danielle Manapat, Keelyn Brenna, Kevin Grimm

Journal of Behavioral Data Science · 2026

DOI: 10.35566/jbds/bainmmbg ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Behavioral Data Science, 2026. Verfasst von Catherine Bain und 4 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Psychological assessments are often used to make classification decisions (i.e., identify whether an individual meets a set of criteria for a psychological diagnosis). A psychometric approach, like item response theory (IRT), first estimates a latent trait score from item responses and then compares that score to a cut-point to determine classes. In contrast, a machine learning approach, such as random forest (RF) predicts class membership directly from item responses. While both methods show promise for diagnostic classification, their relative robustness to differential item functioning (DIF) is unclear. This study compared the classification performance of an IRT- and a RF-based approach to classification under conditions varying DIF presence and severity, along with other sample and scale characteristics via Monte Carlo simulation. Single-group IRT served as a baseline representing standard psychometric practice, as it is widely used for diagnostic classification but assumes item invariance across groups. Results indicated that when DIF was absent or minimal, both approaches yielded comparable classification metrics. However, as DIF severity increased, IRT-based classification performance declined, whereas RF maintained robust performance across conditions. These findings suggest that RF may maintain more stable classification performance than IRT-based classification when DIF is present but not explicitly accounted for in the model, making RF a viable alternative for diagnostic classification when DIF is suspected but its source or structure is unknown, unmeasured, or complex. Strengths and limitations of each approach are discussed, with particular attention to the trade-off between interpretability and classification robustness in applied assessment contexts.

Zitieren

APA

Bain, C., Manapat, P. D., Manapat, D., Brenna, K., & Grimm, K. (2026). When DIF Goes Unmodeled: Assessing the Viability of Random Forest for Diagnostic Classification. Journal of Behavioral Data Science. https://doi.org/10.35566/jbds/bainmmbg

Vancouver

Bain C, Manapat PD, Manapat D, Brenna K, Grimm K. When DIF Goes Unmodeled: Assessing the Viability of Random Forest for Diagnostic Classification. Journal of Behavioral Data Science. 2026. doi:10.35566/jbds/bainmmbg

BibTeX
@article{Bain2026,
  author = {Bain, Catherine and Manapat, Patrick D. and Manapat, Danielle and Brenna, Keelyn and Grimm, Kevin},
  title = {When DIF Goes Unmodeled: Assessing the Viability of Random Forest for Diagnostic Classification},
  journal = {Journal of Behavioral Data Science},
  year = {2026},
  doi = {10.35566/jbds/bainmmbg}
}
Kostenlos starten →