The Enduring Value of Theory: Comparing HLM and MLLR in Multilevel Behavioral Research



Author Information

Jordan Epistola, University of Maryland, United States
Paul Hanges, University of Maryland, United States
Tiffany Hansbrough, Binghamton University, United States

Abstract

The integration of machine learning (ML) into behavioral research offers promise for advancing theory by capturing complex patterns traditional methods often overlook. However, critiques (Vowels, 2023) caution against the use of ML for theory-building, as such methods can obscure causality and misrepresent theoretical relationships. This study addressed these concerns by comparing Hierarchical Linear Modeling (HLM), a theory-driven approach, and Multilevel Logistic Lasso Regularization (MLLR), a regularization-based ML technique, in predicting episodic ("Remember") versus semantic ("Know") memory judgments in leadership assessments.

Using a multilevel dataset comprising 11,833 observations nested within 614 participants, the study evaluated predictive accuracy, coefficient stability, and the broader implications of these methods for theory-testing in organizational research. HLM consistently outperformed MLLR across key metrics, including AUC, sensitivity, specificity, and balanced accuracy, affirming its utility for theory-driven prediction. Both approaches identified key predictors such as item sentiment and leader liking with consistent directional effects, but MLLR penalized theoretically significant predictors aggressively. For example, item concreteness, the strongest predictor per HLM, and reading level, which was small but significant, were attenuated under MLLR.

These findings highlight the limitations of ML methods in multilevel research, marking one of the first comparisons of MLLR and HLM in organizational science. While MLLR shows potential for exploratory analyses and abductive research, its limitations highlight the need for advancements in computational psychometrics to better address multilevel data. Conversely, HLM’s superior performance and interpretability reinforce the value of theory-driven approaches for both theory-testing and prediction, reinforcing critiques of overreliance on ML in social science research.


Paper Information

Conference: ACP2025
Stream: Industrial Organization and Organization Theory

The full paper is not available for this title


Virtual Presentation


Comments & Feedback

Place a comment using your LinkedIn profile

Comments

Share on activity feed

Powered by WP LinkPress

Share this Research

Posted by James Alexander Gordon