Detecting Model Misfit in Structural Equation Modeling with Machine Learning-A Proof of Concept.

Area of Science:

• Psychometrics
• Statistical Modeling

Background:

• Structural equation modeling (SEM) is widely used in psychological research.
• Evaluating SEM fit using fixed index cutoffs is problematic due to nuisance parameters.
• Researchers often rely on these cutoffs, risking incorrect model acceptance or rejection.

Purpose of the Study:

• To develop a machine learning (ML)-based method for evaluating multi-factorial measurement model fit.
• To create a broadly applicable method that minimizes dependence on nuisance parameters.

Main Methods:

• Trained an ML model using 173 features from 1,323,866 simulated datasets and confirmatory factor analysis models.
• Evaluated ML model performance on 1,659,386 independent test observations.

Main Results:

• The ML model demonstrated high accuracy in detecting model (mis-)fit across various conditions.
• The ML approach outperformed traditional fixed fit index cutoffs.
• Minor misspecifications, like single residual correlations, were challenging for the ML model.

Conclusions:

• Machine learning offers a promising avenue for improving SEM model fit evaluation.
• The developed ML method shows superior performance compared to conventional techniques.
• Further research is needed to address detection of subtle model misspecifications.