Latent Variable Forests for Latent Variable Score Estimation.

Area of Science:

• Psychometrics
• Machine Learning
• Statistical Modeling

Background:

• Latent variable models are crucial for understanding complex constructs.
• Confirmatory Factor Analysis (CFA) is widely used but can be sensitive to parameter heterogeneity.
• Existing methods may produce biased scores in the presence of subgroup differences.

Purpose of the Study:

• To develop a novel algorithm, Latent Variable Forest (LV Forest), for unbiased latent variable score estimation.
• To address parameter heterogeneity in CFA models with mixed response variables.
• To improve the interpretability and accuracy of latent variable scores.

Main Methods:

• LV Forest combines parametric CFA with nonparametric tree-based machine learning.
• It employs parametric model restrictions and a tree ensemble approach.
• Handles ordinal and/or numerical response variables.

Main Results:

• LV Forest provides unbiased latent variable score estimates.
• The algorithm accounts for parameter heterogeneity across population subgroups.
• Demonstrated improved score estimation accuracy on simulated and real survey data.

Conclusions:

• LV Forest offers a robust method for latent variable score estimation in the presence of heterogeneity.
• It enhances the reliability and interpretability of scores derived from CFA models.
• This approach facilitates better understanding of covariates' influence without introducing bias.