On the Relation Between the Linear Factor Model and the Latent Profile Model.

Area of Science:

• Statistics
• Psychometrics
• Econometrics

Background:

• Linear factor models and latent profile models are statistical tools used for analyzing complex data structures.
• Both models decompose covariance matrices, but their estimation procedures can lead to different results.
• Understanding their relationship is crucial for accurate data interpretation in various fields.

Purpose of the Study:

• To investigate the relationship between linear factor models and latent profile models.
• To analyze discrepancies in covariance estimates between these two modeling approaches.
• To explain why a 2-class latent profile model can underestimate observed covariances when data fits a unidimensional factor model.

Main Methods:

• Maximum likelihood estimation was employed, focusing on the joint distribution of manifest variables.
• The study analyzed the conditions under which covariance estimates differ between the two models.
• Mathematical derivations explored the link between unconditional covariances, model fit, and multivariate kurtosis.

Main Results:

• While both models yield equivalent covariance decompositions, they generally do not produce equivalent unconditional covariance estimates.
• A 2-class latent profile model with Gaussian components was found to underestimate observed covariances, but not variances, when data aligns with a unidimensional Gaussian factor model.
• The study identified relationships between unconditional covariances, latent profile model goodness-of-fit, and excess multivariate kurtosis.

Conclusions:

• The findings highlight important distinctions in covariance estimation between linear factor and latent profile models.
• The identified underestimation phenomenon in latent profile models provides critical insights for model selection and interpretation.
• The analysis suggests potential parameter restrictions for improved model symmetry and accuracy.