Ordinal Factor Analysis with Robust Estimation and Predictive Models

Abstract

Ordinal survey indicators are common in behavioral and social science research. Still, they violate continuous normal assumptions when category spacing is unequal, response distributions are skewed, or categories are sparse. This study evaluates a practical workflow for ordinal factor analysis that combines robust confirmatory factor analysis (CFA) estimation with complementary predictive modeling. Simulated five-category ordinal datasets and an empirical Malaysian green consumption dataset were analyzed using WLS, WLSMV, and DWLS estimators based on polychoric correlations. CFA performance was examined through conventional fit indices, parameter recovery, and bootstrap stability of standardized loadings. Predictive models, including random forests, support vector machines, gradient boosting, dense neural networks, convolutional neural networks, and recurrent neural networks, were assessed using matched 5-fold cross-validation. The revised comparison separates measurement validation from prediction, reports estimator stability through resampling, and clarifies the conditions under which robust ordinal estimators and nonlinear predictive models are useful. The study contributes a reproducible benchmark and reporting framework for researchers analyzing Likert-type ordinal data in psychometric and behavioral applications.