Senior Statistical Researcher and PhD Candidate in Statistics with 17+ years of international experience in statistical analysis, institutional research, performance measurement, survey design, and evidence-based decision-making. Current doctoral research at Universiti Utara Malaysia focuses on ordinal factor analysis and the integration of robust confirmatory factor analysis (CFA), machine learning, and deep learning for more accurate, stable, and interpretable measurement. Experienced in official statistics, predictive modeling, KPI development, data governance, executive reporting, and advanced statistical consulting across government, higher education, and research institutions.

Hybrid Factor Analysis for Ordinal Survey Data: Fusing Confirmatory Psychometrics with Machine Learning and Deep Learning under the HFA Framework AbstractOrdinal Likert-type survey data represent the dominant measurement paradigm in the social and behavioral sciences, yet Confirmatory Factor Analysis (CFA) relies on linearity assumptions that distort parameter estimates when applied to ordered categorical indicators. Robust estimators such as Weighted Least Squares Mean and Variance Adjusted (WLSMV) address threshold irregularities via polychoric correlations but remain constrained by linear functional forms, leaving systematic nonlinear item-level structure uncaptured. Modern Machine Learning (ML) and Deep Learning (DL) architectures excel at modeling high-dimensional nonlinearities; however, they sacrifice psychometric interpretability. This study introduces the Hybrid Factor Analysis (HFA) framework -- a unified pipeline fusing the parametric rigor of robust ordinal CFA with nonparametric feature representations from Gradient Boosting (XGBoost) and Variational Autoencoders (VAE). HFA employs an adaptive weighting parameter, omega, optimized via cross-validated grid search, integrating structural measurement weights with computational item-importance scores. The hybrid weight is a convex combination of within-factor normalized CFA loadings and within-factor normalized ML/DL saliency, preserving the theorized item-to-construct mapping while adaptively correcting for residual nonlinear structure. HFA was validated through multi-stage Monte Carlo simulations (n = 250, 500, 1,000) under both even and skewed threshold conditions, and applied empirically to the Green Consumption Behavior dataset from Malaysia (n = 375; five latent constructs; 1,000 bootstrap replications). Under small-sample skewed conditions (n = 250), HFA reduced factor loading mean squared error by 35.2% relative to standalone WLSMV. Empirically, the optimal weight omega = 0.44 captured a substantial nonlinear residuum ignored by purely linear models, yielding CFI = 0.965 and RMSEA = 0.048 versus CFI = 0.942 and RMSEA = 0.068 under WLSMV, narrowing bootstrap confidence-interval widths by 28%, and improving behavioral variance explanation to R-squared = 0.56. HFA bridges the accuracy-interpretability divide, offering quantitative researchers a transparent, adaptive tool for latent construct validation in intelligent data environments. Keywords: hybrid factor analysis; ordinal psychometrics; confirmatory factor analysis; machine learning; deep learning