A Novel Hybrid Latent Class Analysis –Machine Learning Framework for Multiclass Prediction of Complex Outcomes: Application to Marital Wellbeing among Middle Eastern Families

Background : Traditional statistical profiling of complex human outcomes faces a fundamental theoretical dichotomy: generative models excel at explanatory inference but impose restrictive linear constraints, while discriminative machine learning algorithms capture non-linear relationships but obscure unobserved population heterogeneity. To bridge this divide, we propose a novel, two-stage statistical paradigm that synergistically integrates likelihood-based finite mixture modeling with non-parametric ensemble learning. This architecture reframes the task into the discovery of a multidimensional latent construct, followed by a flexible mapping from observable covariates to latent group memberships using tree-based universal function approximators. Methods : The methodological framework relies on four sequential principles, instantiated using a Middle Eastern family wellbeing dataset (n=736) as a proof of concept. First, Latent Class Analysis (LCA) serves as the generative engine to model the target outcome as a multinomial latent variable. Second, the framework transitions to a discriminative approach, treating the discovered profiles as the target variables for non-parametric Ordinal Forest and Ordered Forest ensemble algorithms. Third, the principle of minimal sufficient complexity is operationalized via Recursive Feature Elimination (RFE) with cross-validation to satisfy structural risk minimization. Fourth, we deploy a multi-criteria evaluation ontology, utilizing one-vs-rest multiclass AUC for separability, macro-averaged F1-scores for subgroup equity, and the Ranked Probability Score (RPS) as a proper scoring rule for feature importance and calibration. Results : Empirically, the LCA identified three distinct, meaningful subpopulation profiles with prevalence of 25%, 56%, and 19%, validating the presence of unobserved target heterogeneity. Adhering to structural risk minimization, reduced-feature models achieved multiclass AUC values (0.775–0.793) statistically identical to full-feature architectures, confirming the power of parsimonious representations. Feature importance derived via RPS revealed that psychological constructs (depression, extraversion, and neuroticism) were the primary drivers of profile differentiation, dramatically outperforming demographic or socioeconomic metrics. Conclusion : We present a theoretically grounded, generalizable statistical framework that reconciles the explanatory depth of generative latent variable modeling with the predictive flexibility of non-parametric machine learning. This paradigm offers a principled solution for multiclass profiling in high-dimensional, heterogeneous data, with broad theoretical applicability extending well beyond the illustrative case of family wellbeing.