Individual-level preference heterogeneity and latent segmentation critically shape the choice patterns observed in discrete choice data. However, existing modeling approaches often (i) impose restrictive assumptions on the distribution of heterogeneit...
Individual-level preference heterogeneity and latent segmentation critically shape the choice patterns observed in discrete choice data. However, existing modeling approaches often (i) impose restrictive assumptions on the distribution of heterogeneity, (ii) fix the number of segments a priori, or (iii) face persistent difficulties in jointly interpreting how covariates drive segmentation and which covariates are substantively important. These limitations are particularly consequential in decision-making contexts such as public policy and energy transitions, where conflicts and value trade-offs are salient and an analytical framework is required that secures not only predictive accuracy but also transparent segmentation rationales and reliable probabilistic statements. To address this need, this dissertation proposes SCDP-Logit (Sparse Covariate-Dependent Dirichlet Process Logit). SCDP-Logit learns the number of latent clusters and their occupancy structure from the data via a covariate-dependent logistic stick-breaking nonparametric mixture, while simultaneously enhancing interpretability and mitigating overfitting through hierarchical sparsification (variable selection) that classifies high-dimensional covariates into global, local, and noise components. Stable inference under the complex hierarchical structure is achieved by an efficient MCMC sampling strategy coupled with a relabeling procedure designed to alleviate label switching. The proposed framework is validated through simulations spanning small-, medium-, and large-scale heterogeneity scenarios and through an empirical application. Across simulation settings, SCDP-Logit consistently outperforms benchmark models (e.g., MNL, HB-MNL, LC-MNL, and F-MON) on multiple dimensions, including cluster recovery, parameter estimation, variable-selection reliability, and predictive accuracy and calibration of choice probabilities. The empirical analysis applies the model to a conjoint choice experiment on Korea’s coal phase-out policy, yielding multi-layered heterogeneity and a covariate-driven segmentation mechanism. Relative to competing approaches, improvements are observed in out-of-sample prediction and probability calibration, while stable recovery of individual-level preferences and choice probabilities is maintained even under complex heterogeneity. Although some identification challenges remain for sparse tail segments, the framework produces probabilistic statements that avoid overconfident acceptance predictions across policy scenarios, providing practical value for policy risk assessment. Moreover, the unified probabilistic decomposition of covariate effects on cluster assignment and preference structure transforms segmentation outcomes into explainable criteria and clarifies the operating pathways of policy levers. This enables differentiated policy package design that reflects segment-specific preferences over costs, compensation, location, and transition timing, moving beyond one-size-fits-all prescriptions based on average effects. Selective identification of key covariates in high-dimensional settings further improves interpretability and predictive performance while reducing the operational burden in applied use. Overall, this dissertation integrates nonparametric segmentation, covariate-dependent cluster formation, and interpretable sparsification into a single discrete choice modeling framework, providing quantitative foundations for policy evaluation, targeting, and risk management grounded in structural recovery of heterogeneity and well-calibrated probabilistic prediction.