Latent-Class Choice Segmentation

Two-step approaches first cluster on observed traits and then fit a choice model within clusters, which can produce segments that do not actually differ in behavior and biases the response estimates because the clustering ignored choice. Latent-class choice models treat segment membership as a latent variable and estimate the segments and their choice coefficients simultaneously by maximum likelihood. As Kamakura and Russell showed, this yields segments defined by genuine differences in preference and price sensitivity and recovers the elasticity structure as part of the same model, making the segmentation directly relevant to pricing and competition rather than merely descriptive.

Estimate the model for a range of segment counts and compare penalized information criteria — BIC is the most common, often alongside AIC and CAIC — which balance improved fit against added parameters. Wedel and Kamakura frame segment retention as a formal model-selection problem, and recommend supplementing the criteria with entropy-based separation measures and substantive interpretability, avoiding solutions with tiny or meaningless classes. Because the likelihood can be multimodal, each candidate model should be estimated from multiple random starts so that the comparison reflects well-converged solutions.

Choose hierarchical Bayes when you believe preferences vary continuously across consumers rather than falling into a few discrete types, or when you need individual-level part-worths for every respondent, for example to score each customer for targeting. Latent-class models excel when heterogeneity is genuinely discrete and when you want a parsimonious, interpretable set of segments with their own elasticities. The two are complementary, and analysts sometimes compare them or use mixtures of normals to bridge discrete and continuous heterogeneity; the choice depends on whether the managerial question calls for named segments or for person-level estimates.