Conjoint segmentation groups respondents with similar trade-off patterns into distinct market audiences so product and pricing teams can design targeted tiers rather than building for a non-existent “average” buyer. The two standard approaches work at different stages of the modelling pipeline: k-means clustering groups respondents after individual-level Hierarchical Bayes (HB) utilities have been estimated, while Latent Class (LC) analysis estimates segment membership and segment-level choice models simultaneously directly from the raw choice tasks.
Ordeen runs both methods side by side on the same study (for 2 to 6 segments), pairing reliability-weighted k-means++ with full Bayesian latent-class mixtures so you can verify whether your segments reflect genuine market structure or an artefact of one algorithm. See how both are computed in our methodology.
How the two methods differ mathematically
1. Post-hoc k-means on Hierarchical Bayes utilities
In a two-stage HB + k-means workflow, Hierarchical Bayes first estimates a continuous distribution of part-worth utilities β_n for every respondent n. K-means++ then partitions those N vectors into K clusters by minimising within-cluster Euclidean distance.
The naive pitfall: Standard k-means treats every utility column and every respondent with equal weight. If an attribute was poorly identified in the survey, its noisy posterior draws can dominate Euclidean distance and pull cluster centroids toward estimation noise.
How Ordeen fixes it: Ordeen clusters on a three-block signal-weighted feature representation. Attribute importance ranges, effects-coded level utilities (including reconstructed reference levels), and walk-away (
None) engagement are combined, and each utility dimension is scaled post-standardisation by its Spearman-Brown reliability ratio τ_p² / (τ_p² + σ_p²) — the ratio of true between-respondent heterogeneity (τ_p²) to within-respondent estimation noise (σ_p²). High-signal attributes drive the segments; noisy attributes are damped automatically.
2. Model-based Bayesian Latent Class
Latent Class does not start from individual HB utilities. Instead, it models the population as a finite mixture of K classes with mixing weights π_1, ..., π_K and class-specific utility vectors β_1, ..., β_K:
$$\log p(y_n) = \log \sum_{k=1}^{K} \pi_k \prod_{t} P(y_{nt} \mid \beta_k)$$
The naive pitfall: In Bayesian MCMC sampling, mixture likelihoods suffer from label switching — Class 1 in Chain A may be called Class 3 in Chain B, scrambling posterior averages if combined naively.
How Ordeen fixes it: Ordeen aligns class identities across chains before summarising, checks convergence with a label-invariant r-hat diagnostic, and reports the Bayesian Information Criterion (BIC = −2 log L̂ + p log N, penalised on respondent count N), Leave-One-Task-Out cross-validated (LOTO-CV) hit rate, and posterior classification purity for each candidate K.
Side-by-side decision guide
Dimension | Signal-Weighted K-Means on HB | Bayesian Latent Class (LC) |
|---|---|---|
Speed | Sub-second (runs interactively on saved HB draws) | Minutes on GPU (fits a dedicated mixture model per K) |
How heterogeneity is modelled | Continuous individual tastes partitioned into K neighbourhoods | Discrete mixture of K distinct preference archetypes |
Best for small-to-medium samples (N = 300–500) | Stronger — HB partial pooling stabilises individual vectors before clustering | Needs roughly 100–150 respondents per class to identify cleanly |
Segment-size uncertainty | 95% confidence intervals (Wilson-score) on hard cluster counts | Posterior Dirichlet standard deviations on mixing weights π_k |
Quality & separation metric | Calibrated per-segment silhouette score | Posterior classification purity + BIC + LOTO-CV hit rate |
Ordeen's k-means auto-cap uses floor(N / 150) for the maximum K it will search, so on a 600-respondent study it considers K = 2 through 4, and on 1,500 it goes up to K = 10. The same rule of thumb — around 100–150 respondents per class — is a reasonable minimum for LC identification.
Why best practice is checking both on the same study
Because Ordeen persists both segmentations on the same project and profiles them with the exact same cards — showing integer segment sizes that sum to 100% (via Hamilton apportionment), odds-ratio utility deltas, small-cell-corrected demographic skews (Katz log-risk-ratio 95% lower bound), willingness to pay, and market-simulator shares — you do not have to guess which philosophy fits your market better.
Start with k-means for instant exploration across K = 2 to 6 (guided by silhouette and centroid redundancy checks). Then run Bayesian Latent Class to verify that the core commercial archetypes hold up under a formal mixture likelihood, and check the LOTO cross-validated hit rate. When both algorithms recover the same core buyer groups, you can present the segmentation to leadership with high confidence.
Related reading: What is Latent Class analysis? · What is Hierarchical Bayes? · Conjoint sample size guide · Full methodology
Want to see k-means and Bayesian Latent Class running side by side? Book a demo or email im@ordeen.com.