Latent class models for health preference heterogeneity

When patient preferences are not just heterogeneous but genuinely segmented - distinct groups with qualitatively different priorities - latent class models reveal the segments that mixed logit averages over.

This article explains latent class (finite mixture) logit models for health preference studies, how to determine the number of classes, and how to use class membership for clinical and commercial insights.

Ben White, 07.07.2026

When mixed logit is not enough

Mixed logit models describe preference heterogeneity through continuous distributions - the WTP for nausea reduction varies normally across patients. But sometimes the heterogeneity is not continuous but discrete: there are genuinely distinct groups of patients with qualitatively different preferences. Some patients are strongly efficacy-driven and willing to accept significant side effects for better response. Others are side-effect-averse and would rather have a less effective but better-tolerated treatment.

Mixed logit averages over these groups and produces a parameter distribution that fits the data but cannot identify the segments. Latent class models reveal the segments explicitly.

How latent class models work

Latent class (finite mixture) logit models estimate K distinct preference classes within the data, where K is specified by the researcher. Each class has its own set of utility parameters - its own attribute weights and WTP estimates. The model also estimates the probability that each respondent belongs to each class.

The number of classes is chosen using information criteria (BIC, AIC) and substantive interpretability. A 3-class solution that produces meaningfully distinct and clinically interpretable classes is preferable to a 4-class solution with two near-identical classes.

Class membership can be related to observed patient characteristics using class membership functions - covariates that predict which class a patient belongs to. This turns statistical segments into clinically actionable patient profiles.

TLDR Quick links

SurveyEngine health resources

Forum: modelling questions

Open a free account

Estimating latent class models in Apollo

Step 1: Export your DCE data from SurveyEngine. Latent class models use the same data format as MNL and mixed logit.

Step 2: Estimate models with different numbers of classes. Start with 2 classes and increase to 3, 4, and 5. For each model, record the log-likelihood, BIC, AIC, and AIC3 (AIC with a stronger penalty for additional parameters).

Step 3: Choose the number of classes. Select the model with the best BIC (most parsimonious with good fit) that produces substantively interpretable classes. If a 3-class model has only marginally better BIC than a 2-class model but produces a third class with fewer than 10% of respondents, the 2-class model may be preferable.

Step 4: Interpret class profiles. For each class, examine the utility parameters and WTP estimates. Label classes based on their dominant preference pattern: Efficacy-first, Safety-first, Convenience-driven, Cost-sensitive.

Step 5: Estimate class membership functions. Include patient covariates (age, disease severity, treatment history) in the class membership function to identify patient characteristics that predict class membership.

Worked example - 3-class solution for oncology preference study

A latent class analysis of oncology treatment preferences identifies 3 classes: Class 1 (42%): Efficacy-focused - high WTP for tumour response improvement, moderate tolerance of side effects; Class 2 (33%): Side-effect-averse - low tolerance of nausea and fatigue, moderate WTP for efficacy; Class 3 (25%): Convenience-driven - highest WTP for oral vs IV administration and weekly vs monthly dosing, lower sensitivity to efficacy and safety.

Class membership is significantly predicted by: age (older patients more likely in Class 3, convenience-driven), treatment experience (previously treated patients more likely in Class 2, side-effect-averse), and ECOG performance status (lower performance status more likely in Class 3). These findings directly inform patient stratification in the clinical development programme.