OpinionX uses a Hierarchical Bayes Multinomial Logit (HB-MNL) model to estimate individual-level preference utilities from conjoint ranking data, paired with a maximum-utility scenario simulator. These are industry-standard approaches for choice modelling that provide more accurate and stable preference estimates compared to traditional methods.
Model Framework
The HB-MNL model is a Bayesian discrete choice model that estimates part-worth utilities for each attribute level while accounting for preference differences across respondents. The hierarchical structure "borrows strength" across the population, regularizing individual-level estimates toward a population mean — producing more stable preference scores even with smaller sample sizes per respondent.
At the individual level, each respondent's choices are modelled using a multinomial logit specification. Utility is specified as a linear combination of attribute levels using effect coding, where one level per attribute serves as a reference. The utilities of all levels within an attribute sum to zero, with the reference level's utility derived as the negated sum of the others.
At the population level, individual-level utilities are assumed to be drawn from a population distribution with a population mean and variance per attribute. The model does not estimate cross-attribute correlations. Weakly informative priors are used: a normal distribution for population mean preferences and a half-normal distribution for population variance.
The hierarchical structure creates a shrinkage effect — individual estimates are pulled toward the population mean, with more shrinkage applied to respondents who completed fewer choice tasks. This regularization prevents overfitting and improves the reliability of individual-level estimates.
The posterior distribution is estimated via Markov Chain Monte Carlo (MCMC) using the No-U-Turn Sampler (NUTS). After a warmup period, the sampler produces posterior draws from which OpinionX computes the posterior mean for each parameter as the final part-worth utility per respondent.
Derived Outputs
The primary output is a set of part-worth utilities for each respondent and attribute level, representing the preference value associated with each level. Because HB-MNL estimates these at the individual level, OpinionX can provide personalized preference profiles for each respondent — enabling segmentation analysis and targeted product design.
Scenario Simulator
The scenario simulator lets you create hypothetical product profiles and forecast their market share using a maximum-utility (first-choice) rule applied to the individual-level utilities from the HB-MNL model.
For each respondent, the simulator scores every profile by summing their part-worth utilities for that profile's attribute levels, then assigns them to whichever profile scores highest. Ties are resolved by splitting the respondent's vote equally among the tied profiles. Individual choices are then aggregated across all respondents (or a specific segment) to calculate overall market share for each profile.
Model Assumptions
OpinionX's HB-MNL implementation assumes:
Independence of irrelevant alternatives (IIA): The relative odds of choosing between two alternatives are independent of other alternatives in the choice set.
Linear-additive utilities: Utility is a linear combination of attribute levels with no interaction terms.
Independent normal heterogeneity: Individual utilities are drawn from independent normal distributions per attribute at the population level.
No alternative-specific constants: All alternatives are generic products defined solely by their attributes.
These methods represent current best practices in conjoint analysis and are used by leading commercial platforms and academic researchers worldwide.