Features Conjoint & MaxDiff Survey Items Pricing & Kano Product Evaluation Analytics Market Simulation Survey Analysis How It Works Transparency Pricing Log in Start Free

Analytics

Six models for your choice data.
Each answers a different question.

Averages hide the differences that pricing and segmentation decisions depend on. Choicery estimates preferences for every single respondent and gives you six models to read the structure behind them. Pick the one that fits the question in front of you.

  1. Preference Analysis

    Hierarchical Bayes

    What does every single respondent value?

    The industry standard and the engine behind the market simulator.

    You getUtilities and importance per person

  2. Preference Segments

    Latent Class

    Which groups with different tastes make up my market?

    Finds the groups and picks how many there are.

    You getSegments, their sizes and utilities

  3. Segment Profiles

    Concomitant-Variable Latent Class

    Who is in each segment, and why?

    Latent Class with respondent variables, during or after estimation.

    You getMembership model and segment profiles

  4. AI-Driven Profile Analysis

    Extends TasteNet-MNL

    How do preferences depend on who people are?

    A neural network learns the link from profile to preferences.

    You getCovariate importance and preference curves

  5. Neural HB

    Profile Analysis with Bayesian estimation

    Individual utilities and a covariate story at once?

    The Profile Analysis network as a personal starting point, HB-style deviations on top.

    You getIndividual utilities, split by their source

  6. Interaction Effects

    Interaction Search

    Which attributes only work in combination?

    Every pair of attributes tested, with a correction for testing many.

    You getThe pairs that matter, level by level

One study, six models

Run them on the same data

Every model reads the same choice tasks of your project, with nothing to export or re-import in between. A few things work the same way in all of them.

The codingCategorical or linear

Choose per attribute whether every level gets its own utility or one straight line, for example a linear price. Sign constraints keep a linear effect positive or negative.

The covariatesFrom your survey

Pick survey variables such as age, usage or attitude scores, each marked nominal, ordinal or metric. Every model on this page except Interaction Effects can use them.

The checkHoldout

Hold back a choice task or a share of respondents, and see how well the model predicts what it did not see.

The simulatorResults that keep working

HB and Neural HB utilities feed the market simulator, saved segments become its filters, and a Profile Analysis run unlocks Persona.

The guideOne click from each model

Each model has a methodology guide: how it works, what every setting does and how to report the result.

The hand-overExports and replication

Excel, CSV and JSON exports, and a replication package for every run: Julia for HB, Python for the other models.

Model 1 of 6 · Preference Analysis

Hierarchical Bayes Industry standard

What does every single respondent value?

Every respondent gets their own choice model, and above them sits the population they come from. The sampler moves between both levels: a respondent with few or noisy answers borrows strength from everyone else, a respondent with clear preferences keeps their own signal. Add covariates, and the population level also shows how survey variables such as age, usage or attitudes shift preferences, and how much of the differences between people they explain.

In the recording: each respondent’s own choices on the lower level, the population on the upper level, and the sampler moving between them until the individual utilities settle.

You set

  • Burn-in, draws and the number of chains
  • Covariates: survey variables for the population level, nominal, ordinal or metric
  • Sign constraints on linear attributes, so a price effect can be held negative
  • A random holdout task or a validation split to test prediction

You get

  • Utilities and importance for every respondent, and their averages
  • Convergence and fit at a glance: R-hat, effective sample size, RLH and acceptance rate
  • Covariate effects with 95% credible intervals, their effect on importances and the share of differences they explain
  • Holdout hit rate, an Excel workbook, CSV and a Julia replication package

Checked against benchmarks. Compared with independent CBC/HB benchmark results, Choicery’s individual utilities correlate at roughly 0.96 to 0.99, depending on the conjoint type. The utilities feed the market simulator, willingness to pay, price elasticity, reservation price and sensitivity.

Model 2 of 6 · Preference Segments

Latent Class

Which groups with different tastes make up my market?

Latent Class assumes the market consists of a few groups, each with its own utilities. The EM algorithm alternates between two questions: how likely does each respondent belong to each group, and what does each group prefer? Choicery fits every number of groups in your range, with one pooled group as the baseline, and recommends the solution with the lowest BIC.

In the recording: the same choices as in HB, a few classes with their own utilities, respondents moving between classes, and BIC choosing how many to keep.

You set

  • The range of segment counts to try, from 1 to 10
  • Random starts, so a solution does not hinge on one starting point
  • An optional respondent holdout and a split-half stability check
  • An optional weight variable for respondents

You get

  • Model comparison across all segment counts: BIC, entropy, classification error and a “Ready to report?” verdict
  • Size, utilities and importance of every segment
  • Each respondent’s segment and how certain that assignment is
  • Warnings for unstable or very small segments, and segments saved as a variable for the simulator’s filters

Compare segments by importance. Each segment has its own scale, so raw utilities cannot be compared across segments. Exports come as XLSX, CSV and JSON, with a Python replication package.

Model 3 of 6 · Segment Profiles

Concomitant-Variable Latent Class Rarely available

Who is in each segment, and why?

Latent Class with respondent variables, in one of two ways. Set a variable to active, and it enters the estimation: a person’s characteristics shape their chance of belonging to each segment, estimated together with the choices, so a new person can be placed in a segment from their profile alone. Set it to passive, and the segments are fitted without it and described afterwards with a bias-corrected three-step method (BCH). Very few platforms offer either.

You set

  • Respondent variables, each switched to active (membership model) or passive (profile only)
  • Optional interactions between membership variables
  • Everything else as in Preference Segments

You get

  • A profile of every segment: weighted means and shares of each variable
  • The membership model with odds ratios and chi-square tests, corrected for testing many (Benjamini-Hochberg)
  • BCH-corrected profiles with pairwise tests for passive variables
  • All tables as XLSX

Active or passive. Active variables switch off the holdout and split-half checks, and BCH profiles are computed for runs without active variables or weights.

Model 4 of 6 · AI-Driven Profile Analysis

Extends TasteNet-MNL Only on Choicery

How do preferences depend on who people are?

A neural network, an extension of TasteNet-MNL (Han et al., 2022), turns each respondent’s profile into their part-worth utilities, and a standard logit model checks them against the choices people made. Non-linear effects and interactions between characteristics are learned without anyone specifying them. Because the utilities come from the profile alone, the model can predict a person who never took the survey.

In the recording: a respondent’s profile enters the network, the network returns that person’s utilities, and a standard choice model checks them against the observed choices.

You set

  • The covariates the network learns from
  • A preset chosen by sample size, or your own network size, learning rate and regularization
  • Smooth effects for numeric covariates
  • Restarts and the share of respondents held out for the check

You get

  • A SHAP ranking of which characteristics drive preferences, and in which direction
  • Preference curves per covariate, and each respondent’s individual contributions
  • Group comparisons, probability calibration and covariate interactions
  • Persona in the market simulator, three Excel workbooks and a Python replication package

The honest check. Held-out respondents are predicted from their profile alone. That is a harder test than HB’s fit, so the two numbers cannot be compared directly.

Model 5 of 6 · Neural HB

Profile Analysis with Bayesian estimation Only on Choicery

Individual utilities and a covariate story at once?

An estimator built by Choicery. The Profile Analysis network gives every respondent a personal starting point. Each respondent then deviates from it where their own choices demand, and the deviations are pulled toward zero the way classic HB pulls toward one shared mean. The result reads like an HB run, with a non-linear covariate explanation on top.

You set

  • The Neural HB switch in the Profile Analysis form; covariates become optional
  • How strongly deviations are held back, with 1.0 as the neutral setting
  • Final refit epochs and up to 40 restarts

You get

  • Individual utilities that feed the simulator and willingness to pay
  • A split of each preference into what the covariates explain and what is individual
  • Holdout fit, posterior-predictive and as a point estimate
  • A results workbook in the layout of the HB report

How it is estimated. A MAP fit with a local Gaussian approximation and a neural prior mean, not full MCMC sampling. Bootstrap confidence intervals are available when you run the package on your own computer.

Model 6 of 6 · Interaction Effects

Interaction Search

Which attributes only work in combination?

The baseline is one pooled logit model with main effects only. Choicery adds each pair of attributes on its own and runs a likelihood-ratio test against that baseline. A pair counts only when it stays significant after a false-discovery-rate correction and improves the model by more than your practical threshold, so testing many pairs does not manufacture findings.

In the recording: every pair of attributes gets its own model and a test against the main-effects baseline, and a pair counts only when it clears both hurdles.

You set

  • The confidence level: 90, 95 or 99%
  • The practical threshold: how much a pair must improve the model, 1% by default
  • Whether the none option takes part (DR-CBC and BW-CBC)

You get

  • Every pair with chi-square, p and q values and its improvement
  • A level-by-level explorer that shows where a combination pays off
  • The main effects of the pooled model
  • XLSX, CSV, JSON and a Python replication package

A screening step. The test runs on one pooled model, so differences between people can look like an interaction, and HB itself estimates main effects only. Available for CBC, CBC Shop, DR-CBC and BW-CBC.

Run your next conjoint study in Choicery

Try the full platform for free, and upgrade when your study needs more respondents.

Start Free See the engine

No credit card required · Up to 50 respondents free · No fixed term

Explainer