Load demo data

2. Scoring method

3. Column mapping

Choose the structural columns, then select which columns are attributes.

Chosen value rule

Rows count as chosen when the mapped chosen column is greater than or equal to this value.

Attribute columns
Mapped structural columns are automatically excluded.
Segment columns
Choose respondent descriptors such as customer type or region. Each value used by at least two respondents becomes an audience; segment columns are excluded from attribute coding. Up to 30 audiences can be estimated per run.
Attribute level order and name

Select an attribute or level to rename it. Drag a level to reorder it.

Conjoint analysis overview

No model run yet
Load a CSV and run a model to see notes.

Part-worth utilities

Attribute importance

Levels and scores

Attribute Level Utility Zero-centered utility Importance %
Load demo data

2. Scoring method

3. Column mapping

Map the structural fields used by the MaxDiff scorer.

Selection threshold

Best and worst flags count as selected when the value is greater than or equal to this threshold.

Model summary

No model run yet
Load a MaxDiff CSV and run a model to see notes.

Item scores

Best vs worst percentages

Items and scores

Item Appearances Best selection rate (%) Worst selection rate (%) Descriptive best − worst (%) Best-minus-worst score (%)
Scenario planning

Preference simulator

Build competing profiles and estimate their share of preference using a completed choice-based conjoint Hierarchical Bayes model. Results describe stated preference within the tested offer space—not expected market share or sales.

Run a Conjoint HB model or load a saved Tradeoffs.app HB model to begin.

Scenarios and profiles

Every product profile must use one tested level from each attribute.

No HB model is available yet.
Simulator limitations: preference share is conditional on the profiles entered and the study’s tested levels. Simulation intervals reflect posterior uncertainty, preference heterogeneity, and finite simulation—not total market-forecast error. The simulator does not account for awareness, availability, distribution, competitive offers not entered, survey-to-market calibration, or purchase incidence. Do not describe these results as forecast market share without external calibration and validation.
Business user guide

Understanding Hierarchical Bayes (HB) estimates

HB turns repeated choices into estimates of which product features people prefer. It combines each respondent’s choices with the overall pattern across the sample, producing more stable results when each person answers only a limited number of tasks.

Scope: This guide describes the choice-based conjoint (CBC) model estimated with Hierarchical Bayes in the Conjoint tab. The maximum difference (MaxDiff) tab uses a separate HB estimator.

The 30-second explanation

1
Observe the trade-offs

The model reviews which profile was selected in every task and which alternatives were rejected.

2
Estimate individual preferences

It estimates a preference pattern for each respondent using all of that person’s tasks.

3
Borrow strength from the sample

Clear individual signals remain distinctive; limited or noisy signals are pulled toward the population pattern.

4
Report the market-level result

Tradeoffs.app averages the retained population estimates across the independent chains to produce the displayed part-worths.

Why this helps: estimating each respondent independently would be unstable with only a few tasks. Estimating one aggregate model would ignore meaningful differences between people. HB balances both sources of information.

How to read the results

Part-worth utility

A relative preference score for a level. Higher is more preferred within the study; lower is less preferred. The distance between levels matters more than the absolute number.

Zero-centered utility

Utilities are shifted so the levels within each attribute average to zero. Positive means above the attribute average and negative means below it. Zero does not mean “no value.”

Attribute importance

The utility range within an attribute as a share of all attribute ranges. It shows differentiation across the tested levels—not a universal percentage of the real-world decision.

None alternative-specific constant (ASC)

A separate opt-out constant after accounting for the profiles shown. It is not an attribute level and is excluded from zero-centering and importance.

Safe interpretation: “Within the offers and levels tested, respondents preferred X to Y, and Attribute A created more preference differentiation than Attribute B.”

What the model accounts for

Included in the model

  • All mapped attributes and levels, evaluated together
  • Competition among the alternatives in each task
  • Repeated tasks completed by the same respondent
  • Differences in preferences across respondents
  • A separate tendency to select None when present

Not automatically included

  • Demographics, segments, or other respondent characteristics
  • Interactions such as “the bonus matters more at a high price”
  • Task order, fatigue, learning, or survey-speed effects
  • Differences in response consistency or choice scale
  • Awareness, availability, competitive context, and market execution

The current CBC model estimates main effects only. All non-None attributes are treated as categorical levels; it does not impose a linear price curve or a monotonic relationship.

What the HB settings control

Most settings change calculation reliability and runtime, not the underlying business question.

SettingBusiness meaningPractical guidance
ChainsIndependent attempts to reach the same answerUse four for reporting. One chain cannot assess chain agreement.
IterationsRetained estimates per chainMore draws improve precision only when the sampler is mixing adequately.
Burn-inEarly estimates discarded while the model settlesThe default is a reasonable starting value; more is not automatically better.
Population covarianceWhether preference differences may be correlatedDiagonal is the stable default. Full covariance needs more choice tasks per respondent and stronger diagnostics.
ThinningKeeps every nth estimateKeep every draw unless memory is constrained.
Proposal controlsHow the calculation explores plausible answersKeep the calibrated defaults unless diagnostics identify a sampling problem.
Important: a longer run can reduce simulation noise, but it cannot repair a weak design, poor fieldwork, incorrect column mapping, or an unrepresentative sample.

Model-health checks in plain English

HB level exposure
For each person, how often did every tested level actually appear? Six or more appearances per level is the preferred target for Hierarchical Bayes (HB). When exposure is lower, individual estimates depend more heavily on information shared across the population.
Pairwise balance
How evenly did levels of different attributes appear together? 95–100% is optimal. Also review matrix rank, expected precision, design restrictions, and comparisons with alternative designs.
Chain agreement
Do independent runs reach essentially the same answer? Robust R-hat at or below 1.01 is preferred.
Effective sample size
How much independent information remains after accounting for repetition between draws? At least 400 for both the middle and tails of every reported coefficient is preferred.
Acceptance rates
Is the sampler moving efficiently? These are sampler-specific tuning checks, not proof of convergence. Robust R-hat and effective sample size carry more weight.
Monte Carlo standard error (MCSE) / posterior standard deviation (SD)
How much uncertainty comes from finite simulation? At or below 0.05 is preferred.
Root likelihood (RLH)
How well the estimated utilities reproduce the observed choices relative to chance. It is in-sample fit, not proof of future market accuracy.

Green model health means the calculation is stable. It does not guarantee that the sample, experiment, or business assumptions are valid.

Limitations to include in reporting

  • Stated preference is not purchase behavior. Results describe choices in the study context.
  • Utilities are relative. They should not be compared directly across different studies or samples.
  • Importance depends on the tested range. Changing the included levels can change importance.
  • Main effects can miss combinations. The model does not estimate attribute interactions.
  • None is context-sensitive. Wording, realism, category involvement, and offer quality can alter opt-out behavior.
  • Convergence is necessary, not sufficient. A stable model can still be based on weak or incorrectly mapped data.
  • Current results emphasize averages. Respondent preferences are estimated internally, but the Conjoint results display the population-average part-worths.
  • Uncertainty is not shown per level. The current utility table reports posterior means rather than level-level credible intervals.
  • Validation is in-sample unless holdouts exist. RLH alone does not establish out-of-sample predictive accuracy.
  • The sample defines the market estimate. Unrepresentative recruiting or weighting gaps carry into the results.

Before presenting the findings

  1. 1.Confirm column mapping, None detection, chosen-value rules, attributes, and levels.
  2. 2.Review pooled exposure, the six-appearances-per-level HB diagnostic, pairwise balance, within-task overlap, and matrix rank.
  3. 3.Check respondent count, completed tasks, exclusions made before upload, and sample representativeness.
  4. 4.Require acceptable chain agreement, effective sample sizes, acceptance behavior, and simulation error.
  5. 5.Look for implausible or contradictory utilities that could signal coding, design, or respondent-quality issues.
  6. 6.Frame conclusions within the tested attributes, levels, sample, and choice-task context.
Say this

“Within the tested range, respondents showed a stronger relative preference for X than Y.”

Avoid this

“X will cause a specific increase in sales” unless separate market evidence supports that claim.

Help center