Quick answer
Conjoint analysis estimates how people trade product or service attributes by asking them to evaluate or choose among experimentally varied profiles. Choice-based conjoint is common because respondents select between alternatives, often including a none option. Define a real roadmap, package or pricing decision; choose attributes that are important, understandable, actionable and sufficiently independent; create realistic levels and prohibitions; generate an efficient randomized design; recruit the relevant choice population; and estimate part-worth utilities using an appropriate logit, latent-class or hierarchical Bayesian model. Validate data, design, holdout tasks and external behavior, then use a constrained simulator for scenarios. Preference shares are not literal market shares without availability, awareness, distribution, competition and calibration.
What conjoint analysis is
Conjoint analysis is a family of stated-preference methods that decomposes evaluations of multi-attribute alternatives into estimates of the value associated with attribute levels. It was introduced to marketing as a way to quantify joint effects and trade-offs rather than asking importance directly.
Choice-based conjoint, also called a discrete choice experiment in many contexts, presents sets of profiles and asks respondents to choose. The design varies levels systematically so a model can estimate how each contributes to choice, conditional on the attributes and alternatives shown.
The output is utility on an arbitrary model scale, not satisfaction points or money by default. Utilities support comparisons, importance within the studied level ranges and simulations when interpreted with the model's assumptions.
Use conjoint for consequential trade-offs
Use conjoint when a team can vary features, service, package, price or other offer attributes and needs to understand trade-offs. It can support roadmap, product-line, bundle and pricing scenarios before every alternative exists in market.
Do not use it when the concept is poorly understood, attributes cannot be represented credibly or the real decision is driven by unmeasured awareness, availability or social process. Qualitative research should define the choice before experimental profiles quantify it.
Write the decision and simulator scenarios first. This prevents stakeholders from adding attributes because they are interesting. Every attribute consumes respondent attention and creates levels the design must vary enough to estimate.
Build a conjoint study
Begin with qualitative discovery, competitive evidence and operational constraints. Select attributes that matter to choice, can be understood, can vary independently enough and can lead to action. Define levels spanning a realistic decision range without making dominance obvious.
Create an experimental design with level balance, enough independent variation and plausible combinations. Block tasks across respondents, include a none or status-quo alternative when realistic, and pilot cognitive burden and interpretation.
Field to the relevant choice population, estimate a model suited to heterogeneity, validate held-out and external choices, then simulate only feasible scenarios. Document what the model excludes and how results depend on specification.
Decision
Define the product, package or pricing choice and the market scenario to simulate.
- What will management change?
- Which alternatives do buyers actually face?
Attributes
Select understandable, actionable and sufficiently independent attributes and realistic levels.
- Can the business deliver each level?
- Will respondents believe combinations?
Design
Construct efficient, balanced choice tasks with plausible alternatives and an outside option.
- Can effects be identified?
- Is respondent burden acceptable?
Estimate
Check respondent and design quality, then estimate utilities and preference heterogeneity.
- Which trade-offs explain choices?
- Are segments stable?
Validate and simulate
Test predictions and run bounded scenarios with real-world calibration and constraints.
- Does the model predict held-out choices?
- What does it omit from market share?
Choose attributes and levels carefully
Attributes should use customer language and represent distinct offer dimensions. Avoid overlapping a broad benefit with the feature that creates it, or splitting one favored feature into several attributes that artificially increase its apparent importance.
Levels must be specific, mutually interpretable and actionable. Use realistic price ranges and preserve logical order. Prohibit impossible combinations, but too many prohibitions can reduce design efficiency and confound effects, so consider alternative-specific designs where the market truly differs.
Price should vary independently enough from features to estimate trade-offs, while profiles remain believable. Brand and price interactions may matter. If respondents assume a low price signals poor quality, the model should not blindly interpret every price coefficient as pure economic sensitivity.
- Decision and simulator use written
- Choice population defined
- Qualitative discovery completed
- Attributes distinct and actionable
- Levels realistic and sufficient
- Price range decision-relevant
- Prohibitions minimized
- None or status quo considered
- Design efficiency checked
- Task burden piloted
- Holdouts reserved
- External calibration planned
Design realistic choice tasks
Choice tasks usually show several alternatives described by the same attributes. Efficient algorithms select combinations that improve parameter information while maintaining balance and low correlation. Random versions distribute tasks and reduce systematic context effects.
Too many attributes, alternatives or tasks cause simplification, fatigue and random choice. Partial-profile or adaptive methods can help complex decisions but add assumptions. Pilot completion time, comprehension and whether respondents use plausible decision rules.
Include holdout tasks not used for estimation to test internal prediction. Keep some respondents or external market choices for stronger validation. A repeated survey holdout remains closer to the research exercise than a real purchase.
Conjoint analysis example
The meal-planning study defines features that can actually be packaged and keeps a no-subscription option. The model can estimate whether dietary support changes preference and how that trade-off varies with price within the tested range.
The simulator does not call its output market share. The company constrains plans to feasible combinations and validates a selected package in a live offer, because awareness, channel, billing and competitor action lie outside the conjoint task.
A hypothetical meal-planning service must choose packaging for families and decide whether grocery integration, dietary support or expert help justifies a higher monthly price.
Interviews identify realistic trade-offs and language. The team excludes features that every plan must contain and levels it cannot actually deliver.
Each task shows three credible plans plus no subscription, varying household seats, dietary support, grocery integration, expert help and price under technical prohibitions.
Recruit category decision makers, pilot task comprehension and duration, randomize design blocks and include quality checks without deleting inconvenient price-sensitive respondents.
Use hierarchical Bayes for individual-level part-worth distributions and latent class as a strategic segmentation check, reporting uncertainty and scale caveats.
Compare feasible packages against current and competitor proxies, test holdouts and price sensitivity, then calibrate with a live offer experiment before launch.
The example simulator estimates stated preference under included attributes. It does not include awareness, acquisition, availability, billing friction or competitor response unless those are separately modeled.
Estimate utilities and heterogeneity
Multinomial logit estimates average effects under assumptions about choice errors and substitution. Hierarchical Bayesian methods borrow information across respondents while estimating individual-level utility distributions. Latent-class models identify groups with different preference patterns.
Choice utility scale reflects both preference and error variance, so raw coefficients should not be compared casually across samples or models. Attribute importance depends on the level range tested; widening price or feature range changes the calculated share of importance.
Report parameter uncertainty and inspect signs, monotonicity, interactions and respondent quality. Do not delete people solely because they dislike the sponsor's feature or choose none frequently. Define exclusions from evidence such as inattention, impossibly fast completion or failed comprehension.
Use market simulators as conditional scenarios
A simulator combines utilities for specified alternatives and applies a choice rule to estimate share of preference. It is valuable for comparing packages, feature changes and prices within the experimental support, especially when the competitive set is specified consistently.
Shares of preference omit awareness, distribution, marketing, availability, inertia, channel and supply unless modeled separately. Calibration can align simulations with observed choices, but calibration factors should be documented and tested out of sample.
Use sensitivity analysis across model, scenario, competitor and price assumptions. Optimize profit only after adding cost, capacity and cannibalization. A utility-maximizing concept can be operationally impossible or economically weak.
Validate design, data and predictions
Check design balance, overlap, prohibited combinations and exposure counts before fielding. During fieldwork monitor device display, response time, straight patterns, dominant alternatives and none use. Cognitive interviews explain whether respondents interpret levels as intended.
Evaluate holdout prediction, test-retest reliability where appropriate and posterior predictive behavior. Compare alternative specifications and segment stability. Internal holdouts test the survey model more than real market validity.
Whenever possible, compare against choices from a pilot, sales data or randomized offer test. Investigate disagreement through awareness, availability, omitted attributes, hypothetical bias and changing competition rather than forcing a calibration that preserves the preferred package.
Archive the exact profiles, design seed, exclusions, model specification and simulator rules. Reproducibility matters because small specification changes can alter scenario rankings that appear commercially decisive.
Limitations and common mistakes
Conjoint is a hypothetical choice environment. Respondents may simplify, ignore attributes, overstate consideration or behave differently when spending real money. The model assumes the included profiles and choice rule represent enough of the decision.
Common mistakes include too many attributes, overlapping features, unrealistic levels, no outside option, weak price variation, excessive prohibitions, reporting utility as willingness to pay, calling preference share market share and optimizing outside observed support.
The method is powerful because it forces trade-offs, not because it predicts the future automatically. Pair it with qualitative discovery, cost and feasibility, competitive evidence and behavioral validation.
Direct importance questions let every feature win. Conjoint earns insight by making respondents sacrifice, then modeling the pattern under explicit choice assumptions.
Frequently asked questions
What is conjoint analysis?
It is a stated-preference method that estimates the contribution of attribute levels by asking people to evaluate or choose between experimentally varied product or service profiles.
What is choice-based conjoint?
It is a conjoint method in which respondents choose among several multi-attribute alternatives, often including a none or status-quo option, across repeated tasks.
What are part-worth utilities?
They are model coefficients representing relative preference for attribute levels on an arbitrary utility scale, conditional on the study design and model.
Can conjoint analysis calculate willingness to pay?
It can derive price trade-offs under assumptions, but estimates can be unstable and sensitive to price coding, level range, scale and hypothetical bias. Report uncertainty and validate behavior.
Are conjoint preference shares market shares?
No. They omit or simplify awareness, distribution, acquisition, availability, inertia, execution and competitor response unless those are separately modeled and calibrated.
Sources and further reading
- Green and Rao: Conjoint Measurement for Quantifying Judgmental Data ↗Foundational Journal of Marketing Research exposition, applications and limitations
- Sawtooth Software: What Is Choice-Based Conjoint? ↗Publisher technical overview of tasks, designs, utilities, heterogeneity and simulation
- Sawtooth Software: Design of Conjoint Experiments ↗Publisher technical resources on balance, efficiency, blocking and choice-task design
- Sawtooth Software: Including Holdout Choice Tasks ↗Technical paper on internal validation with fixed holdout choice tasks and limitations