Conjoint Analysis: How to Choose Features and Prices
Conjoint analysis measures preferences by asking people to evaluate products or services described by several attributes together. Statistical analysis separates the contribution of those attributes to the overall evaluation, using choices, ratings, or rankings, as explained in the ISPOR conjoint analysis checklist.
The method measures stated preferences elicited through surveys. These differ from revealed preferences inferred from observed behavior, including purchases.
How conjoint analysis works
Conjoint breaks an offer into attributes and levels, combines them into profiles, and asks respondents to evaluate those profiles. Sawtooth’s introduction to conjoint analysis describes the sequence from constructing combinations to estimating preference scores.
| Term | Meaning |
|---|---|
| Attribute | A dimension of the offer, such as brand, price, or capacity. |
| Level | A specific value or option within an attribute. |
| Profile or concept | A combination of levels describing an offer. |
| Choice task | A question asking the respondent to select a preferred profile from a set. |
| Part-worth utility | An estimated preference score for an attribute level. |
Across repeated tasks, the combinations change. The analysis uses the responses to estimate how changes in the attributes affect preference for the whole offer. This supports research into features, pricing, packaging, and branding.
The central distinction is between evaluating individual features and evaluating offers containing several features together. In choice-based conjoint, respondents select among competing profiles; their choices provide the evidence for estimating trade-offs.
Main types of conjoint analysis
The method determines what respondents do and what kind of data the analysis receives:
- Choice-based conjoint (CBC): Respondents choose a preferred profile from a set. The approach is also called a discrete-choice experiment, or DCE.
- Traditional rating- or ranking-based conjoint: Respondents score or order complete profiles. Full-profile refers to presenting all the study’s attributes together.
- Adaptive conjoint: The interview changes in response to earlier answers. Adaptive choice-based conjoint (ACBC) combines an initial configuration exercise, screening, and subsequent choices. Adaptive conjoint analysis (ACA) is a separate approach and should not be treated as interchangeable with ACBC.
- Menu-based choice: Respondents make selections from a menu of options, allowing analysis of configurations or bundles assembled through multiple selections.
MaxDiff is a related method for prioritizing items. Its usual task asks respondents to identify the best and worst items in a set. Sawtooth’s method comparison distinguishes that task from evaluating products built from multiple attributes. Choosing between these approaches starts with the decision being studied: prioritizing a list of items and comparing complete feature-and-price combinations are different research questions.
How to design a conjoint study
Define the decision and the population
Specify the feature or pricing decision, the people whose preferences matter, and the alternatives they face. The ISPOR checklist calls for a defined research question and study perspective.
Record these choices before writing the survey. The study brief should identify what a result would change: a product configuration, a price range to investigate, or the offers selected for further testing.
Select attributes and levels
Use preliminary research to identify relevant attributes, then define their levels in clear language. The ISPOR guidance on attribute selection recommends evidence for inclusions and omissions; important attributes can also be held constant across profiles.
Document the reason for each attribute and level. For a pricing study, specify the currency, billing period, and what the stated price includes. For a feature decision, distinguish configurations available for implementation from configurations included to explore preferences.
Construct the experimental design
The design determines which profiles appear and which alternatives are compared. According to ISPOR’s experimental-design report, the design must provide enough variation to identify the effects the model will estimate. Attributes that always move together cannot have their separate effects identified from those comparisons.
Showing every possible combination is not generally necessary. A fraction of the complete design can support main effects or selected interactions. The required comparisons depend on the model. Restrictions on implausible combinations should be checked against that requirement.
The number of attributes, alternatives, and tasks also creates a burden for respondents. Experimental-design guidance treats statistical efficiency and respondents’ ability to evaluate the questions as a balance.
Pretest the questionnaire
Pretesting checks whether the intended population understands the descriptions and the task. The research guide Pretesting Discrete-Choice Experiments organizes this work around content, presentation, comprehension, and preference elicitation.
Its interview template asks participants to explain attributes in their own words, describe how they made a choice, identify missing information, and discuss difficulty or fatigue. Use that feedback to revise the complete survey, including instructions and the display on the devices used to answer it.
Plan the sample
There is no respondent count that establishes adequacy for every conjoint study. The practical guide to DCE sample sizes ties requirements to the hypotheses, desired statistical power, assumed effect sizes, analysis model, and experimental design.
The same paper explains that small effects can require larger samples to detect. A result without statistical significance does not by itself establish that an effect is absent.
Specify the comparisons required before recruitment, including any subgroup comparisons. Report the achieved sample for each analysis alongside the recruitment criteria, rather than reporting only the survey’s total completions.
Estimate the preference model
Analysis connects profile evaluations to attribute levels. For CBC, common approaches include multinomial logit, hierarchical Bayes, and latent-class models. Hierarchical Bayes can estimate respondent-level utilities; latent-class models can identify groups with different preference patterns.
A model can also include interactions between attributes, including brand and price. An interaction means that the effect of changing one attribute depends on another attribute’s level. Specify the effects needed for the research question and ensure the design supports them.
How to interpret conjoint results
Part-worth utilities
Utilities describe estimated preference for levels. The useful comparison is the change in utility between levels. Sawtooth’s guide to interpreting utilities explains why a single score for one attribute should not be compared directly with a single score for another.
For an additive model, profile utility is obtained by summing the relevant part-worths. Comparing complete profiles therefore uses differences across all their included levels; the utility interpretation guide demonstrates this calculation.
Attribute importance
A common relative-importance calculation uses the range between the highest and lowest utilities within each attribute. The calculation is:
Attribute importance (%) = attribute utility range / sum of all attribute utility ranges × 100
The tested levels determine the ranges, so importance depends on the study’s level selection. When respondent-level estimates are available, Sawtooth recommends calculating individual importances before averaging them, rather than calculating importance from average utilities.
Willingness to pay
Including price allows analysis of the monetary trade-off associated with a change in another attribute. Willingness to pay is measured relative to a reference level, rather than as an unconditional price for a feature.
One simulation-based approach starts with an offer’s preference share, changes a feature, then raises its price until the share returns to its starting value. The price difference estimates willingness to pay in that competitive setting.
Competition and the ability to decline matter. Sawtooth’s WTP discussion warns that estimates can be overstated when alternatives and the option to walk away are ignored. It also cautions against adding separately estimated feature premiums together. A purchase decision that permits declining every offer calls for an explicit opt-out design decision.
What a conjoint simulator shows
A choice simulator uses estimated utilities to project preference among specified alternatives. Products are configured from study attributes and levels; a none option can receive preference when it was included in the study. The output is commonly called share of preference.
These shares depend on the alternatives entered into the simulation. Sawtooth’s simulator documentation recommends defining a baseline that includes relevant competition before comparing configurations or prices.
The documentation explicitly warns that preference shares generally should not be interpreted as market shares. Awareness, distribution, advertising, availability, and other market conditions can influence actual sales. It also warns that extrapolating beyond tested levels can be unreliable.
Keep the comparison set, tested levels, respondent population, and uncertainty visible beside simulation results. Compare feasible offers within those conditions before extending the model to additional prices or features.
Using the results to choose features and prices
Preference and profitability answer different questions. Sawtooth’s pricing-research explanation notes that a simulator needs production-cost information to estimate profit. A preferred configuration still requires a separate assessment of cost and the ability to deliver it.
Use the study to identify configurations and prices worth testing. For each candidate, record its features, reference offer, price, estimated preference, uncertainty, and delivery cost. Compare candidates using the decision defined at the start of the research.
For a price decision, the same pricing-research discussion identifies historical sales analysis and in-market price tests as sources of behavioral evidence. Where practical, compare the conjoint recommendation with purchases or controlled tests of the actual offer. Record differences in audience, competing alternatives, and offer terms when interpreting any gap between survey estimates and observed results.