AI Glossary / Statistics and probability / Intermediate
Conjoint analysis
Definition
Conjoint analysis is a family of methods for studying how people evaluate alternatives described by several attributes. Respondents assess profiles that vary in features such as price, performance, or service terms. Their choices, ratings, or rankings are analyzed to estimate preferences and trade-offs. In business research, it helps investigate which combinations people value, while recognizing that stated preferences do not automatically equal real purchasing behavior.
What it means
Asking whether customers want low prices, strong performance, and good support often produces an unsurprising answer: they want all three. Conjoint tasks present combinations that require trade-offs. For example, a participant may choose between a faster service at a higher price and a slower service at a lower price.
Attributes describe dimensions of an offer, and levels are the values those attributes can take. Part-worth utilities summarize the contribution of particular levels within a fitted preference model. They are relative quantities on a model-defined scale, not direct measures of happiness or guaranteed revenue.
How it works
Define the decision and the population whose preferences matter. Select relevant attributes and realistic levels using preliminary research. Avoid profiles that respondents cannot understand or that combine features in impossible ways. Too many attributes or tasks can reduce attention and response quality.
Construct an experimental design that varies attributes sufficiently to identify the intended comparisons. Choice-based conjoint asks respondents to select among profiles; other approaches use ratings or rankings. Include an appropriate outside or none option when it is important to the decision, rather than forcing a purchase interpretation onto every choice.
Estimate a model suited to the responses, potentially allowing preferences to vary across people. An additive utility model is common, but interactions may matter. Repeated responses from the same person need appropriate treatment in uncertainty estimates.
Use the fitted results to compare defined scenarios and validate predictions on holdout tasks or relevant behavior. Randomized attribute variation can support causal estimates of effects on stated responses under suitable assumptions. Translating these effects into actual market outcomes requires additional evidence about awareness, availability, competition, and behavior. Attribute importance also depends on the range of levels tested; it is not a universal ranking detached from the study design.
Three practical examples
- A software company studies trade-offs among subscription price, response speed, and support options.
- A transport provider compares passenger preferences for journey time, transfer count, and ticket price.
- An AI service team tests whether customers prefer stronger privacy controls, broader functionality, or lower cost across realistic service packages.
Think of it this way
Think of conjoint analysis as learning preferences from a series of carefully arranged shopping shelves. Each shelf changes the available combinations, making trade-offs visible. The exercise reveals choices within those shelves; it does not automatically describe what happens in a real shop with advertising, stock shortages, habits, and competing demands.
Worked example
Consider an original simplified utility model for a subscription. Fast service contributes 0.8 utility units, standard service contributes zero, and live support contributes 0.4. Price contributes -0.05 utility units per currency unit. These are stipulated model parameters, not estimates from a real survey.
Package A offers fast service and live support at price 12. Its utility is 0.8 + 0.4 - 0.05 times 12 = 0.6. Package B offers standard service and live support at price eight. Its utility is 0 + 0.4 - 0.05 times 8 = 0.
Under a two-alternative multinomial-logit rule with the assumed utility scale, A's modeled choice probability is exp(0.6)/[exp(0.6)+exp(0)], approximately 0.646. B's is approximately 0.354. These probabilities apply to this specified choice set, with no outside option.
The utility gain from fast service is 0.8. With the stipulated linear price coefficient, its compensating price difference is 0.8/0.05=16 currency units. That ratio is a model-based trade-off, not proof that real customers will pay 16 more. Different price specifications, uncertainty, and choice contexts can change the interpretation.
Illustrative case study
In a fictional company, Willow Cloud plans an AI subscription and initially asks customers to rate each proposed feature's importance. Almost every feature receives a high rating, offering little guidance about trade-offs.
The research team develops understandable package profiles, pilots the tasks, and estimates preferences from choices. It tests predictions on held-out tasks and examines differences between customer groups. A simulator compares specified packages, but the team labels its outputs as modeled preferences rather than promised market shares.
The company combines the study with delivery costs and a small real-world pilot. This fictional case illustrates how conjoint evidence can contribute to a decision without replacing behavioral validation or commercial judgment.
Common misconceptions and mistakes
Conjoint analysis is not simply asking people which single feature they like. Utility values are not directly comparable across arbitrary scales or separately fitted models. Simulated preference share is not automatically actual market share. Attribute importance depends on the levels included. A positive feature utility does not establish profitability. Adding many unrealistic attributes can damage the exercise. Willingness-to-pay calculations require careful treatment of the price model, uncertainty, and competitive context.
Business use
Businesses use conjoint analysis to compare product configurations, service packages, and pricing trade-offs. It can focus investment on combinations that customers value and reveal differences across segments. Keep cost, feasibility, and actual behavior alongside preference estimates. Report the sample, attributes, levels, task format, model, and scenario assumptions so that users know the boundaries of the findings.
People also ask
Must respondents choose rather than rate?
No. Choice-based conjoint is common, but ratings and rankings are also used.
Can conjoint estimate causal effects?
Randomized attribute designs can support causal conclusions about stated responses under appropriate assumptions; actual purchasing effects are a separate question.
Does the highest utility identify the best business product?
Not alone. Costs, constraints, competition, and strategic goals also matter.
Check your understanding
Why does the worked example give package A utility 0.6?
Answer: Its feature utilities total 1.2 and its price contribution is -0.6.
Does its modeled 64.6% choice probability guarantee that market share?
Answer: No.
Why might widening the tested price range change measured importance?
Answer: Importance depends partly on the utility variation across the included levels.
Related terms
- Choice-based conjoint
- stated preference
- attribute level
- part-worth utility
- discrete choice model
- multinomial logit
- willingness to pay
- preference share
- experimental design
- holdout validation
Learn next
Learn experimental design and discrete choice models, then practice constructing realistic profiles. Compare utility-based scenario predictions with observed choices. Next study preference heterogeneity, price modeling, and the assumptions connecting stated survey choices to decisions outside the survey.
More in Statistics and probability
- Conformal prediction - Conformal prediction is a framework for constructing prediction sets or intervals with finite-sample coverage guarantees under specified conditions...
- Confounded - Variables or effects are confounded when their contributions cannot be distinguished adequately in the comparison or design being used.
- Confounding - Confounding is a source of bias that can distort the interpretation of an observed association as a causal effect.
- Constellation model - A constellation model is a probabilistic, part-based model for recognizing object categories in images.
- Contingency coefficient - The contingency coefficient, usually meaning Pearson's contingency coefficient C, is a nonnegative measure of association between categorical variables...
- Contingency table - A contingency table organizes observations by combinations of categorical variables.
Sources
Hainmueller, Hopkins, and Yamamoto, Causal Inference in Conjoint Analysis, original methodological research: https://doi.org/10.1093/pan/mpt024 Sawtooth Software official documentation, Market Simulator Models: https://sawtoothsoftware.com/help/lighthouse-studio/manual/market-simulator-models.html Sawtooth Software official documentation, Understanding Willingness to Pay: https://sawtoothsoftware.com/help/lighthouse-studio/manual/understanding-willingness-to-pay.html
Previous: Confusion matrix Next: Conjugate gradient