Market research method
Choice-based conjoint analysis for survey researchers
Choice-based conjoint (CBC) shows respondents a few product profiles at a time — each a combination of attribute levels such as brand, battery life and price — and asks which they would choose, or none. A multinomial logit model turns those choices into a part-worth utility for every level, from which you get attribute importance, willingness to pay and simulated market shares.
How results are checked
Named statistical review pending. Selected automated reference cases are published on the validation page.Analyze a conjoint — free, no signup
Open your choice data (one row per alternative shown) and get aggregate logit and individual HB part-worths, importances, willingness to pay and a what-if market simulator. Free and browser-local.
Analyze a conjoint →How a CBC study works
You define the attributes of the product and the levels each can take. An experimental design combines the levels into profiles and groups them into choice tasks. Each respondent answers a series of tasks; in each, they pick the profile they would buy or, if the design includes it, “none of these”.
Because people choose between whole products, they have to trade attributes off against each other and against price — the way they shop — instead of saying every feature is important.
The method
Formula
P(choose j | task) = e^(Vⱼ) / Σₖ e^(Vₖ), Vⱼ = Σₐ βₐ(level of a in j)
where
- βₐ(level)
- the part-worth of a level, zero-centred within each attribute
- importance
- an attribute's part-worth range divided by the sum of all ranges, in percent
- WTP
- willingness to pay for a level change = −(change in part-worth) ÷ (part-worth per currency unit of price)
- none
- the none option gets its own constant, so the simulator can show who would buy nothing
Worked example
The synthetic earbuds study that the conjoint tool loads as sample data has 300 respondents, each answering 10 tasks with 3 products and a none option. Four attributes vary: brand, battery life, noise cancelling and price, which is estimated as one linear coefficient. Respondents chose none in 17.7% of tasks. The aggregate logit (McFadden R² = 0.164) gives:
| Attribute: level | Part-worth | Std. error |
|---|---|---|
| Brand: Aurora | 0.396 | 0.043 |
| Brand: Nimbus | 0.032 | 0.045 |
| Brand: Pulse | -0.092 | 0.046 |
| Brand: Zenith | -0.336 | 0.047 |
| Battery life: 6 hours | -0.566 | 0.040 |
| Battery life: 10 hours | 0.107 | 0.037 |
| Battery life: 14 hours | 0.459 | 0.036 |
| Noise cancelling: No | -0.349 | 0.027 |
| Noise cancelling: Yes | 0.349 | 0.027 |
| Price (USD): $49 | 0.989 | 0.038 |
| Price (USD): $79 | 0.330 | 0.013 |
| Price (USD): $109 | -0.330 | 0.013 |
| Price (USD): $139 | -0.989 | 0.038 |
| Attribute | Importance |
|---|---|
| Price (USD) | 44.6% |
| Battery life | 23.1% |
| Brand | 16.5% |
| Noise cancelling | 15.7% |
Synthetic earbuds conjoint · 300 respondents · the same data as “Use sample data”.
Price (USD) matters most (44.6% of the total range), then battery life (23.1%). Each extra dollar costs 0.0220 utility, so moving from 6 hours to 14 hours of battery is worth about $46.60 and adding noise cancelling about $31.72 to the average respondent in this sample.
In the simulator, a premium product (Aurora, 14 hours, noise cancelling, $139) against a value product (Zenith, 10 hours, no noise cancelling, $49) with the none option included takes 35.2% against 42.8%, and 22.0% choose neither.
Using it well
When to use it
Use it when
- Use: Pricing and packaging: how much a feature or a brand is worth, and what happens to share when the price moves.
- Use: Product design: which combination of levels wins against a defined set of competitors.
- Use: Trade-off questions where direct importance ratings would say everything matters.
Not the right tool when
- Avoid: Ranking a long list of claims or benefits that do not combine into products: use MaxDiff.
- Avoid: More than about six to eight attributes, which overloads respondents; split the problem or use a different design.
- Avoid: Reading simulator shares as forecast market share. They are shares of preference among the products you entered, before distribution, awareness and availability.
How to interpret it
Rule of thumb
Compare part-worths only within an attribute; the zero point of each attribute is arbitrary. Importance depends on the levels you tested — a wider price range makes price more important — so report it alongside the level ranges, and use the simulator for decisions.
Designing the tasks
A good design is level-balanced (each level appears about equally often), has little overlap within a task, and keeps attributes uncorrelated, so each part-worth can be estimated precisely. Remove impossible combinations with prohibitions rather than dropping levels. A typical study shows 8–15 tasks of 3–4 products each.
The free conjoint design generator builds minimal-overlap, balanced-overlap or random designs, applies prohibitions, adds holdout tasks and reports D-error and the expected standard errors for your sample size before fieldwork.
Aggregate logit or hierarchical Bayes?
The aggregate logit estimates one set of part-worths for the whole sample and is the quickest way to see the structure. Hierarchical Bayes (HB) estimates part-worths for each respondent, which captures differences in taste and gives more realistic simulator shares. The Crosstabs conjoint tool runs both in the browser; HB runs in a background worker.
Questions
Frequently asked questions
What is the difference between choice-based conjoint and traditional conjoint?
Traditional (ratings- or ranking-based) conjoint asks respondents to score individual profiles. Choice-based conjoint asks them to choose one profile from a set, or none, which is closer to real purchase behaviour and lets the model include a no-purchase option.
What are part-worth utilities?
Part-worths are the estimated value of each attribute level. They are on an interval scale and centred within each attribute, so compare them within an attribute: the difference between two levels tells you how much more one is preferred.
How is attribute importance calculated in conjoint?
For each attribute, take the range of its part-worths (highest minus lowest level), then divide by the sum of the ranges of all attributes. Importance depends on the levels tested, so a wider price range makes price look more important.
How do you calculate willingness to pay from conjoint?
Estimate price as a linear coefficient (utility per currency unit). Willingness to pay for a level change is the change in part-worth divided by the absolute price coefficient. It is an average across the sample and is sensitive to the price range tested.
How many respondents does a choice-based conjoint need?
It depends on the number of levels, tasks and products per task. Many studies use several hundred respondents; a design generator that reports expected standard errors for your sample size is the best way to check before fieldwork.
Sources and next steps
References & further reading
- [1]Louviere, J. J., & Woodworth, G. (1983). Design and analysis of simulated consumer choice or allocation experiments: An approach based on aggregate data. Journal of Marketing Research, 20(4), 350–367.
- [2]Train, K. E. (2009). Discrete Choice Methods with Simulation (2nd ed.). Cambridge University Press.
- [3]Orme, B. K. (2010). Getting Started with Conjoint Analysis (2nd ed.). Research Publishers.
Analyze a conjoint — free, no signup
Open your choice data (one row per alternative shown) and get aggregate logit and individual HB part-worths, importances, willingness to pay and a what-if market simulator. Free and browser-local.
Analyze a conjoint →