Market research tool
Choice-based conjoint analysis
How results are checked
Named statistical review pending. Selected automated reference cases are published on the validation page.Most important attribute
44.6%importance
Price (USD), then Battery life (23.1%)
Aggregate effects-coded logit · noise cancelling worth 31.72 USD
Choice-based conjoint (CBC) shows respondents a few products at a time — each a mix of attribute levels such as brand, features and price — and asks which they would choose, often with a "none" option. A multinomial logit model turns the choices into part-worths for every level, from which you get attribute importance, the money value of a feature (willingness to pay) and simulated market shares for any line-up.
Run a conjoint analysis on your data
Open a long-format file — one row per alternative shown, with respondent, task, attribute levels and a 0/1 chosen flag — and pick the attributes. The aggregate logit appears at once; hierarchical Bayes runs on request in a background thread of this tab. Free, no account, and the file is never uploaded. "Use sample data" loads a synthetic wireless-earbuds study: 300 respondents, 10 tasks each, 3 products plus "none". Still planning the study? Build the questionnaire design with the CBC design generator; its simulated or merged response files open here directly.
CSV, Excel, SPSS, Stata, SAS, R, Parquet, SQLite or JSON. The file is read in this browser tab and never uploaded. For Triple-S, select the .sss or .xml metadata file and its data file together.
Effects coding, linear price and hierarchical Bayes
Categorical attributes are effects-coded: an attribute with L levels gets L − 1 parameters and the last level is minus their sum, so part-worths are zero-centred within each attribute. A numeric attribute such as price can instead be estimated as one linear slope, which is what willingness to pay needs: WTP = part-worth difference ÷ utility lost per currency unit. A "none" alternative (a row with every attribute blank, or flagged in a none column) gets its own constant.
The aggregate logit describes the average respondent. Hierarchical Bayes estimates part-worths for every respondent with the same Metropolis-within-Gibbs engine as the MaxDiff tool, and the market simulator then averages over each respondent's saved draws — which is what lets similar products share preference realistically and makes first-choice simulation meaningful.
The method
Formula
P(choose j | task) = exp(xⱼ·β) / Σₖ exp(xₖ·β)
where
- xⱼ
- effects-coded levels of alternative j (plus a none indicator)
- β
- part-worths; one vector per respondent under HB
- Importance
- range of an attribute's part-worths ÷ sum of ranges across attributes × 100
- WTP
- (part-worth of level − part-worth of first level) ÷ −(utility per price unit)
Worked example
In the synthetic sample, 300 respondents made 3000 choices and picked "none" in 17.7% of them. The aggregate logit (McFadden R² 0.164) makes Price (USD) the most important attribute at 44.6%, then Battery life at 23.1%.
Price was estimated as linear: each extra dollar costs 0.0220 utility. So noise cancelling is worth about $31.72 and moving from 6 to 14 hours of battery about $46.60 to the average respondent.
A premium earbud (Aurora, 14 hours, noise cancelling, $139) against a value one (Zenith, 10 hours, no noise cancelling, $49) takes 35.2% versus 42.8%, with 22.0% choosing neither (logit shares from the aggregate model; run HB in the tool for respondent-level shares).
Using it well
When to use it
Use it when
- Use: Pricing and product design: which features to bundle, and what they are worth in money.
- Use: Forecasting preference shares for new line-ups against competitors.
- Use: Two to about eight attributes with two to fifteen levels each, and eight or more choice tasks per respondent.
Not the right tool when
- Avoid: You only need to rank a list of claims or features — MaxDiff is simpler for respondents.
- Avoid: Ratings or rankings of full profiles rather than choices — use ratings-based conjoint (SPSS CONJOINT).
- Avoid: Shares are read as market forecasts without calibration: they ignore distribution, awareness and promotion.
How to interpret it
Rule of thumb
Compare part-worths only within an attribute; importance depends on the level ranges you tested, so a wide price range makes price look more important. Treat simulated shares as shares of preference among the products you define, not as volumes, and prefer HB-based simulations when products are similar.
Questions
Frequently asked questions
What data layout does the conjoint tool need?
Long format: one row per alternative shown, with a respondent ID, a task ID, one column per attribute holding the level shown, and a chosen column (1 = chosen, 0 = not). An optional alternative column orders the rows. A 'none' option is a row whose attribute cells are all blank, or a row flagged 1 in a none column you select.
How are attribute importances calculated?
For each attribute, the range of its part-worths (highest minus lowest level) is divided by the sum of ranges across attributes. Under HB this is done for every respondent and then averaged, which is why HB importances can differ from importances computed from average part-worths.
When is willingness to pay shown?
When you choose a price attribute estimated as linear and its slope is negative. Each level's value is its part-worth difference from the attribute's first level divided by the utility lost per unit of price, using mean utilities (a ratio of means under HB).
What is the difference between logit and first-choice shares?
Logit (share of preference) gives each product exp(utility) ÷ the sum over products for every respondent and averages. First choice gives the whole vote to the highest-utility product separately for each of a respondent's saved HB draws and averages over those draws, so a respondent near a tie splits their vote (close to randomized first choice). First choice needs individual utilities, so it is available after HB runs.
Is my data uploaded?
No. The file is read and analysed in this browser tab, HB runs in a Web Worker on your device, and exports are created locally. No account is needed.
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.doi.org
- [2]Train, K. E. (2009). Discrete Choice Methods with Simulation (2nd ed.). Cambridge University Press.doi.org
- [3]Rossi, P. E., Allenby, G. M., & McCulloch, R. (2005). Bayesian Statistics and Marketing. Wiley.doi.org
Try it on your own data — free, no signup
Upload a CSV, Excel, SPSS, Stata, SAS, R, Parquet or JSON file. Everything runs in your browser; your file never leaves your device.
Open the workspace →