Market research tool
CBC conjoint design generator
How results are checked
Named statistical review pending. Selected automated reference cases are published on the validation page.Relative D-efficiency vs. random
137%of random D-error
D-error 0.2293 for 9 part-worths
Minimal overlap · 30 versions × 10 tasks × 3 concepts · zero-prior MNL
A CBC design decides which product concepts appear together in each choice task. A good design shows every level about equally often, pairs levels evenly, avoids repeating a level within a task when it does not have to, and never shows a combination that cannot exist. This generator builds reproducible designs with minimal overlap, balanced overlap or random construction, and scores them by D-efficiency for the multinomial logit model you will analyse them with.
Generate a conjoint design
Type up to 8 attributes with up to 7 levels each, choose up to 6 concepts per task, the number of tasks and up to 300 versions, add prohibited pairs, holdout tasks or a test–retest task, and press Generate. The design, its diagnostics, a respondent preview and the CSV exports are all computed in this tab: nothing is uploaded and no account is needed. "Use sample data" loads the wireless-earbuds attributes of the conjoint analysis sample.
Attributes and levels
2 to 8 attributes with up to 7 levels each. Type one level per line.
Design settings
Versions are alternative questionnaires; respondents are spread evenly across them. The seed makes the design reproducible: the same settings and seed always give the same design.
Minimal overlap: within a task each level appears as evenly as possible, and across the design the least-used levels and level pairs are preferred.
From design to analysis
The design file has one row per concept with version, task, task type, concept number and one column per attribute — the layout survey platforms import for CBC exercises. The response template lists respondents with the version each should see. When fieldwork is done, open the export (one row per respondent, one column per task with the option chosen) in the "Merge real answers" panel to get a long file that the conjoint analysis tool opens directly; it recognises the version and task-type columns and leaves holdout tasks out of estimation.
Before fielding, download simulated responses: choices generated from random part-worths through the same logit model. Analysing them checks that the design recovers the part-worths and shows the precision to expect. They are test data, not answers from people.
The method
Formula
D-error = det(M)^(−1/p), M = (1/V) Σᵥ Σₜ (1/C) Σⱼ (xⱼ − x̄ₜ)(xⱼ − x̄ₜ)ᵀ
where
- xⱼ
- effects-coded levels of concept j (L − 1 columns per attribute, last level −1)
- C, V
- concepts per task and number of versions; the sum runs over each version's design tasks
- p
- number of part-worth parameters, Σ (L − 1)
- SE
- √(diagonal of M⁻¹ ÷ n) for n respondents, with all part-worths zero
- Relative D-efficiency
- mean D-error of 10 random designs of the same size ÷ D-error of this design × 100
Worked example
The sample uses the earbuds attributes — brand (4 levels), battery (3), noise cancelling (2) and price (4) — with 3 concepts and a "none" option in 10 tasks and 30 versions. Minimal overlap shows every brand 225 times, never repeats a brand within a task, and keeps the two-way balance at 0.95 or better (smallest ÷ largest pair count). Noise cancelling has only two levels for three concepts, so it necessarily overlaps in every task.
Its D-error is 0.2293 for 9 parameters, against 0.3143 for random designs: a relative D-efficiency of 137%. Balanced overlap scores 121%, and a random design built with the same seed 102% with a worst pair balance of 0.55.
With 300 respondents the expected standard errors of the part-worths are 0.019–0.034 for minimal overlap and 0.022–0.040 for the random design; Johnson's rule of thumb asks for at least 67 respondents for main effects.
Using it well
When to use it
Use it when
- Use: Planning a choice-based conjoint study before programming the survey.
- Use: Checking whether a planned number of tasks, concepts and respondents gives precise enough part-worths.
- Use: Comparing minimal overlap, balanced overlap and random designs, or the cost of prohibitions.
Not the right tool when
- Avoid: Designs with priors, interaction effects or alternative-specific attributes — this generator targets main effects with zero priors.
- Avoid: Adaptive CBC or menu-based conjoint, which build tasks during the interview.
- Avoid: Ratings-based conjoint with a fractional factorial of full profiles — use an orthogonal plan instead.
How to interpret it
Rule of thumb
Lower D-error is better; relative D-efficiency above 100% means more precision than a random design of the same size. Level balance close to 1 and little overlap beyond what the number of levels forces are signs of a sound main-effects design. Prohibitions, few versions and very few tasks lower efficiency; more versions mostly help when tasks per respondent are few. Standard errors assume every part-worth is zero, so treat them as a lower bound for the precision of an aggregate logit.
Questions
Frequently asked questions
What is the difference between minimal and balanced overlap?
Minimal overlap repeats a level within a task only when there are more concepts than levels, which maximises the information about main effects. Balanced overlap allows one extra repeat per level, so respondents see some identical levels side by side; that costs a little main-effect efficiency but gives more information on interactions and can make trade-offs feel more natural.
How many versions do I need?
Each version is a different questionnaire, and respondents are spread evenly across them. More versions spread the level combinations over more respondents, which matters most when each respondent sees few tasks. The tool reports D-efficiency for the number you choose, so you can see when adding versions stops helping.
Why is the none option left out of D-efficiency?
The none alternative has no attribute levels, so it carries no information about part-worths in a zero-prior design. D-error and the standard errors are computed for the product concepts; the none option is still included in the respondent preview, the simulated responses and the merged file.
What are holdout and test–retest tasks for?
Fixed holdout tasks are shown identically to every respondent and kept out of estimation, so you can check how well the model predicts choices it did not see. A test–retest task repeats the first holdout at the end with the concepts in reverse order, to measure how consistently respondents answer.
Is anything uploaded?
No. The design is generated in this browser tab from your settings and seed, and every export, including the merge of real answers, is created on your device.
Sources and next steps
References & further reading
- [1]Huber, J., & Zwerina, K. (1996). The importance of utility balance in efficient choice designs. Journal of Marketing Research, 33(3), 307–317.doi.org
- [2]Kuhfeld, W. F., Tobias, R. D., & Garratt, M. (1994). Efficient experimental design with marketing research applications. Journal of Marketing Research, 31(4), 545–557.doi.org
- [3]Orme, B. K. (2010). Getting Started with Conjoint Analysis: Strategies for Product Design and Pricing Research (2nd ed.). Research Publishers.
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 →