1. Design the survey
Level balance
Version 1, first three tasks
Build the screens in any survey or form tool: assign each respondent one version, ask its tasks in order, and record the option number picked. Then paste the design and a sheet of answers (respondent, version, then one column per task) below.
2. Analyse the answers
My answers are one row per respondent (wide): merge them with the design
Result
Attribute importance
Part-worth utilities
Willingness to pay
Each level against its attribute's first level, in units of the price attribute. Stated-choice WTP tends to run above what people pay in a real market: use it to rank features and set a range to test, not as a price.
Market simulator
Describe up to eight products (yours and rivals') and see the share of preference each would get. Prices may sit between the tested levels; other attributes must use tested levels.
Coefficients (multinomial logit)
The example: does the tool get back what was put in?
There is no public, freely licensed set of real conjoint answers to bundle, so the example is simulated, and that is what makes it a test: the true utilities are known. A meal-kit subscription is described by price per serving ($7 to $13), recipes per week (2 to 5), menu, delivery and packaging; 300 respondents each answer 12 tasks of three kits plus "none", and every respondent has their own utilities drawn around known means (the script is simulate.mjs; the page's example button runs it). The fit: price coefficient −0.392 per dollar, hit rate 55.1% against 25% by chance, McFadden R² 0.205. Price takes 60.0% of the importance and packaging 3.6%. People would pay $1.09 more per serving to pick their delivery day and $0.86 for the chef's premium menu.
Against the truth: the true mean price coefficient is −0.45, and across 200 repeated studies the estimate averages −0.381. That is not a bug but what one set of utilities for everyone does when people differ: every coefficient of clear size shrinks, here by 7 to 15 percent (the small vegetarian-menu effect and the none constant, which varies most between people, moved the other way). Willingness to pay is a ratio, so much of the shrinkage cancels: the delivery-day WTP averages $0.90 across studies against a true $0.89, and the others land within $0.01 to $0.10 of the truth, the none constant aside. The model-based standard errors understate the spread of estimates across studies by 5 to 19 percent; the cluster-robust ones, which the intervals here use, are within 6 percent (README, "Verification").
Method and citation
Design. Each version draws every attribute's levels from a shuffled deck that refills when empty, taking distinct levels within a task whenever the attribute has enough of them (balanced overlap, close to minimal overlap) and redrawing identical concepts; of 30 candidate versions it keeps the one with the largest determinant of the information matrix at zero utilities (D-efficiency; Kuhfeld, Tobias and Garratt 1994; Huber and Zwerina 1996). The expected standard errors are from the inverse of that information matrix divided by the number of respondents; at real, non-zero utilities they are somewhat larger.
Model. The conditional (multinomial) logit of McFadden (1974), as used for choice experiments since Louviere and Woodworth (1983): each concept's utility is the sum of its levels' part-worths (dummy coded against each attribute's first level, or a slope for a numeric attribute centred at its mean), plus a constant for "none", and the chance of each pick is the softmax of the utilities in its task. Fitted by Newton-Raphson with step halving to the exact maximum likelihood (Train 2009). Standard errors: model-based from the Hessian, and cluster-robust by respondent (sandwich, with G/(G−1)), which the intervals use because one person's twelve answers are not independent. Importance, willingness to pay and simulated shares get percentile intervals from a bootstrap that resamples respondents.
One segment. Aggregate logit fits one set of utilities for everyone. It has the IIA property in the simulator, and under heterogeneity it shrinks coefficients toward zero (the example shows by how much). Hierarchical Bayes, mixed and latent-class logit are not here yet.
Not here (yet): prohibited level pairs in the design, adaptive CBC, dual-response none, MaxDiff (the quick-pick survey has a best-worst mode), and individual-level utilities.
Checked against: statsmodels 0.15.0 ConditionalLogit with its own dummy coding, and a numpy cluster-robust sandwich, on four simulated studies (forced choice, five concepts with two numeric attributes, with and without heterogeneity): coefficients to 1e-7, standard errors to 1e-6, log-likelihood to 1e-9; a hand-worked two-concept study; and a coverage simulation in which the 95% intervals covered the true utilities in 95.0% (model-based) and 94.8% (cluster-robust) of 2,400 intervals (README, "Verification").
@incollection{mcfadden1974conditional,
author = {McFadden, Daniel},
title = {Conditional logit analysis of qualitative choice behavior},
booktitle = {Frontiers in Econometrics}, editor = {Zarembka, Paul},
publisher = {Academic Press}, address = {New York}, pages = {105--142}, year = {1974}}
@article{louviere1983design,
author = {Louviere, Jordan J. and Woodworth, George},
title = {Design and analysis of simulated consumer choice or allocation experiments: an approach based on aggregate data},
journal = {Journal of Marketing Research}, volume = {20}, number = {4}, pages = {350--367}, year = {1983},
doi = {10.1177/002224378302000403}}
@article{huber1996utility,
author = {Huber, Joel and Zwerina, Klaus},
title = {The importance of utility balance in efficient choice designs},
journal = {Journal of Marketing Research}, volume = {33}, number = {3}, pages = {307--317}, year = {1996},
doi = {10.1177/002224379603300305}}
@article{kuhfeld1994efficient,
author = {Kuhfeld, Warren F. and Tobias, Randall D. and Garratt, Mark},
title = {Efficient experimental design with marketing research applications},
journal = {Journal of Marketing Research}, volume = {31}, number = {4}, pages = {545--557}, year = {1994},
doi = {10.1177/002224379403100408}}
@book{train2009discrete,
author = {Train, Kenneth E.},
title = {Discrete Choice Methods with Simulation}, edition = {2nd},
publisher = {Cambridge University Press}, year = {2009},
doi = {10.1017/CBO9780511805271}}
Questions
What is choice-based conjoint?
Each respondent sees a few screens, each with two to five versions of a product described by the same attributes (price, size, plan, delivery) at different levels, and picks the one they would buy, or none. Because every screen forces a trade-off, the choices reveal how much each level is worth relative to the others, which a rating question cannot. It is the most common form of conjoint analysis in market research.
How many respondents and tasks do I need?
A common rule of thumb (Johnson and Orme) is respondents times tasks times concepts per task, divided by the largest number of levels of any attribute, of at least 500, and 300 or more respondents for a study that will be acted on. The design step here also shows the expected standard error of every coefficient for a given number of respondents, and how many respondents bring them all under a target such as 0.05.
What are part-worths and attribute importance?
A part-worth is the utility a level adds, on the logit scale. They are shown centred so the levels of each attribute average zero, which makes them easy to compare within an attribute; only differences matter. Importance is each attribute's range of part-worths (best level minus worst) as a share of the sum of ranges. It depends on the levels you tested: a wider price range makes price look more important.
How is willingness to pay computed?
Mark price as numeric so it gets one coefficient per unit, then the willingness to pay for a level is its utility difference from the attribute's first level divided by minus the price coefficient. Stated-choice WTP usually runs higher than what people pay in a real market, and when the price coefficient's interval comes near zero the WTP interval becomes very wide; the page warns when it does.
What does the market simulator show?
Shares of preference: for the products you describe, and optionally the none option, the logit probability that a respondent picks each one. They are not sales forecasts: they ignore awareness, distribution and how often people buy. An aggregate logit also has the IIA property, so a new product takes share from every rival in proportion to its share, even from rivals it does not resemble.
Aggregate logit or hierarchical Bayes?
This page fits one set of utilities for everyone (aggregate multinomial logit). When people differ, that shrinks every coefficient toward zero; in the simulation on this page the clearly sized ones by 7 to 15 percent, while ratios such as willingness to pay moved less. Its standard errors are cluster-robust by respondent so they stay honest under that heterogeneity. Hierarchical Bayes or latent-class models estimate individual or segment utilities and are not here yet.
Is my data uploaded anywhere?
No. The design, the answers and every estimate are computed by JavaScript in this browser tab. There are no accounts, cookies or analytics vendors.