Course · 2. Choice-based conjoint (CBC)
The model
Chosŏn fits three models to the same data and shows them side by side.
| tastes differ across people | switching learned from data | |
|---|---|---|
| Aggregate logit | no | no (proportional to share) |
| Standard HB logit | yes | no (proportional within each person) |
| Our model | yes | yes |
Our model in plain terms
- Tastes: each person has their own part-worths, spread around an average. The spread for each level is estimated; a large spread means some people like what others dislike.
- Switching structure: the in-the-moment part of each choice is split into pieces shared by products with the same brand, the same flavor, the same can size, and a piece unique to each product. The shares of each piece are estimated. You choose which attributes can carry switching on the setup screen.
- One person, all their choices: the model scores each respondent's whole set of choices together, so it uses that people are consistent across tasks.
Technically this is a multinomial probit with random tastes, fitted by maximum likelihood with a standard simulation method for the probabilities. The settings that matter to you are on the setup screen; the defaults suit most studies.
Settings
- How tastes vary: everyone the same; each level varies on its own (default); or varies with correlations (premium), which can show patterns like "people who value Berry are less price sensitive".
- Learn switching along: the attributes whose shared levels can make products closer substitutes.
- Holdout: leave out each respondent's last task to check predictions (default).
- Compare with standard HB logit: on by default in the tutorial, optional elsewhere.
Compute
Studies within the free size run on our servers' CPUs, typically in one to two minutes. Larger studies run on GPUs on the premium plan.
Try it in the tutorial: model setup
Next: Reading the results
Tutorial links open the example project in the app; it is free and needs a free account.