Course · 2. Choice-based conjoint (CBC)
When our model helps, and when it doesn't
We test every claim on simulated data where the truth is known, and only claim what those tests show.
Where it helps (shown in tests)
- Switching between similar products. Learned from the survey, no decision tree drawn up front. This is what changes launch, cannibalization and range answers.
- Near-duplicates. Variants of the same product split their share instead of stacking it.
- Forecast level. The predicted volume or profit level is closer to the truth, not just the ranking. Whether that changes go / no-go decisions against a hurdle has not been tested yet.
- Full shelves. Studies with 20 products per task fit in minutes on GPU. Tested at 20; partly tested at 60.
Partly shown
- Small studies. Some evidence of better answers at around 50 respondents; not yet conclusive. On very small studies (a few dozen people, a few tasks each) the model can do worse than a plain logit, and Chosŏn says so.
Where it makes little difference
- Feature priority alone. Importances and willingness to pay come out close to the standard models'.
- Products with no switching pattern. If products differ in no systematic way, our model agrees with the logit and does not invent similarity.
- Confidence ranges and range optimization are available in standard tools too; we provide them, but they are not why to choose this model.
End of chapter 2. Next: Menu-based choice