Course · 2. Choice-based conjoint (CBC)

Reading the results

The results open on an overview: the main findings in words, each with its 95% range. The tabs behind it:

• Utilities and importance: part-worths per level, centred on zero within each attribute, and each attribute's share of the importance. Price shows as a slope: utility per unit of money.
• Willingness to pay: each level's value against the attribute's reference level, in money, at average tastes.
• Tastes: how much people disagree about each level, and the share of people who prefer each level to the reference. A level can be liked on average and still disliked by a third of people; that is where segments live.
• Segments (on the Tastes tab): groups of respondents with similar estimated tastes, found by k-means. A separation score says whether the groups are distinct or mostly split a continuum; with about ten choices per person, treat segments as tendencies. Each segment shows what stands out about it and its importances.
• Switching: the switching structure. "Shared flavor 33%" means a third of the in-the-moment variation is shared by products with the same flavor, so they are close substitutes. "Each product on its own" is the part a logit assumes is everything.
• Diagnostics: fit, holdout scores against the two logits with ranges, and our average part-worths next to HB's.
• Export: everything as JSON or CSV, including each respondent's estimated tastes.

A note on comparing part-worths across models

Raw part-worths depend on each model's scale, so our numbers and HB's are not directly comparable one for one. Compare what they imply: willingness to pay, importances and simulated shares.

Try it in the tutorial: key findings

Next: The simulator

Tutorial links open the example project in the app; it is free and needs a free account.