What Is Each Feature Actually Worth? Against Price

How it works

The module detects which of two conjoint shapes arrived. With a 0/1 preference column and a choice-task column it fits a conditional (multinomial) logit by maximising the conditional-logit log-likelihood directly with a quasi-Newton optimiser and an analytic gradient, taking standard errors from the observed information matrix. Otherwise it fits ordinary least squares of the preference on dummy-coded attribute levels. Either way the coefficients are then re-expressed as sum-to-zero part-worths within each attribute — an exact linear transform, so the covariance matrix carries through and every level including the reference gets an interval. Attribute importance is each attribute's part-worth range as a share of the total range. Willingness-to-pay comes from the slope of the price part-worths on the numeric price levels, which is itself a linear combination of the part-worths, so the delta method uses the full covariance between a level and the slope. The market simulation sums part-worths over configurations built from the tested levels, adding logit shares only when the choice model identifies the scale. The heterogeneity check subtracts the other attributes' aggregate contributions and reads each respondent's own favourite level off what remains.

Use it on conjoint or discrete-choice survey data where respondents evaluated whole product profiles — rating them, ranking them, or choosing between them — and you need the relative weight of each attribute and the exchange rate between a feature and money.

Not for revealed-preference data such as actual transactions (use a choice or regression model on the observed prices), not for rating single features one at a time in isolation, and not when you need individual-level utilities for segmentation — that needs a hierarchical model this module does not fit.

What data do you need?

One row per profile a respondent evaluated, with a column per attribute and the stated preference. For choice-based conjoint, add a choice-task column and make the preference a 0/1 chosen flag:

Minimum 20 rows · Best with 300-8,000 rows (roughly 50 or more respondents evaluating 8 or more profiles each)

What's in the report?

Standard-library analysis: what is each product feature worth relative to price? Map the preference column and the attribute columns from a conjoint survey — ratings/rankings of whole profiles, or choice-based conjoint with a chosen flag and a choice-task column — and get part-worth utilities for every tested level with confidence intervals, attribute importance reported against the level ranges that were actually tested, willingness-to-pay derived from the price slope when the price coefficient supports it, a market simulation over product configurations built from the tested levels, and a respondent-level check for whether the aggregate is masking opposed segments.