One of the first charts people ask for after a conjoint study is attribute importance: a bar chart adding up to 100% that answers "did buyers care more about brand, price, speed or warranty?"
Most conjoint tools calculate it with the range method. Its weakness has been known for decades: you can change the answer just by testing a wider price range or adding an extreme level.
How the range method works, and why it breaks
A conjoint model gives every level of every attribute a utility score (a part-worth). The range method takes each attribute's best level minus its worst, and divides by the total across all attributes:
$$\text{Range importance}_a = \frac{\max(\beta_a) - \min(\beta_a)}{\sum_k \big(\max(\beta_k) - \min(\beta_k)\big)}$$
Sawtooth Software's own documentation makes the point directly: "a price attribute ranging from $100 to $200 will appear less important than one ranging from $100 to $300."
Two problems follow:
One extreme level dominates. Add a very unpopular level (a £500 price, or "no warranty") and the attribute's worst utility drops sharply. That one level inflates the whole attribute's importance, even if real choices happen between two middle levels with similar utilities.
More levels, more importance. Attributes tested with more levels tend to show a wider range than attributes with two levels, all else equal (Wittink, Krishnamurthi & Reibstein, 1990).
What a Shapley decomposition does instead
The Shapley value comes from cooperative game theory (Shapley, 1953): it divides a shared result fairly between the players who produced it. Lipovetsky and Conklin (2001) adapted it to measure the contribution of predictors in regression, and the same idea applies to a choice model, with attributes as the players.
To measure the importance of one attribute, say brand:
Take every possible group of the other attributes: price alone, price and warranty, speed and warranty, none at all, and so on.
For each group, measure how much better the model predicts choice when brand is added.
Average that gain over all the groups, then scale the results so all attributes add up to 100%.
In Ordeen, "how well the model predicts choice" is measured by how sharply it separates the likely choice across a grid of product profiles (an entropy measure).
Why it is a safer summary
It reflects how much an attribute moves choice, across all its levels, rather than letting the two most extreme levels decide.
It adds up properly. Shapley values are the unique way of sharing the total that is fair in a precise mathematical sense, so the contributions sum exactly to the whole.
Importance is still a summary
Even a better importance figure is one number for a whole attribute. Before a specific decision, such as moving the Pro plan from £79 to £99, use the market simulator to see the change in share directly.
Related reading: Methodology · Using a market simulator · Conditional pricing in conjoint
Sources
Lipovetsky, S., & Conklin, M. (2001). Analysis of regression in game theory approach. Applied Stochastic Models in Business and Industry, 17(4), 319–330.
Sawtooth Software, Discover help: CBC utilities and attribute importance (checked October 2026).
Shapley, L. S. (1953). A value for n-person games. In Contributions to the Theory of Games II, 307–317.
Wittink, D. R., Krishnamurthi, L., & Reibstein, D. J. (1990). The effect of differences in the number of attribute levels on conjoint results. Marketing Letters, 1(2), 113–123.
