Description
Build a tree from similarity ratings, fit branch lengths to paired-choice data, and test whether it beats the constant-ratio model, capturing that similar options compete more with each other.
A preference tree is elimination by aspects restricted to aspects that nest in a tree, which cuts the parameters to at most 2n - 2. It captures the similarity effect that simple logit or constant-ratio models miss: adding an option takes share mainly from options similar to it. A trinary inequality on paired-choice odds tests the tree structure; across celebrities, political parties and academic disciplines, 78 to 96 percent of triples met it, and the tree fitted far better than the constant-ratio model. Trees built from similarity ratings and from choices largely agreed. The root does not matter for similarity but does for preference, so alternative roots are compared. The method gives a transparent, testable model of substitution among options.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Extinct
- Stage
- Strategy, Definition
- Data
- Quantitative
- Effort
- Weeks
Steps
- Collect pairwise similarity ratings and paired choices (and, if possible, choices from triples) for the same set of options.
- Build a tree from the similarity data, for example with an additive tree algorithm such as ADDTREE.
- Fit branch lengths to choice probabilities by maximum likelihood, trying alternative roots.
- Compare fit with the constant-ratio (logit) model using a likelihood-ratio test.
- Report the share of triples that meet the trinary inequality.
- Use the fitted tree to predict shares when options are added or removed.
Sources
- Preference, Belief, and Similarity (Tversky, 2004)
- Tversky, A., & Sattath, S. (1979). Preference trees. Psychological Review, 86
Also known as
- Pretree analysis
- Preference tree fitting
- Hierarchical elimination model
Categories
- Analytics & Behaviour
- Statistics & Modelling