Description
Model a respondent's full ranking of options as a sequence of logit choices, first from all options and then from those remaining, to estimate attribute utilities from ranking questions.
When respondents rank alternatives rather than choose one, the probability of the full ranking under logit is the product of successive choices: the top option chosen from all, the second from those left, and so on. Each ranking of J alternatives is 'exploded' into J - 1 pseudo-observations, with already-ranked options removed from each later choice set, and estimated with standard logit software. This uses all the information in a ranking rather than only the first choice. Because a person's pseudo-observations share their tastes, Train recommends a mixed logit estimated as panel data; with probit, rankings become utility-difference inequalities simulated by GHK. The product form depends on independent extreme-value errors, so lower ranks, which respondents give less carefully, can bias estimates if treated as equally informative.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Goal
- Prioritise, Measure attitudes
- Stage
- Strategy, Definition
- Data
- Quantitative
- Effort
- Days
Steps
- Collect rankings of alternatives described by attributes, or a choose-then-choose-again sequence.
- Explode each ranking of J alternatives into J - 1 pseudo-observations, repeating the explanatory variables.
- For each pseudo-observation remove alternatives already ranked higher and code the next-ranked one as chosen.
- Estimate an ordinary conditional logit on the pseudo-observations.
- Prefer a mixed logit that treats each person's pseudo-observations as a panel so their correlation is modelled.
- Check whether lower ranks behave consistently with top ranks, for example by fitting the top two ranks only and comparing.
Sources
- Discrete Choice Methods with Simulation (Train, 2009)
Also known as
- Rank-ordered (exploded) logit for ranking questions
- Exploded logit
- Rank-ordered logit
- Ranked-data logit
- Plackett-Luce model
Categories
- Analytics & Behaviour
- Statistics & Modelling