Ordinal Tests of Conjoint Measurement Axioms

Description

Check from the rank order of a factorial design's cells whether factors can be scaled to combine additively or by another simple rule, and if so build interval scales.

Conjoint analysis assumes that part-worths add; functional measurement tests that assumption numerically through parallel lines. Conjoint measurement asks the prior question using order information only: does the ranking of all factor-level combinations admit any additive representation? Krantz, Luce, Suppes and Tversky show that it does if the ordering is a weak order and satisfies independence (each factor's levels keep their order whatever the other factors are) and double cancellation, with joint independence for three or more factors. Independence alone is not enough: a design can pass it and still fail double cancellation. Patterns of reversed or flat orders (sign dependence) diagnose whether a multiplicative, distributive or dual-distributive rule fits instead. When axioms hold, scales are fitted by solving the implied inequalities, minimising inversions, or by a monotone transformation that minimises interaction (Kruskal's MONANOVA).

Attributes

Loss type
Unmigrated
Effort
Days

Steps

  1. Use a full factorial design in which every level of each factor is combined with every level of the others.
  2. Obtain an ordering of all cells from ranks, choices, median prices or mean ratings.
  3. Check transitivity, then independence: each factor's level order must be the same at every level of the other factors; a clear crossing rules out additivity.
  4. Check double cancellation in every 3 x 3 sub-matrix (if ax >= fq and fp >= bx then ap >= bq) and, with three or more factors, joint independence.
  5. If orders reverse or flatten across another factor's levels, tabulate sign dependence and identify which simple polynomial (multiplicative, distributive, dual distributive) is compatible, using three-factor tests to separate them.
  6. Fit scales by solving the inequalities; with fallible data minimise inversions (maximise Kendall's tau) by linear programming or monotonically transform cell means to minimise interaction, then test the remaining interaction with adequate power.

Sources

  • Foundations of Measurement, Vol. 1: Additive and Polynomial Representations (Krantz et al., 1971)

Also known as

  • Additive conjoint measurement
  • Axiomatic conjoint measurement
  • Cancellation axiom tests
  • Double cancellation test
  • Thomsen condition
  • Polynomial conjoint measurement
  • Sign-dependence diagnosis
  • Testing additivity of ratings
  • Minimum-inversion additive scaling

Categories

  • Analytics & Behaviour
  • Survey & Psychometrics