Ordinal Tests of Spatial Similarity Model Assumptions

Description

Use factorial stimuli and pairwise dissimilarity judgements to check, pattern by pattern, whether perceived similarity behaves like distance in a dimensional space before trusting an MDS map.

Multidimensional scaling will fit any symmetric dissimilarity data in enough dimensions, so a good fit does not show that people perceive the items spatially. These tests fix the stimulus dimensions in advance through a factorial design and check ordinal properties that any distance model must satisfy. Relative curvature asks whether equal adjacent steps on two attributes give equal end-to-end differences; translation invariance asks whether a difference on one attribute is judged the same at every level of the other; the corner inequality, derived from the triangle inequality, asks whether the route through a corner of a 3 x 3 layout is never shorter than the route through the centre. Counting satisfying and violating patterns per respondent shows whether a Euclidean, city-block or feature model is warranted, and whether attributes are integral or separable.

Attributes

Loss type
Unmigrated
Stage
Discovery, Design
Effort
Days

Steps

  1. Build stimuli from a factorial design on attributes chosen in advance, for example icon size by colour saturation.
  2. Collect dissimilarity or similarity judgements for all pairs from each respondent (or confusion or classification data).
  3. Test relative curvature: for triples varying on one attribute, find cases where both adjacent steps on attribute 1 are judged at least as large as on attribute 2, and check the end-to-end comparison agrees.
  4. Test translation invariance: compare the same difference on one attribute at different shared levels of the other.
  5. Test the corner inequality in each 3 x 3 sub-design, classifying configurations as corner-shorter, centre-shorter or uninformative.
  6. Tabulate satisfying and violating patterns per respondent and choose a spatial metric, a feature model or separate treatment of attributes accordingly.

Sources

  • Foundations of Measurement, Vol. 2: Geometrical, Threshold, and Probabilistic Representations (Suppes et al., 1989)

Also known as

  • Axiom tests of multidimensional scaling
  • Corner inequality test
  • Relative curvature test
  • Integral versus separable attribute test
  • Ordinal tests of the assumptions of spatial similarity models

Categories

  • User Research
  • Cognitive Psychology & Perception