Description
Before trusting an MDS map of similarity or card-sort data, computes how concentrated and how mutual nearest-neighbour links are; values above geometric bounds show the map misleads.
For each item, find its nearest neighbour in the raw data (highest similarity, most co-occurrence or most confusions). Centrality (C) is the mean of the squared counts of how often each item is someone's nearest neighbour; reciprocity (R) is the mean rank of each item in its own neighbour's ordering. Geometry constrains both: in two dimensions no more than five items can share a nearest neighbour, and for points sampled from a continuous space both statistics stay below about 2. Across 100 published data sets, perceptual stimuli mostly stayed within bounds, but conceptual sets including a category name or prototypical member reached centrality of 15.7 while their 2D maps showed about 1.8. Trees tracked the statistics; 3D maps did not. So high C and R in the data indicate hub structure that spatial maps cannot represent, and a low C to R ratio points instead to outliers.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Design, Evaluation
- Data
- Analytical / expert
- Effort
- Hours
Steps
- For each item, identify its nearest neighbour in the data, with a stated rule for ties.
- Compute centrality C as the mean of squared nearest-neighbour counts across items, and reciprocity R as the mean rank of each item in its neighbour's similarity order.
- Compute the same statistics from the distances in the fitted MDS map.
- If the data exceed about 2 and the map does not, prefer a tree, cluster or hybrid representation over the spatial map.
- Name the hub items and check whether they are category labels, umbrella features or prototypical members.
Sources
- Preference, Belief, and Similarity (Tversky, 2004)
- Foundations of Measurement, Vol. 2: Geometrical, Threshold, and Probabilistic Representations (Suppes et al., 1989)
Also known as
- Nearest neighbour analysis
- C and R statistics
- Hub detection for perceptual maps
Categories
- Analytics & Behaviour
- Statistics & Modelling