MerQuR+Statistics Glossary

Statistics Glossary → Metric and non-metric MDS (SMACOF)

Metric and non-metric MDS (SMACOF)concept

The two forms of multidimensional scaling, which displays distances or dissimilarities between objects on a low-dimensional map, preserving either the distances themselves (metric) or only their rank order (non-metric).

Multidimensional scaling (MDS) places objects in a two- or three-dimensional space so that the distances on the map match the given dissimilarities as closely as possible. Metric MDS requires the map distances to be proportional (or linearly related) to the dissimilarities. Non-metric (Kruskal) MDS preserves only the ordering: two more dissimilar objects must be further apart on the map, but the size of the difference does not matter; this is achieved with monotone regression. For data such as ranked similarity judgements, non-metric MDS is more appropriate.

SMACOF (de Leeuw) is an algorithm that reduces the stress criterion step by step and can be used for both forms. Fit is judged by stress: for Kruskal's stress-1, roughly below 0.05 is good, around 0.1 fair and above 0.2 poor, though the expected stress changes with the number of dimensions and objects. Because the solution can get stuck in a local optimum, several starting configurations are tried. The axes have no intrinsic meaning; the map can be rotated and reflected.

An everyday example

A group of people is asked how similar each pair of ten fruits is. Non-metric MDS uses only rankings such as 'an apple is more like a pear than a cherry is like a watermelon' to place the fruits on a map; the citrus fruits gather in one corner and the stone fruits in another.

In MerQur

The 'Metric MDS (preserves distances)' option of Multidimensional Scaling: when checked, SMACOF is applied on a ratio scale; when unchecked, only the order is preserved (non-metric, monotone MDS).

Analyses where this concept appears

Multidimensional Scaling (MDS)

Related concepts

Stress, fit measures and the Shepard diagram (MDS)Distance measure (Euclidean, Manhattan, Gower, great-circle)