Multidimensional Scaling (MDS)
Multidimensional Scaling · Also known as: MDS, metric MDS, non-metric MDS, proximity scaling
Multidimensional scaling maps objects described only by pairwise similarities or dissimilarities into a low-dimensional geometric space so that distances in that space reflect the original proximity structure as faithfully as possible. It is widely used to visualize the hidden structure of psychological, social, and behavioral data.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
+2 more
When to use it
Use MDS when your data take the form of pairwise proximities — similarity ratings, confusion frequencies, correlation matrices, or any symmetric matrix of distances — rather than multivariate measurements on variables. It is the natural choice for perceptual mapping, brand positioning, and judged similarity tasks where respondents cannot rate objects on named attributes but can compare pairs. Non-metric MDS should be preferred when the proximities are only ordinal (ranks). Do NOT use MDS when you have raw multivariate data and want to explain variance (use PCA or EFA instead), when group membership must be incorporated in the geometry (use discriminant analysis), or when the number of objects is very small (fewer than about 10) because the solution becomes trivially perfect with few constraints. MDS also requires adequate pairwise coverage; if many pairs have missing similarity judgments, the configuration becomes unreliable.
Strengths & limitations
- Works directly on pairwise proximity data without requiring raw variable-level measurements.
- Non-metric MDS is robust to the measurement scale of the proximities — it needs only their rank order.
- Produces an intuitive visual map that reveals clusters, gradients, and outliers in similarity structure.
- Makes no distributional assumptions; it is a purely geometric, optimization-based procedure.
- Applicable across very diverse domains: psychophysics, marketing, linguistics, ecology, and network analysis.
- The number of dimensions k must be chosen by the analyst; there is no formal significance test for dimensionality.
- The axes of the solution are arbitrary up to rotation, reflection, and dilation, requiring post-hoc substantive interpretation.
- Local minima in the Stress optimization mean different starting configurations can yield different solutions, especially with few objects or many dimensions.
- Degeneracy can occur in non-metric MDS, particularly with many tied proximity values, producing artificially low Stress.
- Large n (many objects) make it computationally intensive and the resulting map visually cluttered and hard to interpret.
Frequently asked
What is the difference between metric and non-metric MDS?
Metric MDS treats the input proximities as interval- or ratio-scale values and tries to make Euclidean distances proportional to them. Non-metric MDS uses only the rank order of the proximities, fitting a monotone transformation to them rather than a linear one. Non-metric MDS is more flexible and appropriate when similarity judgments are ordinal or when the exact numeric values of the proximities are uncertain.
How many dimensions should I use?
Plot Stress against the number of dimensions and look for an elbow — the point where adding another dimension no longer produces a meaningful drop. Two dimensions are preferred for visual presentation. A Stress value below 0.10 in the chosen dimensionality is a conventional threshold for an acceptable fit.
How is MDS different from PCA or factor analysis?
PCA and EFA start from a data matrix of observations-by-variables and decompose its covariance or correlation structure. MDS starts from a proximity matrix — pairwise similarities or distances — and finds a coordinate configuration that reproduces those proximities geometrically. If you have raw multivariate data, PCA/EFA are more appropriate; if you have only pairwise comparisons, MDS is the natural tool.
What Stress value indicates an acceptable solution?
Kruskal's guidelines: Stress below 0.05 is excellent, 0.05–0.10 is good, 0.10–0.15 is fair, 0.15–0.20 is poor, and above 0.20 is unacceptable. However, these thresholds depend on n and k; with many objects even a moderate Stress can represent a good fit relative to chance.
Can MDS handle missing pairwise similarities?
Standard MDS algorithms assume a complete proximity matrix. Missing values can be handled via imputation before analysis or by using specialized algorithms (such as constrained or partial MDS), but extensive missingness makes the recovered configuration unreliable and should be avoided.
Sources
- Kruskal, J. B. (1964). Multidimensional scaling by optimizing goodness of fit to a nonmetric hypothesis. Psychometrika, 29(1), 1–27. DOI: 10.1007/BF02289565 ↗
- Cox, T. F. & Cox, M. A. A. (2001). Multidimensional Scaling (2nd ed.). Chapman & Hall/CRC. ISBN: 978-1584880943
How to cite this page
ScholarGate. (2026, June 3). Multidimensional Scaling. ScholarGate. https://scholargate.app/en/statistics/multidimensional-scaling
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Cluster AnalysisStatistics↔ compare
- Correspondence AnalysisStatistics↔ compare
- Discriminant AnalysisStatistics↔ compare
- EFAStatistics↔ compare
- Latent Class AnalysisStatistics↔ compare