Description
Rotate one set of factor loadings or map coordinates to match a target as closely as possible, and use the remaining misfit to judge how similar two solutions are.
Factor, component and multidimensional scaling solutions are defined only up to rotation, so two solutions can look different while describing the same structure. Procrustes analysis finds the transformation T that brings matrix A as close as possible to target B in least squares. With T constrained to be orthonormal, Schonemann's solution uses the singular value decomposition of A'B; centring and scaling may be added. The residual sum of squares, and congruence coefficients per dimension, measure how similar the solutions are once arbitrary rotation is removed. Generalised Procrustes analysis fits several configurations to a common consensus at once, and oblique and maximum-congruence variants exist. Jackson uses it to check whether simplified rounded vectors capture a real solution; Krzanowski used it to choose variable subsets. The main caution is that flexible transformations can make unrelated matrices fit, and no standard significance test exists, so permutation baselines help.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Siloed
- Stage
- Discovery, Evaluation
- Data
- Analytical / expert
- Effort
- Hours
Steps
- Obtain the two configurations to compare (loadings, component scores or MDS coordinates) with matched rows.
- Choose the target: a hypothesised pattern, the original-language solution, or the full-variable solution.
- Centre, and if wanted scale, both matrices, then compute the orthonormal rotation from the SVD of A'B.
- Apply the rotation and compute the residual sum of squares and the congruence coefficient for each dimension.
- Judge misfit against a baseline, for example by repeating the fit with randomly permuted rows.
- Use generalised Procrustes for more than two configurations, or oblique versions when dimensions may correlate.
Sources
- A User's Guide to Principal Components (Jackson, 1991)
- Fractals, Random Shapes and Point Fields (Stoyan & Stoyan, 1994)
Also known as
- Orthogonal Procrustes analysis
- Target rotation
- Generalised Procrustes analysis
- Maximum congruence rotation
- Oblique Procrustes
- Procrustes rotation
Categories
- Analytics & Behaviour
- Statistics & Modelling