Statistics Glossary → Covariance and correlation matrix
Covariance and correlation matrixconcept
A square table holding the variances of several variables on its diagonal and the covariance (or correlation) of each pair of variables off the diagonal.
For p variables the covariance matrix is p × p; each diagonal element is a variable's variance, the off-diagonal (i, j) element is the covariance of variables i and j, and the matrix is symmetric. The correlation matrix is obtained by dividing each covariance by the product of the two standard deviations; its diagonal consists of 1s and it does not depend on units of measurement. Most multivariate methods (MANOVA, discriminant analysis, principal components, factor analysis, structural equation modeling, the Mahalanobis distance) work on this matrix.
If principal component analysis is run on the covariance matrix, variables with large variances dominate the components; for variables measured in different units the correlation matrix (standardized data) is used. For the matrix to be invertible, no variable may be an exact linear combination of the others, and the number of observations must be sufficiently larger than the number of variables; a determinant very close to zero signals multicollinearity. Equality of covariance matrices across groups (Box's M test) is an assumption of MANOVA and linear discriminant analysis; when it fails, quadratic discriminant analysis uses a separate matrix for each group.
An everyday example
A sports club measures its players' height, weight and jump height. In the covariance matrix the height-weight covariance is 45 (cm·kg) and the height-jump covariance 12 (cm·cm); because the units differ these numbers cannot be compared directly. Switching to the correlation matrix gives r = 0.70 for height-weight and r = 0.25 for height-jump, making the strength of the relationships comparable.
In MerQur
In discriminant analysis 'Additional outputs: Covariance matrices (pooled, group, total)' prints these matrices; LDA builds its boundary with the pooled covariance and QDA with the per-class covariance. Principal component analysis offers 'Covariance-matrix PCA', Cronbach's alpha analysis 'Additional outputs: Inter-item correlation/covariance matrix + summary', and exploratory factor analysis 'Additional output: Correlation matrix, significance, determinant, anti-image'; variable clustering works on the correlation matrix. In MANOVA and mixed-design ANOVA 'Box M' tests the equality of the group covariance matrices, and in mixed-design ANOVA the sphericity ε values are computed from the pooled within-group covariance; in ANCOVA the SE of adjusted mean differences comes from the full covariance matrix, and in GEE the QIC uses the robust covariance.
Analyses where this concept appears
Mixed-Design ANOVA, Split-PlotMANOVAANCOVAVariable Clustering (VARCLUS)Generalized Estimating Equations (GEE)Discriminant Analysis (LDA / QDA)Principal Component Analysis (PCA)Cronbach's Alpha and Reliability AnalysisExploratory Factor Analysis (EFA)
Related concepts
Box's M test (equality of covariance matrices)Eigenvalue and eigenvector