PCOMPS Subcommand (MANOVA: Multivariate command)

PCOMPS requests a principal components analysis of each error matrix in a multivariate analysis. You can display the principal components of the error correlation matrix, the error variance-covariance matrix, or both. These principal components are corrected for differences due to the factors and covariates in the MANOVA analysis. They tend to be more useful than principal components extracted from the raw correlation or covariance matrix when there are significant group differences between the levels of the factors or when a significant amount of error variance is accounted for by the covariates. You can specify any of the keywords listed below on PCOMPS.

COR. Principal components analysis of the error correlation matrix.

COV. Principal components analysis of the error variance-covariance matrix.

ROTATE. Rotate the principal components solution. By default, no rotation is performed. Specify a rotation type (either VARIMAX, EQUAMAX, or QUARTIMAX) in parentheses after the keyword ROTATE. To cancel a rotation specified for a previous design, enter NOROTATE in the parentheses after ROTATE.

NCOMP(n). The number of principal components to rotate. Specify a number in parentheses. The default is the number of dependent variables.

MINEIGEN(n). The minimum eigenvalue for principal component extraction. Specify a cutoff value in parentheses. Components with eigenvalues below the cutoff will not be retained in the solution. The default is 0; all components (or the number specified on NCOMP) are extracted.

ALL. COR, COV, and ROTATE.

  • You must specify either COR or COV (or both). Otherwise, MANOVA will not produce any principal components.
  • Both NCOMP and MINEIGEN limit the number of components that are rotated.
  • If the number specified on NCOMP is less than two, two components are rotated provided that at least two components have eigenvalues greater than any value specified on MINEIGEN.
  • Principal components analysis is computationally expensive if the number of dependent variables is large.