Computing Partial Correlations from the Data Matrix.

Abstract : The usual way of computing partial correlations is based on the formation of the covariance matrix, that amounts to squaring the data matrix, thus inviting a potential loss of numerical accuracy. This paper recommends the determination of partial correlations from the data matrix: the QR decomposition of the data matrix is computed and plane rotations are applied to the resulting upper triangular matrix, which is the Cholesky factor of the covariance matrix. It is shown that if rotations are applied to the triangular matrix so as to leave the number of its zero entries invariant, the sines of the rotation angles are partial correlations. Different ways of organizing the computations are presented for extracting any set of partial correlations.