We investigate the Oja sign covariance matrix (Oja SCM) for estimating partial correlations in multivariate data. The Oja SCM estimates directly a multiple of the precision matrix and is based on the concept of Oja signs, a multivariate generalisation of the univariate sign function, which obey some form of affine equivariance property. Simulations show that the asymptotic distribution gives a good approximation of the exact finite-sample distribution already for samples of moderate size. We find it to outperform the classical sample partial correlation in case of heavy-tailed distributions. The high computational costs are its main disadvantage.
|Title of host publication||Compstat 2008: Proceedings in Computational Statistics. Vol. II.|
|Subtitle of host publication||Proceedings in Computational Statistics|
|Number of pages||9|
|Publication status||Published - 2008|