Normalising covariance matrices

Neil3030

I want to compare two groups on the size of their error ellipses in performing a precision judgement.

The 2 dimensional precision allows me compute covariance matrices for each subject, which i then average for each group, and from these group averaged cov matrices I can extract the eigen values and vectors to construct the ellipses.

However I am concerned that the subjects with worse precision are disproportionately dominanting the mean values, which is leading me towards a normalisation procedure for each subject's cov mat before I average.

For this procedure I was thinking to dot divide each cell of each subject's cov mat with the determinant of their cov mat. Would this be appropriate?

alan4cult

Could you work with the correlation matrix instead?

Neil3030

I don't think so. Scaling all the variances to 1 would result in the analysis losing the ability to differentiate subjects on their precision.