 |
 |
Hi Gaël, Not sure if you'll accept this as a clear mathematical reason (one might validly call it hand-waving) but here is what I think makes this important: In seed-based correlation analysis the quantity of interest (i.e. the quantity that is used to define positive findings) is based on correlation which itself is just a re-scaled covariance which in turn is a second order statistics. SO, given that second-order statistics are implicitly what you're after all pre-processing (such as the process used to come up with a time series that end up being correlated) should preserve second-order statistics in as much as possible. Simple Gedankenexperiment: imaging two voxels with identical but inverted time course. The mean is flat and does not represent either one of the time series whereas the Eigen-timeseries will correctly express all the variance in the data. The mean is a first-order statistics and in this (extreme) example generates a vector that does not express variance at all, so any later correlation analysis becomes meaningless. A common argument for the mean is that noise ect will hopefully average out so you end up generating a 'cleaner ' or 'purer' signal. This 'de-noising' in my mind should be done by explicitly using confound regressors, rather than expecting this to happen implicitly during averaging by magic... both mean and Eigentiemseries are linear combinations of the original data within the region, the only difference being that instead of simply weighing all voxels equally the SVD weighs these so that variance is preserved... And here's the killer argument: the Eigenvector is what you get when using SPM and the ROI tool (I think) so it must be right... ;) Hope that makes sense. cheers Christian On 13 Aug 2009, at 10:28, Gael Varoquaux wrote: > On Thu, Aug 13, 2009 at 10:25:22AM +0100, Christian F. Beckmann wrote: >> Hi >> >> Yes, the description of the method is very inclusive. Only issue I >> have >> is that I'd advice using the Eigen-timeseries rather than the mean >> timeseries within the ROIs for further analysis > > Hi Christian, > > That's an interesting statement. I share your opinion on this > technical > point, but I have not been able to pin point a clear mathematical > reason > behind this. Do you have any justification? > > Thanks, > > Gaël _______________________________________________ Christian F. Beckmann, DPhil Senior Lecturer, Clinical Neuroscience Department Division of Neuroscience and Mental Health Imperial College London, Hammersmith Campus Rm 419, Burlington Danes Bldg, Du Cane Road, London W12 0NN, UK Tel.: +44 (0)20 7594 6685 --- Fax: +44 (0)20 7594 6548 Email: [log in to unmask] http://www.imperial.ac.uk/medicine/people/c.beckmann/ Senior Research Fellow, FMRIB Centre University of Oxford JR Hospital - Oxford OX3 9DU Tel.: +44 (0)1865 222551 --- Fax: +44 (0)1865 222717 Email: [log in to unmask] http://www.fmrib.ox.ac.uk/~beckmann