[pymvpa] Is global intensity scaling a problem?

Michael Hanke mih at debian.org
Wed Dec 21 08:28:47 UTC 2011


Hi Michael,

On Tue, Dec 20, 2011 at 10:25:05PM -0800, Michael Waskom wrote:
> Hi Yarick et al.
> 
> In a typical FSL preprocessing pipeline, the timeseries images
> are globally scaled such that the median of all four dimensions (within the
> brain mask) is the same number across scans (usually 10000). To minimize
> duplicated data, it would be nice to take the output of my preprocessing
> and use it for both univariate and multivariate analyses, but I'm wondering
> if this scaling step will introduce any bias. Is this a valid concern? And
> does the answer change if I'm using a leave-one-run-out cross validation
> approach?

I think the typical FSL preprocessing stream is not a problem. As long
as the de-meaning is done using the grand-mean (through time) and you do
not enable the per-volume intensity normalization. Most of the time you
would probably want to run further preprocessing for the multivariate
analysis stream, e.g. scaling features into the same range etc.

Michael

-- 
Michael Hanke
http://mih.voxindeserto.de



More information about the Pkg-ExpPsy-PyMVPA mailing list