[pymvpa] Why to standardize (z-score) a time course?

Yaroslav Halchenko debian at onerussian.com
Sat Jul 31 02:10:58 UTC 2010

primary reason for zscoring is usually to bring data into a reasonable
range of values for a classifier to be sensitive...  usually it implies
that values are primarily around 0 with values from -10 to 10 or so with
the bulk from -2 to 2... z-scoring against baseline also has direct
interpretation of cause

in your case, if they are already betas from GLM -- they might be fine
already... iirc betas are though not scaled by the level of unexplained
variance (i.e. noise) so their values might be somewhat wild, so for
classification it might still be worth z-scoring or any other reasonable
scaling approach

for correlation analysis situation is quite different though, especially
if doing gross z-scoring (instead of z-scoring against baseline).  e.g. if
you within your patterns have some prominent preference, for instance,
anterior higher than posterior for one category,  z-scoring might
damage that effect heavily or obliterate it entirely.

Summary: z-score or not depends on what you want from your analysis and
what is your hypothesis ;-)

On Fri, 30 Jul 2010, Vadim Axel wrote:

>    Hi,
>    When I run classification on raw fMRI data I always make a z-score
>    (subtract the mean and divide by std) for each voxel / scan.
>    Now I tried some sort of correlation analysis, while I run a
>    correlation for beta images (similar to Haxby 2001). So, the z-score is
>    going to be across beta_per_condition values (several values only). The
>    question is: does is make sense to zcore the beta results?
>    I ran some simulations while I add a noise to each voxel. So, I know
>    that there is a real correlation in my data. Whereas without zscore the
>    results indeed show high correlation, after zscoring procedure the
>    correlation is virtually disappeared. I also tried just to subtract the
>    mean across conditions, but the correlations still do not look good.
>    What is the correct way to do run this analysis?
=------------------------------   /v\  ----------------------------=
Keep in touch                    // \\     (yoh@|www.)onerussian.com
Yaroslav Halchenko              /(   )\               ICQ#: 60653192
                   Linux User    ^^-^^    [175555]

More information about the Pkg-ExpPsy-PyMVPA mailing list