[pymvpa] group-level analysis differences vs. pattern classification

Yaroslav Halchenko debian at onerussian.com
Mon Sep 23 13:40:50 UTC 2013

On Mon, 23 Sep 2013, Vadim Axel wrote:

>    Hi,
>    There are two commonly used approaches to analyze the data of the
>    experiment below:
>    Simple design with two conditions (A and B), which �both activate large
>    network of well established regions (e.g. conjunction analysis of A >
>    baseline and B > baseline). The question is whether we can find neural
>    correlates of difference between the two. 

the answer I guess is: yes --  we should be able to if "conditions are
right" (power, etc)

> Direct group-level analysis
>    comparison between A and B results in small activations (~5% of volume
>    comparing to commonality of conjunction analysis) and these activations
>    are located mostly outside the main network, all over the brain.

Remembering that statistics is there only to help us to support/reject
our hypotheses, not really to be treated as "the ground truth", you
might have set up your analysis to include only the "differential"
activations which are within the main network, since that is where you
believe activity or relevance is.

> Given
>    that the result is dependent on p-value threshold, it looks like a
>    classical blobology. �Another approach is to select (independently) the
>    ROIs of the common network nodes and to run MVPA. 

or even run MVPA on full brain happen you data has enough power to cope
with such large initial feature space.

> With this analysis I
>    successfully discriminate between the two conditions. So, two people
>    analyzing the same data can draw absolutely different conclusions: one
>    would say, that small regions X, Y, Z are the regions, which discriminate
>    between conditions A and B;

which given your results above would be sensible conclusion imho besides
that I would have clarified that this set of regions is not necessarily
exhaustive (thus "the regions" statement might be a bit too strong)

>  the other, in contrast, would say that since
>    both A and B activate common network

depending on what is implied by "activate common network" I might argue
because it would be hard (if not impossible) to prove null hypothesis
here that the network is the same for both A and B.

> , the difference between the two lies
>    within this network (different patterns of activity). �

>    What approach is more reliable in your opinion?

as I stated above (if I got the question right), the 2nd approach would
require (much) more analysis to support itself.

Yaroslav O. Halchenko, Ph.D.
http://neuro.debian.net http://www.pymvpa.org http://www.fail2ban.org
Senior Research Associate,     Psychological and Brain Sciences Dept.
Dartmouth College, 419 Moore Hall, Hinman Box 6207, Hanover, NH 03755
Phone: +1 (603) 646-9834                       Fax: +1 (603) 646-1419
WWW:   http://www.linkedin.com/in/yarik        

More information about the Pkg-ExpPsy-PyMVPA mailing list