[pymvpa] Help

[Christopher Markiewicz]

Hi Jorge,

Yarik and Nick are definitely going to be better resources than I will on
this issue, but it occurs to me to ask whether your input datasets consist
of a single beta per condition (per run), or a beta per event? I've found
that classification accuracies are much higher than chance across the whole
brain when using a single beta per condition, while a beta estimate per
event tends to give a distribution (across voxels) with a mean very close
to chance. I'm not familiar with the hyperalignment algorithm, but, if it
doesn't account for the spatial distribution of classification accuracies,
you might need to add a correction step to adjust the distribution to fit
the expected distribution.

Chris

On Mon, Nov 6, 2017 at 12:03 PM, Jorge H <jhevia at hotmail.com> wrote:

> Thank you for the fast answer back!
>
>
> When I say very high accuracies I mean .95 for brain areas related with
> the social task we applied (for example, rTPJ) and .95 for brain areas not
> related with my task (for example, primary motor areas). On the other hand,
> in the run that we are using for the cross-validation, we have the same
> number of targets for each of the conditions. Finally, I will randomise the
> targets in different positions in the run. I will let you know as soon as I
> get the results on next days.
>
>
> I just want to make sure that I'm not getting potentially "fake news"!
>
>
> Thanks a lot.
>
>
> Jorge
>
>
>
> *Dr. Jorge Carlos Hevia Orozco*
> *Unidad de Analisis de Imagenes, Instituto de Neurobiologia, UNAM campus
> Juriquilla*
>
>
> ------------------------------
> *De:* Pkg-ExpPsy-PyMVPA <pkg-exppsy-pymvpa-bounces+jhevia=
> hotmail.com at lists.alioth.debian.org> en nombre de Nick Oosterhof <
> n.n.oosterhof at googlemail.com>
> *Enviado:* lunes, 6 de noviembre de 2017 10:21 a. m.
> *Para:* Development and support of PyMVPA
> *Asunto:* Re: [pymvpa] Help
>
> Greetings,
>
> On 6 November 2017 at 16:52, Jorge H <jhevia at hotmail.com> wrote:
>
> Hi my name is Jorge and I'm running a searchlight hyperanalignment however
> when I run the analysis using anatomical masks of brain areas completely
> not related with the task, I got very high accuracies! I would appreciate
> if somebody give a quick look to my script and let me know if detect
> something wrong.
>
> At first sight I didn't see anything wrong. Things to consider:
>
> - what do you mean by 'very high accuracies' ?
> - is your design balanced, i.e. an equal number of samples for each
> combination of target and chunks?
> - if you randomise the targets manually (so that there should be no
> information about sample conditions), do you still observe 'very high
> accuracies'?
>
>
>
>
>
>
> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20171106/5d2c29b0/attachment.html>