[pymvpa] Q: Time-offset of category labels in Haxby 2001 dataset. Was this ever properly resolved?

[Rajeev Raizada]

> Date: Wed, 9 Mar 2011 10:19:17 -0500
> From: Michael Hanke <michael.hanke at gmail.com>
> Subject: Re: [pymvpa] Q: Time-offset of category labels in Haxby 2001
>        dataset. Was this ever properly resolved?
>
> To add another tiny bit of evidence for a shift you could take a look at
> the peristimulus plot at http://www.pymvpa.org/tutorial_eventrelated.html
>
> The block onset as encoded in the label vector is the t=2 volume/timepoint in
> the plot. You'll see that this volume already reflects a deviation from
> baseline for both plotted conditions. Therefore, I'd say that, on
> average, the labels are shifted by one volume -- considering all
> temporal uncertainty introduced by the timing of the study that Yarik
> mentioned before.

Dear Yarik and Michael,

Many thanks indeed for your replies on this.
Although the exact labelings from 2001 may be lost in the mists of time,
the two facts that the study uses long-duration blocks
and that the block- and rest-durations are not multiplies of the 2.5s TR
both imply that even approximate labelings should be good enough.

Yarik's description of how he did the data-conversion,
and Michael's pointer to the event-related tutorial plot,
are both consistent with the conclusion that the labels
are probably already offset by one TR,
meaning that they already reflect stim-evoked neural responses,
and that they don't require any further time-offsetting.

In further confirmation of that, my experience is that
using the labels as they are, without any additional offsetting,
to select the neural responses assigned to each category
yields classification results that look very reasonable.

Although it would be nice in principle to have a more iron-clad answer,
the conclusion that the labels are "good enough, and seem to be correct"
is really not that bad, especially given that it is a block design.

It would definitely be worth adding to the docs and README.txt
a statement saying that the labels do not require any additional time-offset,
such that any TR with the label, e.g. "shoes", attached to it
can be understood as containing evoked-HRF responses to shoes.
(Just saying "the labels do not require any additional time-offset",
but without any further explanation, might not be clear to readers
who have not seen this discussion in the list-archives).

Another piece of info that can be gleaned from this list's archives
but which is also currently not in the docs is that the 9th run of subj5
contains only the label "rest", due to excessive head motion on that run.

Thanks again for putting this dataset online,
and for your input on helping to clarify the labels issue.

Raj