Hi Dan,
Thanks for this reply - it's particularly helpful as it concentrates
exactly on the point I was trying to make at the beginning of the
thread, and that Federico restated recently.
> Although I didn't reply to Federico, I felt his second point
> ("non-thresholded maps are the only available mean of attempting
> dissociation inference") was potentially very harmful. You can't make
> inferences about dissociations from single unthresholded maps either.
I think this gets right at the key point. Let's consider two regions
A and B, and an activation map for task X.
Thresholded map: A survives thresholding, B does not. Evidence that
A is more activated than B = very near zero.
Unthresholded map: A is in area outlined as significant, B has zero
or negative signal on unthresholded map. Evidence that A is more
activated than B = moderate.
Neither of course constitutes a statistical test of region A vs region
B. It is just that the unthresholded map offers considerably more
evidence than the thresholded map.
Do we agree on that?
See you,
Matthew
> In my view, a better restatement of this second point would be: direct
> statistical comparison is the only available means of testing a
> dissociation between regions; single maps are useful only in
> generating hypotheses. If we're talking about multiple maps (i.e.,
> across studies), then clearly unthresholded maps contain more
> information, although the statement is still false. Unthresholded
> maps contain more information about localization than thresholded
> maps, it's not all or none. At the same time, they facilitate a few
> orders of magnitude more inter-region comparisons.
>
> > > Without at least a power analysis ... one is really in no position to
> > > conclude anything at all about brain regions that show no significant
> > > activation.
> >
> > I don't think there's any way round this one for the single subject
> > analysis; the thresholded map is useless for drawing conclusions about
> > differences between brain areas.
>
> As I'm sure you know by now, I agree with this. What I don't
> understand is why you think unthresholded maps are any different in
> this respect. For a specific inter-region comparison, single maps
> contains one observation per region, so it's not generally possible to
> test the reliability of any observed difference. The sentence of
> Tom's that you quoted is equally applicable to unthresholded maps.
>
> Just for reference, the message you're quoting was in the context of a
> rebuttal of something I posted. I believe Tom interpreted something I
> wrote to suggest that I thought a null hypothesis of no difference
> could be affirmed in sub-threshold regions. In case this clears
> anything up, of course I don't believe so. If it helps, I would be
> happy to go on record as stating that anyone who holds such a belief
> is completely unqualified to do research in this area. I'm not just
> saying this to word my point strongly, I really don't want to see
> reporting requirements shaped by the need to cater to people who don't
> have basic familiarity with the field. (Besides which, obviously if
> you have someone who's inclined to misinterpret data, throwing more
> data at them is not generally helpful.)
>
> > The argument for comparison across studies is simply an extension of
> > the single subject case; if you have almost no evidence for
> > localisation in one study, you merely have very little evidence for
> > localization across 10 studies. If you have the continuous map for
> > 10 studies, and they all show zero or negative activation for area A
> > and strong activation for area B, you have a much stronger case for
> > "A is more involved than B" than you could possibly have from 10
> > thresholded maps, which have simply thrown away this information.
>
> Of course there are some purposes for which unthresholded maps would
> be better, maybe even most purposes. Unthresholded maps obviously
> contain more information than thresholded maps. The reason I didn't
> originally think unthresholded maps should be administratively
> required in favor of thresholded maps in all cases is not because I
> think they are always worse. It's because I think they have both
> advantages and disadvantages, and that their inclusion should be a
> matter of judgment and not one of policy. The most serious
> disadvantage in my view is that they encourage reading deep meaning
> into noise, especially in smoothed data. A second disadvantage is
> that in the cases we're talking about (in which there's no direct
> inter-region comparison reported), it's extraneous to the purpose of
> the report. In my view, a figure should be included in an article if
> it supports some argument that the authors want to make (or that the
> reviewers think they should have wanted to make).
>
> My defense of the thresholded map's role in localization, remember,
> was just a response to your asking if thresholded maps contain
> localizing information. I thought you were arguing that they contain
> no such information, so I thought it would be helpful to provide a
> thought experiment that proves otherwise. This is of course not the
> same thing as claiming that localizing inferences can be drawn from
> single maps. In any case, my argument about why I thought thresholded
> maps might be more useful was mostly academic. In practical terms,
> when it comes to localization, I think we're talking about two levels
> of useless. To the extent the argument for unthresholded maps depends
> on their utility for localization, I think it would be a harmful
> practice.
>
> dan
>
|