Dear Rik, Elizabeth, List,
From her email, I assumed Elizabeth wants to model visual stimuli of different
durations (and therefore varying visually evoked neural input functions)
correctly, since apparently the visual stimuli disappeared when subjects gave
a response, when I understand her email correctly. As you also mentioned, and
I also wanted to point out to her, using derivatives might not be the best
way to model this, since you do actually know the durations. And wouldn't
indeed longer visual stimuli evoke larger HRFS (assuming longer stimuli evoke
longer neural activity, see point below)? They would when you believe that
neural input convolved with HRF is a proper way to calculate expected BOLD
signals.
I do agree with you that interpreting the results is tricky and depends on
your experimental questions. But when indeed Elizabeths visual stimuli
disappear when a subject presses a button, I would think modeling variable
durations is a good way of explaining the signal in visual areas as well!
Indeed, adding derivatives as well would catch additional variations in HRF
timing/duration that is not explained by the stimulus durations.
Maybe a more problematic point here is the neurophysiology of visual neurons.
When I remember well visual neurons show vigorous bursts when a stimulus
appears, and only a lower sustained activity level during the time the
stimulus is present (any visual neurophysiologist around to affirm this? I am
not a visual neuroscientist). This low sustained activation might be
irrelevent from a hemodynamic point of view, especially compared to the
onset-burst. That might make modeling the stimuli as events with duration 0
a much better option after all. Of course BOLD does not necessarily reflect
neural output but merely neural input. But what goes up, must come down,
right?
Just my 0,02 eurocents ...
Wish you all a nice weekend,
Bas
Op zaterdag 2 april 2005 09:24, schreef Rik Henson:
> Elizabeth / Bas
>
> > I would simply measure your response times and put the real stimulus
> > durations in your model, instead of letting a Taylor series
> > approximations such as temporal/dispersion derivatives estimate your
> > varying response. I think series expansions are only useful when you
> > don't now the variability in HRF duration/timing, and want to catch it.
> > Since you do know the real stimulus durations (by recording reactions)
> > and hence the expected changes in HRF shape, you could use that a priory
> > knowledge to create your GLM instead of adding derivatives. Or perhaps I
> > am missing the point here?
>
> Unfortunately, I don't think it is this simple.
>
> The derivatives of the canonical HRF are primarily to capture
> *haemodynamic* variability (eg, across the brain), rather than *neural*
> variability (though in principle, of course, we cannot tease these two
> sources apart using BOLD only).
>
> The reason that the HRF derivatives will not help much for durations ~0-4s
> is that the dominant effect of increasing the duration is to increase the
> *height* of the response; it's shape changes little. In other words, the
> parameter estimates for the temporal and dispersion derivatives will change
> little compared to the change in the parameter estimate for the canonical
> HRF. This is because the HRF effectively integrates the total neural
> activity over periods of seconds.
>
> There is nothing *wrong* with modelling trials with different durations
> (rather than using "events", all with duration 0). It can give quite
> different results (e.g, if your trial durations vary considerably within a
> condition). It does however dramatically change the interpretation of the
> parameter estimate:
>
> For trials modelled by their duration, the parameter estimate reflects the
> response *per unit time*.
>
> For trials modelled as events (or by the same duration), the parameter
> estimate reflects the response *per trial*.
>
> To see the difference, imagine that stimulus is a visual stimulus presented
> for 200ms, and the RT of its motor response varies from 1000-2000ms across
> trials:
>
> For an area like V1, the duration of neural activity, and hence size of the
> BOLD response, is likely to be constant across trials (assuming it only
> cares about the duration of the visual flash). In this case, the
> "constant-duration" (eg "event") model will fit better.
>
> For an area like M1, the duration of neural activity may vary with the RT
> (perhaps), hence size of the BOLD response will vary across trials. In this
> case, the "varying-duration" model will fit better.
>
> So which is a better model depends on how you think the neural processing
> is affected by trial duration - *and this may differ for different
> regions/processes!*
>
> For the latter reason, I prefer to model trial durations (provided they are
> less than ~4s) as a parametric modulation. That way, you can get the best
> of both worlds, identifying regions like V1 in the main effect of the
> trial, and regions like M1 in the main effect and parametric modulation.
> (Though if you want to allow for different mean RTs *across*, rather than
> *within* conditions, then using a parametric modulation, rather than
> differing durations, can get a bit more complicated: see my previous
> SPMlist emails on this topic)
>
> Finally, you can of course model varying duration *and* convolve with the
> canonical HRF and its derivatives.
>
> Rik
>
> ---------------------------------------------------------
>
> DR RICHARD HENSON
>
> MRC Cognition & Brain Sciences Unit
> 15 Chaucer Road
> Cambridge, CB2 2EF
> England
>
> EMAIL: [log in to unmask]
> URL: http://www.mrc-cbu.cam.ac.uk/~rik.henson
>
> TEL +44 (0)1223 355 294 x522
> FAX +44 (0)1223 359 062
> MOB +44 (0)794 1377 345
>
> ---------------------------------------------------------
> -----Oorspronkelijk bericht-----
> Van: SPM (Statistical Parametric Mapping)
> [mailto:[log in to unmask]]Namens Elizabeth Chua
> Verzonden: vrijdag 1 april 2005 22:56
> Aan: [log in to unmask]
> Onderwerp: [SPM] variable durations and dispersion derivatives
>
>
> Hello SPMers-
>
> I'm using a self-paced event-related design where the stimulus offsets are
> controlled by the subjects responses. There are differences in reaction
> times between different conditions which are typically on the order of 1-2
> seconds, but occasionally as large as 3 seconds. There is also varability
> with a condition for the stimulus duration, typically on the order of 1
> second.
>
> My understanding is that modeling the durations models different "heights"
> for the hemodynamic response. Is this correct? If so, does this work
> within a condition, between conditions, or both? Is this recommended?
>
> Also, my reading of the dispersion derivative is that this accounts for
> differences in the width/ length of the hemodynamic response. Is this
> something I should use with my reaction time differences? or will it
> remove differences between conditions that might be meaningful?
>
> Thank you in advance,
> Elizabeth
|