Thanks, Tim and Eugene, for the patient explanations. I think I've got it, now.

So, now the question is this:

What's the *right* way to design this experiment? I want to compare
activation involved in "hard" problem solving to activation involved
in "easy" problem solving, expecting to get the usual suspects --
parietal cortex and DLPFC, and curious to see what else emerged from a
whole-brain analysis. The reason I'd gone with the original
alternating easy/hard design was to equate the motor responses and
visual stimulation between the two conditions, with the idea that by
including the RT in the model, I could remove the "time on task"
concern.

I now see that that design is not going to be good for me. One option
is to go with a block design, where subjects solve as many problems as
possible in each block. When I pilot this version, it gives me the
activation I expect, but it also gives me the motor and visual
activations that I was hoping to avoid.

Ideas?

Thanks!
Todd


On Thu, Feb 18, 2010 at 3:04 AM, Tim Behrens <[log in to unmask]> wrote:
> Hi again
>
>>
>
> The effect you point out lies behind the reason for the differences in the
> copes - if the true BOLD responses to the different duration trials are in
> fact more similar than is assumed by the HRF linear convolution model, then
> the shorter duration trial regressors will need much larger betas to fit the
> responses. This does make comparisons of trial types that have brief and
> variable durations trickier to interpret - the effect could be due to
> greater neural activity or a non-linearity of the BOLD response to brief
> stimuli.
>
> Eugene is being subtle :)
> What he means to say is that you simply cannot expect to compare a 1s trial
> to a 3s trial.
> There are two reasons for this:
> i) The linearity assumptions of the hrf are not good in this range.
> ii) Your modelling assumptions of neural activity are not good. For example,
> what if there were a region that exhibited phasic activity to the onset of a
> trial (instead of continued activity throughout the trial)? The neural
> activity would be identical in each case. Presumably the BOLD response would
> be identical, but your betas will be 3x bigger for the short than the long
> trial - it will look like a massive difference in activity, even though the
> underlying activity is identical.
>
> T
>
>
>  As Tim mentioned, this is not as much of an issue if trials are longer than
> around 5 secs.
> Eugene
>
>
>
>>
>> Thanks again,
>> Todd
>>
>>
>>
>
>>
>> On Wed, Feb 17, 2010 at 4:59 PM, Tim Behrens <[log in to unmask]>
>> wrote:
>> > Hi
>> > On 17 Feb 2010, at 20:53, Todd Thompson wrote:
>> >
>> >> If the cope is reflecting the "scaling factor" by which you have to
>> >> multiply the regressor to re-construct the raw signal data, shouldn't
>> >> these copes be equal in order to reconstruct the 2.5% signal change?
>> >
>> >
>> >
>> > No -  The model goes like this
>> > Y=Xb
>> >
>> > so if X(n) is 8 times smaller than X(n-1), then b(n) has to be 8 times
>> > bigger to make the same Y  (and Var(b(n)) has to be 64x bigger!!)
>> >
>> > In fact X(n) will not be 8 times smaller, because hrf convolution
>> > doesn't
>> > carry on adding height after a certain number of seconds (roughly 5, I
>> > would
>> > guess off the top of my head, but it is a guess - it would be easy to
>> > figure
>> > it out in matlab), but it could easily be 4 or 5 times smaller - hence
>> > you
>> > have vastly different copes and varcopes for the same %signal change.
>> >
>> > T
>> >
>
>
>