Hi Rick,

rick reynolds Wrote:
-------------------------------------------------------
> Hi Joseph,
>
> Using AM1 simply applies the modulators to the
> individual
> responses, rather than allowing for a mean
> response along
> with modulation responses. The single regressor
> in AM1
> should be the sum of the modulation regressors
> (including
> the mean, so 3 in your case).
>

I see, so if I wanted baseline * mod1 and baseline*mod2 fits, which is a way of faking what I want, I would enter two separate AM1 timeseries (but both of these would contain the activity found in the mean).

> Note that when the stimulus events are so very
> long (2-3
> minutes?), the main regressor would likely eat up
> much
> more variance than the modulation terms, though
> that
> depends on the strength of the modulation
> response, of
> course. Also, be very careful with the baseline
> model.
> Having only 3 long events in a run makes
> distinguishing
> baseline from task more difficult.

Yes the design is suboptimal in many ways, but it is also set in stone. I'm just trying to polish the stone, so to speak. We have 9 trials of task A and 6 of task B, so there is a contrast I left out of my description for clarity. However your prediction exactly plays out in the data (and makes sense to me): we have people watching movies in the scanner for the duration of the block, and surely any region that is going to be parametrically modulated by the content of the video is going to generally saturate its BOLD signal over the course of a ~3 minute video. However these are the same videos as 2 previously published papers using this task, so I must be doing somthing differently to those who came before me.

>
> Getting to the implied question, I am not sure
> what you
> might want with only 2 regressors. Would one be
> the
> mean and the other a combined modulation? To
> compare
> with the SPM result, how many betas did the other
> one
> generate, 1 or 2?
>
> - rick

I'm in the unfortunate position of trying to interpret this from a very condensed paper. They report ONLY a 'parametrically modulated' (SPM language) set of betas (single group and then between group contrast), and they use an incidental behavioural measure as a contrast of no interest (to show their brain-behaviour relationship isn't actually just related to motor events). They never report a non-modulated finding, which suggests to me that what they were actually showing was a mean*modulator result (something like I would get with AM1).

To be clear, they show some regions that 'modulate' with the task, but NOT a contrast between regions that are consistently activated by the task and those that are parametrically modulated by the task... which is what is making me suspicious.

I'm looking at the SPM documentation now (pg 259 from here: SPM12 manual) and it actually looks like SPM computes these amplitude modulations in the same way as AM2.

However I know that 3ddeconvolve uses marginal statistics -- is there any chance you know whether SPM does/does not use marginal stats? Would my AM1 strategy above get be to a place where I 'sort of' get around these marginal statistics? I know it is right/appropriate to use marginal stats, I just need to demonstrate to others what is going on with the data.

Thanks for your help with this!

Joseph