Hi Phil,

> My first question is, is this an accurate description of how to create this regressor?

Yes, it looks reasonable to me.

> My second question is, if this is an accurate description of the approach, why is it
> necessary to deconvolve the seed time course rather than convolving the condition
> code with the canonical HRF prior to multiplying by the seed? I suspect there’s a
> good reason, since the latter approach seems more intuitive/easier, but it’s bugging
> me that I don’t see the reason deconvolving the seed is preferred.

There are two aspects involved here. The first one is that the interaction between the seed and a target region presumably occurs at neuronal, not BOLD, level, and this is why the multiplication is typically executed between the condition coding and the deconvolved seed time course.

The second perspective can be viewed from the mathematical formulation. The standard process of creating the interaction is like this

Con ( conditioning X Decon ( S(t) ) ),

where Con() and Decon() are the convolution and deconvolution operators with the presumed BOLD impuse response HRF, and S(t) is the seed time course. Your suggestion can be expressed as

Con ( conditioning ) X S(t).

The two approaches are not the same simply because convolution and multiplication are not commutative. Having said that the difference for block designs is probably small due to the accumulative effect of consecutive stimuli in a block, so you could adopt the second approach. However the difference for event-related designs could be substantial.

Hope this helps.

Gang



Edited 2 time(s). Last edit at 10/11/2012 11:25PM by Gang.