Print

Print


Thanks Jonathan (& everyone)

More feedback would be very much appreciated!  I'm sure this is relatively straightforward and would love someone to shed some light.

I should have made my question more explicit (and clearer).  I have stimulus types A and B, both broken into correct & incorrect responses.  Type B is presented in 3 different colors.  I.E.:
A-correct, A-incorrect, B1-correct, B1-incorrect, B2-correct, B2-incorrect, B3-correct, B3-incorrect.

I am only interested in A-correct and B-correct [collapsing across color] but don't know how to model it.

Regarding the first model type with regressors
A-correct; A-incorrect;
B-correct; B-incorrect;
Color 1; Color 2; Color 3
I think the main problem here is that it's degenerate - B-correct + B-incorrect = Color 1 + Color 2 + Color 3

Regarding the second:
A-correct; A-incorrect;
B1-correct; B1-incorrect;
B2-correct; B2-incorrect;
B3-correct; B3-incorrect;
I agree I can ask the question I want but my power is drastically reduced, which is far from ideal.


Does this make sense?  It seems like the best way is to ignore the effect of color altogether - if I model only type A and B and ignore color it would be best in terms of power, but I'm ignoring predictable effects I should model.  It seems like a lot of people routinely do this in one form or another, however (e.g., model stimuli but ignore responses; do not model feedback trials etc.).  In a different study which is quite similar at a practical level I've seen someone do the equivalent of compare type A to type B-color 1 and ignore B responses in the other colors altogether (which I think is very questionable, since they share the same cognitive process).

I'd really love some feedback - I'd hate to hit problems when I eventually try to publish.

Thanks for your time

Christopher



On 18/7/09 1:41 PM, "Jonathan Peelle" <[log in to unmask]> wrote:

Hi Christopher

On Fri, Jul 17, 2009 at 2:11 AM, Christopher
Benjamin<[log in to unmask]> wrote:
> Hi, I'm modelling a task in which participants perform the same cognitive task on a number
> of stimuli which differ only in their color.  I'm having trouble working out the best way to
> model this and would love feedback.
>
> I have stimuli of four different colors, with responses to each color split into correct and
> incorrect responses.

What question are you interested in asking of the data?  That will be
helpful in informing your choice of model.


> A. If I consider modelling the process and then the color separately [i.e., regressors (1)
> correct, (2) incorrect, (3-6) color], I think I'm likely to end up with a degenerate design.

In this setup, the first regressor will model the activity in common
across all correct trials, regardless of what color items were
presented in.  The color regressors, in this model, would then tell
you how activity associated with each color differed from the common
activity modeled by regressor 1.

Note that as you have it here the color regressors will pick up
activity for both correct and incorrect trials, which may not be what
you want....you might want to split the colors by correct/incorrect.


> B. If I model correct and incorrect responses by color [(1)blue-correct, (2)blue-incorrect,
> (3)purple-correct...(8)] it seems intuitively wrong - I'm splitting variance for a cognitive
> process that is almost identical between four different regressors (for both correct and
> incorrect responses).  I also think I lose a lot of power in each regressor's estimate.

On the one hand you are correct, in that, assuming that blue-correct
and purple-correct share a lot of activity, you're splitting up a
single process, and giving yourself less data for each regressor.
However, note that in this case, doing a contrast across several
columns of your design matrix (i.e. the four columns coding for
correct responses) should give you the same answer (i.e. the average
activation for this process) as approach A above.  The only caveat
would be if the number of correct/incorrect trials differs much as a
function of color, then you will get slightly different weightings
depending on which approach you take.

Hope this helps,
Jonathan