Dear Jennifer,
> This question involves first-level analysis in FSL using the Full
> Model Setup for statistics.
>
> I’ve been trying to figure out a way of comparing 1 conditions
> activation, against the summed activation of 2 other conditions.
> Right now each condition is represented by an original EV.
> Initially, I was going to do this by collapsing the conditions in
> timing files (make a timing file for conditions 2 and 3, rather than
> 2 separate files). But after looking through the FSL documentation,
> I think it can be done via contrasts.
>
> Let’s say I have 9 conditions. I want to compare Condition 1 to the
> activation in condition 2+3. This is to find where the activation in
> condition 1 differs from only conditions 2+3 (ignoring the other
> conditions 4-9).
>
> The documentation seems to suggest it would be set-up by specifying
> the contrast this way [1 -1 -1 0 0 0 0 0 0]. So I set to zero all
> the EVs that I don’t care about (conditions 4-9). Set to -1 the
> conditions that I’m comparing against and set-to 1 the condition I
> want the activation for (that is different than the activation
> across the -1 EVs).
>
> The entire design matrix would be as follows (y axis =contrast, x
> axis=original EV)
> 1 -1 -1 0 0 0 0 0 0
> -1 1 -1 0 0 0 0 0 0
> -1 -1 1 0 0 0 0 0 0
> 0 0 0 1 -1 -1 0 0 0
> 0 0 0 -1 1 -1 0 0 0
> 0 0 0 -1 -1 1 0 0 0
> 0 0 0 0 0 0 1 -1 -1
> 0 0 0 0 0 0 -1 1 -1
> 0 0 0 0 0 0 -1 -1 1
I think you mean "set of contrasts" here rather than "design matrix"
which is something quite different.
Your first contrast above does indeed compare the activation in the
first condition to the sum of the activations in conditions 2 and 3.
Such a contrast makes sense for looking at e.g. an interaction effect
where task A and task B are performed in conditions 2 and 3
respectively and where _both_ tasks A and B are performed in condition
1.
I am not sure though of the set of your three first contrasts can all
make sense at the same time.
> Though, perhaps the -1 should actually be specified as -.5 (since
> there are 2 EVs involved).
That compares the mean activation in 2 and 3 versus the activation in
1.
> My worry is that what I’ve found in the documentation doesn’t
> involve ignoring some EVs. See this excerpt:
> “An important point to note is that you should not test for
> differences between different conditions (or at higher-level,
> between sessions) by looking for differences between their separate
> individual analyses. One could be just above threshold and the other
> just below, and their difference might not be significant. The
> correct way to tell whether two conditions or session's analyses are
> significantly different is to run a differential contrast like [1
> -1] between them (or, at higher-level, run a higher-level FEAT
> analysis to contrast lower-level analyses); this contrast will then
> get properly thresholded to test for significance.”
This is a very different matter. It basically says that two
thresholded contrasts can _look_ very different while in reality the
pertinent un-thresholded maps are very similar. It basically just says
not to assess e.g. group differences by eye balling the individual
groups thresholded stats, but instead perform a formal test (a task-by-
group interaction).
I hope that answers your questions.
Jesper
|