Hi all,Say you are modelling a task using a first-level model in SPM. The task consists of three conditions: congruent, incongruent, and control. When setting up the contrasts for this model, how would one properly weight the conditions for the following contrast: [congruent + incongruent] > control? Would this be 0.5 0.5 -1? On a related note, would the contrast [congruent + incongruent] > baseline be 0.5 0.5 0, or 1 1 0? I am confused as I've read that positive (and negative) contrasts should sum to 1, but I've seen people use contrasts such as 1 1 0 0 etc. for certain contrasts, which obviously do not sum to 1. Any input appreciated!
Hi
If you want to model the contrast (congruent +incongruent)>baseline using a Tcontrast you could use either the vector 1 1 -2 or 0.5 0.5 -1 and if you want to model the opposite (i.e. baseline>(congruent+incongruent)) you have to change all signs in the expression (i.e. -1 -1 2 or -0.5 -0.5 1). When you see contrasts as 1 1 0 or something like that, they could be F contrasts that are a little bit different and are used to test a hypothesis about general effects, independent of the direction of the contrast!
Hope this helps!
I agree with Roberta. Just an additional note: for t-contrasts like the one you are describing, all weights should sum up to 0, not to 1 (as in her examples).
Cheers,
Ant
Thank you both for your input!
I thought I'd share some other information that I stumbled upon regarding this issue. It seems that using e.g. 0.5 0.5 -1 is preferable over 1 1 -2. To my understanding this is due to how one is effectively scaling the data, which could have implications for e.g. calculating % signal change (see: mumford.bol.ucla.edu/perchange_guide.pdf).
Best,
Lasse
Reading your question, I understood you do not model the baseline trials but instead leave them as implicit baseline in the GLM, which is of course fine.
What I do not understand is why you would want to do a ((cong + incong) > baseline) contrast. Is it to identify regions responsive both to cong and incong events? In that case, a contrast is not the best solution, as it can (conceptually) average across activations to both tasks which could results in false positives. For instance, if activity in a specific regions is highly compatible with the cong regressor (beta = 10, just for illustration) but not at all (beta=0) with the incong regressor, you could still get significant activity in the (cong + incong) contrast.
A preferable approach is the conjunction approach (see Friston et al., 2005, NeuroImage).
Thank you Niv for raising this point!
I omitted some details in the original post for simplification, as my question was a general one regarding contrast weights. Actually I am interested in between-group differences in my study (i.e. patients vs. controls). I am mainly interested in investigating between-group differences on the Incongruent > Congruent contrast.
Prior studies using a similar task paradigm have also compared groups on the [Incongruent + Congruent] > baseline, so I thought that I would do the same (to be able to compare my results with these studies). But you do raise an important point regarding the use of such contrasts; i.e. one would be unable to know if any significant between-group effects was mainly due to incongruent or congruent trials. I guess this could be clarified by extracting % signal change or raw beta weights from any sig clusters.
- Badges
- Science topic
- Similar topics
- Correction