Dear Helmut,
thank you very much!

Indeed, you point out a number of additional issues that are most relevant. I know that with regards to correlation between regressors, the design is far from optimal. As that's the way it has been done, I still try to figure out the best way of analyzing the data now. We'll try your suggestion (1). I also understand your point regarding (2) - nothing to subtract - Thanks! Regarding your question (3): yes, there are 3 different instructions... (the actual design is even more complex - but never mind...)

Apart from that, I still have a much more basic question:

What happens to regressors in the model estimate when they are included, but marked with 0 in the contrast definition? Will they be fully ignored on all levels?
But why then would we enter zeros for the movement regressors?

Consider the following examples:

1)
Names/onsets: fixation, instruction, picture_snake, picture_spider, R1, R2, R3, R4, R5, R6
Contrast: [0, 0, -1, 1, 0, 0, 0, 0, 0, 0]

Vs.
2)
Names/onsets: picture_snake, picture_spider, R1, R2, R3, R4, R5, R6
Contrast: [-1, 1, 0, 0, 0, 0, 0, 0]

Vs.
3)
Names/onsets: picture_snake, picture_spider
Contrast: [-1, 1]

What is the difference? And if there is no difference between 1 and 2, why would I add R1 to R6 anyway?

Hope my question becomes more clear now.
Thanks very much again!
Stefan

-----Ursprüngliche Nachricht-----
Von: H. Nebl [mailto:[log in to unmask]] 
Gesendet: Dienstag, 12. Mai 2015 13:30
An: [log in to unmask]; Stefan M. Schulz
Betreff: Re: which regressors to add to the model for controlling different aspects of experiment

Dear Stefan,

If you don't have any "fixation only" trials then you probably can't differentiate between activations due to stimuli and those due to fixation cross (except maybe if there's a very long and jittered fixation period, for which you assume constant activations), as the regressors are probably highly correlated, which is bad. Note that the design orthogonality plot might be misleading, as it only shows the regressors, not the possible combinations (like pic1 + pic2 + pic3 vs. fix). It might be the same issue for the instructions if they were presented on a trial-by-trial basis.

(1) Thus for a first attempt I would go with a model specification with three regressors corresponding to the three types of pictures plus the motion parameters. But see (3).

(2) There's nothing to adjust for that contrast. (spider - box only) - (snake - box only) results in spider - snake. It's the difference between spider and snake you're interested in, so it doesn't matter whether you subtract something on both sides or not. 

> but I’m not sure if this will not mess up controlling for movement 
This doesn't have to do anything with movement. You just have to make sure to set up the correct contrast vectors. If the order of the predictors is spider snake box R1 R2 R3 R4 R5 R6 go with [1 -1 0 ... ] to test for spider > snake.

(3) Are there really three different types of instruction (and not an unspecific one pointing to trial onset)? Then one would probably want to explicitely model these instructions, as they might well result in relevant differences, especially when it comes to phobia. Due to possible correlations with the picture regressors it might require some adjustments to the model, e.g. trying to model an epoch covering instruction and picture. But this depends on the design. Was an instruction followed by several trials of the same type?

Best

Helmut