Hi Helmut,

I’ve actually had similar thoughts about the task design and the long task blocks. Both with regard to behavioral data and activations. I’ve done a preliminary check using percent signal change, and looks like activations do not change too much within a trial. It is certainly a good point.

I think I want to start with a single subjects analysis to look at how behavior across trials is related to activation across trials.

A) Regarding the purpose of my study - my primary goal is to be able to show how changes in behavioral performance across trials are related to changes in activation across trials. From my understanding, I do this by:

Setting up ONE condition
have behavioral data as the first PM. Am I correct that I would input my PM values as [.2050, .5370, .1150, .2290, .2300, .6290, .9130, .8540, .4170, .644] (i.e. my behavioral data for the 10 trials), and choose first-order (1st order because I want the values unchanged).

B) My second goal is to also model task versus rest to show that there is an activation during the task compared to during the rest. How can I do this?

I’m just not sure how to practically implement these in SPM. I’m totally new to this.

I really appreciate your guidance :)
Thanks,
Joelle

ps. the design matrices, these are something created by the SPM program after running an analysis, right?

On Wed, Jun 3, 2015 at 8:44 PM, H. Nebl <[log in to unmask]> wrote:

Dear Joelle,

Something with the formatting seems to have gone wrong in the last message, and it should read "within long blocks" instead of "within look blocks" ;-) Anyway. Find attached some design matrices for illustrative purpose. Except otherwise noted they are based on a single task condition with onsets 0:138:1242 s and duration = 124 s.

In summary:
1) It is very important whether to go with orthogonalization at all, and in case you do, whether you go with serial ortho. (in which the order of the regressors is important, this would also hold for a PM reflecting the behavioral score) or whether you just ortho. the different PM regressors onto the unmodulated task regressor, as this will affect the time courses of the predictors and also the interpretation. E.g. with serial ortho. and a PM for linear changes included any type of PM testing for quadratic changes will result in the same time course, be it based on e.g. n^2, (n-5.5)^2 or (n-2.25)^2 with n = trial no. 1:1:10. This will not be the case with ortho. onto the task regressor only or with the ortho. disabled completely.
2) Another aspect, high-pass filter. Due to your long blocks it is questionable whether it makes sense to test for linear changes depending on trial no. at all. With HPF removing the low frequencies from the predictor the linear change becomes especially evident for time points between the trials. The low frequencies are also removed from the data, so this shouldn't be a problem as such, but it is questionable whether you still test for / are able to detect any linear changes in activation while the stimulus is "on". This also holds if the HPF is adjusted to something larger than twice the cycle of 62 + 7 TR, and it will also hold if you don't use any HPF by setting it to "Inf" and instead adding drift regressors to account for linear, quadratic, ... drifts over time. Probably this problem extends to other types/orders of PMs. Interestingly this has not yet properly been investigated and/or the problem tends to be neglected for paradigms in which long blocks are common (like in your learning paradigm). In these instances activations reported to reflect "linear changes" are probably just artefacts ;-) Note that if you go with short durations (same onsets, but e.g. duration = 10 s instead of 124 s) then the predictor has lots of signal within the high frequencies, which are not affected by the HPF (see attachment).

Best

Helmut