Hi everyone,

I know this issue has been addressed in previous posts. However, I am quite confused about the main effects as confusion in the denoising step.
I frequently model "non-interest" conditions like instruction reminder, short or long rest periods, etc.
I am not interested in any connectivity from them.
Does this mean I shouldn't add their main effects as confusion in the denoising step? Or maybe I should add them but leave them in the 1st level analyses?

What is the rule of thumb in this case?

Best, 

Alex

Hi Alex,

That's a good question. In general any external factor/timeseries that could have a "task-activation" effect on the BOLD signal (i.e. where the BOLD signal in some areas may correlate with this external factor) can potentially influence your connectivity measures. In very general terms, you could always:

Case #1) if you believe this factor does not affect the BOLD signal, or if it does you prefer not to "control" for its effect (i.e. you would not mind if shared BOLD signal fluctuations in two distant regions where the BOLD signal in both areas covaries with this external factor appear/count as "connectivity" between those two regions). In this case you would not enter this factor as a condition and you would not include this factor in the list of potential confounding terms during denoising

Case #2) if you believe this factor may affect the BOLD signal, and you would prefer to "control" for its effect (i.e. when estimating connectivity between two areas you want to estimate connectivity strength during an imaginary scenario where this external factor was constant across time). In this case you would not enter this factor as a condition but you would include it in the list of potential confounding terms during denoising

Case #3) if you believe this factor may affect not only the BOLD signal directly but it may also act as a modulatory term, where the strength of connectivity between two areas may change as a function of this factor (e.g. connectivity may be stronger when this factor is high/low compared to when this factor is low/high). In this case you would enter that factor as a condition (e.g. if it is a block or event design) and run gPPI or weighted GLM analyses, or you would enter that factor as a 1st-level covariate (e.g. if it is a continuous timeseries) and run temporal-modulation analyses (and in both cases this factor should also be included, as before, in the list of confounding terms during denoising). 

Coming back to your case, I imagine that factors that represent the main events in your experimental design (e.g. the subject is listening to instructions during this block, the subject is resting in this block, etc.) can all be expected to fall in cases #2 or #3 above, while factors that represent more subtle or "non-interesting" variability (e.g. differentiating vs. long vs short resting blocks, differentiating the form of the instructions given to the subjects, etc.) may fall in case #1 above. Of course the difference between these cases is often simply a matter of how you are interpreting and would like to analyze the potential effect of that factor, so multiple/different approaches are perfectly reasonable (as each approach will answer somewhat different questions). 

Hope this helps
Alfonso
Originally posted by alexandre :

Hi Alfonso, 

I was not aware of a case #3 possibility.
Thanks!

Alex