Hi Rick,
Thanks for the reply. Regarding the necessity of the PSC - yes, I'd like to have it for my subsequent analyses. What I'm struggling to wrap my head around is the idea of what I actually want for my subsequent analyses - I need some sort of "scaled" and "clean" signal; the appeal of the second idea was the logic of: 1. Extract task-related signal change only, then 2. Work out the PSC per condition as a percentage of its OWN mean signal change. Now that I've written it out, it seems a little more dodgy.
Further, I tried subtracting the baseline, and worked out the mean of those results. They weren't 0, but the numbers did vary a lot. And of course, once I divided the subtracted results by the mean of the subtracted results, I got very odd numbers. I think it's just flawed logic on my part?
And the decision to leave scaling till later was mainly about not wanting to use the raw mean, instead opting for something more "accurate" (given concerns about trend).
To that end, I have another question. I want to know how well the model (especially the baseline aspects) fits the data. In an earlier version of some of my later analyses, I was calculating a PSC for time points of interest by dividing by the mean of the first 2 time points. I thought this was redundant and so removed that, but if the baseline I divide by now for PSC is in some sense, inaccurate, it may be worthwhile to use the earlier method of establishing a mean per trial.
I used a polort value of 4, and "-bout" with 3dDeconvolve. I got a whole lot of extra t-statistics for each polort, for each run. How do I interpret them? Using run#1pol#0_coef and tstat (linear trend?), most of the voxels (basically all) are highly significant. Am I to assume that this means this trend is more or less fully accounted for in the model? Fewer voxels are highly significant using run#1pol#1, etc. I know I can also use the "-cbucket" output in the grapher to check the fit. Are there any other methods? I suppose my question here is, how small a number of significant voxels for a pol# would I have to see to justify dropping my polort value?
My data was quite messy, and since I am unable to re-collect data, I am trying to get the best AFNI-processed results before running ROC/SVM/MVPA type analyses.
Thanks a bunch,
Mahen