I am working on a task based fMRI analysis (still stuck at single subject level), where I am using GAM function to model regressors I am not interested in, and TENTzero to model regressors of interest (this is because I had previously a collinearity problem when using TENT for all regressors).
In order to get a better understanding, I ran two separate analysis, in one I used TENT and in another TENTzero for the regressors of interest. I was hoping to find not big differences when looking at a GLT contrast afterwards. However, what I see is that there are cases where I see some significant clusters when using TENT, which are not there when using TENTzero. I looked at the estimated HRF for my two regressors that I use in the contrast in both TENT and TENTzero analysis and it seems to be that whenever the first and last betas that are set to be 0 in TENTzero, differ a lot from the estimated first and last betas in the TENT analysis, a significant cluster would be only visible in the TENT analysis.
My question is: Is the usage of TENTzero at this point a standard and is always recommended over TENT? Or is there an objective criteria, which I can use to decide now with which one I should go? My collinearity issue is resolved either way, so no collinearity when using TENT or TENTzero.
Is the usage of TENTzero at this point a standard and is always recommended over TENT?
Or is there an objective criteria, which I can use to decide now with which one I should go?
The questions are better framed from a different perspective: Do you want to capture interpretable (and presentable) BOLD response? Or do you want to achieve a better model fit? The difficulty is when these two goals are not always aligned all the time as you’ve already observed; when such a scenario occurs, you need to decide which one is the priority: there is no standard answer. If you believe that your subjects didn’t have any anticipations for the tasks, it may make sense to fix the BOLD response at 0 at the stimulus onset. On the other hand, simply because you could see more clusters by modeling the BOLD effect at the stimulus onset time and render a better model fit, it does not necessarily mean that’s a meaningful BOLD estimate you intended to capture: for example, the model could catch something unrelated to the task of interest, especially when the experiment was not well designed to fine-tune the BOLD response estimates. Also look at those brain regions where you see BOLD estimates at the stimulus onset time: do the BOLD curves look reasonable to you? Similarly do the same for the differences between the two basis functions regarding the BOLD response tails.
Once you know your priority, it is only a technique issue to choose between TENTzero and TENT. For example, even with TENTzero(b,c,n), you can still capture the anticipation effect by setting the ‘b’ parameter to be negative (e.g., -TR). By the same token, if you believe that your current TENTzero setting does not catch the tail long enough, increase the ‘c’ and ‘n’ parameters.
I choose to capture interpretable BOLD signal and I do not expect my subjects to have anticipations for the different task conditions.
I also took a closer look at the beginning and the end tails of the BOLD curves and unfortunately the result of this inspection is inconclusive (see attached images).
Commend for attachment: The significant clusters/voxels are from a comparison of two conditions, CR and FA, and I am looking at the contrast of the summed betas from CR - summed betas from FA. Image 1 shows a cluster that was significant in TENT analysis for this contrast, but not TENTzero, and image 2 shows a cluster that was significant in TENTzero, but not TENT. The upper images show the HRF for CR from the TENT analysis, and the bottom images from the TENTzero. I think that in image 1 the shapes look more reasonable for TENT, and in image 2 for TENTzero. However, in general I am not sure, if those shapes are considered “reasonable” . Some do not look like the traditional HRF at all.
Does this mean that there is something wrong with my analysis? (it was a very fast paradigm, stimulus presentation for 3 sec, 1 sec ITI, 5 different conditions. When looking through my significant clusters, I can see HRF’s that look pretty reasonable and others that don’t.
What does it tell me about those significantly activated clusters? Is the signal not “real” and I should not consider those clusters?
Would it be better or worse to specify my GLT’s with [[x…y]] and look at significant clusters for individual betas from the TENTS, when I have estimated HRF’s that don t look reasonable?
Thank you very much for answering my questions in advance
Carolin
I am looking at the contrast of the summed betas from CR - summed betas from FA
No, that is not a good testing strategy because summing over all the betas is less sensitive especially when there are undershoots (negative betas). Use the F-test through the [[x…y]] specification.
after switching my testing strategy to [[x…y]] and comparing the F-statistic between TENT and TENTzero, it was obvious that TENTzero is the better choice. TENT seemed to pick up some motion (as seen by significant voxels around the edge of the brain and ventricles).
The estimated HRF’s look reasonable for most voxel clusters, but I do have some that look weird (see attachment). They don t really look like “traditional” HRFs. What does it mean for my analysis? Should I assume that significant clusters that have such weird HRF shapes did not pick up signal relevant to the experiment? Do I consider those clusters in my further analysis?
it was obvious that TENTzero is the better choice. TENT seemed to pick up some motion (as seen by significant voxels
around the edge of the brain and ventricles).
Good to hear that!
I do have some that look weird (see attachment). They don t really look like “traditional” HRFs.
What does it mean for my analysis? Should I assume that significant clusters that have such
weird HRF shapes did not pick up signal relevant to the experiment? Do I consider those clusters
in my further analysis?
They don’t look like HDR to me. So I tend to believe that in this case the model picked up something unrelated to HDR. This shows another important aspect of estimating the HDR instead of presuming a shape-prefixed HDR because it plays a role of filtering out false detection based on the HDR shape. Is the cluster in the region you’re looking for?
yes, the cluster is in one of the regions where I would expect some activation based on the literature.
I also performed an analysis using GAM instead of TENT and the analysis using GAM shows a significant cluster at this location as well.
I checked the HDR in a bigger set of voxels in this cluster. Only some look a little more reasonable (see attachment).
At this particular region it’s a little bit hard to use the TENTzero results alone as evidence, but I suggest that you take the results from GAM (and the literature if available) as supporting evidence.
The
National Institute of Mental Health (NIMH) is part of the National Institutes of
Health (NIH), a component of the U.S. Department of Health and Human
Services.