I have a question about connectivity analysis done with fMRI time series collected during a block-design task (30 sec on, 30 sec off). I have a group of controls and a patients group with n=20 in each. I want to do functional connectivity (correlation) or effective (eg Granger causality) or even dynamic FC.
My issue is that the fMRI time series are much longer for the controls than the patients, about 2 times longer (~480 vs ~220 volumes). This is because subjects performed a fatiguing task and the patients fatigued much earlier and so there are less volumes collected. So, the correlation for the controls would be computed with more time points than the correlation of the patients.
-1-
So, my question is whether the correlation that are calculated with a different number of time points for each group can still be compared between groups with a t-test and be meaningful?
-2-
If this an issue, is there a way out? Maybe up-sampling the patient time series or some other methods? In a prior analysis, I have selected portions of the time series so the groups have equal number of time points, but that results in less total time points and discontinuity.
Interesting question. As a clarification, is everyone starting at the same level of fatigue in the beginning?
Without knowing the answer to that, my general feeling would be to crop the runs so everyone has the same number of time points.
The reason for that is that 1) yes having more time points can absolutely impact the inter-class correlation (ICC) among (many) other things; and 2) as a reviewer I would be wary if someone "invented" new data.
Subjects start fresh and performing more blocks induce fatigue; we measured grip force to index fatigue.
The controls need about 12 blocks (group ave) to reach a fatigue level
The patients need about 5 blocks (group ave) to reach a fatigue level.
Scenario 1
This is what I'm currently doing
I select specific blocks (eg five blocks) in which performance is matched between groups (as you suggested). But this reduces the number of time points in the analysis and creates discontinuities (I would have liked to get away from simple correlation and try effective connectivity like Granger but discontinuities can be a problem with GC).
Scenario 2
This is what I was hopping for
I use all the blocks from the start until fatigue is reached and I can track fatigue development pretty well. But, I have different number time points per group .
It's a predicament for sure! Do you have a measure of relative fatigue?
I'm not sure how I would apply Granger to that type of data, but one thing you could do is estimate node-based connectivity within each block and then try and examine correlation changes over time. Then you could potentially growth (inverse growth) model the change in node-correlations over time for each group with interactions between. Another place to look would be edge time-series analyses for the same type of approach just with fancier math.
Hi the edge time-series analysis seem an interesting approach
What do you mean by "relative fatigue"? We have grip force data for the whole task so I can see when performance change
An edge time series approach would be to z-score each node's time series, and then, element-wise multiple pairs of these z-scored signals. And you can just repeat this operation for all the pairs of nodes (i.e., edges) you'd like to look at. No fancy math required ; ) The resulting signals document the frame-wise instantaneous similarity (the mean of these data == correlation), which could be interesting to ya. Unfortunately I'm not sure it solves the differential # of timepoints issue though... that's tricky!
hi
thanks for the ideas of edge ts...
I think I'll have to keep the same # of time points per group; it's just too obvious/easy to critic (and it's a valid critic!!!)
best
-p
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.