For my fixel-based analysis study I am trying to visualize the mean FDC at particular tracts as percent reductions from the control group mean in a bar plot, similar to what Zarkali et al. did for their 2020 Neurology paper, and what Mito et al. did in their 2018 Brain paper. I had a question about how the authors extracted the data for making these plots.
From my understanding and based on how I set up my design and contrast matrices I first get the control group mean from adding the beta0 and beta1 files from the fixelcfestats output. I then calculate the percent effect for FDC in the patient group at each tract by dividing the absolute effect by the control group mean and multiplying by 100. Then I threshold this percent effect by each tract mask that I obtained from registering an atlas to my population template space, performing masked tckgen at the tract of interest, and using tck2fixel to get the corresponding fixel mask at that tract, and then I output the mean percent effect. I was wondering if this is the correct way to go about plotting these mean percent reductions, because I wasn’t sure about how I could calculate the confidence intervals at each tract with the percent effect for all patients relative to the control group like what Mito et al and Zarkali et al did.
Thanks for all your help!
From my understanding and based on how I set up my design and contrast matrices I first get the control group mean from adding the beta0 and beta1 files from the fixelcfestats output.
I have always disliked the documented instructions regarding generation of derivative quantities based on the GLM beta coefficients. Those instructions are predicated on specific details that aren’t really communicated properly. I’d say that this is satisfactory as long as:
The first column is a global intercept containing all 1’s;
The second column is a group classification, with 1 being the control group and -1 being the experimental group;
All other columns in the design matrix are demeaned.
Even then, it’s only actually exactly the control group mean if there is no group-wise difference in your nuisance regressors.
These instructions originate from back when
fixelcfestats was responsible for fixel-fixel connectivity matrix generation / fixel data smoothing / permutation testing. The beta coefficients were the only way to get at the smoothed fixel data. However now that these components are spread across three separate commands, you could just take your smoothed fixel data and directly calculate the group mean using
I then calculate the percent effect for FDC in the patient group at each tract by dividing the absolute effect by the control group mean and multiplying by 100.
If your design matrix has a column where 1 indicates control and -1 indicates experimental group, and your t-test contrast extracts the value of that beta coefficient (so the value of the “absolute effect” is equivalent to the value of that beta coefficient), then that absolute effect actually encodes half of the difference between those two groups, since what that beta coefficient encodes is the predicted magnitude of the change in FDC for a unitary change in the “explanatory variable”, but the magnitude of the group difference in the “explanatory variable” is actually 2.
Again, the instructions in that document only apply in specific circumstances, and that’s really not communicated well…
I wasn’t sure about how I could calculate the confidence intervals at each tract with the percent effect for all patients relative to the control group like what Mito et al and Zarkali et al did.
I’m guessing that confidence intervals were calculated based on an inferred normal distribution and the fixel-wise standard deviation provided by the GLM. I would think that upper and lower bounds would be calculated per fixel and then these would be averaged across each tract mask, but you would need to ask the authors to confirm.