Hi Dr. Winkler and FSL experts,
Thank you for your work on the PALM package, I’m a first time user but after doing some reading I think PALM has the ability to handle a number of analytic problems I have encountered while working on my master's project.
I have resting state data from a case-control study where subjects were sampled within a developmental period of interest. I am utilizing graph theoretical metrics to examine developmental differences in network (and nodal) segregation and integration between a clinical and control group. ROIs are defined a priori based on a relatively detailed nodal parcellation (421 nodes). These data have proved difficult because the distribution of a few of the graph metrics I am interested in (5 in particular: degree, eigenvector centrality, betweenness centrality, within-module degree z-score, participation coefficient) are non-normal and a few do not play well with simple transformations. Running some residual diagnostics on more straightforward parametric LME models were discouraging, which brought me to the world of permutation testing.
Further complicating the analytic picture, I have hoped to avoid a thresholding bias in graph construction and have thresholded and binarized adjacency matrices across 20 FC thresholds (there are reasons I chose FC thresholding other than proportional thresholding). So, essentially, I have a very large number of comparisons that need to be made (at each node, across thresholds). Making things even more complicated is the fact that my data are not exchangeable and to my understanding will need to include some blocked structure that represents both subjects and nodes.
I am looking for some guidance on how to get an analysis like this off the ground using your NPC approach. In previous investigations I stumbled upon a great R package called permuco that implements a number of techniques you’ve written about over the past few years. It generally works well when I run analyses at the level of the overall RS networks using TFCE (there are only 7 tests, so I am less concerned about a multiplicity of tests, the reason it makes sense to use TFCE on this data is to treat our thresholds as having a sort of "spatial structure" [i.e. a true effect should be conserved across multiple thresholds, similar to Drakesmith et al’s (2015) MTPC approach]). But at the node level, since we have 421 nodes I think a joint test as you suggest in the NPC paper is a much more principled approach.
Since the data I would like to input to PALM is probably closer to non-imaging data after all is said and done, I had a tough time finding other posts that explicitly address what I am hoping to do with my data. Essentially, I’d like to run one large model using NPC that simultaneously models main effects of groups and continuous age, and group x age interactions at each node (which I plan to extend to network-level analyses). I see that the modified NPC algorithm has the capability to compute spatial statisitcs such as TFCE, but am a bit unclear as to how this is best specified in a palm call. I was also getting a bit confused if in specifying a call to palm if it makes sense to set each threshold as a separate “modality” and include 20 separate columns with a given connectivity metric per subject and per node on the rows (i.e. 421 nodes x 84 subjects = 35,364 rows by 20 columns representing thresholds) OR if it’s correct to just include one column with thresholds, nodes, and subjects on the rows. If the latter case is correct if I am interested in computing spatial statisitcs how does the NPC call know how to group my thresholds?
Right now my best guess in specifying a proper call is:
palm -i nodes_by_subjects_by_thresholds.csv -d design.csv -t contrasts.csv -eb blocks.csv -vg auto -npcmethod Fisher -npccon -T -corrcon
Where this call would be run 5 separate times (one for each graph theoretic metric). As I mentioned, I guess I’m not sure what the format of the -i call should look like since this is going to be “non-imaging” (csv) data. in the documentation it says: “Each row constitutes a subject or observation, and each column represents a measurement.” But I am not sure of the proper translation to my case. Depending on this, the design, contrasts and blocks will look different. I also specified -npccon since the contrasts are what need to be estimated simultaneously, with a set of dummy codes to capture the effects of interest.
Thanks so much for your time (and apologies for the novel, hope this was clear enough!).
To unsubscribe from the FSL list, click the following link: