Dear CONN users,

We want to conduct a multi-voxel pattern analysis (MVPA) on a resting-state fMRI data from a group of patients. We have two clinical profiles. In our analysis we want to investigate the functional sources of these profiles. Basically we have three seed regions: the nucleus accumbens, the putamen and the caudate. We want to make use of MVPA to identify the different functional connectivity patterns across the profiles. Do you think that this technique would be useful for this question and could you recommend some guidelines about how to conduct the MVPA to answer this question?

All the best!

Dear Eylul,

MVPA is a great solution when you have either no a priori seeds or a priori seeds that are relatively large (and expectedly heterogeneous in their functions or patterns of functional connectivity). In your case, these three regions (nucleus accumbens, putamen, caudate) are relatively/comparatively small, so, at least as an initial approximation, it may be simpler to perform a standard multivariate seed-based connectivity analyses using your three regions as separate seeds, and simply using a [1 0 0; 0 1 0; 0 0 1] contrast across seeds in order to identify differences in functional connectivity patterns with any of these three seeds.

That said, the main limitation of the above approach is that, while it takes into account the differences in connectivity patterns across the three seeds, it still implicitly assumes that within each of these three regions connectivity patterns are relatively homogeneous (e.g. so that looking at the average BOLD timeseries within the caudate is a good representation of the BOLD signal across all voxels within the caudate). MVPA allows you to go beyond that restriction/assumption by analyzing the connectivity patterns within those regions using a multivariate data-driven representation with arbitrary number of dimensions instead of an a priori user-defined parcellation with three components, one per region, separately, as seed-based analyses would use. To run these analyses, simply define a new masked-MVPA first-level analysis (select the 'mask' option in CONN MVPA first-level tab and enter a mask that covers your three regions of interest) and then perform the same second-level analyses as before (e.g. between-group differences) using the resulting components now instead of the three a priori seeds as in the original seed-based analyses. In practice, in this case, given the relatively small ROIs, I still imagine that perhaps only a few components (perhaps 3 or even less) should suffice to cover most (e.g. above 90%) of the variability in connectivity patterns across those three areas, but feel free to explore and/or increase that number (e.g. the files PCAcov_*Component.nii will show how much variance is explained by the first N components; typically it is not recommended to increase the number of components beyond ~1/10th of the number of subjects in order to keep reasonable sensitivity/power of second-level statistics)

Hope this helps
Alfonso
Originally posted by Eylul Turan: