I am currently analyzing rodent connectivity data with MRtrix and I was able to generate whole brain connectivity matrices which I then analyzed in NBS. I now would like to explore individual ROI’s.

Is it acceptable to utilize the whole brain matrices and simply zero out every row except the row of the relevant ROI and then analyze the new matrix in NBS? Or would this give different results than running tckedit with the relevant ROI?

Those are in fact two very different questions, so I need to separate them out:

Is it acceptable to utilize the whole brain matrices and simply zero out every row except the row of the relevant ROI and then analyze the new matrix in NBS?

It’s “acceptable” as far as being possible to do and having a logical interpretation as long as you actually understand what’s going on.

In NBS, following application of a test statistic threshold, the enhancement algorithm finds those supra-threshold edges that form a connected cluster and determines the size of the cluster; it is (the maximum of) this size that goes into the null distribution for shuffled data, and it is this size for non-shuffled data that is compared against the null distribution to form a p-value. If you zero out all edges except for one row, then every single remaining edge shares a node with every single other edge; it is impossible for two supra-threshold edges to not form a cluster. As such, what you think is NBS is actually just “the number of supra-threshold edges”. This is the quantity on which your null distribution will be based; and therefore, if you reject the null hypothesis, your conclusion should be that “the number of supra-threshold edges involving this node is greater than predicted by chance”. That statement does not have connection specificity, and I would encourage interpretation of the result in that way.

Or would this give different results than running tckedit with the relevant ROI?

This question fails to disambiguate the method of statistical inference with the derived quantitative parameter per edge.

Performing whole-brain tractography and then selecting streamlines will yield different streamline counts to performing targeted tracking, regardless of how those data are to subsequently be used. This is true not only of the relative densities of different edges, but also of the scaling between subjects of the sum of “total” connection density across all edges (see inter-subject connection density normalisation).