[color=#000000]Hi Roger,[/color]
[color=#000000]That makes sense. I would suggest, instead of analyzing each component separately (which would require a somewhat conservative multiple comparison correction to account for all those individual analyses), to simply test the first-few components (e.g. the first three or four components; just testing the first two components would be a bit biased by the results that you have already observed). [/color]
[color=#000000]To clarify, typically in MVPA analyses you want to test the first-N components (and use an eye(N) between-measures contrast) to evaluate potential between-condition and/or between-subject differences in connectivity. Yet, the choice of exactly how many components N to best include in your MVPA second-level analyses depends on the actual data. If the differences in connectivity between your groups are somewhat strong/salient (compared to the between-subjects residual differences in connectivity) they will typically be captured by the first few MVPA components, while if they are relatively weak they will typically be captured by higher-order MVPA components. So, ideally, you would want to include as many components as possible to make sure that those between-group differences in connectivity are well captured/represented by your group-agnostic MVPA components. Yet, practically, the power of your second-level analyses to detect those between-group differences in connectivity drops as you include more and more MVPA components in those analyses, mainly from the associated drop in degrees of freedom when simultaneously testing more measures. The number of components extracted by the first-level MVPA analyses (in your case 10 components, this is defined by default by attempting to keep approximately a 1:5 ratio between the number of components extracted and the number of subjects in your study) represents roughly a higher-bound on the number of components you might possibly want to test before the power drops dramatically, but it does not necessarily represent the "optimal" number of components that you would want to test (particularly if the number of subjects in your study is not too low). Typically you need to define a priori how many MVPA components you would like to include in your second-level analyses. To play it conservatively (if you do not have pilot or previous experimental data to base this descision on), it is not a bad idea to plan a couple of N values and associated second-level analyses (e.g. one using only a few components, e.g. 3 or 4, and another using more components, e.g. close to the 1:5 ratio above). In this way you would only need to apply a multiple comparison correction across these two analyses (compared for example to testing each component separately which would require a multiple comparison correction across all those individual components). In my experience 3/4 components is becoming my default choice for many MVPA analyses unless I have considerably large datasets and I know a priori that the effects that I am looking for a relatively subtle/weak (and of course unless I have very few subjects and I cannot even reasonably afford 3/4 components, in which case I use the default 1:5 ratio value), but your experience might vary. Last, again just to clarify since I have not explicitly mentioned this, because MVPA components are obtained using PCA, when one extracts M components in the first-level MVPA analysis step, and then only includes N (N<=M) components in the second-level analyses, that is exactly equivalent to having just extracted N components in the original first-level analyses and then including all of them in the second-level analyses, so the choice of N (how many components to actually include in your analyses) can be delayed to the second-level analysis step as long as you have extracted [i]at least[/i] that many components in your first-level MVPA analyses (which is why the default in CONN is set to the 1:5 ratio rule which roughly represents this higher-bound on the number of components one would typically want to include in their second-level analysis step)[/color]
[color=#000000]Hope this helps[/color]
[color=#000000]Alfonso [/color]
[i]Originally posted by Roger Mateu:[/i][quote]Thanks Alfonso! It helps so much.
I have another important question about MVPA
I'm using Conn MVPA method to extract 10 between-groups ([1 -1]) components (I have 52 subjects), which clusters must be used as a seed in a second-level analysis. I have some questions about it:
As I read in this forum, the proper way to do it is selecting all components (eye(10)) on voxel-to-voxel between-measures second-level analysis step (and [1 -1] between-group selection) , and use each resulting significant cluster as a seed. Is this correct?. This second level Seed-to-voxel analysis showed few low-signification results.
In contrast, using each significant cluster of each component independently, offered high-significant and meaningful results, specially clusters from the two first components. As you explained in this forum, individual MVPA components do not have simple interpretations, but our results suggest that first and second component clusters contain areas with divergent iFC between patients and controls (and high correlation with structural analysis differences). How can I explain this situation in order to statistically validate our results with this last method?
Thanks again
Roger[/quote]