Print

Print
Hi VM,

I would read this as indicating that modalities 1 and 3 aren't adding much information regarding the effect being sought not already provided by modality 2. I would still not argue for not collecting any of the other modalities based on this test alone. A reason is that, in the complete absence of signal in modalities 1 and 3, we would expect the result from NPC to be less strong than with modality 2 alone (that is as if modalities 1 and 3 were "diluting" the signal from modality 2). Another reason is that the effect for that partial test may be too weak, requiring large samples to be seen.

A great case for use of NPC is to combine different views from the same data. For example, using DTI metrics we can compute FA, MD, AD, RD, etc. Some of these may have different sensitivity profiles, yet all reflect some changes in the water diffusion. Combining them with NPC gives that extra strength to find differences that otherwise might have remained unseen.

Another case is to combine cortical thickness with surface area: both may change modestly in opposite directions, leading to no net volume changes. NPC is a good replacement for grey matter volume in this case, for keeping the sensitivity (we have a biorXiv paper on this).

One more case for NPC is to combine results of temporal ICA, investigating whether independent temporal time courses share a common spatial overlap across subjects.

I would argue that NPC shouldn't be used to suggest absence of effects for a partial test on the grounds that that test didn't change the results of the combination.

Hope this clarifies!

All the best,

Anderson


On 6 August 2017 at 11:38, neuroimage analyst <[log in to unmask]> wrote:
Thanks Anderson for the insight. 

I was expecting some additional regions to show up using npc which was not found using independent statistical tests with each modality because then one could make the argument of increase in power due to increase in modalities. 

However, in my case, the cluster that showed up overlapped with the cluster using modality 2 alone with no additional region showing significance.

Given such a scenario, my question is still very basic in terms of whether or not future studies should continue to acquire the 3rd modality as 2nd modality alone seems to lend enough power to find those regions that are different between the groups. 

But, again, may be that is not how it supposed to be interpreted and one could still argue for the acquisition of the 3rd modality. If I have to make the case for the acquisition of the 3rd modality; then how should I go ahead and make it?

I will greatly appreciate your response.

Thanks

Regards

VM 

On Sun, Aug 6, 2017 at 5:00 AM, Anderson M. Winkler <[log in to unmask]> wrote:
Hi VM,

I think the question is a bit ill posed... NPC combines the evidence against the null from the separate (partial) tests. It can be understood as a meta-analysis, except that, instead of summary results from different studies, we use the the actual subject-level data, and further, we use the same subjects, while taking into account, through permutation, the non-independence between the various measurements obtained per subject.

NPC can also be seen as a non-parametric counterpart to MANOVA with some nicer properties, such as the ability to identify the direction of the effects, both jointly or separately, and higher power, particularly as the number of modalities being combined increase.

The exact profile with which each partial test contribute to the final test is known from the combining function used with the test. See Figure 3 from our NPC paper. As you can see, for some functions, even non-significant results can contribute (together with others) to a significant joint effect. This isn't the same as "how much each has contributed" because there are non-linearities in the way these functions become significant, even more so when they aren't independent.

Hope this helps!

All the best,

Anderson


On 5 August 2017 at 13:58, neuroimage analyst <[log in to unmask]> wrote:
Hi,

I would greatly appreciate if anyone could provide me some input on how to interpret the results of multivariate statistics:

Briefly, I have 3 modalities and 2 groups and I ran statistics on each of these modalities separately and identified regions that are significantly different between the groups in each modality separately (FWER corrected).

Then I ran the 3 modalities using -npc in PALM and found a cluster (FWER) that overlaps with the significant cluster obtained using modality 2. 

Is there a way to understand how much has each modality contributed to the cluster obtained to be significant using -npc way? Or can this be interpreted as the future studies could only focus on modality 2 and no need to test for modality1 and modality3?

Thanks

Regards
-VM