Hi Niels,

Thanks for the comments. Please see below:

On 13 September 2017 at 05:23, Niels Bergsland <[log in to unmask]> wrote:

Hi Anderson,
Thanks for the great PALM tool and all the work you continue to put into it!

I have been using NPC for joint inference on three modalities, M1, M2 and M3 comparing two groups of subjects.

I looked at the partial test results and in terms of the number of suptrathreshold voxels, the order is M2 > M1 > M3

When considering the NPC analysis of M2 and M1, there is a very modest increase in the suprathreshold voxels. This was more or less expected since M2 and M1 tend to be highly correlated with one another.

When considering the NPC analysis of M2 and M3, there is a substantial increase in the total number of supra threshold voxels. In my study, this is particularly interesting and so all is good!

However, out of curiosity, I tried running NPC with M1, M2 and M3 together. What I found surprising is that in this case, the number of suprathreshold voxels is actually a little bit smaller than using just M2 and M3 (about a 4% difference). While the overall interpretation remains the same, I'm a bit puzzled about this and was wondering if you could explain why this is happening. All of these are with respect to FWE corrected results.

I'm using the following options: -T -tfce2d -npc -accel tail -n 500 and the default Fisher combining function.

There are two things here: first is that some voxels in M1 may have an effect that goes in the opposite direction than the other two. The second is that TFCE introduces some non-linearities in the statistical map that may distort (for good or for bad) some of the relationships. For example, a less strong effect that is well spread may compensate a local, focal, strong effect. None of this is per se a problem, just a feature of the test statistic. If you believe that this is complicating the interpretation of the joint analysis, have a look without TFCE.

Finally, looking at the uncorrected data, the number of suprathreshold voxels is actually slightly higher when using all three modalities, so it seems that it has to be related to the FWE correction.

This is expected: the corrected results will consider the thousands of voxels and so, less significant findings.

Of course I'm going with the FWE corrected results but looked at these just to try to understand what's happening here. I just want to make sure that I'm not missing/misunderstanding something in reference to the part on the PALM wiki page where it states "...as the number of modalities being combined increases, the power of the test also increases...,"

Yes, power increases provided that there is signal in all modalities. This contrasts with classical tests as MANOVA where there is a point of maximum power after which, even if there is true signal in all modalities, power starts to fall. Also, power increases more steeply as the number of modalities increase than in MANOVA.

Of course, simply adding more modalities doesn't increase power: modalities with no signal, if included, will reduce power (as if diluting the effect).

Hope this helps!

Cheers,

Anderson

I hope that this is all clear and appreciate any feedback you have.

Thanks!
Niels

--
Niels Bergsland
Integration Director / Research Assistant Professor of Neurology
Buffalo Neuroimaging Analysis Center / University at Buffalo
100 High St. Buffalo NY 14203
[log in to unmask]