Print

Print
Hi there,

please let me chime in with a question:
If we’d use palm in the context of dual regression to test multiple ICs using multiple „pseudo“- modalities (using -i for each group IC) and corrmod (and corrcon), should the permutation scheme be synchronized? Should we use npc ?
Thanks!
Cheers,
Andreas

Von: FSL - FMRIB's Software Library <[log in to unmask]> on behalf of "Anderson M. Winkler" <[log in to unmask]>
Antworten an: FSL - FMRIB's Software Library <[log in to unmask]>
Datum: Donnerstag, 22. Juni 2017 um 19:42
An: <[log in to unmask]>
Betreff: Re: [FSL] PALM : Synchronized permutations

Hi Jay,

Thanks for the comments. Please see below:

On 20 June 2017 at 12:00, SUBSCRIBE FSL Jay <[log in to unmask]> wrote:
Hello Anderson Winkler,

I am trying to read and understand your article "Non-Parametric Combination and Related Permutation Tests for Neuroimaging" specifically the idea behind performing synchronised permutations. I came across the following text in the article

"Moreover, non-independence does not need to be explicitly modelled, either between observations, between partial tests, or across space for imaging data, thus making such tests applicable to a wide variety of situations."

1) I understand that dependency between the input modalities need not be modelled between partial tests. Then, why synchronised permutations is performed for all the partial tests? Why not non-synchronised permutations?

It's precisely the synchronised permutation scheme that allows the dependencies to be captured implicitly. By "synchronised" it is meant that however the observations are reordered in a given permutation for one modality, it is done exactly in the same way for the other(s).

One way of stating this is that, as we permute, for each permutation we ask: "what my test statistics would look like if my observed data were like this?". Answering this question needs that the data is rearranged in the same way for all modalities.

The same applies to NPC and to FWER-correction using the distribution of the maximum statistic (which is, in fact, a form of NPC too).


2) Why is the combining function not applied directly on the input modalities? What is the idea behind partial p-values?

The combination is of the evidence for (or against) the null hypothesis, and this is summarised by the test statistic and its p-value. Moreover, not always it is meaningful to directly combine data that come in different units, scales, or represent different physical quantities. How to combine, for example, blood flow measured with PET and fractional anisotropy measured with diffusion MRI? Combining evidence towards an hypothesis is feasible, but these modalities can't trivially combined otherwise.

Hope this helps!

All the best,

Anderson

 

Thank you for your time.
Best wishes
Jay