Cluster-forming threshold in SnPM

593 views
Skip to first unread message

Hill, Donal

unread,
Mar 30, 2016, 10:14:37 AM3/30/16
to snpm-s...@googlegroups.com
Dear experts,

I have been comparing the results of paired t-tests using SnPM and FSLs randomise. In both instances, I am comparing cluster extent results, where I use the “-c <thresh>” option in FSL.

When using FSL, you have to provide a t-statistic threshold as the value of <thresh>. However, within SnPM, the user provides an uncorrected p-value as input. I am therefore having trouble setting my SnPM and FSL analyses up to use exactly the same cluster-forming threshold.

When I run SnPM with a p < 0.001 threshold, I see the initial output:

Initialising...
Working on correct permutation...
Warning: Pseudo-T cluster-forming threshold defined by P-value using Gaussian approximation P=0.001 -> Z=3.0902; actual
Pseudo-T threshold unknown but may be higher than 3.0902.

where the Z value is what I would expect for such a p-value. I don’t know, however, what T value SnPM has settled on using. I would like to know this information, so that I can use precisely the same t-statistic threshold within my FSL analysis.

Any help would be fantastic!

Kind regards,
Donal






Hill, Donal

unread,
Mar 30, 2016, 1:57:46 PM3/30/16
to Hill, Donal, snpm-s...@googlegroups.com
Hi again,

Sorry for the noise! I realise now that SnPM can also take T-stat thresholds as input. Just means I have to rerun everything ;)

Cheers,
Donal

Thomas Nichols

unread,
Mar 31, 2016, 3:33:07 AM3/31/16
to Hill, Donal, snpm-s...@googlegroups.com
That's right! Use either p-value or statistic value to specify the cluster forming threshold.

-Tom
> --
> You received this message because you are subscribed to the Google Groups "Statistical Nonparametric Mapping" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to snpm-support...@googlegroups.com.
> To post to this group, send an email to snpm-s...@googlegroups.com.
> Visit this group at https://groups.google.com/group/snpm-support.
> To view this discussion on the web, visit https://groups.google.com/d/msgid/snpm-support/AA49B3AA-C01F-4BD4-B87A-8CA9D76C8D83%40kcl.ac.uk.
> For more options, visit https://groups.google.com/d/optout.

wang.z...@gmail.com

unread,
Jul 25, 2017, 11:20:43 PM7/25/17
to Statistical Nonparametric Mapping, donal...@kcl.ac.uk
But which value is suitable for the Cluster-forming threshold in SnPM?
0.001 or 0.0001?

Thomas Nichols

unread,
Jul 26, 2017, 2:39:57 AM7/26/17
to wang.z...@gmail.com, Statistical Nonparametric Mapping, donal...@kcl.ac.uk
The great thing about the non-parametric approach is that *any* threshold is valid. However if you use a threshold much below P=0.01 you may find that the clusters (even if significant) are too large to be useful. 

-Tom

__________________________________________________________
Thomas Nichols, PhD
Professor of Neuroimaging Statistics
Oxford Big Data Institute
Li Ka Shing Centre for Health Information and Discovery
Nuffield Department of Population Health
University of Oxford



wang.z...@gmail.com

unread,
Jul 26, 2017, 4:18:25 AM7/26/17
to Statistical Nonparametric Mapping, wang.z...@gmail.com, donal...@kcl.ac.uk
Thanks, Prof. Nichols,

But which threshold of p_unc for cluster forming is suggesttive most?
The website of snpm tells "For cluster-wise inference with p < 0.05 FWE corrected using a cluster-forming threshold of p < 0.0001", that means, 0.0001 is recogmended?

Thomas Nichols

unread,
Jul 27, 2017, 2:32:19 AM7/27/17
to wang.z...@gmail.com, Statistical Nonparametric Mapping, donal...@kcl.ac.uk
No no! That was juat an example. For reference, SPM uses a cluster forming threshold of uncorrected P=0.001 and FSL used to use P=0.01. You may want to look at Woo & Wager that suggest P=0.01 gives clusters that are too large to interpret & instead recommend P=0.001 (but I've also heard push back from users that P=0.001 can lack sensitivity, so it's not a settled issue).


-Tom

__________________________________________________________
Thomas Nichols, PhD
Professor of Neuroimaging Statistics
Oxford Big Data Institute
Li Ka Shing Centre for Health Information and Discovery
Nuffield Department of Population Health
University of Oxford



Reply all
Reply to author
Forward
0 new messages