SnPM - cluster-level inference - two-sample t-test

[nitinw...@gmail.com]

Dear SNPM community,

I have the below problem with cluster-level inference, when running 2-sample t-test, 5000 permutations, Cluster inference: "Yes (slow)". I have no problems getting voxel-wise inference.

I did download the updates in: http://www2.warwick.ac.uk/fac/sci/statistics/staff/academic-research/nichols/software/snpm/distribution/snpm13_updates

cluster-forming threshold I tried were: 0.0001, 0.001, 0.01.
significance level: uncorrected k = 0

Running job #1
------------------------------------------------------------------------
Running 'Inference'

SnPM: snpm_pp
========================================================================
Warning: Pseudo-T cluster-forming threshold defined by P-value using Gaussian approximation P=-1 -> Z=1.325;
actual Pseudo-T threshold unknown but may be higher than 1.325.
> In snpm_pp at 534
In snpm_run_pp at 20
In matlabbatch/private/cfg_run_cm at 26
In cfg_util>local_runcj at 1672
In cfg_util at 953
In cfg_ui>MenuFileRun_Callback at 707
In gui_mainfcn at 96
In cfg_ui at 53
Working on spatial extent...
Computing suprathreshold voxels...done
Loading & conditioning SupraThreshold statistics...Failed 'Inference'
Error using snpm_pp (line 831)
Unable to read MAT-file /imaging/nw04/MARpipeline/lc_grph99_unnrm_p84_stopfilt_transdef_mf_tsss/SnPM_ST.mat

File may be corrupt.
In file "/imaging/rh01/SPM/spm12_fil_devel/toolbox/SnPM13/snpm_pp.m" (???), function "snpm_pp" at line 831.
In file "/imaging/rh01/SPM/spm12_fil_devel/toolbox/SnPM13/config/snpm_run_pp.m" (v1716), function "snpm_run_pp" at line 20.

The following modules did not run:
Failed: Inference

I would be grateful for any advice/help on above.

Thanks,
Nitin

[Thomas Nichols]

Dear Nitin,

Well, as a sanity check, can I confirm that cluster inference with "Yes (fast)" works?

There is something odd with your warning message.. it reports a P-value of "-1" and a Z value of "1.325", which corresponds to a P of 0.0930.   Examining the code, I can't see how you could ever get a P-value of -1.

Can you say exactly what cluster-forming threshold you used to get this error?

-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.
For more options, visit https://groups.google.com/d/optout.

--

__________________________________________________________
Thomas Nichols, PhD
Professor, Head of Neuroimaging Statistics
Department of Statistics & Warwick Manufacturing Group
University of Warwick, Coventry  CV4 7AL, United Kingdom

Web: http://warwick.ac.uk/tenichols

Email: t.e.n...@warwick.ac.uk
Tel, Stats: +44 24761 51086, WMG: +44 24761 50752
Fx,  +44 24 7652 4532

[Nitin Williams]

Hi Tom,

Thank you for your reply.

That warning, with that p-value (-1) and z-value (1.325) occurs for every cluster-forming threshold I tried i.e. 0.01, 0.001 and 0.0001.

We havn't yet tried cluster-inference with Yes (fast) option, so will do this and get back to you. just to clarify, with the Yes (fast) option, the cluster-forming threshold should be set to NaN and significance level can be set to: uncorrected k=0?

please find attached the job-file we used (snpm_batch_job.m).

the code we ran was:

jobfile = {'/imaging/nw04/MARpipeline/scripts/snpm_batch_job.m'};

jobs = repmat(jobfile, 1);

inputs = cell(3, 1);

inputs{1} = {sprintf('/imaging/nw04/MARpipeline/lc_grph%d_unnrm_p%d_%s_%s',PctCut,sel_ord,FiltName,basename)};

inputs{2} = {lcP{1:(Nsub/2)}};        % 2 Groups: Two Sample T test; 1 scan per subject: Group 2 scans - cfg_files

inputs{3} = {lcP{((Nsub/2)+1):Nsub}}; % 2 Groups: Two Sample T test; 1 scan per subject: Group 2 scans - cfg_files

spm_jobman('run', jobs, inputs{:}); 

any advice/suggestions much appreciated, especially on the job-file.

Regards,

Nitin

[Thomas Nichols]

Hi Nitin,

I'll have to dig into this and try to reproduce it... though I won't be able to get to this until next week.

I'll tell you where I'll start:  I'll try to create this analysis using the batch interface and see if I get the error.  If you can try to replicate the error via the batch interface, please let me know.

That warning, with that p-value (-1) and z-value (1.325) occurs for every cluster-forming threshold I tried i.e. 0.01, 0.001 and 0.0001.

OK, weird.

 

We havn't yet tried cluster-inference with Yes (fast) option, so will do this and get back to you. just to clarify, with the Yes (fast) option, the cluster-forming threshold should be set to NaN and significance level can be set to: uncorrected k=0?

No. The variable job.ST.ST_U has to be set to the P-value or statistic value cluster-forming threshold, and make sure that job.ST.ST_later doesn't exist.  

Actually, I can't even assure you that that will capture all the changes.  The safest thing is to load the batch job into the batch system and use the interface to select 'fast', and then you can set a cluster-forming threshold (it defaults to 2.3).

I'll get back to you as soon as I can, but do keep me posted if you discover anything.

-Tom

[Nitin Williams]

Hi Tom,

I've started cluster-inference (fast) through the batch interface, with cluster-forming threshold as default (2.03) - should be finished in around 3 days, so will let you know if it works fine then.

Thanks,

Nitin

[Thomas Nichols]

3 days!?  Can I ask what is your sample size and number of voxels? I've never had such a slow runtime since the 90's. 

-Tom