SnPM for SPM5, SnPM5b, released July 10th, 2006!: See release notes and download information
SnPM3, the latest SnPM toolbox for SPM2 is available.: See below for details.
A guide for entering FSL data into SnPM has been added to the web-documentation.: Click the link for FSL users here or below left.

SnPM: Statistical nonParametric Mapping

The Statistical nonParametric Mapping toolbox provides an extensible framework for voxel level non-parametric permutation/randomisation tests of functional Neuroimaging experiments with independent observations. The SnPM toolbox provides an alternative to the Statistics section of SPM. SnPM uses the General Linear Model to construct pseudo t-statistic images, which are then assessed for significance using a standard non-parametric multiple comparisons procedure based on randomisation/permutation testing. It is most suitable for single subject PET/SPECT analyses, or designs with low degrees of freedom available for variance estimation. In these situations the freedom to use weighted locally pooled variance estimates, or variance smoothing, makes the non-parametric approach considerably more powerful than conventional parametric approaches, as are implemented in SPM. Further, the non-parametric approach is always valid, given only minimal assumptions.

Because the appropriate permutations of the data labellings is highly dependent on the experimental design, different designs must be treated individually. This is handled within SnPM using a PlugIn architecture for "design modules". Currently, PlugIn modules are available for:

SingleSub: Two Sample T test; 2 conditions, replications
SingleSub: Simple Regression (correlation); single covariate of interest
MultiSub: One Sample T test on diffs/contrasts; 1 condition, 1 scan per subject
MultiSub: Simple Regression (correlation); single covariate of interest, 1 scan per subject
MultiSub: Paired T test; 2 conditions, 1 scan per condition
MultiSub: Randomized Design Paired T test; 2 conditions, replications
MultiSub: Within Subject ANOVA; multiple scans/subject
2 Groups: Test diff of response; 2 conditions, 1 scan per condition
2 Groups: Two Sample T test; 1 scan per subject
>2 Groups: Between Group ANOVA; 1 scan per subject

Additional PlugIns will be made available as they are built.

The approach is described in full in Holmes (1994); Holmes (1996) and Nichols & Holmes (2001). (See references below.) These pages give a brief overview of the methodology, and describe the software, it's installation and use.

Statistical nonParametric Mapping refers to the enterprise of making statistical inferences on volumetric statistic images with minimal assumptions using non-parametric statistical techniques. SnPM refers to an implementation of Statistical nonParametric Mapping by Andrew Holmes and Tom Nichols.

Statistical nonParametric Mapping - A quick primer

SnPM uses a permutation test to determine the signifiances of voxels and clusters of supra-threshold voxels. A basic understanding of the permutation test is essential for knowledgeable use of the SnPM toolbox. Fortunately, the basic concept is simple and intuitive: For example, consider a simple two sample problem, with one observation being made on each member of the two groups. If there is really no difference between the two groups, we would be fairly surprised if most of the group one observations were larger than the group two observations. Permutation tests provide a formal mechanism for quantifying this ``surprise'', leading to significance tests.

SPM - assumptions & low df problems...

The parametric approach embodied in SPM uses the General Linear Model to produce a statistic at every voxel, and assesses the resulting statistic image for significant regions using distributional approximations for a continuous random fields with the same marginal distribution and smoothness. This explicitly assumes that the data are derived from strictly stationary homogeneous discrete Gaussian random fields. Further, it is implicitly assumed that the statistic image is sufficiently smooth that it's properties may be approximated by a continuous random field. However t (& F) statistic images with low (denominator) degrees of freedom are very noisy (due to the unreliability of voxel variance estimates from low numbers of observations). These noisy t & F-statistic images have such low estimated smoothness that a continuous random field with the same smoothness would have features at sub-voxel resolution. The result is that these features affect the distribution of the maximal statistic of the continuous field, against which the voxel values of the statistic image are compared for significance, resulting in a conservative test (i.e. significance is under-estimated).

SnPM - a primer...

In contrast to the multitude of assumptions and approximations of the parametric approach, the non-parametric approach exploits the design of the experiment (randomisation test), or makes simple intuitive distributional assumptions, such as symmetry (permutation test). The thinking is simple: If there is really no experimental effect, then the experimental labels, whether A's and B's (for an activation study) or a set of real numbers (for a covariate study), are arbitrary. Any reallocation of the labels to the scans would lead to an equally plausible statistic image. So, considering the statistic images associated with all possible re-labellings of the data, we can derive the distribution of statistic images possible for this data. We can then test our hypothesis of no experimental effect by comparing the statistic for the actual labelling of the experiment with this distribution. Effectively we quantify our surprise at the observed findings on the basis of what we would expect were there no experimental effect! If, out of N possible relabellings the actual labelling gives the r^th most extreme statistic, then the p-value is r/N. The details are worked out in the references given below.

Assumptions...

SPM makes many strong assumptions about the nature of your data, while SnPM makes a few weak assumptions: SPM assumes that your data, at each voxel, are normally distributed and, across voxels, are derived from continuous random fields with a stationary covariance structure. SnPM assumes that, under the null hypothesis of no experimental effect, you can switch the experimental labels of your data and always expect the same (null) result.

Variance smoothing...

The freedom of SnPM from parametric distributional assumptions allows us to pool variance estimates over neighbouring voxels, under the mild assertion that the true variance image is smooth, giving additional degrees of freedom. The Pseudo-t statistics computed with such a smoothed variance image don't exhibit the noise of low df t-statistic images, This variance smoothing circumvents the low df noisy statistic image problem discussed above, giving the non-parametric approach greater power than it's parametric counterpart.

Next...

Understanding the assumptions of SPM requires an understanding of normality, independence and the basics of Gaussian random field theory. Understanding the assumptions of SnPM requires an understanding of exchangability. See the example page for a brief introduction to exchangability and the literature for more detailed discussion.

...to top

SnPM5b - released 10^th July 2006

This is the SnPM toolbox for SPM5. While it has been tested in-house, we have received limited feedback from external users and so it is still considered a beta release.

SnPM5b has all the features of SnPM3 plus the same image read and write functionality as SPM5. Specifically, it will read Analyze and NIfTI images and write NIfTI images.

If you have previously used SnPM2 but haven't used SnPM3, please read the notes below on changes from SnPM2

Currently SnPM5b uses the sequential type interface of SPM2. We are working to make it use the same hierarchical job manager as SPM5.

...to top

SnPM3

The final release of SnPM3 was made on 6^th April 2006, a toolbox for use with SPM2. While bugs will be fixed, no further development will be made on SnPM3.

What's new in SnPM3

Improved cluster size inference
Previously in SnPM, cluster size inference was very slow and resulting in a gigantic SnPM_ST.mat (2GB or more!) being created. Now, a cluster-defining threshold can be set before the computation stage, avoiding the gigantic data file. The result is the capability of analyzing much larger images and faster 'Results' processing.
Note, however, if you prefer, you can still use the old method; in the 'Configure' step, if you elect to "Collect Supra-Threshold stats" and choose not to "Define threshold now", then an SnPM_ST.mat file will be created and you may set the threshold in the Results step.
Nonparametric uncorrected cluster-level P-values
Previously SnPM only had corrected P-values for cluster size inference. Based on an assumption of stationarity, or homogeneous smoothness, SnPM now offers cluster-level uncorrected P-values.
Parameter estimates and statistic values saved in images
A significant improvement of SnPM3 is that all parameter estimates and statistic images are written out as images instead of being stored in .mat files. This makes it easier to take the results of SnPM and interrogate the results in other programs.
'Results' questions reorganized
With the introduction of uncorrected P-values, there are now two distinct images upon which inference can be based: T images and P-value images. Conventionally, in parametric statistics, there is a single 1-to-1 mapping between T- and uncorrected P- values ; with non-parametric P-values determined locally at each voxel, this is not true. For example, voxel A may have a bigger T-value than voxel B, but Voxel B may have the more significant P-value! This is because we do not assume all voxels have a common null distribution, and, here, voxel A's null distribution indicates that large values are more typical under the null than for voxel B.
To help navigate through the the possible options (T/P image, cluster/voxel, FDR/FWE/uncorrected), we have changed the sequence of questions asked in 'Results' step.
The following is an outline ... In the following bold texts refer to questions asked by the interface, while italics are the possible answers.
Results for which image?

T image
Inference Method?
Voxel-level inference
Corrected Threshold?
FWE
Interp: Chance of one or more false positive voxels controlled
Uncorrected

Cluster-level inference (Only if supra-threshold stats have been collected)
Corrected Threshold?
FWE
Interp: Chance of one or more false positive clusters controlled
Uncorrected

P image

Voxel-level inference
Corrected Threshold?
FDR
Interp: Long-run fraction of false positives among detected voxels controlled
Uncorrected

Distribution and installation instructions are given below.

Please report bugs to <snpm-authors at umich.edu>. Peculiarities and general queries should be raised on the SPM email discussion list. We will report patches to the list and make updates available if necessary. See the SnPM3 bugs & fixes section for previously reported bugs and fixes.

SnPM2 - released 16^th March 2005

What was new in SnPM2

SPM2 Compatibility
In addition to using SPM2 library routines, the results output format is closer to the SPM2 format. Also, SnPM2 has the 'click-to-see' feature of SPM2's results page.
Nonparametric uncorrected voxel-level P-values
Previously, SnPM99's uncorrected P-values were based on parametric theory. So, for example, there were no uncorrected P-values for the smoothed variance pseudo-t images, since no parametric theory exists for pseudo-t.
SnPM2 provides nonparametric uncorrected P-values at each voxel. An image of nonparametric uncorrected P-values is always written out, and they are displayed on the results page. The distribution of these P-values across the image is shown before the results page.
FDR-corrected (voxel-level) P-values
FDR-corrected P-values are now shown in the results page, and a FDR plot is shown (a P-P plot). They are computed using the Benjamini-Hochberg method, using nonparametric uncorrected P-values.
Combining toolbox
Enhanced cluster size inference can be obtained by combining voxel intensity with cluster volume. A voxel-cluster combining toolbox has been created for the SnPM toolbox. This sub-toolbox impliments cluster mass [Bullmore, et al., HBM 7:38-48, 1999], and a generalization of Poline's voxel-cluster test [Poline et al., NI 5:83-96, 1997].
These methods are characterized and described in "Combining Voxel Intensity and Cluster Extent with a Permutation Test Framework," Hayasaka & Nichols, NeuroImage 23(1):54-63, 2004.
Eventually these mothods will be fully integrated with SnPM.

SnPM distribution & installation

The SnPM toolbox is a small suite of MatLab functions and M-files that integrate with SPM2 routines. Thus, to use SnPM3 you must have already installed SPM2, and to use SnPM5 you must have already installed SPM5. See the SPM distribution page for SPM hardware & software requirements, downloading & installation instructions, and conditions of distribution. You should be familiar with SPM before using SnPM.

The software is available via FTP, but we ask you to complete a brief registration form prior to downloading. Having completed the form, you will be directed to the download location, via a keyword enabled FTP URL. You should also periodically check the SnPM3 bugs & fixes area, and replace older files shipped in the standard distribution.

Note SnPM99 and SnPM2 are still availble. Use the same registration form to request the SnPM99 or SnPM2 FTP URL.

Installation is relatively straightforward. Download the gzipped tar file and unzip it in a location where you want the snpm3 directory. Then, add the full pathname of this directory to your MATLABPATH. (You can check this by typing which spm in MatLab and examining the path, or try spm ver.)

For step by step instructions on downloading and installing SPM distributions, refer to the SPM distribution page.

...to top

Documentation

In addition to the main SPM documentation, you are encouraged to read

The appropriate peer reviewed articles, referred to below.
The SnPM manual page, duplicated in the SnPM online help facility.
The example page.
Basic non-parametric statistical texts, such as Good (1994) & Edgington (1980) will help clarify the underlying concepts of permutation/randomisation testing.

References

Statistical Issues in functional Brain Mapping Holmes AP (1994) Doctor of Philosophy Thesis, University of Glasgow, December 1994. Non-Parametric Analysis of Statistic Images From Functional Mapping Experiments Holmes AP, Blair RC, Watson JDG, Ford I (1996) Journal of Cerebral Blood Flow and Metabolism 16:7-22 Nonparametric Analysis of PET functional Neuroimaging Experiments: A Primer Nichols TE, Holmes AP (2001) Human Brain Mapping, 15:1-25. Holmes & Watson, on ``Sherlock'' We reply to Halber et al.'s ``Performance of a Randomization Test for Single-Subject ¹⁵ O-Water PET Activation Studies'' published in the Journal of Cerebral Blood Flow and Metabolism 171033-1039. Halber et al assert that our non-parametric approach (their implementation of which they dub `Sherlock') is less powerful than a ``standard'' analysis. This conclusion is at variance with our findings, which we consider is simply due to the fact that the ``standard analysis'' they compare to does not strongly control experimentwise Type~I error. Randomization Tests Edgington ES (1980) Marcel Dekker, New York & Basel Permutation tests: A practical guide to resampling methods for testing hypotheses Good P (1994) Springer-Verlag, New York: Holmes AP (1994)
Doctor of Philosophy Thesis, University of Glasgow, December 1994.
Non-Parametric Analysis of Statistic Images From Functional Mapping Experiments Holmes AP, Blair RC, Watson JDG, Ford I (1996) Journal of Cerebral Blood Flow and Metabolism 16:7-22 Nonparametric Analysis of PET functional Neuroimaging Experiments: A Primer Nichols TE, Holmes AP (2001) Human Brain Mapping, 15:1-25. Holmes & Watson, on ``Sherlock'' We reply to Halber et al.'s ``Performance of a Randomization Test for Single-Subject ¹⁵ O-Water PET Activation Studies'' published in the Journal of Cerebral Blood Flow and Metabolism 171033-1039. Halber et al assert that our non-parametric approach (their implementation of which they dub `Sherlock') is less powerful than a ``standard'' analysis. This conclusion is at variance with our findings, which we consider is simply due to the fact that the ``standard analysis'' they compare to does not strongly control experimentwise Type~I error. Randomization Tests Edgington ES (1980) Marcel Dekker, New York & Basel Permutation tests: A practical guide to resampling methods for testing hypotheses Good P (1994) Springer-Verlag, New York: Holmes AP, Blair RC, Watson JDG, Ford I (1996)
Journal of Cerebral Blood Flow and Metabolism 16:7-22
Nonparametric Analysis of PET functional Neuroimaging Experiments: A Primer Nichols TE, Holmes AP (2001) Human Brain Mapping, 15:1-25. Holmes & Watson, on ``Sherlock'' We reply to Halber et al.'s ``Performance of a Randomization Test for Single-Subject ¹⁵ O-Water PET Activation Studies'' published in the Journal of Cerebral Blood Flow and Metabolism 171033-1039. Halber et al assert that our non-parametric approach (their implementation of which they dub `Sherlock') is less powerful than a ``standard'' analysis. This conclusion is at variance with our findings, which we consider is simply due to the fact that the ``standard analysis'' they compare to does not strongly control experimentwise Type~I error. Randomization Tests Edgington ES (1980) Marcel Dekker, New York & Basel Permutation tests: A practical guide to resampling methods for testing hypotheses Good P (1994) Springer-Verlag, New York: Nichols TE, Holmes AP (2001)
Human Brain Mapping, 15:1-25.
Holmes & Watson, on ``Sherlock'' We reply to Halber et al.'s ``Performance of a Randomization Test for Single-Subject ¹⁵ O-Water PET Activation Studies'' published in the Journal of Cerebral Blood Flow and Metabolism 171033-1039. Halber et al assert that our non-parametric approach (their implementation of which they dub `Sherlock') is less powerful than a ``standard'' analysis. This conclusion is at variance with our findings, which we consider is simply due to the fact that the ``standard analysis'' they compare to does not strongly control experimentwise Type~I error. Randomization Tests Edgington ES (1980) Marcel Dekker, New York & Basel Permutation tests: A practical guide to resampling methods for testing hypotheses Good P (1994) Springer-Verlag, New York: We reply to Halber et al.'s ``Performance of a Randomization Test for Single-Subject ¹⁵ O-Water PET Activation Studies'' published in the Journal of Cerebral Blood Flow and Metabolism 171033-1039.
Halber et al assert that our non-parametric approach (their implementation of which they dub `Sherlock') is less powerful than a ``standard'' analysis. This conclusion is at variance with our findings, which we consider is simply due to the fact that the ``standard analysis'' they compare to does not strongly control experimentwise Type~I error.
Randomization Tests
Edgington ES (1980)
Marcel Dekker, New York & Basel
Permutation tests: A practical guide to resampling methods for testing hypotheses
Good P (1994)
Springer-Verlag, New York; Halber et al assert that our non-parametric approach (their implementation of which they dub `Sherlock') is less powerful than a ``standard'' analysis. This conclusion is at variance with our findings, which we consider is simply due to the fact that the ``standard analysis'' they compare to does not strongly control experimentwise Type~I error.
Randomization Tests
Edgington ES (1980)
Marcel Dekker, New York & Basel
Permutation tests: A practical guide to resampling methods for testing hypotheses
Good P (1994)
Springer-Verlag, New York
Randomization Tests Edgington ES (1980) Marcel Dekker, New York & Basel Permutation tests: A practical guide to resampling methods for testing hypotheses Good P (1994) Springer-Verlag, New York: Edgington ES (1980)
Marcel Dekker, New York & Basel
Permutation tests: A practical guide to resampling methods for testing hypotheses Good P (1994) Springer-Verlag, New York: Good P (1994)
Springer-Verlag, New York

...to top

Support

SnPM is being made freely available to the functional Neuroimaging community in the same spirit as SPM. As with SPM, SnPM may be considered academic shareware. It is supplied as is, and no formal support or maintenance is provided or implied.

General SnPM issues and idiosyncrasies should be raised on the SPM email discussion list, at <spm_at_jiscmail.ac.uk>, which the authors monitor. Direct queries to the authors will usually be answered to the whole list.

The SnPM authors may be contacted at <snpm-authors at umich.edu>. Please report bugs to the authors direct. Suggestions for improvement are welcome, particularly those substantiated with code! Requests for specific design PlugIns will be considered, but necessarily as invitations for collaborative work.

See the SPM documentation & support page for further resources.

...to top

SnPM5: Bugs & Fixes

This section describes the bugs that have been reported, along with the appropriate fixes. Updated versions of appropriate SnPM functions are available from our anonymous FTP server in the snpm5_updates directory: http://www.sph.umich.edu/ni-stat/SnPM/snpm5_updates

`snpm_ui.m` v4.19

Synopsis: A variable is not defined; also, F-contrasts cause error "CAT arguments dimensions are not consistent".

Fix: Merge of two bug fixes, from 07/01/13 and 08/11/07.

Thanks: Thanks to Shary and Satoru Hayasaka for reporting the first problem, Ken Rando for the latter one.

Download: snpm_ui.m.

-Tom 08/11/07

`snpm_cp.m` v1.2

Synopsis: change based on the update of snpm_STcalc.m and record suprathreshold statistics.

Download: snpm_cp.m.

-Hui 07/06/23

`snpm_pp.m` v1.3

Synopsis: some minor changes to improve the display

Thanks: Thanks to Darren Gitelman, Department of Neurology, Northwestern University for those changes.

Download: snpm_pp.m.

-Hui 08/01/10

`snpm_STcalc.m` v1.3

Synopsis: changes to avoid double calculations or indexing in for loop. Speeds calculation.

Thanks: Thanks to Darren Gitelman, Department of Neurology, Northwestern University for those changes.

Download: snpm_STcalc.m.

-Hui 08/01/10

`snpm_mip.m` v1.1

Synopsis: Add snpm_mip.m in SnPM5 toolbox based on updates of SPM5. In addition, we also update snpm_pp.m and snpm_combo_pp.m because of the change. The users can download them from this link.

Download: snpm_mip.m.

-Hui 07/07/12

`snpm_combo_pp.m` v1.2

Synopsis: pre-allocate CIinfo matrix

Thanks: Thanks to Darren Gitelman, Department of Neurology, Northwestern University for those changes.

Download: snpm_combo_ppc.m.

-Hui 08/01/10

...to top

SnPM3: Bugs & Fixes

This section describes the bugs that have been reported, along with the appropriate fixes. Updated versions of appropriate SnPM functions are available from our anonymous FTP server in the snpm3_updates directory: http://www.sph.umich.edu/ni-stat/SnPM/snpm3_updates

`snpm_defaults.m` v4.4: Decreased nMax4DefVol to 9.

Synopsis: The maximum number of scans for which volumetric mode is default is set in spm_defaults, nMax4DefVol. Previously this was set to 16, but some very large anatomical datasets will crash even with such small sample sizes.

Fix: This variable has been now set to 9, so for any analysis with 9 or more scans, you will now be prompted to use "Volumetric" or planar mode.

Thanks: Thanks to Satoru Hayasaka for reporting this problem.

Download: snpm_defaults.m.

-Tom 06/04/27

Additional Plugins

There are no additional plugins at this time

...to top

Thanks

We are indebted to everyone who has tried the software and reported bugs. Special thanks goes to Jun Ding, University of Michigan Biostatistics, for work on the new SnPM3 version. We also would like to thank Yanjun Xu of the Mental Health Research Institute, University of Michigan; Yanjun did important work on porting SnPM96 to Matlab 5.

...to top

Next SnPM page: SnPM manual

SnPM by Andrew Holmes of the Astra Zeneca
& Tom Nichols of University of Michigan Biostatistics department

SnPMauthors at <snpm-authors at umich.edu> Last modified: Fri Nov 7 14:25:51 EST 2008

SnPM

Statistical nonParametric Mapping

A toolbox for SPM

SPM - assumptions & low df problems...

SnPM - a primer...

Assumptions...

Variance smoothing...

Next...

Improved cluster size inference

Nonparametric uncorrected cluster-level P-values

Parameter estimates and statistic values saved in images

'Results' questions reorganized

SPM2 Compatibility

Nonparametric uncorrected voxel-level P-values

FDR-corrected (voxel-level) P-values

Combining toolbox

snpm_ui.m v4.19

snpm_cp.m v1.2

snpm_pp.m v1.3

snpm_STcalc.m v1.3

snpm_mip.m v1.1

snpm_combo_pp.m v1.2

snpm_defaults.m v4.4: Decreased nMax4DefVol to 9.

`snpm_ui.m` v4.19

`snpm_cp.m` v1.2

`snpm_pp.m` v1.3

`snpm_STcalc.m` v1.3

`snpm_mip.m` v1.1

`snpm_combo_pp.m` v1.2

`snpm_defaults.m` v4.4: Decreased nMax4DefVol to 9.