[pymvpa] group level test on decoding accuracy

Nick Oosterhof n.n.oosterhof at googlemail.com
Mon May 11 13:50:28 UTC 2015

On 10 May 2015, at 15:17, Christopher J Markiewicz <effigies at bu.edu> wrote:

> There's also a much more intensive non-parametric test used by Stelzer
> et al. (2012) <https://dx.doi.org/10.1016/j.neuroimage.2012.09.063>, but
> it requires a lot of computing time and at least temporary storage space.

If you feel comfortable in Matlab / GNU Octave, you may also consider an implementation in CoSMoMVPA* [1] which supports both the Stelzer approach [2] and a permutation approach [3]. The second option is much faster, but also a bit conservative (and becomes more conservative with stronger effects). 

This implementation also supports Threshold Free Cluster Enhancement [4] for both approaches, which has several advantages over fixed-threshold clustering:
- does not require setting an uncorrected threshold (reducing experimenter’s degrees of freedom);
- addresses "problems of smoothing, threshold dependence and localisation in cluster inference” [4];
- is more robust against to nonstationarity of the data [5].

There is an example [6] for surface-based analyses, but it also works for volume-based approaches. 

* disclaimer: I am the main developer for this package
[1] https://github.com/CoSMoMVPA/CoSMoMVPA/blob/master/mvpa/cosmo_montecarlo_cluster_stat.m
[2] http://www.ncbi.nlm.nih.gov/pubmed/23041526
[3] http://www.ncbi.nlm.nih.gov/pubmed/17517438
[4] http://www.sciencedirect.com/science/article/pii/S1053811910012917
[5] http://www.ncbi.nlm.nih.gov/pubmed/20955803
[6] http://cosmomvpa.org/contents_demo.html#demo-surface-tfce

More information about the Pkg-ExpPsy-PyMVPA mailing list