plot_cluster_stats_spatio_temporal

✉️ Mailing List Archive (read-only)
1

In the example online, plot_cluster_stats_spatio_temporal.py, X is passed
in to ttest 0 and 1 in the forth dimension: X = X[:, :, :, 0] - X[:, :, :, 1
] # make paired contrast.
That subtraction seems strange to me since it seems to lose the variance
unique to each condition. I would have expected something more like
...1samp_test(X[:, :, :, 0], X[:, :, :, 1]). What am I missing here?

Also, my ultimate goal is 1samp and 2samp spatio-temporal clustering with
TFCE with MEG. ANOVA for groupxcondition would be great as well. In
spatio_temporal_cluster_1samp_test, the doc says TFCE will be used if a
dict is passed in for threshold=, but what is that dict suppose to look
like?

Much thanks for the excellent package,
Matt

P.S. Anaconda was really important for getting this working in centos5
without root access. Spyder is working as well which is great for
transitioning from matlab.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20140107/530722b5/attachment.html

2

In the example online, plot_cluster_stats_spatio_temporal.py, X is passed in
to ttest 0 and 1 in the forth dimension: X = X[:, :, :, 0] - X[:, :, :, 1] #
make paired contrast.
That subtraction seems strange to me since it seems to lose the variance
unique to each condition. I would have expected something more like
...1samp_test(X[:, :, :, 0], X[:, :, :, 1]). What am I missing here?

you're right. We assume here equal variance and then the paired t test
is just a one sample test on the difference.

suggestions of improvement are really welcome.

Also, my ultimate goal is 1samp and 2samp spatio-temporal clustering with
TFCE with MEG. ANOVA for groupxcondition would be great as well. In
spatio_temporal_cluster_1samp_test, the doc says TFCE will be used if a dict
is passed in for threshold=, but what is that dict suppose to look like?

have a look at the TFCE example

http://martinos.org/mne/stable/auto_examples/stats/plot_cluster_methods_tutorial.html

and especially the variable:

threshold_tfce

Much thanks for the excellent package,

thanks

P.S. Anaconda was really important for getting this working in centos5
without root access. Spyder is working as well which is great for
transitioning from matlab.

that's indeed what we recommend to install to get started

Best,
Alex

3

Is there a way currently to do the 1samp test without assuming equal
variance?

4

Hey Matt,

The statistical functions are designed to accommodate different statistical
measures on their inputs. See if you can use the 2-sample version, changing
the statistical function (`stat_fun`) being called to one of your own (by
default it uses `f_oneway`):

https://github.com/mne-tools/mne-python/blob/master/mne/stats/cluster_level.py#L1267

Sorry this function isn't in the documentation -- it should be! I fixed it
in master, so the function should show up in the developmental version of
the documentation in the next couple days:

http://martinos.org/mne/dev/python_reference.html#statistics

In the meantime you can see the example using it in the docs:

http://martinos.org/mne/stable/auto_examples/stats/plot_cluster_stats_spatio_temporal_2samp.html#example-stats-plot-cluster-stats-spatio-temporal-2samp-py

If this function won't work for your use case, open an issue on GitHub and
we can talk about how to expand or modify the API (and hopefully add an
example along the way).

Cheers,
Eric

5

I actually already have that 2samp example working on my data, but I used
two conditions in one group instead of two groups of subjects just to get
it working. I started with the 2samp because the structure was what I was
familiar with coming from matlab's ttest functions, i.e. pass both
conditions instead of a subtraction. Is there a way to easily modify the
2samp example to be valid with a 1samp situation? I tried
stat_fun=scipy.stats.ttest_rel, but it looks like I need to create a series
of my own functions as you have with f_oneway.

I should mention that I've done everything in matlab/fieldtrip so far, and
I am learning python in order to get these cluster methods working on our
data (in source space), so it will take some effort to put together my own
stat_fun.

I am happy to open a issue on github. The primary goal of the issue would
be to get cluster with TFCE with "hat" working in a 1 or 2 sample
situation, so I can use the same method for my group and condition
comparisons if that's possible. Should I add that to the issue page with an
ENH tag?

Are there any resources (such as powerpoint slides from a talk) that
explain the clustering used here from the ground up?

thanks,
Matt

6

The only stats "tutorial" type thing I have is the example on TFCE + hat in
the docs, which you've presumably seen. It's possible Alex or Denis has
given some sort of tutorial or knows about some slides...?

If you want reading that's well grounded mathematically, I know that Hari
Bharadwaj has some slides available from a lecture he gave:

http://nmr.mgh.harvard.edu/whynhow/HB_WhyNHow_Stats_Feb7_2013.pdf

It's realistically outside the scope of `mne-python` to provide users with
access to all statistical functions, since then we'd be replicating
something like R(py) or SPSS in our software. We've tried to provide some
of the most broadly useful examples (ttest_1samp and f_oneway) to get
people started, with mechanisms that allow them to build arbitrary
contrasts of their own (hopefully).

Building your own `stat_fun` would make a good introduction into the python
world. Go ahead and open an issue on GitHub and one of us can probably help
you along if you get stuck. If your code / `stat_fun` ends up being a
broadly applicable enhancement to the package, then we can work on
including it. The issue can also serve as the proper venue for discussing
inclusion, lest we bore the mne_analysis readership :slight_smile:

Eric