[FieldTrip] 2x3 anova in cluster analysis

Thu Dec 8 17:33:45 CET 2011

Hi Jan-Mathijs and Fieldtrippers,

I have a follow up question refering to statistical inference about
monotonic trends in the permutation framework.

I have four conditions A, B, C, D and I would like to test whether there is
a monotonic change in the ERP/TF among the conditions (i.e. A<B<C<D but
possibly also A < B <= C < D). I guess one way would be to run multiple t
tests. However it seems a bit unconvenient as one needs to correct for
multiple comparisons. Furtheremore, the second case (A < B <= C < D) might
not be well captured.

I found a thread from 31 Aug 2011 ("[FieldTrip] depsamplesregrT help" by
Jonas Obleser) suggesting that depsamplesregrT will be the way to go.
However, your reply to Nico somehow destroyed my confidence. If I got the
reply correctly this is not that easy as I would be testing a particular
linear model which is a bit of a problem. Thus, the question:

How would you suggest testing for such a monotonic effect in the
permutation framework?

Best,
Marcin

>
> 2011/12/5 jan-mathijs schoffelen <jan.schoffelen at donders.ru.nl>
>
>> Hi Stan et al,
>>
>> Actually FieldTrip allows in a very straightforward way to plug-in one's
>> favourite statistic. The idea is the following: when specifying
>> cfg.statistic you need to specify a string that under the hood points to a
>> function, e.g. 'indepsamplesT' eventually causes a function to be called,
>> answering to the aptly chosen name: 'statfun_indepsamplesT' which is doing
>> the legwork. As long as the statfun_younameit has properly defined input
>> and output arguments, you can let it do whatever you want. FieldTrip
>> contains a bunch of 'statfuns' in the statfun directory but there's no
>> reason not to implement any yourself (and ideally contribute them to the
>> repository).
>> Another discussion altogether is the question whether the test statistic
>> of interest is an appropriate one that allows for statistical inference
>> using the permutation framework. Over the past years there have been
>> various threads on this mailing list regarding the possibility to test for
>> interactions using a permutation test. You can look this up in the archive.
>> Strictly speaking this is statistically not possible this way. The
>> permutation-framework tests the null-hypothesis of exchangeability of data
>> across allocated conditions, whereas when one wants to test for an
>> interaction effect tests one tests for a particular linear (parametric)
>> model of the dependent variables explaining variance in the dependent
>> variable. Inferring that exchangeability across conditions is unlikely
>> (i.e. obtaining a small p-value through permutation) does not necessarily
>> lead to the conclusion that there is an actual interaction effect. Though
>> opinions among the statisticians about this seem to vary, there may be a
>> way out: one could either use bootstrapping to obtain confidence intervals
>> for the interaction F under the null-hypothesis. The recipe for this would
>> be to do a cell-specific demeaning of the data to impose the
>> null-hypothesis of no interaction, and then through bootstrap resampling
>> obtain a reference distribution of the interaction F. Alternatively,
>> although I haven't thought this one through for any case other than a 2x2
>> anova, one could reduce the interaction effect to a two-condition contrast
>> where observations belonging to the 'main diagonal' of the 2x2 conditions
>> design are labeled as condition 1, and the observations belonging to the
>> other diagonal are labeled as condition 2. One could then proceed with
>> using a T-statistic as a descriptive statistic across these two
>> 'conditions' on which inference can be done.
>>
>> Cheers and a happy Sinterklaas to you all,
>>
>> Jan-Mathijs
>>
>>
>> On Dec 5, 2011, at 8:17 PM, Stanley Klein wrote:
>>
>> Nico,
>>
>> Your question gives me the chance to ask something I've been curious
>> about. It seems to me that one of the nifty things about permutation
>> cluster analysis is that it allows you to do anything that seems reasonable
>> as long you decide on the process ahead of time and don't do any future
>> tweaking other than fixing coding errors. I'm curious whether the Fieldtrip
>> code make it easy to insert new statistics modules.
>>
>> I presume your actual question had to do with whether Fieldtrip already
>> has the possibility of doing that 2x3 ANOVA.  I suspect it doesn't have
>> that capability. However, it would be nice if the permutation cluster
>> analysis could be made sufficiently modular that one could make use of all
>> the complicated clustering and permuting and all, but enable the user to
>> substitute their own statistical method.  I haven't actually watched my
>> students doing the nitty-gritty use of Fieldtrip's version of the
>> permutation test (other than knowing that is is very nice and well
>> documented) so I don't know how easy it is to stick in one's own favorite
>> statistic, but I'm hoping it is easy.
>>
>> Stan
>>
>> On Mon, Dec 5, 2011 at 4:58 AM, Nico Alexander Willi Kremers <
>> nkremers at uni-bonn.de> wrote:
>>
>>>
>>> Hi,
>>>
>>> I followed the discussion about implementing a 2x2 within subjects
>>> design into a cluster statistic with much interest. I want to implement a
>>> 2x3 repeated measure anova in a cluster analysis using fieldtrip. I think
>>> for this type of analysis subracting two conditions from each other and
>>> then calculate a ttest is not appropriate. Is there another way of
>>> calculating the interaction effect between the two factors and generate
>>> clusters for real and permutation data? What specifications must be set and
>>> how?
>>>
>>> Thank you very much for your answer,
>>>
>>> Nico
>>> ______________________________**_________________
>>> fieldtrip mailing list
>>> fieldtrip at donders.ru.nl
>>> http://mailman.science.ru.nl/**mailman/listinfo/fieldtrip<http://mailman.science.ru.nl/mailman/listinfo/fieldtrip>
>>>
>>
>> _______________________________________________
>> fieldtrip mailing list
>> fieldtrip at donders.ru.nl
>> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
>>
>>
>>    Jan-Mathijs Schoffelen, MD PhD
>>
>> Donders Institute for Brain, Cognition and Behaviour,
>> Centre for Cognitive Neuroimaging,
>> Radboud University Nijmegen, The Netherlands
>>
>> Max Planck Institute for Psycholinguistics,
>> Nijmegen, The Netherlands
>>
>> J.Schoffelen at donders.ru.nl
>> Telephone: +31-24-3614793
>>
>>
>> _______________________________________________
>> fieldtrip mailing list
>> fieldtrip at donders.ru.nl
>> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
>>
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20111208/fe38b462/attachment.html>