[FieldTrip] ft_freqstatistics to find significantly different time samples

[Thomas Lusignan]

Hi FieldTrippers!

Quick question: Is it possible to* use ft_freqstatistics* to compare two
groups of time-frequency representations,* time-sample by time-sample* (and
for each frequency of interest)?

If is is possible, *how should my design matrix look like* for comparing,
say: two conditions, one with five trials and one with 6 trials, each trial
having 100 time-samples?

Thank you all !

Regards,

Thomas Lusignan
Master's student
Université de Montréal

Le lun. 11 nov. 2019 à 15:24, Thomas Lusignan <thomas.lusignan at gmail.com> a
écrit :

> Dear List,
>
> I am Thomas, a Master's student in Paul Cisek's lab in Montréal. I work on
> monkey local field potentials and spike data recorded during a
> decision-making task. This means as of now I have only one subject, but
> hundreds of sessions.
>
> I figured out how to plot time-frequency representations of each condition
> (i.e. trial types). However, I would like to statistically compare
> conditions together to see if each time-frequency point is significantly
> different between conditions. I've already used ft_preprocessing and
> ft_freqanalysis on the data to get power spectra (with cfg.keeptrials =
> 'yes').
>
> *I'm looking to use ft_freqstatistics to do the following:*
>
> *1-* For one channel, process each time-point at every frequency of
> interest, taking the power at that moment for each trial (e.g.
> data.powspctrm(1:53, 1,1,1), then data.powspctrm(1:53, 1,1,2), etc.).
>
> *2-* Use permutations to perform a t-test to see if the average of the
> points in one condition is different from the average of the second
> condition.
>
> *3-* Repeat for each time-point and frequency, correcting for multiple
> comparisons.
>
> *4.* Plot the resulting heatmap and mask.
>
>
> So, dear List,* here are my questions:*
>
> *A: *Is what I want to do possible with ft_freqstatistics, or some other
> function in FieldTrip?
> *B: *If it is, what am I doing wrong? I'm guessing it has to do with the
> design matrix, but I don't know how to fix it.
> *C: *If it works, can I compare distributions of different sizes (i.e.: a
> condition with 53 trials vs a condition with 85 trials)? I *could* randomly
> select only a certain number of trials so that sample sizes are equal, but
> would rather not.
>
>
>
> *Following is how the input data is structured, the cfg and the result of
> the call.*
>
> My *data *looks like this, for each condition :
> data:
>         label: {1×15 cell}            % 15 channels with data
>        dimord: 'rpt_chan_freq_time'
>          freq: [1×57 double]          % from 4 to 60 Hz
>          time: [1×1000 double]        % one second of data
>     powspctrm: [53×15×57×1000 double] % 53 trials. Not constant between
> conditions.
>           cfg: [1×1 struct]
>
> The *cfg *I use for my call to ft_freqstatistics is the following:
>
>              channel: 'Ch:1 PMd idES:377'
>              latency: 'all'
>            frequency: 'all'
>          avgoverchan: 'no'
>          avgovertime: 'no'
>          avgoverfreq: 'no'
>           resampling: 'permutation'
>            parameter: 'powspctrm'
>               design: [2×106 double]
>               method: 'montecarlo'
>     numrandomization: 1000
>            statistic: 'depsamplesT'
>                alpha: 0.025
>                 ivar: 1
>             correctm: 'bonferroni'
>
> The *design *looks like this, where the first row represents condition
> type and the second, trials:
>
>      1   1   1   ...    1     1     1     2     2     2   ...    2     2
>   2
>      1   2   3   ...   51    52    53     1     2     3   ...   51    52
>  53
>
> I'm pretty sure this is not the right design for my needs, as permutations
> only switch between "corresponding" trials (e.g. index 1 with 54, 2 with 55
> and so on).
>
>
> When I call ft_freqstatistics, the result looks like this:
>
> the call to "ft_selectdata" took 0 seconds
> using "ft_statistics_montecarlo" for the statistical testing
> using "ft_statfun_depsamplesT" for the single-sample statistics
> constructing randomized design
> total number of measurements     = 106
> total number of variables        = 2
> number of independent variables  = 1
> number of unit variables         = 1
> number of within-cell variables  = 0
> number of control variables      = 0
> using a permutation resampling approach
> repeated measurement in variable 2 over 53 levels
> number of repeated measurements in each level is 2 2 2 2 2 2 2 2 2 2 2 2 2
> 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
> 2 2
> computing statistic
> estimated time per randomization is 0.00 seconds
> computing statistic 1000 from 1000
>
> performing Bonferroni correction for multiple comparisons
> the returned probabilities are uncorrected, the thresholded mask is
> corrected
> the call to "ft_freqstatistics" took 4 seconds
>
> stat =
>
>   struct with fields:
>
>        prob: [1×57×0 double]
>     cirange: [1×57×0 double]
>        mask: [1×57×0 logical]
>        stat: [1×57×0 double]
>         ref: [1×57×0 double]
>      dimord: 'chan_freq_time'
>        freq: [1×57 double]
>       label: {'Ch:1 PMd idES:377'}
>        time: [1×0 double]
>         cfg: [1×1 struct]
>
> Since stat is basically an structure of empty matrices, I figure
> something's wrong.
>
>
>
>
> Thank you all very much,
>
> Regards,
>
> Thomas Lusignan
>
>
>
> P.S.: My neurophysiological data is stored in a custom SQL database, which
> means that I have no .plx files to import (and therefore no headers, which
> makes using FieldTrip a little less straightforward). My data structures
> therefore come from a function I built which gets the data and organizes it
> correctly (or so I think, but I might be wrong).
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20191120/6d03bc14/attachment.html>