Statistical comparisons in virtual electrode spectrograms

[Piers Cornelissen]

Dear Eric thanks for your thoughts.

I would like to try: "a neighbourhood structure as a cell array (in your
case, of length 1) of structs". But could you clarify a little more what
the syntax would be for this cfg?

Re- the hypothesis tests, when I put the script together I hadn't really
thought through the hypothesis testing carefully. I was just trying to
get something to work at all.

But you are quite right, it doesn't make sense to statistically compare
E versus E+I

However, I do want to apply a test, separately, for E baseline v E
active, *and* E+I baseline versus E+I active.

Might it make sense, do you think, to compare ([E+I]-E) baseline versus
([E+I]-E) active, to try to work out whether there is indeed any
significant induced activity?

Best regards

Piers
> Hi Piers,
>
>
>
> In my first reply to your email, I didn't notice that you were analyzing
> virtual channels. In that case, the suggestion in my previous email will not
> work. As I see it now, you have only a single virtual channel. In other
> words, you do not have a spatial dimension in your data. Nevertheless,
> clusterrandanalysis wants a spatial neighbourhood structure. In the case of
> virtual channels, you have to construct this neighbourhood structure by
> hand. Read the help info of clusterrandanalysis, and you will find out how
> to specify a neighbourhood structure as a cell array (in your case, of
> length 1) of structs.
>
> I also read that you want to compare evoked and (evoked + induced) power
> spectra. Are these power spectra calculated on the same raw data? If this is
> the case, does it make sense to test the null hypothesis that they come from
> the same probability distribution (which is what clusterrandanalysis tests)?
> This null hypothesis would imply that the pure induced power spectra are
> zero. Do I see this correctly?
>
>
> Greetings,
>
> Eric
>
>
>
>
>> -----Original Message-----
>> From: FieldTrip discussion list [mailto:FIELDTRIP at NIC.SURFNET.NL] On
>>
> Behalf Of
>
>> Piers Cornelissen
>> Sent: Wednesday, November 08, 2006 3:19 PM
>> To: FIELDTRIP at NIC.SURFNET.NL
>> Subject: [FIELDTRIP] Statistical comparisons in virtual electrode
>>
> spectrograms
>
>> Dear all
>>
>> Can anyone advise on the following:
>>
>> I have 6 sets of virtual electrode data (from CTF based MEG recordings)
>> from a total of 9 subjects, each of whom carried out 4 conditions of
>> interest.
>>
>> So far, for each virtual electrode I have succeeded (I hope) in computing
>> group level spectrograms for each condition, where each spectrogram
>> reflects either evoked or evoked+induced power changes.
>>
>> What I would like to do is to a) compare spectrograms, and also b)compute
>> statistics for each group level spectrogram to ask the question whether
>> the power changes in the active window (0 - 700ms) are significantly
>> different from the those in the passive/baseline window (-250 - 0 ms).
>>
>> So far I have had alot of help from Jan, and, on his suggestion, I have
>> copied out my query below, plus provided the matlab scripts which cause
>> the problem:
>>
>> The attached script E_versus_EI.m shows the code for
>> analysing one condition from one particular virtual electrode. This gives
>> the evoked as well as the evoked+induced spectrograms (as far as I
>> understand it).
>>
>> To carry out the comaprison between the evoked versus evoked+Induced test,
>> I thought I should be able to modify the freqgrandaverage step for
>> generating both iefgranfavg and efgrandavg by setting cfg.keepindividual
>> to 'yes'. In doing so, I was hoping that this would retain the 9 separate
>> powerspectra in each case, which it seems to have done. I was then hoping
>> that this would be sufficient input to the second script test_clusrand.m
>> (which is just a minor modification of  E_versus_EL.m). Certainly the
>> powspctrm field for each dataset appears as: [4-D double] which I think is
>> correct.
>>
>> However, when I ran clusterrandanalysis on iefgranfavg and efgrandavg, I
>> got the following errors:
>>
>>
>> cfg15 =
>>
>>          statistic: 'depsamplesT'
>>        alphathresh: 0.0500
>>       makeclusters: 'yes'
>>          minnbchan: 2
>>    clusterteststat: 'maxsum'
>>             onetwo: 'twosided'
>>              alpha: 0.0500
>>         nranddraws: 100
>>            channel: {'V0'}
>>
>>
>>>> %cfg15.latency = [0 0.5]  % increases sensitivity if prior temporal
>>>>
>> information is  knowm
>>
>>>> cfg15.frequency = [5 40];
>>>>
>>>> %frange = 'all_freqs';  % file addition for saving
>>>>
>>>> [clusrand] = clusterrandanalysis (cfg15, ind_efgrandavg,
>>>>
>> ind_iefgrandavg);
>> Selecting and formatting the data.
>> selected 1 channels
>> selected 191 time bins
>> selected 176 frequency bins
>> Calculating the neighbourhood structure of the channels.
>> ??? Error using ==> clusterrandanalysis (getneighbgeometry)
>> Did not find gradiometer or electrode information.
>>
>> Error in ==> /home/willis/plc/matlab/clusterrandanalysis.m
>> On line 473  ==> [cfg,data] = getneighbgeometry(cfg,data,varargin{1});
>>
>>
>> I noticed also that the way I used freqgrandaverage in order to compile
>> the two datasets appeared to throw out the gradiometer information. At
>> least there was a warning to this effect. So, I am guessing that using
>> freqgrandaverage in this way is incorrect. Therefore please could anyone
>> suggest how I compile my 2 sets of nine spectrograms in such a way that
>> they provide appropriate input to clusterrandanalysis. Or, if as Jan seems
>> to suggest, that clusterrandanalysis can't cope with virtual electrode
>> data, does anyone have any other clever solutions.
>>
>> Please!
>>
>> Thanks in anticipation
>>
>> Piers C
>>
>>
>>
>>
>>
>>
>>
>
>
>
>