Hi Panagiotis,
I asked some colleagues for advice about your problem and one idea
that came up is that perhaps the wavelets have mean different from
zero. I checked it and it is indeed the case. The actual mean values
are quite small (like 1e-7) but for data that is in T they can
actually be large. In the attached version I added mean-correction of
the wavelets. I'm not sure it'll completely solve the problem as still
you get something like 1e-18. Alternatively you can try converting the
data to fT as I suggested before.
If this is really the reason it might be a serious argument for
switching to fT at conversion. Let me know what you get.
Best,
Vladimir
On Tue, Nov 16, 2010 at 2:01 PM, Panagiotis Tsiatsis
<[log in to unmask]> wrote:
> Dear Vladimir, dear all,
>
> On 11/15/2010 12:04 PM, Vladimir Litvak wrote:
>
>> This sounds like something worth looking into.I can think about the
>> following:
>>
>> 1) Make sure you really use exactly the same trials, just to rule out
>> this factor.
>
> I can try this in the near future and report the results, although the sets
> of trials in the two cases should be overlapping extremely
>
>> 2) If you use robust averaging, don't use it for this testing.
>
> No, I was not using robust averaging.
>>
>> 3) Perhaps try several different estimation methods and see if this
>> phenomenon is common to all of them.
>
> That is a good point - so, thinking naively, I assumed that it might have to
> do something with the fact that when you correct for baseline, the waveform
> in the baseline is much closer to its mean that the rest of the trials and
> that at the point that the baseline ends, the waveform is less restricted
> (not bc-corrected) and this maybe enable some higher frequencies to occur
> because of the transition between the baseline period and the rest of the
> trial. Of course this might be stupid, but I could not find any rigid
> arguments to exclude this posssibility.
>
> So then I tried to set the baseline period to actually be the whole trial,
> that is from each trial I subtract its mean (and not just the mean of a
> prestimulus period). This way I could test if what I was writing before
> could be the reason. But when I compared the TF data of this (the
> whole-trial base-correction) to the data where no baseline correction had
> been applied at all, I still found significant differences. (of the order of
> [10^-25, 10^-24])
>>
>> 4) This might have something to do with numeric issues. The numbers
>> for power in MEG become very small as you mentioned, smaller than
>> Matlab's epsilon. I've been planning to start changing units to fT at
>> conversion, but I'm waiting for Fieldtrip to provide better generic
>> support for determining what the units are. Maybe you should try to
>> multiply your data by 1e15 before computing TF. But then also change
>> the units to 'fT' because otherwise you'll get really large numbers in
>> the exported images.
>>
>
> I would also be worried for numerical issues, that is mainly for rounding /
> underflow problems that might appear somewhere along the computational chain
> (especially if matrix inversions are involved and the matrices become (close
> to) singular). On my machine I get the following concerning precision:
>
>>> eps(0)
>
> ans =
>
> 4.9407e-324
>
>>> eps(realmin)
>
> ans =
>
> 4.9407e-324
>
>>> realmin/eps(realmin)
>
> ans =
>
> 4.5036e+015
>
> so at a first glance it looks ok - but as I said before these numbers are
> not a guarantee that numerical errors do not propagate / get amplified
> across the computational chain.
>
> I am still confused about this differences - I will try to test further
> whether numerical errors (as it seems to be the most possible scenario) are
> to blame.
>
> Any insights would be highly appreciated.
>
> Thanks and best,
> Panagiotis
>
>
>> Best,
>>
>> Vladimir
>>
>>
>>> Dear Panagiotis,
>>>
>>> On Sat, Nov 13, 2010 at 4:51 AM, Panagiotis Tsiatsis
>>> <[log in to unmask]> wrote:
>>>>
>>>> Hello dear Vladimir, hello dear all,
>>>>
>>>> Let me come back to this issue - first of all I absolutely agree that
>>>>
>>>> The slow drifts will only affect the lowermost frequency bin (if it
>>>> includes the DC) so baseline correction in the time domain does not
>>>> rescale all the frequencies or anything of that sort.
>>>>
>>>>
>>>>
>>>> but the funny thing (and the main reason why I send the previous e-mail)
>>>> is
>>>> that after processing the same data once with baseline correction and
>>>> once
>>>> without, the Time Frequency analysis of the mean trials differ even in
>>>> frequencies as high as 10 - 20Hz and this difference can be (at least)
>>>> in
>>>> the range (-2,2)*10^-25. ( I calculated the contrast of the means of the
>>>> TF
>>>> data with and without baseline correction). This is one order of
>>>> magnitude
>>>> less that my strongest activations in average TF (~4*10^-24) but
>>>> comparable
>>>> to the contrast values among conditions in TF. I understand that
>>>> baseline
>>>> correction affects the artifact rejection process as well but to me the
>>>> effect seems far than being small and insignificant. I also have to note
>>>> that I have more than 150 trials per conditions whether I apply baseline
>>>> correction or not and this number is really similar in each case (+-5
>>>> trials). The baseline duration that I used for testing was 100 ms.
>>>>
>>>> I would absolutely expect to see the very same thing that you wrote in
>>>> your
>>>> previous email - but this is not the case. Any intuitions?
>>>
>>> Thanks and best,
>>> P.
>>> On 11/11/2010 6:09 PM, Vladimir Litvak wrote:
>>>>
>>>> Dear Panagiotis,
>>>>
>>>> On Thu, Nov 11, 2010 at 4:21 PM, Panagiotis Tsiatsis
>>>> <[log in to unmask]> wrote:
>>>>>
>>>>> 'Baseline correction is no longer done automatically by
>>>>> spm_eeg_filter.
>>>>> Use
>>>>> spm_eeg_bc if necessary.'
>>>>>
>>>>> Dear All,
>>>>>
>>>>> I 've got a naive question concerning filtering and baseline correction
>>>>> in
>>>>> MEG data. When applying high-pass filtering in the data, the following
>>>>> message appears:
>>>>>
>>>>> 'Baseline correction is no longer done automatically by spm_eeg_filter.
>>>>> Use
>>>>> spm_eeg_bc if necessary.'
>>>>>
>>>>> 1st Question: I suppose this means that the filtering functions does
>>>>> not
>>>>> subtract the mean of the trial / continuous file, that is the zero
>>>>> coefficient of the fourier transform, right?
>>>>>
>>>> Yes, the filtering function used to subtract the baseline in SPM5 so
>>>> that warning is there for historical reasons.
>>>>
>>>>> 2nd Question: Would it be neccessary to apply Baseline Correction in
>>>>> MEG
>>>>> data? That is, are there any DC compponent biases that might differ
>>>>> across
>>>>> subjects or "strong", very slow drifts in the recorded activity across
>>>>> time?
>>>>> I guess it should be neccessary for EEG data where there are amplifier
>>>>> offset and slow conductance drifts, but I am not totally sure if this
>>>>> is
>>>>> the
>>>>> case for MEG recordings
>>>>>
>>>>> 3rd Question: I am mainly asking the above questions because I want to
>>>>> compare the difference in activity in the Time-Frequncy domain among
>>>>> conditions (difference in power across various frequncy bands in time),
>>>>> and
>>>>> I think that in one sense applying baseline correction in the time
>>>>> domain
>>>>> and then transforming it to the Time - Frequency domain kind of
>>>>> normalizes
>>>>> the power of activity across the different frequency bands according to
>>>>> the
>>>>> baseline, which might eventually smear out the effect (difference in
>>>>> frequency amplitude in time) that I want to see. In that sense I think
>>>>> that
>>>>> applying or not Baseline corrections is a matter of what I want to
>>>>> check
>>>>> for
>>>>> (relative/absolute power differences). The bottom-line question then
>>>>> would
>>>>> be whether or not it is absolutely neccessary to apply baseline
>>>>> correction
>>>>> in MEG (time / time-frequency) data because for example there would be
>>>>> DC
>>>>> biases that would be different for different recordings.
>>>>>
>>>> There are slow drifts in the MEG that in most cases necessitate
>>>> baseline correction of high-pass filtering if you want to look at
>>>> ERFs. However, this is not relevant for your time-frequency analysis.
>>>> The slow drifts will only affect the lowermost frequency bin (if it
>>>> includes the DC) so baseline correction in the time domain does not
>>>> rescale all the frequencies or anything of that sort. The only problem
>>>> might be that large DC offsets in the data confuse some TF estimation
>>>> methods so I'd at least subtract the baseline or the mean before doing
>>>> TF.
>>>>
>>>>> 4th Question (irrelevant to the others): I know it would be
>>>>> computationally
>>>>> extremely heavy, but is there a way to transform continuous data in the
>>>>> Time
>>>>> - Frequency domain? It would be useful as then i.e. I would not have
>>>>> to
>>>>> apply TF every time that I reepoch the data and I would have no
>>>>> "edge-effects" when converting single trials in TF. Plus, it would be
>>>>> helpful in eyeballing spontaneous activity data
>>>>>
>>>>>
>>>> This is possible in principle but SPM functions will have great
>>>> difficulties handling this kind of data. If you want to do it for 275
>>>> MEG channels you'll have huge data arrays and can run into memory
>>>> problems. So if you want to do it you need to write your own code
>>>> possibly using Fieldtrip functions and only convert to SPM format once
>>>> you extract your epochs. What you can do to avoid edge effects is to
>>>> pad your epochs with extra data. There is now a function called
>>>> spm_eeg_crop (I think it was added after the latest public release but
>>>> I can send it to you) that you can use to later remove that padding
>>>> from your TF dataset.
>>>>
>>>> Best,
>>>>
>>>> Vladimir
>>>>
>>>>> I would really appreciate your opinion on these matters. I know that
>>>>> they
>>>>> might be really basic questions, but I still don't feel absolutely sure
>>>>> about the answers.
>>>>>
>>>>> Thanks and best, and apologies for the long e-mail - I tried to explain
>>>>> my
>>>>> questions as clearly as I could.
>>>>>
>>>>> Panagiotis
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> Panagiotis S. Tsiatsis
>>>>> Max Planck Institute for Biogical Cybernetics
>>>>> Cognitive NeuroImaging Group
>>>>> Tuebingen, Germany
>>>>>
>>>
>>> --
>>> Panagiotis S. Tsiatsis
>>> Max Planck Institute for Biogical Cybernetics
>>> Cognitive NeuroImaging Group
>>> Tuebingen, Germany
>>>
>>>
>
>
> --
> Panagiotis S. Tsiatsis
> Max Planck Institute for Biogical Cybernetics
> Cognitive NeuroImaging Group
> Tuebingen, Germany
>
|