[FieldTrip] TF transform, data length + padding

Pomper, Ulrich Ulrich.Pomper at charite.de
Wed May 28 12:50:31 CEST 2014

Hey Julian,
Thanks a lot for your help! Both explanations seem quite plausible. 
I think I will try to avoid the padding and stick with the longer data segments.

From: fieldtrip-bounces at science.ru.nl [fieldtrip-bounces at science.ru.nl] On Behalf Of Julian Keil [julian.keil at gmail.com]
Sent: Wednesday, May 28, 2014 12:36 PM
To: FieldTrip discussion list
Subject: Re: [FieldTrip] TF transform, data length + padding

Hi Ulrich,

one possible explanation for this could be the subtraction of the mean when calling ft_freqanalysis. If you don't specify cfg.polyremoval explicitly, the mean is removed. Of course, changing the time-dimension will change the mean and therefore result in a different value used in the ft_preproc_polyremoval - step.

I tried this with a dataset, setting cfg.polyremoval to -1, thereby avoiding the demean results in virtually identical time-freq results for a short (800ms) or long (1000ms) interval. However, setting cfg.polyremoval to 0 (the default) will change the output.

Also, setting the padding to 5, as you did in your code, will increase the frequency resolution which again might influence the way the low frequencies contribute to your output.

If someone has a more in-depth explanation or if I got this completely wrong, please correct me!



Dr. Julian Keil

AG Multisensorische Integration
Psychiatrische Universitätsklinik
der Charité im St. Hedwig-Krankenhaus
Große Hamburger Straße 5-11, Raum E 307
10115 Berlin

Telefon: +49-30-2311-1879
Fax:        +49-30-2311-2209

On Tue, May 27, 2014 at 3:18 PM, Pomper, Ulrich <Ulrich.Pomper at charite.de<mailto:Ulrich.Pomper at charite.de>> wrote:
Dear community,
I have a question regarding  time-frequency transformation.
The data I want to analyse last from -800 to -200 ms before stimulus onset, I'm using a 400 ms taper.
Importantly, want to make sure that no stimulus-evoked activity leaks into my prestim analysis.

I first tried to cut the segment down to  -1000 to 0 ms and used (zero) padding (see figure attached, top row). This would mean that the taper, at the last samplepoint of interest (-200ms), would include data up to 0ms
Suprisingly, I get quite different looking results when using a longer data segment (-1200 to 0 ms) (fig. bottom row) without padding.

Can someone explain to me what is going on here, and which approach would be more correct?
Either, the padding in the first variant introduces artificial low frequency activity, or the taper in the second variant uses more than 400ms, at least for the low frequencies.
Note that the difference between conditions (left coloumn) look almost identical, so whatever happens, it happens to both conditions very similarly.

I hope someone can educate me on that matter.
Cheers, Ulrich

%% TF transform %%

            cfg = [];
            cfg.continuous = 'no';
            cfg.channel      = 'all';
            cfg.output = 'pow';
            cfg.method       = 'mtmconvol';
            cfg.foi          =  2:0.5:35;
            cfg.toi          =  -1:0.01:0;
            cfg.taper   = 'hanning';
            cfg.tapsmofrq = 2;
            cfg.t_ftimwin    = 400
            % cfg.pad        = 5;
fieldtrip mailing list
fieldtrip at donders.ru.nl<mailto:fieldtrip at donders.ru.nl>

More information about the fieldtrip mailing list