New FieldTrip User Questions

Fri Apr 16 23:24:04 CEST 2010

Thank you for clarifying.  One more question below.

1)      With regard to our frequency domain analysis, I've already used FFT
to look at general trends in the sample and delay periods but am interested
in plotting the evolution of those oscillations over trial time.  I believe
that the cluster analysis and MonteCarlo simulation are the best method to
adopt for quantifying difference between conditions, but please do correct
me if you have other suggestions.

The cluster-based permutation tests are for testing differences between
conditions, not for characterizing temporal evolution (unless you mean by
"temporal evolution" a mean power difference between the first part and the
second part of the trials).

By temporal evolution I'm referring to the continuous changes in the TFR map
over the course of our long trials.  It is difficult to categorize these
into discrete epochs (e.g., early, middle, late delay) without diluting some
of the effect, so I would like to analyze the entire TFR without any further
subdivision.  I would like to compare the entire TFR obtained during
different conditions for areas of significant difference.  Am I
understanding correctly that the cluster-based permutations are the best way
to approach this question?

Yes.

I use wavelets to compute the TFRs as below.  Can I then use multitapers for
further smoothing?

Thank you so much again!

Best,

Allen

____________________________________________________________________________
______

Allen Ardestani    Email: aardesta at ucla.edu

Phone: (310) 825-5528

Medical Scientist Training Program

David Geffen School of Medicine at UCLA

Semel Institute for Neuroscience and Human Behavior

760 Westwood Plaza

Los Angeles, CA 90095-1759

USA

From: FieldTrip discussion list [mailto:FIELDTRIP at NIC.SURFNET.NL] On Behalf
Of Eric Maris
Sent: Thursday, April 01, 2010 1:51 PM
To: FIELDTRIP at NIC.SURFNET.NL
Subject: Re: [FIELDTRIP] New FieldTrip User Questions

Hi Allen,

I just now came across the FieldTrip toolbox and have a few questions - I
apologize in advance if I am misusing this discussion forum or if my
questions are too obvious.  I am working with a very large dataset of
simultaneous LFP, unit, and NIRS signal from macaques for the purpose of
examining time and frequency dynamics of the signals in various cognitive
conditions.  I would like to implement the clustering/bootstrapping
methodology outlined in [Journal of Neuroscience Methods 164 (2007) 177-190]
to identify significant changes in synchronization and phase-locking.

The specific data in question are LFPs recorded while the monkeys perform a
working-memory task, which consists of a 2s visual sample, 20s delay, and
subsequent choice period.  A 20s baseline precedes the sample (t=0), which
is time-locked to the animal's foveation on the visual sample.  TF plots are
computed using Morlet wavelets for each trial, averaged across trials, and
then normalized with respect to the average baseline.  The first question
I'd like to address is: what time-frequency regions exhibit significant
difference between Correct (left plot) and Incorrect (right plot) trials.

cid:image002.jpg at 01CAA0E0.B7CEC3E0 cid:image003.jpg at 01CAA0E0.B7CEC3E0

I have a number of questions about the appropriateness of applying
bootstrapping to our specific dataset:

1)      Because we use wavelets for spectral decomposition (with
frequency-dependent changes in spectral and temporal resolution) is it valid
to simply cluster across different frequencies?

Yes it is. However, with your 20 sec. trials, I would not go for the high
temporal resolution that he wavelet transform gives you. I would do one
Fourier-based analyses for the first 2 seconds (encoding stage) and another
one for the rest of the trial (retention stage).

2)      We are interested in comparing different conditions, where each is
computed relative to its own baseline.  Is there anything wrong with using
the baselines of the same dataset for the null distribution as opposed to
using a different set of baseline data?

Do not baseline correct your data. Compare the raw power spectra for the
correct response trials with those of the incorrect response condition.

3)      The data have been recorded at 1KHz, and this oversampling results
in high spatial-frequency noise.  Do I need to do any downsampling/smoothing
before applying the statistics?

No. However, with such long trials, I would definitely smooth in the
frequency domain using multitapers (also available in Fieldtrip).

4)      For evaluating the relationship between channels, we compute the
phase-locking value (PLV), which has no meaningful single-trial
measuremetnt.  Is there any way to apply statistical analyses directly to
the average TF plots without resorting back to the individual trials?

Have look at the paper by Maris, Schoffelen, and Fries (2007, JNM) that
describes cluster-based permutation tests for coherence differences. This
may be what you need.

Best,

____________________________________________________________________________
______

Allen Ardestani    Email: aardesta at ucla.edu

Phone: (310) 825-5528

Medical Scientist Training Program

David Geffen School of Medicine at UCLA

Semel Institute for Neuroscience and Human Behavior

760 Westwood Plaza

Los Angeles, CA 90095-1759

USA

----------------------------------