<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
  <head>
    <meta http-equiv="content-type" content="text/html;
      charset=ISO-8859-1">
  </head>
  <body bgcolor="#ffffff" text="#000000">
    Dear Field-trippers<br>
    First of all, I would like to thank the Fieldtrip mentors, and all
    the contributors.<br>
    I find this toolbox more than a toolbox. The website and the active
    mailing list makes it stimulating and definitively instructive.<br>
    Which motivates me to share with you some reflections and questions.<br>
    <br>
    I did some statistics on data from implanted electrodes (ECoG) in
    human. For the purpose of this analysis, I mainly looked at the time
    frequency space, so I first ran the following script:<br>
    (with data being the output of the <i>ft_preprocessing</i>
    function)<br>
    <br>
    <font color="#3333ff">load data;<br>
      trig = [3 4 5];<br>
      for cond = 1:length(trig)<br>
          cfg = [];<br>
          cfg.method = 'wavelet';<br>
          cfg.output = 'fourier';<br>
          cfg.foi = 2:2:50;<br>
          cfg.toi = -0.5:0.02:0.05;<br>
          cfg.keeptrials = 'yes';<br>
          cfg.keeptaper = 'yes';<br>
          cfg.width = 5;<br>
          cfg.trials = find(data.trialinfo(:,1) == trig(cond));<br>
          TF_Mwlt_fourier{cond} = ft_freqanalysis(cfg, data);<br>
      end;<br>
    </font><br>
    I used Morlet wavelet because a previous post from Robert that
    recommended not to use multitapering for PLF <br>
    (<a class="moz-txt-link-freetext"
href="http://mailman.science.ru.nl/pipermail/fieldtrip/2006-March/000446.html">http://mailman.science.ru.nl/pipermail/fieldtrip/2006-March/000446.html</a>).<br>
    And also to facilitate comparison with other studies.<br>
    <br>
    The output being Fourier, I computed power and phase concentration
    (aka PLF or ITC),<br>
    (both calculated at single trial level for stats and then averaged
    for <i>ft_multiplot</i><i>TFR</i> )<br>
    <br>
    <font color="#3333ff">powplf_data.pow                  =
      abs(data.fourierspctrm) .^2;<br>
      powplf_data.powspctrm      =
      abs(mean(squeeze(powplf_data.pow),1));<br>
      powplf_data.plf                     =
      data.fourierspctrm./abs(data.fourierspctrm);<br>
      powplf_data.plf_average    = 
      abs(mean(squeeze(powplf_data.plf),1));</font><br>
    <br>
    <br>
    The first statistics I wanted to run was a comparison of the power
    and the PLF for each condition against their respective baseline
    period.<br>
    To do so I applied the following method (based on Delorme et Makeig
    2004) for a given channel:<br>
    - draw a value within the baseline period for each trial
    (independently for each time point and frequency).<br>
    - average along the trial dimension<br>
    - repeat those steps thousand time<br>
    - use those thousand repetition to construct the distribution<br>
    - count the percentage of values above (/below) the observed
    post-onset value from the data (at a given latency and frequency).<br>
    Define significance for power using two tails (p < 0.025 & p
    > 0.975), and one tail for PLF (p<0.05).<br>
    <br>
    I decided to write my own function because I was not sure that I
    could do it using Field trip.<br>
    I noticed that there is the <i>statfun_actvsblT</i> that can be
    specified in <i>cfg.method</i> field of <i>ft_freqstatistics</i>.<br>
    But I have two concerns about it:<br>
    - I prefer to used a randomization method for distribution reason.
    Indeed even if for power it seems to be ok with my data (seems to be
    normally distributed),<br>
    it is by definition not the case with PLF (which is more like a
    gamma or F distribution; because values are more concentrated near
    to zero, rare value toward 1).<br>
    - this function average  over the specified baseline time period.
    This average step makes more sense to me in the case of ERP
    analysis, but less with time frequency.<br>
    Especially with PLF, since the average will have the tendency to
    compress the values toward 0.<br>
    <br>
    I guess that an alternative would be to use <i>statfun_diff_itc</i>
    with one condition being post-onset period and the other "fake"
    condition being the baseline.<br>
    But in this case the length (duration) of the two pools should be
    identical, as the time points would be "paired". Am I correct about
    this or is it more flexible ?<br>
    <br>
    Here are the points I would like to discuss:<br>
    1) I think I did my analysis the correct way, but it might not be
    the case, so any comments about the method are very welcome.<br>
    2) when someone is interested in the phase, is there a "better"
    method  to compute the time frequency transform?<br>
    (or any method is good as long as there is no frequency smoothing)<br>
    3) can we imagine a future extra/new option in the <i>statfun_actvsbl</i>
    that would allow for choosing between averaging and taking a random
    time point within the specified time period ?<br>
      Maybe this makes less sense for T-stats than in the case of a
    randomization test ? and even less when this is not done at the
    level of single trials ?<br>
    <br>
    Thanks in advance for your comments.<br>
    <br>
    Manuel<br>
    <br>
    <br>
    <br>
    <pre class="moz-signature" cols="72">-- 
Manuel Mercier, PhD
Research Fellow

Cognitive Neurophysiology Laboratory,
Children’s Evaluation and Rehabilitation Center (CERC),
Departments of Pediatrics
Albert Einstein College of Medicine,
1225 Morris Park Avenue
Bronx , New York, NY 10461

phone: +1 (718) 862 1824
fax: +1 (718) 862 1807</pre>
  </body>
</html>