coherence normalization

Tom Holroyd tomh at KURAGE.NIMH.NIH.GOV
Fri Oct 22 18:25:17 CEST 2004


> Sounds interesting. But why are you running an ANOVA? The coherence in
> a voxel will not be normally distributed.

You'd think that, wouldn't you?  :-)

[By the way, that's an American expression meaning roughly, I'm
just as surprised as you are.]

Here's a random sample, from subject 6, condition 4:

  #Magnitude          Freq      Cum_Freq
   -2.124902            21            21
   -1.874683            47            68
   -1.624463            74           142
   -1.374243           134           276
   -1.124023           180           456
   -0.873804           208           664
   -0.623584           249           913
   -0.373364           323          1236
   -0.123144          2052          3288
    0.127076           343          3631
    0.377295           264          3895
    0.627515           219          4114
    0.877735           161          4275
    1.127955           114          4389
    1.378175            72          4461
    1.628394            47          4508
    1.878614            33          4541
    2.128834            16          4557
    2.379054             2          4559
    2.629274             1          4560

These coherence values have already been z-scored.  That's
something I forgot to mention before -- since the coherence
values turn out to be not too badly distributed, and since I'm
only interested in the differences anyway, I z-scored 'em.

> Furthermore, it does not allow for a decent multiple comparison
> correction (except bonferoni), since GRFT does not apply.

That's true, but my experience has been that normalizing the
distributions (so that individual differences in power levels
between subjects are eliminated) has almost exactly the same
effect as a random permutation analysis (which I have done, but
not on these data).  The beamformers apparently provide quite a
lot of independence already (you know, it's all coming from the
same CSD matrix, but each voxel has a large prior in the form of
the forward solution -- somebody should write a paper about
that).  And since the coherence is normalized anyway, it's fine.

I can always do a permutation analysis later, by shuffling the
conditions and rerunning the ANOVAs a few thousand times (they
are very fast).

> Did you try sourcestatistics with method randomization or
> randcluster? They do multiple comparison correction.

You mean sourcedescriptives?  No, I haven't, but I will.  I was
in a hurry (Neuroscience is next week, you know).  Did I mention
it took 24 hours to calculate all the volumes?  (Yes, I know, use
the Beowulf cluster. NIH's cluster is called "Biowulf".  Ha ha.)

> Sofar no one ever asker me for the phase, but it would be interesting
> to do some more postprocessing of the source.avg.csd matrices to make
> something like phase easier to plot. I would like to hear suggestions.

Let me think about that for a while.

Dr. Tom Holroyd
"A man of genius makes no mistakes. His errors are volitional and
are the portals of discovery." -- James Joyce



More information about the fieldtrip mailing list