[Eeglablist] Significance testing

[Stavros Zanos]

Hi all.

We are analyzing ECoG data from subjects performing m different cognitive
tasks. Each task is repeated n times (trials) for each cognitive task, so
we end up with a population of mxn signals for each channel. We want to
perform statistical significance testing between these signals; e.g. is
there a difference in voltage between any two time points within the same
signal? or between two respective time points from signals from different
tasks?

This seems at first like a problem of multivariable analysis, with
task,channel,trial and time as the independent variables, and the voltage
(or any transformation of it in the freqency/time domain) as the dependent
variable. Nonparametric tests (like Wilcoxon) need to be corrected for
multiple comparisons, yielding impossibly low significance levels. MANOVA
might be an appropritate method, but it is too computationally demanding
for such big cluster sizes.

I saw a couple of papers utilizing Generalized Estimating Equations (GEE)
for this kind of comparisons, based on the notion that ECoG are
correlated. There are a couple of theoretical reasons for which GEE might
not be the optimal solution; GEE assume that independent units (trials)
come from independent sources, and since in our case they come from 1
patient, the variability in responses might be deceptively small (even
with a low signal:noise ratio), and we could get biased results. Another
problem is the discrepancy between cluster sizes (very big) and
observations (or trials) per cluster (relatively small), which is not the
typical GEE case.

I was wandering if anyone has any personal experience on these issues,
especially to what be a valid method for the comparisons.

Thanks in advance for any insights.

Stavros Zanos, MD