I am running a phase-locking analysis on MEG/EEG combined source data on the single trial level.
And I have a quick question for this mailing list
I was wondering whether the number of trials used for computing the Phase Lock Value (PLV) will affect the synchrony results?
I meant if I am comparing 2 conditions with different numbers of trials in each (let's say N=50 vs N=100) do you think that the synchrony will change because of different signal to noise ratio?
I am asking this because I have done such comparison, and strangely this is the condition with the less number of items that is showing increasing synchrony compared to the condition with the larger number (in the gamma band).
It will be very helpful if somebody already came across this issue, and/or if you can point me toward a relevant paper discussing this problem.
To my understanding, the PLV should not depend on the number of
trials. However, here are two tests you could do:
1. Pick N = 50 trials from the N = 100 condition a few times and check
whether the PLV is the same as when using all the trials.
2. Using bootstrapping, you could calculate the variability
(confidence intervals) of PLV in each condition to see whether this
difference is true or just due to the variability in the data.
PLV does in fact depend on the number of trials because it follows the Raleigh distribution. Think of this like a chi-square distribution. The mean/peak of the distribution will get closer to 0 as there are more trials, so your result is exactly what you would expect. It does not have anything to do with the SNR however.
For an example of this, calculate the PLV for random data with different numbers of trials. You will see the PLV change.
Avniel
Hi All,
I am running a phase-locking analysis on MEG/EEG combined source
data on the single trial level.
And I have a quick question for this mailing list
I was wondering whether the number of trials used for computing the
Phase Lock Value (PLV) will affect the synchrony results?
I meant if I am comparing 2 conditions with different numbers of
trials in each (let's say N=50 vs N=100) do you think that the
synchrony will change because of different signal to noise ratio?
I am asking this because I have done such comparison, and strangely
this is the condition with the less number of items that is showing
increasing synchrony compared to the condition with the larger
number (in the gamma band).
It will be very helpful if somebody already came across this issue,
and/or if you can point me toward a relevant paper discussing this
problem.
Thx to all for reading!
To my understanding, the PLV should not depend on the number of
trials. However, here are two tests you could do:
1. Pick N = 50 trials from the N = 100 condition a few times and check
whether the PLV is the same as when using all the trials.
2. Using bootstrapping, you could calculate the variability
(confidence intervals) of PLV in each condition to see whether this
difference is true or just due to the variability in the data.
PLV does in fact depend on the number of trials because it follows
the Raleigh distribution. Think of this like a chi-square
distribution. The mean/peak of the distribution will get closer to
0 as there are more trials, so your result is exactly what you would
expect. It does not have anything to do with the SNR however.
For an example of this, calculate the PLV for random data with
different numbers of trials. You will see the PLV change.
Hi all,
So my guess was wrong. The obvious remedy then to pick N = 50 random
trials from the N = 100 condition to make the two comparable, right?
Yes, picking 50 randomly would be a good and easy remedy.
I think the Raleigh distribution could also be used to "rescale" the PLV values, though I have to think about whether this would be valid from a statistical point of view (off the top of my head, you could transform your PLV values into probability values and compare the probability values, the problem is that I do not believe that probability values follow a normal distribution [particularly near 0 or 1], so there may be a further normalization step required).
Another reasonably easy route if you want to use all trials is to derive
the Null distribution non-parametrically by mixing all the 150 trials and
drawing many permutations of 50 and 100 trials respectively. There might
be some subtleties because variances are different for 50 and 100 trials
but its reasonable to ignore those higher order effects.