Dear All,
I am using SPM TF-rescale (logR option, with a baseline [-300 0]ms) to
rescale each frequency bin of my TF. I however noticed that this leads
to some events, on some channels, to be perfectly 0 after correction.
Looking at the baseline, the scale looks smaller then for another
(randomly chosen) event, but it is for sure non-null (see attached: in
blue baseline of event leading to 0 trial, in red, baseline of previous
trial, on one channel, for one frequency bin).
Looking a little deeper, it seems that this is because the log10 is
performed on the whole baseline, before averaging, hence leading to a
-Inf value for the baseline. However, if the opposite was performed
(i.e. averaging across the baseline before taking the log10), the
baseline for that channel would have a value of 0.6341
(log10(mean(xbase)) instead of mean(log10(xbase))).
Having 'null' trials throws off all univariate tests I perform later on.
Worse, the channels with 'null' events have to be discarded from
multivariate analysis (in a specific case).
Also, I discovered this pitfall completely by coincidence, as I do not
review all events at each channel and each frequency bin. There should
be a warning and maybe marking those events as bad if that happens.
Can I perform the average before log10? How can this be solved?
Thanks!
Best,
Jessica
|