[FieldTrip] baseline correction - Cluster Based Permutation Test

eleonora parrotta eleonora_p at hotmail.it
Mon Oct 7 18:12:30 CEST 2019

Dear Fieldtrippers,

I am performing analysis of EEG data. I have two conditions of interest: A and B, and I have performed spectral decomposition with the wavelet method to obtain the power (frequencies from 1 to 50 Hz).

I performed a cluster-based permutation test to assess whether there are significant power differences between the two conditions of interest in the time-interval from the onset of the stimulus to 1 sec after:

%% clusterbased analysis

cfg = [];

cfg.latency = [0 1]; % Time of Interest

cfg.frequency = [8 12]; % Frequency of interest

cfg.channel =  {'PO3', 'P3', 'O1', 'P7','O2','P8','PO4','P4'};

cfg.neighbours = neighbours;

cfg.avgoverfreq = 'yes';

cfg.avgovertime = 'no';

cfg.avgoverchannel = 'no';

cfg.method = 'montecarlo';

cfg.statistic = 'ft_statfun_depsamplesT';

subj = 18;

design = zeros(2,2*subj);

for i = 1:subj

  design(1,i) = i;


for i = 1:subj

  design(1,subj+i) = i;


design(2,1:subj)= 1;design(2,subj+1:2*subj) = 2;

cfg.design   = design;

cfg.uvar     = 1;

cfg.ivar     = 2;

cfg.correctm         = 'cluster';

cfg.clusteralpha     = 0.05;

cfg.clusterstatistic = 'maxsum';

cfg.minnbchan        = 2;% specifies with which sensors other sensors can form clusters

cfg.tail             = 0;

cfg.clustertail      = 0;

cfg.alpha            = 0.05;

cfg.numrandomization = 1000;

[alpha] = ft_freqstatistics(cfg, avg_pred, avg_ran);

When I perform the cluster-based analysis on raw power I obtain more than one negative cluster, with highly significant differences between my two condition (p=0.01). The problem is that when I run the cluster-based analysis on baseline-corrected power the analysis is not even able to find any cluster of electrodes that behave differently in one condition compared to the other one (no matter which kind of baseline I applied, I tried with both ‘absolute’ and ‘relchange’ and even different time-window).

I am aware of the potential loss of power that can occur when we use baseline-correction (http://datacolada.org/39)

At the same time I am not sure how correct is to perform stats on non-baseline corrected power, as without applying this procedure drifts and offsets potentially present in the signal could affect the statistical test, executed on two non-normalized datasets.

I was wondering whether I am not aware of potential ways to specify that a baseline-correction has been applied to the data, or if there is a correct way of doing it.

So far I couldn’t find any reference that explains it clearly, but if anyone has an explanation I would really appreciate.

Thanks in advance,


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20191007/715e5b7a/attachment.html>

More information about the fieldtrip mailing list