<div dir="ltr"><div><div>Hi all.<br><br>My question is: to know if I understand correctly the non parametric process using permutations, and to understand the process for the clustering part in it.<br><br>I wanted to make sure I understand correctly the meaning of the results I'm getting when using the non parametric approach<u><b> for multiple subjects</b></u> and the cluster statistics. i.e. I compare an 'Active' and 'Control' data sets that each subject has (for example, for 8 subject I have 16 averages over trials)<br>
<br>
As far as I understand:<br>Each subject has a subject-specific-average that is calculated for each condition (i.e. one vector (1X248 channels) for each condition that contains the average over selected time, freq and trials).<br>
Then, randomly permute the subject-specific averages of the two conditions, independently for every subject (i.e. in each permutation there will be one average selected from each subject) and calculate the statistic results for each channel.<br>
</div><div>We repeat this step many times as to construct a "permutation distribution" from the statistics value from each permutation.<br></div><div>finally, we look at the real "active vs control" statistics to see where their value lies in this permutation distribution, and that is the value we will save for that channel in the XXX.stat field. (we have 248 values and permutation distributions, one for each channel).<br>
</div><div><br>When we want the "clusters" - I am not totally understand the process here. <br>In the paper by Maris and Oostenveld , 2007 they talk about cluster in the following manner:<br>A Cluster based permutation test, is a test in which for each partition we cluster an <u><b>adjacent time samples</b></u> that exhibit a similar difference (in sign and magnitude). This is done by selecting all samples whose t-values is larger than some threshold, cluster the selected sample in connected set <u><b>on the basis of temporal adjacency</b></u> and calculate cluster level statistics by taking the sum of t-values within a cluster. Finally, for each partition we take the largest cluster statistics as a value for the permutation distribution. This distribution will serve us for testing the observed largest cluster statistics as was explained earlier.<br>
<br></div><div>When I use <b>cfg.correctm = 'cluster'; </b>Do we look for adjacent channels that are significant? so the clustering process is a spatial one (and not temporal?)<br><br></div><div><br></div><div>Now for what I did:<br>
I applied the statistic test using the following definitions and [stat] = ft_freqstatistics(...):<br><br>sublist = {'2','3','8','11','15','17','18','19'};% 8 subjects...<br>
cfg = [];<br> cfg.channel = {'MEG'};<br> cfg.latency = [1.8 1.95];<br><b> cfg.method = 'montecarlo';</b><br> cfg.frequency = 'all';<br><b> cfg.statistic = 'depsamplesT';<br>
cfg.correctm = 'cluster';</b><br> cfg.clusteralpha = 0.05;<br> cfg.clusterstatistic = 'maxsum';<br> cfg.minnbchan = 2;<br> cfg.tail = 0;<br> cfg.clustertail = 0;<br>
cfg.alpha = 0.025;<br> cfg.numrandomization = 500;<br> cfg_neighb.method = 'distance';<br> cfg.neighbours = ft_prepare_neighbours(cfg_neighb,alphaAvgSub3note2);<br> <br> Nsub = length(sublist);<br>
cfg.design(1,1:2*Nsub) = [ones(1,Nsub) 2*ones(1,Nsub)];% condition number<br> cfg.design(2,1:2*Nsub) = [1:Nsub 1:Nsub];% subject number<br> cfg.ivar = 1;<br> cfg.uvar = 2;<br> % cfg.avgovertime = 'yes';<br>
cfg.avgoverfreq = 'yes';<br> <br> [stat] = ft_freqstatistics(cfg, alphaAvgSub2note3, alphaAvgSub3note3, alphaAvgSub8note3, alphaAvgSub11note3,... <br> alphaAvgSub15note3, alphaAvgSub17note3, alphaAvgSub18note3, alphaAvgSub19note3,...<br>
alphaAvgSub2note2, alphaAvgSub3note2, alphaAvgSub8note2, ... <br> alphaAvgSub11note2, alphaAvgSub15note2, alphaAvgSub17note2, alphaAvgSub18note2, alphaAvgSub19note2);<br><br></div>I then plot the results using ft_clusterplot(...)<br>
<br></div><div>Thanks a lot<br><br></div><div>Inbal<br></div><div><br></div></div>