Tags: madrid2019 eeg-sedation

Cluster-based permutation tests on resting-state EEG power spectra

Introduction

This tutorial gives an introduction to the statistical analysis of resting state EEG power spectra by means of cluster-based permutation tests.

We will sketch the background of permutation tests and apply it to different experimental questions. In this tutorial we will continue with the Chennu et al. dataset. We will use data that has already been preprocessed and spectrally analyzed, but if you are interested in the raw data from all subjects, you can download it from our download server. Please note that you do not have to download all subjects for this tutorial.

In a step-by-step fashion, this tutorial will show how to

  1. Compute within-participant contrasts
  2. Compute between-participant contrasts

Toward the end of the tutorial there are some more challenging exercises:

  1. Compute a multivariate ANOVA to test the drug effect on the entire power spectrum
  2. Compute a 2x2 interaction
  3. Compute a correlation between an external variable and the power spectrum

Overview

EEG data in the study by Chennu et al., 2016 was recorded from multiple participants in multiple experimental conditions, i.e. levels of sedation. Data is recorded in in blocks (runs) of 10 mins, in which the participants received varying amounts of Propofol, an anesthetic drug, aimed to produce a relaxed but still responsive behavioral state. For more details on the dataset click here.

For every subject, average power spectra are computed over all segments of data belonging to each of the sedative state. Thus, for every subject, the data are summarized as condition-specific power spectra for each of the channels. The question of interest is whether the data is different between the sedative states. The permutation test that will be described in the following informs us about the following null hypothesis: the probability distribution of the condition-specific power averages is not different between the sedative states.

Preparing the dataset

To test the difference between the baseline and moderate sedative states, we use data for all subjects that has been preprocessed and for which the power spectra have been computed. You can download freq_resting.mat from our download server. The MATLAB file contains four data structures, one for each sedation level.

% load averages for each individual subject, for each condition
load freq_resting.mat

Defining of regions of interest

occipital_ROI = {'E50','T5','E59','E60','Pz','E65','E66','E67','O1','E71','E72','Oz','E76','E77','O2','E84','E85','E90','E91','T6','E101','E51','E97'};
frontal_ROI   = {'E3','E4','E5','E6','E7','Fp2','E10','Fz','E12','E13','E15','E16','E18','E19','E20','Fp1','E23','F3','E27','E28','E29','E30','E105','E106','E111','E112','E117','E118','E123','F4'};

cfg = [];
cfg.keepindividual = 'yes';
base_sedation = ft_freqgrandaverage(cfg, base_sedation{:});
mild_sedation = ft_freqgrandaverage(cfg, mild_sedation{:});
mode_sedation = ft_freqgrandaverage(cfg, mode_sedation{:});
reco_sedation = ft_freqgrandaverage(cfg, reco_sedation{:});

Get the numerical indices of the regions of interest to compute averages

sel_oROI = match_str(base_sedation.label, occipital_ROI);
sel_fROI = match_str(base_sedation.label, frontal_ROI);

Also load the electrode positions

elec = prepare_elec_chennu2016(base_sedation.label);

In the frequency analysis tutorial we learnt different ways to normalize the power spectra, here we can employ one of them. We will normalize the power spectra for each of the individuals, using the mean power over a frequency range defined in freq_norm.

freq_oi   = [8 15];   % frequency range of interest
freq_norm = [0.7 40]; % frequency range used to normalize the spectrum
foi_norm  = nearest(base_sedation.freq, freq_norm);

common_denominator = mean(base_sedation.powspctrm(:,:,foi_norm(1):foi_norm(2)),3);
base_sedation.powspctrm_b = bsxfun(@rdivide, base_sedation.powspctrm, common_denominator);     mild_sedation.powspctrm_b = bsxfun(@rdivide, mild_sedation.powspctrm, common_denominator);
mode_sedation.powspctrm_b = bsxfun(@rdivide, mode_sedation.powspctrm, common_denominator);
reco_sedation.powspctrm_b = bsxfun(@rdivide, reco_sedation.powspctrm, common_denominator);

To simplify the results, let us collapse the data using the ROIs and frequency ranges defined in the original paper.

cfg = [];
cfg.channel     = frontal_ROI;
cfg.avgoverchan = 'yes';
cfg.frequency   = freq_oi;
cfg.avgoverfreq = 'yes';
cfg.parameter   = {'powspctrm','powspctrm_b'};
base_sedation_fROI = ft_selectdata(cfg, base_sedation);
mild_sedation_fROI = ft_selectdata(cfg, mild_sedation);
mode_sedation_fROI = ft_selectdata(cfg, mode_sedation);
reco_sedation_fROI = ft_selectdata(cfg, reco_sedation);

cfg.channel     = occipital_ROI;
base_sedation_oROI = ft_selectdata(cfg, base_sedation);
mild_sedation_oROI = ft_selectdata(cfg, mild_sedation);
mode_sedation_oROI = ft_selectdata(cfg, mode_sedation);
reco_sedation_oROI = ft_selectdata(cfg, reco_sedation);

Collect the data to plot it using the plotSpread function, which is specific for this tutorial and available from the download server.

data_raw_fROI    = {base_sedation_fROI.powspctrm...
  mild_sedation_fROI.powspctrm...
  mode_sedation_fROI.powspctrm...
  reco_sedation_fROI.powspctrm};
data_raw_oROI    = {base_sedation_oROI.powspctrm...
  mild_sedation_oROI.powspctrm...
  mode_sedation_oROI.powspctrm...
  reco_sedation_oROI.powspctrm};

data_between_fROI ={base_sedation_fROI.powspctrm_b...
  mild_sedation_fROI.powspctrm_b...
  mode_sedation_fROI.powspctrm_b...
  reco_sedation_fROI.powspctrm_b};
data_between_oROI ={base_sedation_oROI.powspctrm_b...
  mild_sedation_oROI.powspctrm_b...
  mode_sedation_oROI.powspctrm_b...
  reco_sedation_oROI.powspctrm_b};

figure

subplot(2,2,1);
plotSpread(data_raw_fROI,[],[],{'baseline','mild','moderate','recovery'});
ylabel('abs. power (V^2)');
title('raw PSD Front');

subplot(2,2,2);
h3 = plotSpread(data_between_fROI,[],[],{'baseline','mild','moderate','recovery'});
ylabel('rel. power');
title('between PSD Front');
set(h3{1},'LineWidth',1,'Marker', '.','Color','k','MarkerFaceColor','k')

subplot(2,2,3);
plotSpread(data_raw_oROI,[],[],{'baseline','mild','moderate','recovery'});
ylabel('abs. power (V^2)');
title('raw PSD Occip');

subplot(2,2,4);
h6 = plotSpread(data_between_oROI,[],[],{'baseline','mild','moderate','recovery'});
ylabel('rel. power');
title('between PSD Occip');
set(h6{1},'LineWidth',1,'Marker', '.','Color','k','MarkerFaceColor','k')

Let us make topoplots and power spectra for each ROI in each sedative condition, similar to Fig 5A in original paper.

cfg = [];
cfg.elec             = elec;
cfg.parameter        = 'powspctrm_b'; % you can plot either powspctrm (default) or powspctrm_b
cfg.xlim             = [8 15]; % frequency range to make the topoplot
cfg.highlight        = 'on';

We can use some options to improve the figure cosmetics.

cfg.highlightchannel = {frontal_ROI occipital_ROI};
cfg.highlightsymbol  = {'o','*'};
cfg.highlightcolor   = [0 0 0];
cfg.highlightsize    = 6;
cfg.markersymbol     = '.';
cfg.comment          = 'no';
cfg.colormap         = 'jet';

figure('position',[680 240 1039 420]);
subplot(2,4,1); ft_topoplotER(cfg, base_sedation); colorbar; title('baseline');
subplot(2,4,2); ft_topoplotER(cfg, mild_sedation); colorbar; title('mild');
subplot(2,4,3); ft_topoplotER(cfg, mode_sedation); colorbar; title('moderate');
subplot(2,4,4); ft_topoplotER(cfg, reco_sedation); colorbar; title('recovery');

subplot(2,4,5);loglog(base_sedation.freq,...
  [squeeze(mean(mean(base_sedation.(cfg.parameter)(:,sel_fROI,:),2),1))...
  squeeze(mean(mean(base_sedation.(cfg.parameter)(:,sel_oROI,:),2),1))]);
xlim([0.5 45]);
grid on; hold on;
plot([10,10],[10^-3 10^2],'--k')
ylim([10^-3 10^2]);
legend('Front ROI','Occip ROI','Location','southwest');
xlabel('Frequency (Hz)');
ylabel(cfg.parameter);
title('baseline');

subplot(2,4,6);loglog(mild_sedation.freq,...
  [squeeze(mean(mean(mild_sedation.(cfg.parameter)(:,sel_fROI,:),2),1))...
  squeeze(mean(mean(mild_sedation.(cfg.parameter)(:,sel_oROI,:),2),1))]);
xlim([0.5 45]);
grid on; hold on;
plot([10,10],[10^-3 10^2],'--k')
ylim([10^-3 10^2]);
xlabel('Frequency (Hz)');
ylabel(cfg.parameter);
title('mild');

subplot(2,4,7);loglog(mode_sedation.freq,...
  [squeeze(mean(mean(mode_sedation.(cfg.parameter)(:,sel_fROI,:),2),1))...
  squeeze(mean(mean(mode_sedation.(cfg.parameter)(:,sel_oROI,:),2),1))]);
xlim([0.5 45]);
grid on; hold on;
plot([10,10],[10^-3 10^2],'--k')
ylim([10^-3 10^2]);
xlabel('Frequency (Hz)');
ylabel(cfg.parameter);
title('moderate');

subplot(2,4,8);loglog(reco_sedation.freq,...
  [squeeze(mean(mean(reco_sedation.(cfg.parameter)(:,sel_fROI,:),2),1))...
  squeeze(mean(mean(reco_sedation.(cfg.parameter)(:,sel_oROI,:),2),1))]);
xlim([0.5 45]);
grid on; hold on;
plot([10,10],[10^-3 10^2],'--k')
ylim([10^-3 10^2]);
xlabel('Frequency (Hz)');
ylabel(cfg.parameter);
title('recovery');

1. Compute within-participant contrasts

We now consider experiments involving multiple subjects that are each observed in multiple experimental conditions. Typically, every subject is observed in a large number of trials, each one belonging to one experimental condition. Usually, for every subject, averages are computed over all trials belonging to each of the experimental conditions. Thus, for every subject, the data are summarized in an array of condition-specific averages. The permutation test that is described in this section informs us about the following null hypothesis: the probability distribution of the condition-specific averages is independent of the experimental conditions.

Permutation test

Cluster-level permutation tests for power spectra are performed by the function ft_freqstatistics. This function takes as its input arguments a configuration structure (cfg) and one or multiple data structures. These data structures must be produced by ft_freqanalysis or ft_freqgrandaverage. The argument list of ft_freqstatistics contains one data structure for every experimental condition. For comparing the data structures base_sedation and mode_sedation, you must call ft_freqstatistics as follows:

% do NOT EXECUTE this yet, it is just to introduce the function
[stat] = ft_freqstatistics(cfg, base_sedation, mode_sedation);

Many fields of the configuration (cfg), such as the selection of channels, are not unique to ft_freqstatistics; their role is similar in other FieldTrip functions. We first concentrate on the fields that are unique to ft_freqstatistics.

foi_contrast = [0.5 30];

cfg = [];
cfg.channel          = 'all';
cfg.frequency        = foi_contrast;
cfg.parameter        = 'powspctrm_b';
cfg.method           = 'ft_statistics_montecarlo';  % use the Monte Carlo method to calculate probabilities
cfg.statistic        = 'ft_statfun_depsamplesT';    % use the dependent samples T-statistic as a measure to evaluate the effect at each sample
cfg.correctm         = 'cluster';
cfg.clusteralpha     = 0.05;                        % threshold for the sample-specific test, is used for thresholding
cfg.clusterstatistic = 'maxsize';
cfg.clusterthreshold = 'nonparametric_common';
cfg.minnbchan        = 2;                           % minimum number of neighbouring channels that is required
cfg.tail             = 0;                           % test the left, right or both tails of the distribution
cfg.clustertail      = cfg.tail;
cfg.alpha            = 0.05;                        % alpha level of the permutation test
cfg.correcttail      = 'alpha';                     % see https://www.fieldtriptoolbox.org/faq/why_should_i_use_the_cfg.correcttail_option_when_using_statistics_montecarlo/
cfg.computeprob      = 'yes';
cfg.numrandomization = 500;                         % number of random permutations
cfg.neighbours       = cfg_neigh.neighbours;        % the neighbours for each sensor to form clusters

nsubj                     = size(base_sedation.powspctrm,1);
design                    = zeros(2,2*nsubj);
design(1,1:nsubj)         = 1;
design(1,nsubj+1:2*nsubj) = 2;
design(2,1:nsubj)         = 1:nsubj;
design(2,nsubj+1:2*nsubj) = 1:nsubj;

cfg.design   = design; % design matrix
cfg.ivar     = 1;      % the 1st row codes the independent variable (sedation level)
cfg.uvar     = 2;      % the 2nd row codes the unit of observation (subject)

In the following section we will describe the various options one-by-one.

  • With cfg.method = ft_statistics_montecarlo we choose the Monte Carlo method for calculating the significance probability. This significance probability is a Monte Carlo estimate of the p-value under the permutation distribution.

  • With cfg.statistic = ft_statfun_depsamplesT we choose the dependent samples T-statistic to evaluate the effect (the difference between the baseline and the moderate condition) at the sample level. In cfg.statistic, many other test statistics can be specified. Which test statistic is appropriate depends on your research question and your experimental design. For instance, in a within-UO design (present one), one must use the dependent samples T-statistic (ft_statfun_depsamplesT). And if you want to compare more than two experimental conditions, you should choose an F-statistic (ft_statfun_indepsamplesT or ft_statfun_depsamplesFmultivariate; you can give a try in the Challenging exercise section below).

  • We use cfg.clusteralpha to choose the critical value that will be used for thresholding the sample-specific T-statistics. With a value of 0.05, every sample-specific T-statistic is compared with the critical value of the univariate T-test with a critical alpha-level of 0.05. The value of cfg.clusteralpha does not affect the false alarm rate of the statistical test at the cluster-level. It is a rational threshold for deciding whether a sample should be considered a member of some large cluster of samples, which may or may not be significant at the cluster-level.

  • We use cfg.clusterstatistic to choose the test statistic that will be evaluated under the permutation distribution. This is the actual test statistic and it must be distinguished from the sample-specific T-statistics that are used for thresholding. With ‘maxsum’, the actual test statistic is the maximum of the cluster-level statistics. A cluster-level statistic is equal to the sum of the sample-specific T-statistics that belong to this cluster. Taking the largest of these cluster-level statistics of the different clusters produces the actual test statistic.

  • The value of cfg.minnbchan is a tuning parameter that determines the way the clusters are formed. More specifically, we use it to specify the minimum number of channels from the neighbours that is required for a selected sample (i.e., a sample who’s T-statistic exceeds the threshold) to be included in the clustering algorithm. With the default value, it sometimes happens that two clusters are spatially connected via a narrow bridge. Because they are connected, these two clusters merge into a single cluster. If you want to interpret clusters as reflecting spatially distinct sources, such a combined cluster does not make much sense. To suppress this type of combined clusters, you can choose to ignore all selected samples (on the basis of their T-values) if they have less than some minimum number of neighbors that were also selected. This minimum number is assigned to cfg.minnbchan. This number must be chosen independently of the data.

  • cfg.neighbours is a structure that you need to have previously created using ft_prepare_neighbours.

  • We use cfg.tail to choose between a one-sided and a two-sided statistical test. Choosing cfg.tail = 0 affects the calculations in three ways. First, the sample-specific T-values are thresholded from below as well as from above. This implies that both large negative and large positive T-statistics are selected for later clustering. Second, clustering is performed separately for thresholded positive and thresholded negative T-statistics. And third, the critical value for the cluster-level test statistic (determined by cfg.alpha; see further) is now two-sided: negative cluster-level statistics must be compared with the negative critical value, and positive cluster-level statistics must be compared with the positive critical value.

  • We use cfg.alpha to control the false alarm rate of the permutation test (the probability of falsely rejecting the null hypothesis). The value of cfg.alpha determines the critical values with which we must compare the test statistic (i.e., the maximum and the minimum cluster-level statistic).

If you want to run a two-sided test, i.e. you want to test both the left (negative clusters) and the right (positive clusters) tail of the distribution. you have to split the critical alpha value by setting cfg.correcttail = ‘alpha’; i.e. this effectively does a Bonferroni correction and sets cfg.alpha = 0.025, corresponding to a false alarm rate of 0.05 in a two-sided test.*

  • We use cfg.numrandomization to control the number of draws from the permutation distribution. Remember that ft_freqstatistics approximates the permutation distribution by means of a histogram with a Monte Carlo approximation of the true permutation distribution. In this tutorial, we use cfg.numrandomization = 500 to keep the computational time low. In general you should set this to a higher value (1000 or up). If it turns out that estimated p-value is very close to the the critical alpha-level (0.05 or 0.01), you should increase this number.

  • We use cfg.design to store information about the independent variable (the experimental manipulation) and the units of observation (the subjects). Consider the hypothetical case that 12 participants has been observed in two different conditions. Then the design matrix looks like this:

cfg.design =

Columns 1 through 12
1     1     1     1     1     1     1     1     1     1     1     1...
1     2     3     4     5     6     7     8     9    10    11    12...

Columns 12 through 24
2     2     2     2     2     2     2     2     2     2     2     2...
1     2     3     4     5     6     7     8     9    10    11    12...
  • We use cfg.ivar to indicate the row of the design matrix that contains the independent variable (i.e. condition).

You should be aware that the sensitivity of the statistical test (i.e., the probability of detecting an effect) depends on the width of the frequency interval that is analyzed, as specified in cfg.frequency. For instance, assume that the difference between the two experimental conditions extends over a short frequency interval only (e.g., between 8-12hz). If you would have good reasons a-priori that this short frequency interval is the only interval where an effect is likely to occur, then you should limit the analysis to this interval (i.e., choose cfg.frequency = [8 12]). Choosing a frequency interval on the basis of prior information increases the sensitivity of the statistical test. If there is no prior information, you must compare the experimental conditions over the complete frequency interval. This is accomplished by choosing cfg.frequency = ‘all’.

Now, run ft_freqstatistics to compare base_sedation and mode_sedation using the configuration described above.

stat1 = ft_freqstatistics(cfg, base_sedation, mode_sedation);

The format of the output

The output of ft_freqstatistics has separate fields for positive and negative clusters. For the positive clusters, the output is given in the following pair of fields: stat1.posclusters and stat1.posclusterslabelmat. The field stat.posclusters is an array that provides the following information for every cluste

  • The field stat1.clusterstat contains the cluster-level statistic (the sum of the T-values in each cluster).

  • The field stat1.prob contains the proportion of draws from the permutation distribution with a maximum cluster-level statistic that is larger than clusterstat. The elements in the array stat.posclusters are sorted according to their p-value: the cluster with the smallest p-value comes first, followed by the cluster with the second-smallest, etc. Thus, if the k-th cluster has a p-value that is larger than the critical alpha-level (e.g., 0.025), then so does the (k+1)-th. Type stat.posclusters(k) on the MATLAB command line to see the information for the k-th cluster.

The field stat1.posclusterslabelmat contains indices (i.e. numbers 1, 2, 3, …) that identify the clusters to which the (channel,frequency)-pairs (the samples) belong. For example, all (channel,frequency)-pairs that belong to the third cluster, are identified by the number 3. As will be shown in the following, this information can be used to visualize the topography of the clusters.

For the negative clusters, the output is given in the following pair of fields: stat1.negclusters and stat1.negclusterslabelmat. These fields contain the same type of information as stat1.posclusters and stat1.posclusterslabelmat, but now for the negative clusters.

By inspecting stat1.posclusters and stat1.negclusters, it can be seen that only the first positive and the first negative cluster have a p-value less than the critical alpha-level of 0.025. This critical alpha-level corresponds to a false alarm rate of 0.05 in a two-sided test. You can inspect the details of the first positive cluster like this

stat1.posclusters(1)

ans =
       prob: 0.2136
clusterstat: 68
     stddev: 0.0183
    cirange: 0.0359

And by typing stat.negclusters(1), you should obtain the following:

stat1.negclusters(1)

ans =
       prob: 0.0020
clusterstat: -3046
     stddev: 0.0020
    cirange: 0.0039

It is likely that the p-values in your specific output are a bit different. This is because ft_freqstatistics calculated a Monte Carlo approximation of the permutation p-values: the p-value for the k-th positive cluster is calculated as the proportion of random draws from the permutation distribution in which the maximum of the cluster-level statistics is larger than stat.posclusters(k).clusterstat. The random draws will be different every time you execute the code, hence the distribution will also be slightly different.

Plotting the results

Get the 1st positive and negative cluster

sigposmask = (stat1.posclusterslabelmat==1) & stat1.mask;
signegmask = (stat1.negclusterslabelmat==1) & stat1.mask;

cfg = [];
cfg.frequency  = foi_contrast;
cfg.avgoverrpt = 'yes';
cfg.parameter  = {'powspctrm','powspctrm_b'};
base_sedation_avg = ft_selectdata(cfg, base_sedation);
mild_sedation_avg = ft_selectdata(cfg, mild_sedation);
mode_sedation_avg = ft_selectdata(cfg, mode_sedation);
reco_sedation_avg = ft_selectdata(cfg, reco_sedation);

Choose the cluster you want to see: either positive or negative

base_sedation_avg.mask = signegmask;
mode_sedation_avg.mask = signegmask;

cfg = [];
cfg.elec          = elec;
cfg.colorbar      = 'no';
cfg.maskparameter = 'mask';  % use the thresholded probability to mask the data
cfg.maskstyle     = 'box';
cfg.parameter     = 'powspctrm_b';
cfg.maskfacealpha = 0.5;

figure; ft_multiplotER(cfg, base_sedation_avg, mode_sedation_avg);
title('within-participant BASELINE vs MODERATE');

2. Compute between-participants contrasts

In a between-participant experimental designs, we analyze the data of multiple participants that are observed during different experimental conditions, for example comparing a control group to a patient group. In these cases we want to answer in general the question whether there is a systematic difference in the EEG recorded in the two groups. In the experiment that we are using, each level of sedation has been applied to the same subject. But the way that subjects responded to it differs, i.e. we can split the participants in a drowsy and a responsive group.

First we compute the hit rate of each participant and condition knowing that the number of correct responses in that task is 40. See point 5.

hit_rate = (covariates(:,:,3)./40).*100;

Chennu et al 2016 made a more elaborate behavioral analysis to detect the drowsy group. Here we just used to select arbitrarily 70% threshold to match Figure 1B of the original paper.

We now describe how to statistically evaluate the power spectra difference between the responsive and the drowsy groups. The format for these variables, are a prime example of how you should organize your data to be suitable for ft_freqstatistics. Specifically, each variable is a structure, with each subject’s averaged stored in one cell. To create this data structure we will make copies and we will select the subgroups based on the behavioral performance (see below). Note that this subgroup selection is orthogonal to the EEG contrast we will test.

drowsy_group = find(hit_rate(:,3)<70);
respon_group = setxor(1:size(covariates,1),drowsy_group);

figure;
plot(hit_rate(respon_group,:)','-^b','MarkerFaceColor','b');
hold on;
plot(hit_rate(drowsy_group,:)','marker','^','color',[0 0.5 0],'MarkerFaceColor',[0 0.5 0])
hold off;
ylabel('Perceptual hit rate (%)');
set(gca,'XTickLabel',{'baseline','','mild','','moderate','','recovery'});

copy the datasets and select the relevant subgroups

[base_sedation_respon base_sedation_drowsy] = deal(base_sedation);
[mild_sedation_respon mild_sedation_drowsy] = deal(mild_sedation);
[mode_sedation_respon mode_sedation_drowsy] = deal(mode_sedation);
[reco_sedation_respon reco_sedation_drowsy] = deal(reco_sedation);

cfg = [];
cfg.parameter   = {'powspctrm','powspctrm_b'};

cfg.trials      = respon_group; % cfg.trials will select the 'subj' dimension
base_sedation_respon = ft_selectdata(cfg, base_sedation_respon);
mild_sedation_respon = ft_selectdata(cfg, mild_sedation_respon);
mode_sedation_respon = ft_selectdata(cfg, mode_sedation_respon);
reco_sedation_respon = ft_selectdata(cfg, reco_sedation_respon);

cfg.trials      = drowsy_group; % cfg.trials will select the 'subj' dimension
base_sedation_drowsy = ft_selectdata(cfg, base_sedation_drowsy);
mild_sedation_drowsy = ft_selectdata(cfg, mild_sedation_drowsy);
mode_sedation_drowsy = ft_selectdata(cfg, mode_sedation_drowsy);
reco_sedation_drowsy = ft_selectdata(cfg, reco_sedation_drowsy);

We now perform the permutation test using ft_freqstatistics. The configuration settings for this analysis differ from the previous settings in several fields:

  1. We have to select a different statistic for the sample level effect (in cfg.statistic)
  2. The design matrix is different (i.c., it now contains only one row instead of two)
  3. The independent variable is now the group assignment

The configuration looks as follow:

cfg                 = [];
cfg.channel          = 'all';
cfg.avgovergchan     = 'no';
cfg.frequency        = foi_contrast;
cfg.avgovergfreq     = 'yes';
cfg.parameter        = 'powspctrm_b';
cfg.method           = 'ft_statistics_montecarlo';
cfg.statistic        = 'ft_statfun_indepsamplesT';
cfg.correctm         = 'cluster';
cfg.clusteralpha     = 0.05;
cfg.clusterstatistic = 'maxsum';
cfg.clusterthreshold = 'nonparametric_common';
cfg.minnbchan        = 2;
cfg.tail             = 0;
cfg.clustertail      = cfg.tail;
cfg.alpha            = 0.05;
cfg.correcttail      = 'alpha';
cfg.computeprob      = 'yes';
cfg.numrandomization = 1000;
cfg.neighbours       = cfg_neigh.neighbours;

design = zeros(1,size(respon_group,1) + size(drowsy_group,1));
design(1,1:size(respon_group,1)) = 1;
design(1,(size(respon_group,1)+1):(size(respon_group,1)+size(drowsy_group,1))) = 2;

cfg.design = design;
cfg.ivar   = 1;

We now describe the differences between this configuration and the configuration for a within-trials experiment.

  • Instead of an dependent samples T-statistic, we use the independent samples T-statistic to evaluate the effect at the sample level (cfg.statistic = ft_statfun_indepsamplesT). This is because we are dealing with a between-UO instead of a within-UO design.

  • The design matrix in a between-UO design is different from the design matrix in a within-UO design. In the design matrix for a between-UO design, you have to specify the group to which each unit (subject) is assigned. For example, consider a hypothetical study with two experimental groups, the first group with 4 subjects and the second group with 6 subjects: the design matrix then looks like this: design = [1 1 1 1 2 2 2 2 2 2]. The data from the subjects will be shuffled between the two groups, but the number of subjects per group will remain the same (i.e. 4 versus 6).

Now, use the configuration above to perform the following statistical analysis:

stat2 = ft_freqstatistics(cfg, reco_sedation_respon, reco_sedation_drowsy);

Plotting the results

cfg = [];
cfg.alpha     = stat2.cfg.alpha;
cfg.parameter = 'stat';
cfg.zlim      = [-3 3];
cfg.elec      = elec;
ft_clusterplot(cfg, stat2);

cfg = [];
cfg.elec         = elec;
cfg.zlim         = [1.5 3];
cfg.xlim         = [8 15];
cfg.parameter    = 'powspctrm_b';
cfg.markersymbol = '.';
cfg.comment      = 'no';
cfg.colormap     = 'jet';
cfg.colorbar     = 'no';

figure('position',[680 240 1039 420]);
subplot(2,4,1); ft_topoplotER(cfg, base_sedation_respon); colorbar; title('base Responsive');
subplot(2,4,2); ft_topoplotER(cfg, mild_sedation_respon); colorbar; title('mild Responsive');
subplot(2,4,3); ft_topoplotER(cfg, mode_sedation_respon); colorbar; title('mode Responsive');
subplot(2,4,4); ft_topoplotER(cfg, reco_sedation_respon); colorbar; title('reco Responsive');

subplot(2,4,5); ft_topoplotER(cfg, base_sedation_drowsy); colorbar; title('base Drowsy');
subplot(2,4,6); ft_topoplotER(cfg, mild_sedation_drowsy); colorbar; title('mild Drowsy');
subplot(2,4,7); ft_topoplotER(cfg, mode_sedation_drowsy); colorbar; title('mode Drowsy');
subplot(2,4,8); ft_topoplotER(cfg, reco_sedation_drowsy); colorbar; title('reco Drowsy');

We can also plot power spectra for each group separately, as a function of sedative state. Let us first plot them for the frontal region of interest:

figure;

subplot(2,4,1); loglog(base_sedation_respon.freq,...
  [squeeze(mean(mean(base_sedation_respon.(cfg.parameter)(:,sel_fROI,:),2),1))...
  squeeze(mean(mean(base_sedation_drowsy.(cfg.parameter)(:,sel_fROI,:),2),1))]);
xlim([0.5 45]);
grid on; hold on;
plot([10,10],[10^-3 10^2],'--k')
ylim([10^-3 10^2]);
legend('Front ROI resp','Front ROI drow','Location','southwest');
xlabel('Frequency (Hz)');
ylabel(cfg.parameter);
title('baseline');

subplot(2,4,2); loglog(mild_sedation_respon.freq,...
  [squeeze(mean(mean(mild_sedation_respon.(cfg.parameter)(:,sel_fROI,:),2),1))...
  squeeze(mean(mean(mild_sedation_drowsy.(cfg.parameter)(:,sel_fROI,:),2),1))]);
xlim([0.5 45]);
grid on; hold on;
plot([10,10],[10^-3 10^2],'--k')
ylim([10^-3 10^2]);
xlabel('Frequency (Hz)');
ylabel(cfg.parameter);
title('mild');

subplot(2,4,3); loglog(mode_sedation_respon.freq,...
  [squeeze(mean(mean(mode_sedation_respon.(cfg.parameter)(:,sel_fROI,:),2),1))...
  squeeze(mean(mean(mode_sedation_drowsy.(cfg.parameter)(:,sel_fROI,:),2),1))]);
xlim([0.5 45]);
grid on; hold on;
plot([10,10],[10^-3 10^2],'--k')
ylim([10^-3 10^2]);
xlabel('Frequency (Hz)');
ylabel(cfg.parameter);
title('moderate');

subplot(2,4,4); loglog(reco_sedation_respon.freq,...
  [squeeze(mean(mean(reco_sedation_respon.(cfg.parameter)(:,sel_fROI,:),2),1))...
  squeeze(mean(mean(reco_sedation_drowsy.(cfg.parameter)(:,sel_fROI,:),2),1))]);
xlim([0.5 45]);
grid on; hold on;
plot([10,10],[10^-3 10^2],'--k')
ylim([10^-3 10^2]);
xlabel('Frequency (Hz)');
ylabel(cfg.parameter);
title('recover');

And for the occipital region of interest:

subplot(2,4,5); loglog(base_sedation_respon.freq,...
  [squeeze(mean(mean(base_sedation_respon.(cfg.parameter)(:,sel_oROI,:),2),1))...
  squeeze(mean(mean(base_sedation_drowsy.(cfg.parameter)(:,sel_oROI,:),2),1))]);
xlim([0.5 45]);
grid on; hold on;
plot([10,10],[10^-3 10^2],'--k')
ylim([10^-3 10^2]);
xlabel('Frequency (Hz)');
ylabel(cfg.parameter);
legend('Occip ROI resp','Occip ROI drow','Location','southwest');

subplot(2,4,6); loglog(mild_sedation_respon.freq,...
  [squeeze(mean(mean(mild_sedation_respon.(cfg.parameter)(:,sel_oROI,:),2),1))...
  squeeze(mean(mean(mild_sedation_drowsy.(cfg.parameter)(:,sel_oROI,:),2),1))]);
xlim([0.5 45]);
grid on; hold on;
plot([10,10],[10^-3 10^2],'--k')
ylim([10^-3 10^2]);
xlabel('Frequency (Hz)');
ylabel(cfg.parameter);

subplot(2,4,7); loglog(mode_sedation_respon.freq,...
  [squeeze(mean(mean(mode_sedation_respon.(cfg.parameter)(:,sel_oROI,:),2),1))...
  squeeze(mean(mean(mode_sedation_drowsy.(cfg.parameter)(:,sel_oROI,:),2),1))]);
xlim([0.5 45]);
grid on; hold on;
plot([10,10],[10^-3 10^2],'--k')
ylim([10^-3 10^2]);
xlabel('Frequency (Hz)');
ylabel(cfg.parameter);

subplot(2,4,8); loglog(reco_sedation_respon.freq,...
  [squeeze(mean(mean(reco_sedation_respon.(cfg.parameter)(:,sel_oROI,:),2),1))...
  squeeze(mean(mean(reco_sedation_drowsy.(cfg.parameter)(:,sel_oROI,:),2),1))]);
xlim([0.5 45]);
grid on; hold on;
plot([10,10],[10^-3 10^2],'--k')
ylim([10^-3 10^2]);
xlabel('Frequency (Hz)');
ylabel(cfg.parameter);

The remainder of the tutorial contains some more challenging exercises.

3. Compute a multivariate ANOVA to test the drug effect on the entire power spectrum

The Chennu et al. dataset is very rich and allows us to look at more complex contrasts. Let us consider not only simple contract between two sedative states, but all four states. The power spectrum of very subject is summarized in 4 arrays of condition-specific averages. The permutation test that is described in this section addresses the following null hypothesis: the probability distribution of the condition-specific averages is independent of the four experimental conditions.

cfg = [];
cfg.channel          = 'all';
cfg.frequency        = foi_contrast;
cfg.avgovergfreq     = 'no';
cfg.parameter        = 'powspctrm_b';
cfg.method           = 'ft_statistics_montecarlo';
cfg.statistic        = 'ft_statfun_depsamplesFmultivariate';
cfg.correctm         = 'cluster';
cfg.clusteralpha     = 0.05;
cfg.clusterstatistic = 'maxsum'; %'maxsum', 'maxsize', 'wcm'
cfg.clusterthreshold = 'nonparametric_common';
cfg.minnbchan        = 2;
cfg.tail             = 1; % For a F-statistic, it only make sense to calculate the right tail
cfg.clustertail      = cfg.tail;
cfg.alpha            = 0.05;
cfg.computeprob      = 'yes';
cfg.numrandomization = 500;
cfg.neighbours       = cfg_neigh.neighbours;

nsubj = size(covariates,1);
design = zeros(2,4*nsubj);
design(1,1:nsubj)           = 1;
design(1,nsubj+1:2*nsubj)   = 2;
design(1,nsubj*2+1:3*nsubj) = 3;
design(1,nsubj*3+1:4*nsubj) = 4;
design(2,:) = repmat(1:nsubj,1,4);

cfg.design = design;
cfg.ivar   = 1; % sedation level
cfg.uvar   = 2; % subject number
  • We use the ft_statfun_depsamplesFmultivariate for a repeated-measures (i.e. dependent samples) multivariate ANOVA.

  • The design matrix, the cfg.ivar and the cfg.uvar will be the same as in the within-UO design

We now pass the data from all four conditions as input variables:

stat3 = ft_freqstatistics(cfg, base_sedation, mild_sedation, mode_sedation, reco_sedation);

Plotting the results

cfg            = [];
cfg.frequency  = foi_contrast;
cfg.avgoverrpt = 'yes';
cfg.parameter  = {'powspctrm','powspctrm_b'};
base_sedation_avg = ft_selectdata(cfg, base_sedation);
mild_sedation_avg = ft_selectdata(cfg, mild_sedation);
mode_sedation_avg = ft_selectdata(cfg, mode_sedation);
reco_sedation_avg = ft_selectdata(cfg, reco_sedation);

% copy the mask field to each variable
base_sedation_avg.mask = stat3.mask;
mild_sedation_avg.mask = stat3.mask;
mode_sedation_avg.mask = stat3.mask;
reco_sedation_avg.mask = stat3.mask;

cfg = [];
cfg.zlim          = [0 90];
cfg.elec          = elec;
cfg.colorbar      = 'no';
cfg.maskparameter = 'mask';  % use the thresholded probability to mask the data
cfg.maskstyle     = 'box';
cfg.parameter     = 'powspctrm_b';
cfg.maskfacealpha = 0.1;
figure; ft_multiplotER(cfg, base_sedation_avg, mild_sedation_avg, mode_sedation_avg, reco_sedation_avg);

4. Compute a 2x2 interaction

The interaction effects are usually the most interesting in a multifactorial design. In this dataset we will compute the interaction using the within-participant factor SEDATION (baseline vs moderate) and the between-subject factor GROUP (responsive vs drowsy) based on the strategy outlined in this frequently asked question.

Let us prepare the data

  1. Compute the within-participant contrast for each group: baseline vs moderate
  2. Compute the between-participant contrast of the differences computed in step 1

The complexity here is that we are dealing with a between-within participants design, but the design matrix and statistical pipeline is the one we used in between-participant contrast

cfg = [];
cfg.parameter = {'powspctrm','powspctrm_b'};
cfg.operation = 'subtract';

sedation_respon_d = ft_math(cfg, base_sedation_respon, mode_sedation_respon);
sedation_drowsy_d = ft_math(cfg, base_sedation_drowsy, mode_sedation_drowsy);

cfg = [];
cfg.channel          = 'all';
cfg.frequency        = foi_contrast;
cfg.avgovergfreq     = 'no';
cfg.parameter        = 'powspctrm_b';
cfg.method           = 'ft_statistics_montecarlo';
cfg.statistic        = 'ft_statfun_indepsamplesT';
cfg.correctm         = 'cluster';
cfg.clusteralpha     = 0.05;
cfg.clusterstatistic = 'maxsum';
cfg.clusterthreshold = 'nonparametric_common';
cfg.minnbchan        = 2;
cfg.tail             = 0;
cfg.clustertail      = cfg.tail;
cfg.alpha            = 0.05;
cfg.correcttail      = 'alpha';
cfg.computeprob      = 'yes';
cfg.numrandomization = 1000;
cfg.neighbours       = cfg_neigh.neighbours;

design = zeros(1,size(respon_group,1) + size(drowsy_group,1));
design(1,1:size(respon_group,1)) = 1;
design(1,(size(respon_group,1)+1):(size(respon_group,1)+size(drowsy_group,1))) = 2;

cfg.design = design;
cfg.ivar   = 1;

stat4 = ft_freqstatistics(cfg, sedation_respon_d, sedation_drowsy_d);

Now select the significant sensors and frequencies and plot the interaction.

Get the 1st positive and negative cluster

signegmask = (stat4.negclusterslabelmat==1) & stat4.mask;

Pool all channels and frequencies that in the cluster

chanoineg = match_str(stat4.label,stat4.label(sum(signegmask,2) > 0));
foilimneg = stat4.freq(sum(signegmask,1) > 0);
foilim = [min(foilimneg) max(foilimneg)];

Choose the cluster you want to see

base_sedation_avg.mask = signegmask;
mode_sedation_avg.mask = signegmask;

Choose the cluster you want to see: positive or negative

base_sedation_avg.mask = signegmask;
mode_sedation_avg.mask = signegmask;

cfg = [];
cfg.frequency   = foilim;
cfg.avgoverfreq = 'yes';
cfg.channel     = chanoineg;
cfg.avgoverchan = 'yes';
cfg.parameter   = {'powspctrm','powspctrm_b'};
b_r = ft_selectdata(cfg, base_sedation_respon);
m_r = ft_selectdata(cfg, mode_sedation_respon);
b_d = ft_selectdata(cfg, base_sedation_drowsy);
m_d = ft_selectdata(cfg, mode_sedation_drowsy);

Compute the confidence intervals

parameter = 'powspctrm_b'; % make sure this is the same as the one used in ft_freqstatistics

Compute the means

x_b_r = mean(b_r.(parameter),1);
x_m_r = mean(m_r.(parameter),1);
x_b_d = mean(b_d.(parameter),1);
x_m_d = mean(m_d.(parameter),1);

sem_b_r = sem(b_r.(parameter),1);
sem_m_r = sem(m_r.(parameter),1);
sem_b_d = sem(b_d.(parameter),1);
sem_m_d = sem(m_d.(parameter),1);

figure;

errorbar([1 2],[x_b_r,x_m_r]',[sem_b_r,sem_m_r(1)],'-rs');
ylabel(parameter);
hold all
errorbar([1 2],[x_b_d x_m_d]',[sem_b_r,sem_m_r],'-bs');
ylabel(parameter);
xlim([0 3]);
ylim([-0.5 1.5]);
title(['Interaction ' num2str(foilim(1)) '-' num2str(foilim(2)) 'Hz band'])
set(gca,'XTickLabel',{'','','baseline','','moderate','',''});
legend('Responsive','Drowsy','Location','northwest');

5. Compute a correlation between an external variable and the power spectrum

Here we will test for correlations between other external quantitative variable and the EEG data.

A common perspective on statistical testing starts from the distinction between dependent and independent variables. The EEG signals are typically considered to be the dependent variable. The independent variable can be the experimental conditions, as defined by task instructions, stimulus type, learning history, etc. The label independent variable suggests that it must be under the experimenter’s control. However, this is not necessarily the case, for example in case of reaction times, response accuracy or the drug concentrations in the blood.

Because neither the behavioral accuracy, nor the neural activity are under precise experimental control, it is arbitrary how the roles of dependent and independent variable are assigned. In FieldTrip, we use the convention that the variable with the smallest dimensionality is assigned the role of independent variable. For our example, this implies that accuracy is assigned the role of independent and the neurobiological signal the role of dependent variable. In fact, accuracy is represented by a single number, whereas the neurobiological signal often has a spatial (the channels) and a spectral (the frequencies) dimension. More information can be found in this frequently asked question.

The Permutation Distribution Results From Breaking the Association Between Dependent and Independent Variable

The null hypothesis that is tested by a permutation test involves that the probability distribution of the dependent variable is identical for all possible values of the independent variable. Quantitative independent variables, such us drug concentration, may have an infinite number of values, and in this case it is more difficult to conceptualize the probability distributions within each of these values. However, there is a null hypothesis for quantitative independent variables: statistical independence between the dependent and the independent variable. Testing for this statistical independence is possible in the same way for categorical as for quantitative independent variables: breaking the association between dependent and independent variable by randomly permuting the values of the independent variable. In a between-UO design, these values are permuted across the UOs, and in a within-UO design, they are permuted across the conditions in which the UO has been observed.

It is important to point out that the hypothesis of statistical independence rules out all possible relations between the dependent and a quantitative independent variable, and not only the linear relation. This contrasts with the three test statistics for quantitative independent variables that are implemented in FieldTrip: these are only sensitive to deviations from statistical independence that can be captured as a linear relation between the dependent and the independent variable. Of course, new test statistics can be formulated (and implemented as statfuns) with a sensitivity profile that is optimized for particular non-linear deviations from statistical independence.

Statistical Testing of the Relation Between a Neurobiological and a Behavioral Variable

Now, it is time to prepare the data as follows:

  1. Compute the within-participant contrast for each group: baseline vs moderate
  2. Compute the between-participant contrast of the differences computed in step 1
    cfg = [];
    cfg.channel          = 'all';
    cfg.frequency        = [8 20]; % let us test alpha and low beta bands
    cfg.avgoverfreq      = 'no';
    cfg.avgoverchan      = 'no';
    cfg.parameter        = 'powspctrm_b';
    cfg.method           = 'ft_statistics_montecarlo';
    cfg.statistic        = 'ft_statfun_correlationT';
    cfg.type             = 'spearman'; % type of the correlation (see help corr to know other types)
    cfg.correctm         = 'cluster';
    cfg.clusteralpha     = 0.05;
    cfg.clusterstatistic = 'maxsize';
    cfg.clusterthreshold = 'nonparametric_common';
    cfg.minnbchan        = 2;
    cfg.tail             = 0;
    cfg.clustertail      = 0;
    cfg.alpha            = 0.05;
    cfg.correcttail      = 'alpha';
    cfg.computeprob      = 'yes';
    cfg.numrandomization = 1000;
    cfg.neighbours       = cfg_neigh.neighbours;

    subj = size(mode_sedation.powspctrm,1);
    design = zeros(1,subj);

    % The three columns of the covariates are
    %  - drug concentration
    %  - reaction time
    %  - correct responses

    design(1,:)  = covariates(:,3,1)./1000; % select the column with concentrations
    cfg.design   = design;
    cfg.ivar     = 1;

    stat5 = ft_freqstatistics(cfg, mode_sedation);

    cfg            = [];
    cfg.alpha      = stat5.cfg.alpha;
    cfg.parameter  = 'stat';
    cfg.zlim       = [-3 3];
    cfg.elec       = elec;
    ft_clusterplot(cfg, stat5);