Modeling condition ERP differences using std_envtopo()
Important note: this functionality has been disabled in EEGLAB v9 as it was not compatible with the new design structure. It has been ractivated in EEGLAB 11 though.
A STUDY and component cluster-based version of the ERP plotting function envtopo.m is std_envtopo.m. The std_envtopo.m function allows you to determine the contribution of specific clusters to the grand ERP of the STUDY - the grand mean ERP for all the datasets from each STUDY condition, or to determine which clusters contribute the most to the grand ERP. To find, in the clustered dataset above (before editing), the four clusters that contribute the most to the grand mean ERPs for each condition, use the following command:
>> std_envtopo(STUDY, ALLEEG, 'clusters', ,'subclus',[3 4 6 7 8 ], ... 'env_erp', 'all', 'vert', -1100, 'baseline', [-200 0], ... 'diff', [1 2], 'limits', [-1300 1500 -5 5], 'only_precomp', 'on', ... 'clustnums' , -4,'limcontrib', [300 600]);
The example above searches among all the clusters ('clusters', ) for the four component clusters that contribute most to the grand ERP ('clustnums', -4) in the time window [300 600] ms. In constructing the grand ERPs, five designated artifact component clusters are excluded ('subclus', [3 4 6 7 8 ]). The computation uses all the datasets in the STUDY ('env_erp', 'all') but only the components that were actually clustered ('only_precomp', 'on'). In this example, components with residual dipole model variance above 15% (0.15) were not clustered and thus were also not included in the grand mean ERPs. We also ask for the largest contributing clusters to the difference wave between the two conditions ('diff', [1 2]). For more information type: >> help std_envtopo
Below are the three resulting figures, one for each condition and a third for the ERP difference between the two conditions. The N400 can be easily distinguished in the grand difference ERP on the right.
In the three cases above, the four largest contributing clusters together account for 40%, 77%, and 67% of the respective grand mean ERP variances (pvaf). Three clusters (Cls 5, Cls 9, Cls 12) are among the four largest contributors to both conditions (left and middle images). These three, plus Cluster 10, account for roughly equal parts of the N400 difference (right image), implying that the N400 effect is not spatially focal or unitary, in line with results from invasive recordings.
Note that the percent variance accounted for (pvaf) values for the four clusters add to over 100%, though together they account for only 67% of difference-ERP variance. This is because the scalp projections of the four clusters are spatially correlated, not orthogonal. Therefore, their projections can cancel each other. Here, for example, the positive right frontocentral channel projections of components in Cluster 10 partially cancel the spatially overlapping negative projections of Clusters 9, 12, and 5 at the same scalp locations. In this sense, ICA can re-capture more of the source data than is captured directly in the recorded scalp channel signals.
Some other clustering procedure and a brief discussion and example of independent component clustering, from a chapter in press, is available here.