IC 1 is clearly capturing most of the blink activity, and we established in the previous chapter that blinks are both a major source of noise and a confound in the MMN experiment. It’s therefore clear that we should exclude this IC when we reconstruct the EEG data from the ICs. But what about the other ICs?
To answer this question, we need to go back to the three issues described at the beginning of the previous chapter. First, does the artifactual activity create substantial noise that degrades our data quality? Second, does the artifactual activity vary across groups or conditions, creating a confound? Third, does the artifactual activity indicate a problematic change in the sensory input (e.g., because the direction of gaze has changed)? The third issue is mainly relevant for visual experiments, so we won’t worry about it for the MMN.
In the chapter on artifact rejection, we addressed the issue of noise by looking at whether the SME was improved by rejection. We addressed the issue of noise by looking at the averaged ERPs, especially when we inverted the artifact rejection process and included only the trials that were marked for rejection. We can use the same two approaches with artifact correction.
We already assessed the effects of blink rejection and correction on the SME for our first pass at the ICA decomposition earlier in this chapter, and we found that correction improved the SME. Now let’s assess the effects of correcting for horizontal eye movements by excluding IC 13. Specifically, we'll compare excluding only blinks with excluding both blinks and horizontal eye movements. First, make two copies of the dataset with our improved ICA decomposition (10_MMN_preprocessed_transferredICAweights). Then apply EEGLAB > Tools > Remove components from data as in the earlier exercise to remove the blink IC from one dataset and to remove both the blink and horizontal eye movement ICs from the other dataset. You should then create averaged ERPs for these two datasets (which will require adding an EventList, running BINLISTER, and epoching the data). Make sure to specify a custom data quality window of 125-225 ms. Before averaging, use EEGLAB > ERPLAB > Preprocess EEG > Selective Electrode Interpolation to interpolate F7 and PO4 (Channels 3 and 24), excluding the bipolar EOG channels (32 and 33).
If you look at the SME values for FCz in the 125-225 ms time range, you’ll see that eliminating the horizontal eye movement IC had virtually no impact on the data quality. This is consistent with what we saw with artifact rejection in the previous chapter. Horizontal eye movements just don’t have much impact on midline electrode sites, and most of the eye movements occurred during the break periods.
We also need to consider whether the horizontal eye movements were a confound. With artifact rejection, we did this by making averages from only the trials marked with artifacts. Here, we’ll conduct an analogous procedure, in which we reconstruct the data only from the eye movement component (IC 13) and then make averaged ERPs. To do this, select the dataset with the transferred weights (10_MMN_preprocessed_transferredICAweights), select EEGLAB > Tools > Remove components from data, leave the List of component(s) to remove from data field blank, and enter 13 into the field labeled Or list of components to retain. Name the resulting dataset 10_MMN_preprocessed_IC13only. If you scroll through this dataset, you’ll see how the eye movement potentials spread to the other scalp sites (see, e.g., the highlighted time period in Screenshot 9.8.). These potentials are largest at the lateral frontal electrodes but can also be seen at posterior scalp sites (e.g., P7 and P8). However, the positive and negative sides of the dipole largely cancel out at the midline sites (e.g., Fz, FCz, Cz).
Screenshot 9.8
Now go through the steps needed to create averaged ERPs from this dataset, and then plot the resulting ERPs. You should see that the EEG channels are essentially flat lines. (There is some activity in the bipolar channels, but these are the original waveforms, not reconstructed from IC 13.) From these flat ERP waveforms, which should solely reflect the horizontal eye movements, we can conclude that the horizontal eye movements were not a meaningful confound. In other words, the fact that the averaged ERPs were largely flat in the EEG channels indicates that any horizontal eye movements were approximately equally likely to be leftward and rightward, resulting in opposite-polarity EOG signals that canceled out in the averages. And the fact that there was no difference between standards and deviants, especially at the key FCz site, indicates that there were no meaningful differences in horizontal eye movements between standards and deviants that could confound our MMN results.
So, should we correct for the horizontal eye movements by excluding IC 13 when we reconstruct the EEG data? Probably not. There is little to be gained, and given that ICA is imperfect and IC 13 may contain brain activity mixed with the horizontal eye movements, we have more to lose than to gain. However, IC 13 didn’t produce much activity in the ERPs when we looked only at this IC, so removing it will also have little impact. Before making a final decision about this, I would want to see how horizontal eye movements impact the data in the other participants. If they’re generally not a problem, I would not correct for them. But if they seem problematic (in terms of data quality and/or confounding activity) in several of the participants, I would probably correct for them in all participants (for the sake of consistency).
We also need to consider whether to exclude IC 16, which has a blink-like scalp distribution and shows small deflections for each blink. Repeat the series of steps you conducted for IC 13 but use IC 16 instead. That is, assess the SME values after excluding both IC 16 and IC 1 with the SME values after excluding only IC 1, and then look at the averaged ERPs after reconstructing the data from only IC 16. When I did this, I found that excluding both IC 16 and IC 1 led to a tiny improvement in SME compared to excluding only IC 1, and I found a small but concerning difference between deviants and standards in FCz when the data were reconstructed from IC 16. Given the small confound in the averaged waveforms, and the fact that we’ve already established that blinks differ between deviants and standards, I would exclude IC 16 along with IC 1 in the final analyses.
You can also try this set of procedures with the component that represents the spike potential (IC 3).
Does this sound like a lot of work? It is! But if your goal is to publish a paper in a scientific journal, making a permanent contribution to the scientific literature that others can build on, it’s worthwhile to take the time to deal with artifacts in a careful and thoughtful manner. Also, the scripts provided at the end of the chapter show you how to make the process more efficient by automating some parts of it.
