9.10: Exercise- Was Blink Correction Effective?

Applied Event-Related Potential Data Analysis 2e · LibreTexts

0%

Contents

    Text

    Size
    100%
    Theme
    Typeface
    Line length
    Spacing

    Read aloud

    How

    Speed: 1×
    Light up
    Highlight

    Symbols: Mulberry Symbols, © 2018–2026 Steve Lee, CC BY-SA 4.0 (licence, the set), the site's own copies.

    From LibreTexts, under CC BY-NC 4.0, as the source states it; its equations are written in TeX, which the PDF shows typeset. This page at LibreTexts.

    9.10: Exercise- Was Blink Correction Effective?

    Now that we’ve seen that blinks are a confound in the ERP CORE N400 experiment, we’ll ask whether our ICA-based blink correction procedure was effective at minimizing the voltages produced by the blinks. It will be easiest to determine this by looking at grand average data.

    Two scripts were run to create grand averages that will help us determine if blink correction was successful. The first was named N400_correction_rejection_phase4_rejection.m. For each of our 20 participants, this script took the ICA-corrected data (X_N400_preprocessed_pruned), added an EventList, ran BINLISTER using a bin descriptor file named BDF_N400_Targets.txt to sort the events into bins (bin 1 = related targets; bin 2 = unrelated targets), and epoched the data. Then, artifact detection was applied to mark epochs with C.R.A.P. (which will be described more in the next section). An averaged ERPset was then created from this EEGset for each participant.

    This script also created two new channels, named H-bipolar-corr and V-bipolar-corr. These channels allow us to better see how well the artifact correction worked by creating bipolar HEOG and VEOG signals from the corrected data. When we ran ICA, we corrected the channels we ordinarily use to detect blinks and eye movements (VEOG-lower, FP2, HEOG-left, and HEOG-right). We can therefore reconstruct bipolar channels from these corrected channels to better see whether the artifacts were fully corrected. The H-bipolar-corr channel was created by taking the corrected HEOG-left and HEOG-right channels and making a bipolar channel by computing HEOG-left minus HEOG-right. The V-bipolar-corr channel was created by taking the corrected VEOG-lower and FP2 channels and making a bipolar channel by computing VEOG-lower minus FP2. If blink correction worked well, then VEOG-lower and FP2 should have no blink activity left in them, and the difference between them should be close to zero if we assume that these two sites have equivalent brain activity. In other words, if VEOG-lower and FP2 have the same brain activity and differ only in blink activity, then the difference between VEOG-lower and FP2 should be close to zero if ICA has removed the blink activity.

    Looking at the Corrected Bipolar VEOG in One Participant

    To see this for Subject 1, click Load ERP in the ERPsets panel, navigate to N400_Data > 1, and select 1_N400_correction_plus_rejection.erp. The uncorrected and corrected EOG channels should look like Screenshot 9.25.

    Corrected and Uncorrected EOG for Subject 1

    Screenshot 9.25

    You can see that the corrected bipolar VEOG channel (V-bipolar-corr) is now nearly flat, indicating that the blink voltage has been largely eliminated. You can also see that the large difference between related and unrelated targets in the uncorrected bipolar VEOG channel (VEOG-bipolar) is no longer evident in the corrected bipolar VEOG channel (V-bipolar-corr). That looks like successful blink correction, at least for this one participant.

    Looking at the Corrected Bipolar VEOG in the Grand Average

    To more carefully evaluate the effectiveness of the blink correction, I created and ran an additional script named N400_correction_rejection_phase5.m. This script includes some important steps that will be described in a later section, but it did three things that were important for the present exercise. First, it made a grand average of all the single-participant ERPsets created by the previous script, which is named Grand_N400_correction_plus_rejection.erp. Go ahead and load this grand average using the ERPsets panel. Note that N400_correction_rejection_phase5.m uses the ttest() function, which requires the Statistics and Machine Learning Toolbox (see the previous section for more details). The script has instructions for commenting out the ttest() lines if you don't have this toolbox.

    The uncorrected VEOG-bipolar signal is so large that it’s hard to see the other channels. To reduce this problem, go to the Bin & Channel Selection panel, select channels 1, 2, 5, 9, 13, 16, 31, 33, and 35, and click Apply. This will display the key electrode sites for examining blinks and the N400, as shown in Screenshot 9.26.

    Grand Average Correction+Rejection

    Screenshot 9.26

    As was true for Subject 1, the corrected bipolar VEOG channel (V-bipolar-corr) is now nearly flat, indicating that the blink voltage has been greatly reduced. However, if you look closely, you can still see a difference of a couple microvolts between the related and unrelated targets from approximately 200-450 ms in this channel. That’s very concerning! Note, however, that the difference between the related and unrelated targets is most pronounced from approximately 450-800 ms in the uncorrected bipolar VEOG channel (VEOG-bipolar), but there is no difference between conditions in this time range in the corrected bipolar VEOG channel. So, the correction seems to be working extremely well in the late part of the epoch, but it does not seem to be working as well during the N400 measurement window (which is when we really need it to work).

    Another important thing to note is that the difference between the related and unrelated targets is much larger at the measurement site for the N400 (CPz) than in V-bipolar-corr. We would expect any residual blink activity to get smaller and smaller as we get farther away from the eyes, but the N400 effect is getting bigger. This means that at least most of the N400 effect at CPz must be brain activity rather than blink activity.

    A More Detailed Analysis of the Corrected Bipolar VEOG Signal

    The script also provided more detailed quantitative information about the difference between related and unrelated targets. Specifically, it created an unrelated-minus-related difference wave and measured the mean amplitude from this difference wave during the N400 measurement window (300-500 ms) at the uncorrected VEOG-bipolar channel, at the corrected V-bipolar-corr channel, and at the CPz channel. Then it did a simple statistical analysis to determine whether these amplitudes were significantly different from zero (i.e., whether these amplitudes differed significantly between the related and unrelated targets). Here’s what the script printed in the Command Window:

    Summary of mean amplitude in the N400 latency range:
    Comparison of unrelated-minus-related difference in the uncorrected VEOG-bipolar channel
        mean = 12.89, t(19) = 4.07, p = 0.0007
    Comparison of unrelated-minus-related difference in the corrected VEOG-bipolar channel
        mean = 1.83, t(19) = 1.46, p = 0.1602
    Comparison of unrelated-minus-related difference in the corrected VEOG-bipolar channel
        mean = -7.82, t(19) = -9.74, p = 0.0000

    Without correction, there was a 12.89 µV difference in VEOG-bipolar amplitude between related and unrelated targets in the N400 latency range, which was statistically significant. With correction, this difference was reduced to 1.83 µV and was not statistically significant. At CPz, the difference was 7.82 µV and significant.

    The fact that the effect was small and not statistically significant in the corrected VEOG-bipolar channel provides some reassurance that the blink correction was reasonably effective. However, you can’t “prove the null” with a p value, so these results don’t provide strong evidence that the blink correction completely eliminated the confound. In fact, with all 40 participants, the difference between related and unrelated targets was significant for the corrected VEOG-bipolar channel (Zhang et al., 2024). Thus, it seems as if ICA did not completely remove the blink confound.

    A Potential Alternative Explanation for the Effect in the Corrected Bipolar VEOG Signal

    Given the fact that the difference between related and unrelated targets was large at CPz and became progressively smaller at electrodes near the eyes (after correction), we can conclude with confidence that the N400 effect at CPz was mainly a result of brain activity, not blink activity. However, it still seems possible that some residual blink activity was volume-conducted from the eyes to the EEG electrodes and inflated the observed difference between related and unrelated targets at the EEG sites. 

    However, we should also consider another possibility: Just as blink activity can propagate from the eyes to the EEG electrodes, brain activity can propagate from the EEG electrodes to the VEOG electrodes. In other words, the larger brain response for unrelated targets than for related targets that we see at CPz might be volume-conducted to the FP2 and VEOG-lower electrodes, creating the difference we have been attributing to differences in blinking.

    How can we tell whether some of the effect at CPz is volume-conducted blink activity or whether the entirety of the effect at V-bipolar-corr is volume-conducted brain activity? In our journal article (Zhang et al., 2024), we developed a statistical approach to answering this question using semipartial correlations. That approach is too complicated to demonstrate here, especially because it involves statistical analyses that are not built into eStudio. However, I will give you the bottom line: The statistical analysis provided evidence that the difference between related and unrelated targets at V-bipolar-corr did not consist of volume-conducted N400 activity, whereas a portion of the N400 effect at CPz could be explained by volume-conducted blink activity. 

    In other words, ICA-based blink correction was not sufficient to completely eliminate blink confounds in the ERP CORE N400 paradigm. That’s a sobering conclusion. I’d like to think that we did a pretty good job of implementing ICA-based blink correction, and yet a small but statistically significant blink confound remained. However, I wasn’t really surprised at the failure of ICA to completely eliminate blink confounds. As I’ve noted multiple times in this chapter, ICA is based on assumptions that are known not to be true (most notably that the number of sources is equal to the number of electrodes). As a result, we shouldn’t expect ICA to be perfect. And it is therefore important to assess whether the imperfections in ICA might lead to effects in our data that we might think reflect brain activity but instead reflect blink activity.

    Fortunately, you typically need to worry about this only when the procedures described in the previous section indicate that blinks differ between conditions, as in the ERP CORE N400 paradigm.

    Was the Blink Correction “Good Enough”

    The French Enlightenment author and philosopher Voltaire once wrote something that is often translated (slightly incorrectly) into English as “The perfect is the enemy of the good.” With that in mind, perhaps we shouldn’t require that ICA-based blink correction be perfect and instead be satisfied if it is good enough. But how should we define “good enough”?

    In our journal article (Zhang et al., 2024), we proposed using propagation factors to help us decide what is good enough. Propagation factors tell us how much of the voltage generated by ocular artifacts would be expected to fade by the time they reach a given EEG electrode. In one of my favorite ERP methodology papers, Lins et al. (1993) provided normative propagation factors for a variety of electrode sites. This study indicated that a given blink voltage recorded from a VEOG-bipolar configuration would be expected to produce a propagated voltage of approximately 19% at Fz, approximately 10% at Cz, and approximately 5% at Pz. The 1.83 μV effect we observed in the bipolar VEOG channel after correction would therefore be expected to produce a propagated effect of approximately 0.35 μV at Fz, 0.18 μV at Cz, and 0.09 μV at Pz. Lins did not provide a propagation factor for CPz, but it must be somewhere between the factors for Cz and Pz. Thus, the propagated blink confound at CPz should be somewhere between 0.18 μV and 0.09 μV, which would be less than 2.5% of the 7.82 µV effect observed at CPz. Thus, although ICA did not completely eliminate the confounding eyeblink activity at CPz, the residual activity was quite small. I think this is “good enough” for this study and might even be described as “excellent” given how large blinks are relative to ERPs.

    When we looked at the other ERP CORE experiments, we found that ICA-based blink correction was close to perfect in some paradigms and very good but not perfect in others. That’s good news. It likely means that our ICA approach will be satisfactory in studies using similar paradigms, high-quality EEG recording systems, a cool and dry recording environment, and highly cooperative participants. If your research is dissimilar in one or more ways, or if you don't follow our ICA procedure exactly, you should not assume that our approach will yield satisfactory results in your data. Instead, you should use the procedures described here to provide rigorous evidence about the effectiveness of your artifact correction pipeline.

    In ending this section, I’d like to bring up another quote about perfection, this one from the famous American football coach, Vince Lombardi. When he first became coach of the Green Bay Packers in 1959, he met with his players and explained his philosophy, which was to aim for perfection. He noted that they would never actually achieve perfection, because nothing in this world is perfect. But it is still valuable to strive for perfection. Here’s the actual quote: “Gentlemen, we will chase perfection, and we will chase it relentlessly, knowing all the while we can never attain it. But along the way, we shall catch excellence.” Over the next 9 years, they won 5 NFL championships, including the first two Super Bowls. I have taken Lombardi’s quote and made it my own personal philosophy for conducting scientific research, always chasing perfection in an effort to catch excellence. So, I encourage you to chase perfection in implementing ICA-based artifact correction: ICA will never be perfect, but when carefully applied, it can do an excellent job of minimizing several kinds of artifacts.


    This page titled 9.10: Exercise- Was Blink Correction Effective? was last modified on Thu, 01 Oct 2026 18:52:30 GMT and is shared under a CC BY-NC 4.0 license and was authored, remixed, and/or curated by Steven J Luck.