4.11: Practical Advice
-
- Last updated
- Save as PDF
This chapter has focused on helping you understand how filters actually work and how they can distort your data. I’d like to end with some practical advice about when and how to filter your data. This advice is appropriate for perhaps 95% of perceptual, cognitive, and affective ERP research. If you think your research falls into the other 5%, then you need to be very sure you fully understand how filtering works so that you don’t end up either causing significant distortion of your data by overfiltering or failing to remove enough noise by underfiltering.
Let’s start with the question of when to filter. As mentioned earlier, you should apply your high-pass filter to the continuous EEG to avoid edge artifacts. To further minimize these artifacts, you should use the option labeled "Remove mean value (DC offset) before filtering" when you are filtering continuous EEG data (but not when you are filtering epoched EEG or averaged ERPs).
You can apply your low-pass filter to the continuous EEG, the epoched EEG, or the averaged ERPs. For most researchers, it will be simplest just to apply both the low- and high-pass filters to the continuous EEG. In my lab, we typically apply the low-pass filter only to the averaged ERPs, but that’s mainly for philosophical reasons rather than practical reasons (as described in Luck, 2014).
Now let’s discuss filter parameters. My lab typically uses cutoffs of 0.1 and 30 Hz, with a roll-off of 12 dB/octave. That’s what I’d recommend if you are recording very clean data, especially if you have highly cooperative participants (e.g., neurotypical young adults). If you have a fair amount of high-frequency noise (e.g., 60-Hz line noise or spiky muscle activity), you can increase the roll-off to 48 dB/octave for the low-pass filter and/or drop the cutoff from 30 to 20 Hz. You’ll get a little more distortion, but not enough to matter for most studies. However, I don’t recommend a roll-off of 48 dB/octave for the high-pass filter. If you have a lot of low-frequency noise, which is especially common when the participants move around a lot (e.g., infants or young children), you can raise the high-pass cutoff to 0.2 or even 0.5 Hz and/or increase the roll-off to 24 dB/octave. However, these parameters can cause noticeable distortion of the waveforms. If you mainly have line noise and don’t want to use a low-pass filter at 20-30 Hz, you can use EEGLAB’s cleanline plugin (see Bigdely-Shamlo et al., 2015 for important details). If cleanline doesn't work well for you, you can try the newer Zapline method (de Cheveigné, 2020; Klug & Kloosterman, 2022).
If you want to use a high-pass cutoff of greater than 0.1 Hz or a low-pass cutoff of less than 20 Hz, I strongly recommend that you create artificial ERP waveforms that resemble your data and pass them through the filter (even if you’re filtering the continuous or epoched EEG with your real data). You can then see exactly how the filter distorts your data. If the distortion you see with the artificial data is small compared to the effects you’re seeing in your study, then you don't need to worry. The next section describes how to create and import artificial waveforms.
When my lab measures the onset latency of an ERP component, we will often apply a 10 Hz low-pass filter (12 dB/octave). We find that onset latency measures are highly sensitive to noise, and we get much better statistical power by filtering at 10 Hz. This might seem problematic given that Figure 4.4 shows that a 10 Hz cutoff produces a substantial distortion of the onset latency. However, when we quantify the onset latency of an ERP component, we don’t measure the time that the waveform first deviates from zero. Instead, we measure the time at which the voltage reaches 50% of the peak voltage (see Chapter 9 in Luck, 2014, for a justification of this approach). If you look closely at Figure 4.4, you’ll see that this 50% point is virtually identical for the filtered and unfiltered waveforms. Also, we’re usually comparing the ERPs from two conditions that have both been filtered at 10 Hz, so the effect of the filtering should be equivalent for both conditions. However, we sometimes want to examine the point at which a difference wave first deviates from zero. In these cases, we do minimal filtering or we quantify the amount of latency shift produced by the filter (see, e.g., Bae & Luck, 2018).