What Is an ERP?

What You Are Looking At

A participant is watching a stream of letters. Most are X, the standard; about one in five is O, the rare oddball, and the task is to count the O's. The top trace is one EEG channel at Pz, 20 to 60 µV of ongoing rhythms with a marker at each stimulus. Every stimulus also evokes a response, a few microvolts of P1, N1, and P2 for the standards and a large P3 near 380 ms for the oddballs, but you cannot see it in the raw trace because it is buried in noise five to ten times its size. That response is the event-related potential, the part of the EEG that is time-locked to an event.

Panel B cuts an epoch from −200 to 800 ms around each stimulus and stacks them, and Panel C averages them. The noise is random with respect to the stimulus, positive on one trial and negative on the next at any given latency, so it cancels as more epochs are added. The response is the same on every trial, so it survives. The gold curve is the true waveform that was injected into every trial, and the teal and coral averages settle onto it. The difference wave in violet, oddball minus standard, removes everything the two conditions share and leaves only what changed, a small N2 and the large P3. Panel D plots the noise left in the oddball average against the number of trials; it falls along the 1/√N curve, so doubling the trials improves the signal-to-noise ratio by 41% and quadrupling it doubles the ratio (Luck, 2014, Chapter 1). Every epoch is baseline-corrected first by subtracting its mean prestimulus voltage, so the traces all start at zero.

Try This

1. Press Reset and watch the first ten trials arrive at 1×. Each average is still a wobbly line that looks nothing like the gold curve.
2. Click Add 100 trials. The standard average, with four times as many trials as the oddball, is already smooth; the oddball average still carries visible noise. Add 100 more and compare the residual noise readouts.
3. Push the background noise slider to its maximum (about ±40 µV) and reset. Read off from Panel D how many oddball trials you now need to reach the same noise in the average as before. Twice the noise needs four times the trials.
4. Click Show true ERP to hide the gold curve, and decide, at 20 oddball trials, whether you would trust the P3 you see. Click it again to check.
5. Set the speed to 16× and let it run to the 1000-trial cap. The last few hundred trials barely change the picture; the square-root law gives diminishing returns.

Why It Matters for the Pipeline

Most of the later steps exist to make this average trustworthy. Filtering removes noise that averaging handles badly, artifact rejection throws out trials whose noise is not random, and bins define which events are averaged together. In ERPLAB the average is computed by ERP = pop_averager(EEG, 'Criterion', 'good') after epochs have been cut with pop_epochbin(EEG, [-200 800], 'pre'), and a difference wave is made with pop_binoperator (Lopez-Calderon & Luck, 2014). Components are named by polarity and typical latency (P1, N1, P3, N400) and measured in a window such as 300 to 500 ms. ERPs are very good at timing, at telling which stage of processing changed, and at measuring a process without asking for a response. They are poor at telling where in the brain the activity came from, they are poor at capturing very slow (more than a couple of seconds) or non-time-locked activity, and they need enough trials to average. The ERP CORE resource (Kappenman et al., 2021) supplies standardized paradigms for the P3, N400, and five other components. The task here is a simplified oddball; the ERP CORE P3 task uses five letters with one designated target per block, a 20% target probability, and a button-press response, and the probabilities and timing on this page follow it.

References