ERP Averager

What it does

About 200 single trials from one subject, two conditions, four channels, and a slider for N. In averager mode the widget averages the first N trials of the chosen condition and draws the result over the single trials, with the signal, the residual noise and the resulting signal-to-noise ratio computed from the data and printed underneath, next to the √N reference curve. In sme mode the same trials drive a bootstrap estimate of the standardized measurement error — the standard error of the measurement you would actually report — as a function of trial count, with an analytic reference for the linear measures and a read-out of the trials needed to reach a stated target.

Everything is computed in the browser from the stored trials, so a shared state reproduces the same numbers, and the shuffle is seeded.

Controls

Control	What it sets
N	How many trials are averaged
Shuffle	Advances the seed of the trial order, so the same N draws a different subset
Condition	Which condition is averaged; the rarer one has fewer trials
Channel	Which channel the waveform panel and the measurements follow
Measurement window	The interval the amplitude and latency measures are taken over
Baseline / noise interval	The pre-stimulus interval used as the noise estimate
Polarity	Which extremum the peak measures take
Target SNR	(averager) the signal-to-noise ratio the “trials needed” read-out solves for
Measure	(sme) mean amplitude, peak amplitude or peak latency
Target SME	(sme) the measurement-error threshold the “minimum trials” read-out solves for
Bootstrap replicates	(sme) how many resamples the bootstrap SME uses
Single trials / SEM band / all-trials average	Which reference layers are drawn

What to look for

averager

• At N = 1 the trial is mostly noise and the component is not visible. Raise N and the same deflection appears in the same place every time: averaging did not create it, it removed what was not time-locked to the event.
• The two numbers under the chart move differently. The signal settles early; the noise keeps falling. The signal-to-noise ratio rises because the noise falls, not because the response grows.
• Compare the measured curve with the dashed √N line. Four times the trials buys twice the ratio — so the cost of the next doubling is four times the recording time, and the “trials needed” read-out is that law solved for N.
• Press Shuffle at a fixed N: the path toward the average changes, the average does not. The average is a property of the trial set, not of the order.
• Switch condition. The rare condition has fewer trials, so its average is noisier although nothing about its component changed — an unequal trial count is a difference in measurement precision, not in brain response.

sme

• The SME curve falls as 1/√N: steep on the left, nearly flat on the right. Past a point, more trials buy very little.
• For mean amplitude the bootstrap curve sits on the analytic reference (the standard deviation of the single-trial measurements divided by √N). Switch to peak amplitude and the two separate, because the measurement of the average is not the average of the measurements.
• The “minimum trials” read-out answers the design question directly. Widen the measurement window and the answer falls — then ask what widening the window cost.
• Peak latency has a much larger SME than the amplitudes at small N, and it is the measure whose curve flattens last: a latency needs a peak that is already stable before it can be located.
• The SME is per subject and per condition. Two conditions with different trial counts have different SMEs in the same recording, which is why trial counts belong in a methods section.

Used in

• L3.2 The ERP and its components (averager)
• L3.5 SNR, trial counts and design (sme)

Data provenance

ds-erpcore — ERP CORE (Kappenman et al., 2021), open access, per-subject downloadable. Biosemi ActiveTwo, 30 EEG + 3 EOG electrodes in a 10-20 placement scheme, 1024 Hz, CMS online reference, 60 Hz mains, no software filters, 40 participants per paradigm.

The shipped asset is sub-001’s P3 run: 200 single trials (40 target, 160 standard) at four channels (Pz, Cz, PO7, PO8), stored over −0.199 to 0.801 s and baseline-corrected to (−0.2, 0) s, with a per-trial, per-channel noise estimate — the standard deviation of that trial’s baseline-corrected pre-stimulus window in µV — which is what makes the SNR and SME read-outs computations rather than illustrations. Derived: FIR band-pass 0.1–40 Hz at the native rate, epoched, re-referenced offline to the average of P9 and P10 (the montage has no mastoid sites; P9/P10 are the nearest stand-in), resampled 1024 → 256 Hz. Generated by data/scripts/extract_p2_widgets.py; trials.json is both the widget index and the §4.5 sidecar of trials.bin, and the asset is registered in data/manifest.json. The frame’s provenance line is the authority.

The 40-to-160 trial split is the paradigm’s own: targets are rare by construction, which is exactly the imbalance L3.2 and L3.5 ask you to reason about.

TODO(confirm): the asset sidecar records a licence conflict in the source — the OSF node record for thsqg says CC BY 4.0, the per-paradigm component’s own LICENSE file says CC BY-SA 4.0, and its dataset_description.json says CC0. The site states CC BY 4.0 on the strength of the node record; the author reconciles the three (§10.11 item 8, §13 item 22) and mirrors the entry into data/registry.yaml. TODO(confirm): the dataset DOI and paper DOI are TODO(confirm) in data/directory.yaml. TODO(confirm): the reference used by the dataset’s own published processing pipeline is not recorded, so the P9/P10 stand-in is this site’s choice, not the dataset’s.

Open the code

site/src/components/widgets/w-erp-averager/ — Widget.svelte, index.ts, meta.ts, README.md. Repository link: TODO(confirm) (GitHub org/repo, §13 item 3).