Site, amplifier or cap confounded with group or time

Symptom

A group difference (patients versus controls, old versus young, one country versus another) that coincides with a difference in how the groups were recorded: at different sites, on different amplifiers, with different caps, at different sampling rates, under different mains frequencies, or in different decades. The “effect” is largest in exactly the measures that hardware changes — spectral slope, high-frequency power, absolute amplitude, line-noise residue — and it survives every within-group check because within each group the hardware is constant. A replication on a single system finds nothing, or something smaller.

Cause

Every acquisition choice leaves a fingerprint in the data (L0.3, L1.3): the hardware bandwidth and filters shape the spectrum, the amplifier’s noise floor sets the high-frequency baseline, the cap’s electrode set and reference change amplitudes and topographies, the sampling rate decides which harmonics are visible and how resampling behaves, and the mains frequency decides where the line and its notch sit. When one group is recorded predominantly on one setup and the other on another, the fingerprint difference is perfectly correlated with the group label, and no amount of statistics on the outcome can separate them. Time is a special case: a clinical archive that changed cap systems over twenty years has a confound between recording era and everything else that changed with era (referral patterns, diagnoses, medication).

Detect

• Tabulate hardware per subject (site, amplifier, cap, rate, filters, reference, mains) from the dataset description and the file headers, and cross-tabulate it against the group and against the recording date; any cell imbalance is a confound.
• Compute the hardware-sensitive measures (spectral slope over the full band, power above the physiological range, line-noise residue, noise floor) and test whether they separate the setups before testing whether they separate the groups.
• Fit the group effect with setup as a covariate or stratum, and see how much of it survives; if the design has no within-setup contrast at all, the effect is not estimable.
• Read the catalog’s caveats before analysing a multi-site dataset (L7.6).

Fix

• Design first: balance groups across sites and devices, or record every group on every setup; record era-matched controls in archives.
• Harmonize what can be harmonized (resample to a common rate after proper anti-alias filtering, restrict to the common hardware passband, common reference, common channel subset) and state what cannot (noise floors, cap geometry).
• Model setup explicitly (as a fixed effect, a stratum, or a random effect with enough levels) and report the group effect within setup; when setup and group are fully confounded, say so and do not claim a group effect.
• Report the hardware table with the results; a reader cannot judge a multi-site effect without it.

Example

The site’s own directory documents each variant of this confound (facts from §9.2 and §10.9 of the spec, copied from the catalog):

Dataset	Confound documented in the catalog	Where it bites
ds-aszed	Two amplifiers at 200 and 256 Hz with device-default filters; the channel set must be verified from EDF headers	schizophrenia versus control comparisons; resampling and filter harmonization (L1.1, L2.8)
ds-tdbrain	Two cap systems over about twenty years; 47 healthy controls among 1,274 patients; comorbidity common	any patient-versus-control or era comparison (L7.6)
ds-brainlat	Mains frequency differs by country within one dataset (50 vs 60 Hz)	line-noise handling and any measure near 50–60 Hz across countries (L1.6)
ds-hbn	Each release is a separate accession; the sample is transdiagnostic despite the name	pooling across releases; case–control framing (L7.6, L7.8)

No figure is generated for this entry: the confound lives in the recording metadata, not in a trace. The Level 7 lesson on resting-state biomarkers works through the ds-aszed case with the hardware table above.