Referencing

Every voltage measurement is a difference between two points. When you read “50 µV at Cz,” what you’re actually reading is “50 µV relative to something.” That something is the reference—and which reference you choose changes the apparent amplitude and distribution of every frequency band, every asymmetry calculation, and every connectivity estimate in the entire analysis.

This is not a minor technical detail. The reference scheme is one of the most consequential choices in the pipeline, and one of the least understood by clinicians coming from commercial platforms where the choice was made for them (and often not disclosed).

Why Average Reference

The Coherence Workstation uses the average reference by default:

preprocessing:
  reference: average

Average referencing subtracts the mean signal across all channels from each individual channel at every time point. The result is that no single electrode serves as the reference—the “zero point” is the average electrical potential across the entire scalp.

Why is this preferable? Consider the alternative. A physical reference electrode (typically linked ears, Cz, or a mastoid) injects its own neural activity into every channel’s measurement. If the reference electrode happens to sit over a cortically active region, that activity appears—inverted—at every other electrode. Linked ears, the most common clinical reference, are particularly problematic: the mastoid electrodes pick up temporal lobe activity, which then contaminates the entire montage. Any asymmetry analysis comparing left and right temporal electrodes becomes unreliable, because both are being measured relative to a reference that itself contains lateralized temporal activity.

Average reference avoids this by distributing the reference across all channels. No single electrode’s activity dominates. The mathematical assumption is that the mean electrical field across a sufficiently dense electrode array approximates zero—the positive and negative poles of any cortical dipole roughly cancel when averaged. This assumption holds better with more electrodes (it’s quite good at 64+ channels and approximate at 19), but even at 19 channels, average reference is more principled than any single-electrode alternative.

The 19-channel caveat

Average reference is theoretically ideal with dense electrode arrays (128+ channels) where the spatial sampling approaches uniform coverage. With 19 channels, the average is a coarser approximation—some cortical regions are better sampled than others. Despite this, average reference remains the recommended standard for 19-channel clinical EEG. The alternative—linked ears or Cz reference—introduces known, systematic bias that’s harder to characterize or correct than the sampling limitation of a 19-channel average.

Why Not Linked Ears

Linked ears (A1+A2, or TP9+TP10) was the dominant reference in clinical EEG for decades. There are two reasons the Coherence Workstation doesn’t use it.

First, as noted above, linked ears injects temporal lobe activity into the reference. This creates a baseline contamination that systematically biases any metric derived from temporal electrodes—including asymmetry calculations, connectivity estimates, and coherence values. The bias isn’t random; it’s structured and frequency-specific, which means it looks like a finding rather than an artifact.

Second, linked ears creates an artificial “bridge” between the two hemispheres. Current flows through the linked ear connection, creating volume conduction between hemispheres that doesn’t exist in the brain. This is devastating for connectivity analysis. Any connectivity metric that measures coupling between left and right hemispheres will be inflated by the electrical bridge, not by neural communication. The Coherence Workstation’s dwPLI connectivity analysis (described in the Connectivity page) is designed to be robust to volume conduction—but introducing an artificial conduction path through the reference defeats that robustness.

For these reasons, if the pipeline receives a recording that was collected with a linked-ears reference, it re-references to average before any analysis. The original reference scheme is logged.

REST: The Other Option

The pipeline also supports REST (Reference Electrode Standardization Technique) as an alternative reference scheme. REST uses a mathematical model of the head to estimate what the EEG would look like referenced to a point at infinity—a theoretically ideal reference that no physical electrode can achieve. REST requires a forward model (the same kind used for source localization) and is computationally more expensive than average reference.

REST is available as a configuration option but is not the default. Average reference is simpler, well-understood, and sufficient for clinical analysis. REST may produce marginally better results in some contexts, but the improvement is modest for 19-channel recordings and the additional complexity isn’t warranted as a default.

To use REST, set preprocessing.reference: REST in the configuration. The pipeline will use the bundled forward model assets to compute the REST transformation.

Where Referencing Sits in the Pipeline

Referencing happens after bad channel detection and interpolation, and before ASR and ICA. This ordering matters.

Bad channel detection needs to compare channels against each other, and that comparison works best with the original recording reference (whatever it was). Re-referencing first would change the inter-channel relationships and potentially mask bad channels or create false positives.

ICA, on the other hand, benefits from average-referenced data. The ICA decomposition assumes that sources are mixed across channels—average referencing removes the reference electrode as a confounding source, giving ICA cleaner input for source separation. The ICA page describes this interaction in more detail.

The sequence is: filter → bad channels → re-reference → ASR → ICA. Each step builds on the previous one’s output, and reordering them would change the results.

How Reference Choice Affects Downstream Metrics

Changing the reference changes everything downstream. Here’s what’s affected:

Spectral power changes because the amplitude at each electrode depends on what it’s measured against. Alpha power at O1 measured against average reference will differ from alpha power at O1 measured against linked ears—not because the brain changed, but because the baseline changed.

Asymmetry is particularly sensitive. The log-ratio asymmetry formula (ln(right) - ln(left)) assumes both sides are measured against the same reference. If that reference itself is lateralized (as linked ears often is), the asymmetry estimate is biased.

Connectivity is the most affected metric. Phase relationships between channels depend on the reference scheme. Average reference preserves phase differences between channels while removing common-mode signals; linked ears introduces phase distortions through the electrical bridge. This is why reference consistency matters so forcefully—if you compare two recordings processed with different reference schemes, the connectivity differences you see may reflect the reference change, not the brain.

This is also why the pipeline logs the reference scheme as part of the parameter sidecar. When reviewing a report, you should always know which reference was used.