One of the most widely used reference choices in EEG research is the common average reference, or CAR, which recalculates every channel's value relative to the average of all channels on the scalp.
CAR has a reputation as a noise-cleaning default. It shows up in BCI pipelines, published papers, and open-source toolboxes almost automatically. But a closer look at the available research shows a picture that is more mixed than the reputation suggests.
This piece walks through the math behind CAR, the assumptions it depends on, and the conditions under which those assumptions break down.
What Is the Common Average Reference in EEG?
Every scalp electrode measures a voltage relative to some reference channel, or a small set of reference channels, chosen at the time of recording. Common choices include a single electrode on the earlobe, linked mastoids behind the ears, or a scalp site like Cz.
The problem with a single-site reference is that it is never truly “quiet.” If the reference electrode itself picks up noise or brain activity, that contamination gets subtracted into every other channel, since every channel's signal is defined relative to it.
CAR sidesteps that problem by using a different kind of reference where the average voltage is recorded across the entire electrode array at each moment in time. Instead of subtracting one electrode's value from all others, CAR subtracts the mean of all electrodes from each individual electrode.
In theory, this average acts as a more stable, “quieter” reference point than any single physical electrode could provide, because it draws on information from the whole scalp rather than one location.
CAR in BCI Research
This is why CAR appears so often in brain-computer interface research. For instance, a study published in the International Journal of Engineering and Technology tested CAR as one of twelve re-referencing methods for a P300 speller, a system that detects a specific brain response when a user focuses on a target letter or symbol, and reported CAR as the best-suited technique among those tested.
Furthermore, a 2025 study applied CAR as a standard pre-processing step in a motor imagery classification pipeline, describing its purpose as increasing the signal-to-noise ratio.
How to Calculate the CAR Formula
The mechanics of CAR are simple algebra, not a statistical model that requires data to prove it works mathematically. For an array of N electrodes, each recording a voltage at time t, written as V₁(t), V₂(t), up through Vₙ(t), the CAR-transformed value for any single electrode i is:
V_i(t)^CAR \= V_i(t) - (V_1(t) + V_2(t) + ... + V_N(t)) / N
In plain terms, to find the CAR for a specific electrode, take its original reading at a given millisecond, and subtract the average reading of all scalp electrodes at that exact same millisecond.
Applications of Common Average Reference EEG
Selecting the correct reference approach determines the success of diagnostic assessments and complex research studies.
Clinical setups often prioritize clarity and consistency, ensuring that neurologists can identify markers without interference from reference-induced artifacts. Researchers favor this global approach when mapping how interventions might influence neural activity, as demonstrated in findings on breathing modulated oscillations, where a neutral baseline is necessary for isolating respiratory-linked effects from local neural rhythms.
In both clinical settings and experimental neuroscience, researchers rely on consistent data interpretation that minimizes artificial distortions. This method helps maintain transparency when comparing patient recordings across different sessions or facilities.
By utilizing standard referencing, analysts ensure that observed changes in amplitude reflect biological shifts rather than technical shifts in the baseline. This objective stance supports clearer diagnostic reporting and valid results in broad brainwave measurement applications.
How CAR Reduces Common-Mode Noise
The argument for CAR rests on a concept called common-mode noise. This refers to interference that appears on nearly every electrode at roughly the same strength, rather than noise that is specific to one location.
Classic examples include 50/60 Hz electrical line noise from nearby power sources, muscle activity that spreads across the scalp through tissue conduction, and slow drifts caused by an electrode shifting slightly against the skin.
Because this kind of noise is shared broadly across the array, averaging all channels together should, in theory, produce a reasonable estimate of that shared noise component. Subtracting the average from each channel then removes much of that shared interference while leaving channel-to-channel differences, which are more likely to reflect actual brain activity, largely intact.
The Core Assumptions Behind CAR
CAR's noise-reduction logic only holds up if several conditions are true of the data. These assumptions are described consistently across EEG textbooks and tutorials, though their real-world validation is thin within the available evidence.
The zero-mean assumption. At any instant, the average of all voltages across the head is presumed to be close to zero, meaning positive and negative activity roughly balances out across the scalp.
Dense, even electrode coverage. The array is presumed to cover the head thoroughly enough that the average approximates what a reference point infinitely far from the head, and therefore electrically neutral, would record. Sparse or uneven coverage weakens this approximation.
No single dominant source. No one electrode, bad channel, or large artifact (such as a strong eye blink) should be large enough to skew the average on its own.
When these three conditions hold, the average behaves like a genuinely neutral reference point. When they do not, the average itself becomes distorted, and subtracting a distorted average introduces new problems rather than removing old ones.
Testing CAR's Assumptions With Real EEG Data
Taking a publicly available resting-state EEG recording, for example a standard 64-channel dataset, and computing the global mean waveform before applying CAR often reveals values that deviate from zero, sometimes by a noticeable margin. This deviation is direct evidence of common-mode content sitting in the raw signal, which is exactly what CAR is designed to remove. After CAR is applied, that same global mean is forced to exactly zero at every time point, by definition of the formula.
A more revealing test involves looking at epochs containing large eye-blink artifacts.
Eye blinks generate large voltage swings that are strongest at frontal electrodes but bleed into much of the array. During these epochs, the global mean before CAR often shifts sharply away from zero, because the blink is not evenly distributed but concentrated in one part of the head. When CAR is then applied, this concentrated artifact gets folded into the average and redistributed, in smaller amounts, across every single channel, including ones far from the eyes that were originally clean.
What the Research Says: Mixed Evidence From BCI Studies
The aforementioned study compared twelve re-referencing techniques across three P300 speller datasets, in both offline and online testing conditions, and concluded that the CAR was the best-suited technique among all twelve. However, while the study provides graphical comparisons of classification accuracy and tables detailing average maximum bitrates with standard deviations, it does not report effect sizes or formal statistical significance tests between the methods, which limits how much confidence can be placed in that ranking.
Meanwhile a 2017 study took a different approach with a motor imagery and movement-intention task. Eleven subjects performed and imagined right wrist movements while EEG was recorded from 28 electrodes. The signal was processed using both CAR and Laplacian referencing, a spatial filtering method that emphasizes the difference between a central electrode and its immediate neighbors rather than the whole-scalp average.
Classification accuracy using Laplacian referencing ranged from 63.33% to 100% for imagined movement and 60% to 96.67% for actual movement, with k-nearest neighbor classifiers outperforming quadratic discriminant analysis. Laplacian referencing outperformed CAR overall, though the study does not report CAR's exact accuracy figures for direct comparison. This result suggests CAR may be less suited to tasks involving focal, localized motor-related brain activity.
Lastly, the aforementioned 2025 study embedded CAR as an early pre-processing step inside a larger convolutional neural network pipeline for motor imagery classification, which also included sliding time windows, spectral transformation, and frequency-band extraction. The full pipeline achieved 91.75% accuracy on a competition benchmark dataset. This is a strong result, but because CAR was only one of several processing steps, the study cannot tell us how much of that accuracy is attributable to CAR itself versus the CNN architecture, the windowing technique, or the frequency-band selection.
Taken together, these three studies do not converge on a single conclusion. CAR performed well in a P300 context, performed worse than an alternative in a motor imagery context, and was present but not isolated in a high-accuracy deep-learning context. Thus, the evidence discussed suggests that CAR's standalone benefit remains unclear and appears to depend heavily on the type of brain signal being measured.
When CAR Fails: Artefacts, Sparse Arrays, and Focal Sources
The pattern across these studies lines up with three failure modes that are widely discussed in EEG methodology but only partially supported by direct evidence in the available research.
Large artifacts. A single high-amplitude event, such as a strong eye blink or a muscle spike, can dominate the average calculation if it is large enough relative to the rest of the array. When this happens, CAR does not remove the artifact; it spreads a distorted version of it into every channel, including ones that were originally artifact-free. This is a direct consequence of the CAR formula rather than a separately tested finding, but it follows logically from the demonstration described earlier in this article.
Sparse arrays. CAR depends on the average approximating a neutral reference point, which requires reasonably dense and even coverage of the scalp. With only a handful of electrodes, for example eight to sixteen channels, the average is a much weaker estimate of that neutral point, and the coverage assumption behind CAR is directly violated.
Focal sources. Brain activity that originates from a small, localized region, rather than spreading broadly across the scalp, can behave similarly to a “local” signal that CAR's whole-array averaging is not designed to preserve. Because CAR subtracts a global average, it can partially cancel out signals that are concentrated rather than widespread.
Failure Mode | Key Issue |
|---|---|
Artifacts | Large artifact skews average |
Sparse Arrays | Too few electrodes, weak reference |
Focal Sources | Local signals may be attenuated |
How to Mitigate CAR's Weaknesses
Several adjustments are commonly recommended in EEG practice to address these failure points:
When large artifacts are a concern, identify and interpolate or remove bad channels or artifact‑heavy segments before computing CAR.
When working with a sparse array (e.g., 8–16 channels), avoid CAR and use a fixed physical reference such as linked mastoids.
For tasks targeting focal, localized brain activity, consider Laplacian referencing or source‑space methods that emphasize local gradients rather than a global average.
Is CAR the Right Reference Choice for Your EEG Setup?
CAR remains a widely used reference method with a clear and coherent mathematical basis. It forces the average signal across the array to zero, and this can, in principle, strip out noise that appears broadly and evenly across the scalp. That theoretical appeal explains why it shows up so often as a default step in EEG and BCI pipelines.
Anyone applying CAR to their own data should treat it as a reasonable default rather than a guaranteed improvement. Its assumptions, a roughly zero-mean signal, dense and even electrode coverage, and the absence of a dominating artifact, are worth checking against the specific recording setup and task at hand rather than assumed automatically.
Where those assumptions are unlikely to hold, particularly with sparse arrays or tasks centered on focal, localized brain sources, alternatives such as Laplacian referencing deserve serious consideration.
Why Checking Your EEG Reference Assumptions Outweighs Default Settings
The common average reference is built on the simple mathematical idea of subtracting the whole-scalp average to remove noise shared across electrodes. This works beautifully on paper, but real brain recordings rarely cooperate perfectly. The algorithm always forces the electrode average to zero, but that forced balance does not guarantee a cleaner view of brain activity—only that the numbers add up.
What matters more than the EEG montage itself is whether the recording setup meets the underlying assumptions. Dense, even electrode coverage and the absence of overwhelming artifacts like eye blinks can turn CAR from a risky shortcut into a useful tool. For sparse arrays or tasks that need to capture small, focused brain signals, the same step can spread contamination and blur the very activity a researcher hopes to detect.
The takeaway from the research is not that CAR is good or bad, but that its use demands a deliberate check of the data conditions, not a blind trust in a popular preset.
References
Alhaddad, M. J. (2012). Common average reference (CAR) improves P300 speller. International Journal of Engineering and Technology, 2(3), 21.
Atla, K. G. R., & Sharma, R. (2025). Motor imagery classification using a novel CNN in EEG-BCI with common average reference and sliding window techniques. Alexandria Engineering Journal, 120, 532-546. https://doi.org/10.1016/j.aej.2025.02.001
Syam, S. H. F., Lakany, H., Ahmad, R. B., & Conway, B. A. (2017, December). Comparing common average referencing to laplacian referencing in detecting imagination and intention of movement for brain computer interface. In MATEC Web of Conferences (Vol. 140). https://doi.org/10.1051/matecconf/201714001028
Frequently Asked Questions
What is the common average reference (CAR) in EEG?
CAR is a re-referencing method that subtracts the average voltage of all scalp electrodes from each individual electrode at every time point. This replaces a single physical reference with the whole-scalp average, aiming to create a more stable reference point for the recording.
How does CAR reduce noise in EEG signals?
CAR targets common-mode noise—interference that appears similarly across many electrodes, like power line hum or muscle activity. By averaging all channels and subtracting that average, the shared noise is largely removed while channel-specific brain activity differences are preserved.
What are the core assumptions required for CAR to work well?
CAR assumes the scalp-wide voltage averages near zero at each moment, that electrode coverage is dense and even, and that no single artifact or channel dominates the average. If these don’t hold, the calculated average becomes distorted, and subtracting it introduces errors.
When does CAR fail or introduce artifacts?
CAR can fail with large, localized artifacts like eye blinks, which skew the average and then get spread across all channels. It also struggles with sparse electrode arrays or brain signals that are highly focal, because the global average no longer represents a neutral reference.
What does the available research say about CAR's effectiveness?
Evidence is mixed. One study found CAR worked well for a P300 speller task, but another showed Laplacian referencing outperformed CAR for motor imagery. A third study used CAR in a successful deep-learning pipeline but didn’t isolate its specific contribution, so its standalone benefit remains unclear.
Should I always use CAR as the default reference for my EEG analysis?
Not blindly. CAR is a reasonable default if you have dense, even electrode coverage and the signal is roughly zero-mean without dominating artifacts. For sparse arrays or focal brain activity, alternatives like a fixed physical reference or Laplacian referencing may be more appropriate.
What is Laplacian referencing and how does it compare to CAR?
Laplacian referencing emphasizes the voltage difference between a central electrode and its immediate neighbors, highlighting local brain activity. It outperformed CAR in a motor imagery study, suggesting it’s better suited for detecting spatially focused signals.
How can I mitigate CAR's weaknesses when I do want to use it?
Before computing CAR, identify and remove or interpolate bad channels and large artifacts like blinks. This prevents a single noisy channel or event from distorting the whole-scalp average and contaminating all channels.
What happens when an eye blink occurs in a CAR-referenced recording?
Eye blinks create strong voltage shifts concentrated at frontal electrodes. When CAR is applied, the blink’s influence gets included in the global average and then subtracted, which spreads a smaller but distorted version of the blink into every channel, even those originally clean.
Does CAR actually make the average of all channels zero?
Yes, by definition the CAR transformation forces the sum of all re-referenced voltages to zero at each time point. However, this mathematical property does not guarantee that the resulting signal is a cleaner representation of brain activity—it simply enforces a condition that may or may not match reality.
Emotiv is a neurotechnology leader helping advance neuroscience research through accessible EEG and brain data tools.
Christian Burgos
