Spatial Filtering & Functional Connectivity

Common Average Reference (CAR)

The common average reference re-references every channel by subtracting, at each time point, the instantaneous mean across all electrodes. Its justification comes from the fact that the total current leaving a closed head surface is zero, so with dense and uniform coverage the average potential should approach the (unmeasurable) potential at infinity, making CAR an approximation to a reference-free montage. It is a linear spatial filter and, like all re-referencing, is applied identically to every sample.

The approximation degrades when coverage is sparse or spatially uneven, which is always the case in practice because the lower half of the head is not sampled. This asymmetry injects a spatially structured bias into every channel, and a large-amplitude artifact on one electrode is smeared into all others by the subtraction. CAR is nonetheless a common, cheap default, especially for high-density montages and for ECoG grids where coverage is denser.

Because CAR mixes every channel into every other, an unremoved eye-blink or a single bad electrode contaminates the whole montage; reject or interpolate bad channels before averaging.

Also called
CARaverage reference平均參考