Functional connectivity quantifies statistical dependence between signals: correlated amplitude, consistent phase difference, shared information. Effective connectivity goes further and asks about direction, that is, which signal predicts which.
At the scalp, every source reaches many electrodes at once through volume conduction. Two channels can look strongly "connected" simply because they see the same source with zero lag. Choosing a measure that is insensitive to zero-lag coupling is the central methodological decision.
The measures
| Measure | What it quantifies | Volume conduction |
|---|---|---|
| Correlation | Linear relationship between amplitudes | Sensitive |
| Coherence | Frequency-resolved linear coupling | Sensitive |
| Imaginary coherence | Only the lagged part of coherency | Robust to zero-lag mixing |
| PLV | Consistency of phase difference | Sensitive |
| PLI | Asymmetry of phase difference around zero | Robust; noisy for small lags |
| wPLI | PLI weighted by the imaginary component | Robust and less noise-sensitive |
| Mutual information | Any statistical dependence, linear or not | Sensitive |
| Granger / transfer entropy | Whether one signal's past predicts another | Directed; still affected by mixing |
Connectivity in BCILattice
Feature Extraction has nine connectivity families: correlation, cross-correlation lag, coherence, imaginary coherence, PLV, PLI, wPLI, mutual information and graph metrics. Three modes control the output size. Node Strength gives one column per channel, All Pairs keeps every channel pair (which grows fast; at 64 channels use Node Strength), and Graph Summary reduces the network to five descriptors.
For directed questions, the Causal Connectivity block estimates which channels drive which with Granger causality and transfer entropy, tested against a random baseline. The Channel Graph Builder and Graph Neural Encoder blocks turn electrode relationships into a graph a neural network can learn from.
Cautions
Trial count bias
PLV and coherence are biased upwards with few trials. Compare conditions with equal trial counts.
Reference
The reference electrode is shared by every channel and creates spurious coupling. Re-reference first.
Sensor vs source
Sensor-level connectivity is not brain-region connectivity. Source-space analysis is better done in MNE-Python, Brainstorm or FieldTrip.
Questions
What is the difference between PLV and PLI?
PLV measures how consistent the phase difference between two signals is, including zero lag. PLI only counts consistent non-zero phase lags, which makes it robust to volume conduction but less sensitive to genuine near-zero-lag coupling.
Which connectivity measure is best for EEG?
For scalp EEG, measures robust to volume conduction such as wPLI or imaginary coherence are usually preferred, because coherence and PLV are inflated by sources seen on many electrodes.
Can I use connectivity as a classifier feature?
Yes. Connectivity values per channel or channel pair become columns in the design matrix. Use a compact mode such as node strength, since all pairs on a dense montage creates thousands of features.
Read the reference
Related
EEG Feature Extraction: Which Features to Use
A feature is a reduction of a trial to a number. The right ones depend on where the signal lives (time, frequency, space or the links between channels), and the wrong ones multiply until the model overfits.
EEG Time-Frequency Analysis: ERSP and ITC
A spectrum tells you which frequencies are present. A time-frequency decomposition tells you when, and how power and phase change relative to an event.
EEG Analysis Software
ICA review with components you actually label, channel scoring you confirm, epoching with rejection you can see, and a time-frequency view that computes once and derives seven views from it.
EEG Analysis Tools Compared
The honest answer to "which EEG tool should I use" is that they solve different problems. Here is what each is strongest at, and where BCILattice fits.