MNE Connectivity (grill → analyze → critic)

Functional/effective connectivity of neurophysiology data via the MNE MCP server. This skill is skeptical by design: most connectivity mistakes (volume conduction inflating zero-lag coherence, trial-count/SNR bias, common-reference artifact) run without any error and produce a beautiful heatmap — so the discipline is to grill before computing and critique before believing.

Companion skills: mne-mcp-guard for technical execution safety; mne-methodology-critic for Phase 3. Loaded objects persist in one MNE session. Connectivity needs the `[full]` extra (mne-connectivity; PAC via pactools/tensorpac).

PHASE 1 — GRILL (before computing anything)

Do not compute connectivity until these are answered. If the user can't answer one, propose a sensible default and explicitly flag the open risk — never silently choose.

Design & claim

• What is the hypothesis, and what is the comparison? (group × group, condition × condition,

pre × post) Is the claim undirected (coupling) or directed/effective (who drives whom)?

• Within- or between-subject? Paired or independent? n per cell?
• Confirmatory (seed/edge/band pre-specified) or exploratory (all-to-all, corrected)?

The question that decides validity

• Which measure, and WHY? ⚠️ Volume conduction / field spread makes a single source appear at

many sensors with zero phase lag, which inflates coherence and PLV — they cannot tell true coupling from one spread-out source. For sensor-space EEG/MEG, prefer measures that discard the zero-lag component: imaginary coherence, wPLI, or PLI. (Using coh/PLV to claim genuine sensor connectivity without this caveat is the single most common fatal error here.)

Data & parameters

• Sensor or source space? Anatomical claims ("frontoparietal coupling") need source space;

sensor-space edges are between electrodes, not brain regions.

• Reference? A common reference (and the average reference) injects a shared signal that **inflates

apparent connectivity**; consider source space, current-source-density / Laplacian, or a reference-robust measure.

• Frequency band(s) and width; epoching length (⇒ low-frequency resolution), rejection threshold.
• For PAC: which phase band drives which amplitude band, and over what window?

Inference plan (pin this down NOW, not after seeing results)

• Seed-based or all-to-all? All-to-all is n_channels² pairs → a large multiple-comparison

family; pre-register the seed/edges or plan a network-level/cluster correction.

• Are trial counts matched across conditions? Most connectivity metrics are **trial-count- and

SNR-biased* — fewer/noisier trials lowers* estimated connectivity. Match n, or correct.

• For Granger / directed measures: stationarity and SNR / pre-whitening assumptions, and

sensitivity to differential SNR between the two signals.

• Small n (≲ 15–20)? Then plan permutation / non-parametric, not a normality assertion.

PHASE 2 — ANALYZE

1. Capability + look first. mne_check_status (confirm mne-connectivity is importable);

mne_plot_psd and read the PNG — connectivity is band-specific, so confirm the band of interest actually has power and is line-noise-free.

1. Pick a field-spread-robust measure and compute the channel×channel map with mne_connectivity

(returns a heatmap + strongest pairs):

   mne_connectivity(epochs_name="epochs", method="wpli", fmin=8, fmax=13)   # alpha-band wPLI

(method="imcoh" / "pli" for field-spread-robust; "coh"/"plv" only with an explicit caveat.)

1. Advanced / full control via mne_run_code with the mne_connectivity library:

   from mne_connectivity import spectral_connectivity_epochs
   con = spectral_connectivity_epochs(
       epochs, method=["wpli", "imcoh"], mode="multitaper",
       fmin=8, fmax=13, faverage=True, mt_adaptive=True)
   wpli = con[0].get_data(output="dense")[:, :, 0]      # (n_ch, n_ch) band-averaged

Time-resolved: spectral_connectivity_time(...) for a per-epoch / sliding estimate.

1. Phase-amplitude coupling (cross-frequency). Use pactools or tensorpac (needs the extra):

   from tensorpac import Pac
   p = Pac(idpac=(2, 0, 0), f_pha=(4, 8, 1, 0.5), f_amp=(30, 80, 5, 2))   # MI; theta phase × gamma amp
   pac = p.filterfit(epochs.info["sfreq"], epochs.get_data(picks=ch)[:, 0, :])

Always compare against a surrogate distribution (shuffle phase) — raw PAC values are biased.

1. Granger / directed. Stationarize first (high-pass, possibly difference), match SNR, and

validate model order; report that stationarity/SNR were checked. Prefer source space for any anatomical directional claim.

1. Match trial counts across conditions (subsample to the smaller n) before comparing, and

consider source space for any anatomical interpretation. Archive the equivalent code + figures (the mne-analyst archiving convention).

Best-practice reminders: report per-group rejection rate and retained trial count; state sensor-vs-source explicitly; pre-register seed/edges/band or correct over the real pair family.

PHASE 3 — CRITIC (before believing the result)

Hand the design + result to `mne-methodology-critic` (invoke the skill, or dispatch it as a subagent with references/methodology-checklist.md). For connectivity work it will specifically check:

• zero-lag-sensitive measures (coherence / PLV) used to claim genuine sensor connectivity without

a volume-conduction caveat (FAIL);

• common-reference / average-reference inflation of apparent connectivity;
• trial-count / SNR bias across conditions (unmatched n fakes a difference);
• all-to-all multiple comparisons over channel pairs left uncorrected;
• Granger / directed measures without stationarity & SNR checks;
• sensor-space connectivity interpreted anatomically (electrode edges ≠ region coupling).

Report its BLOCK / REVISE / PASS verdict to the user and act on it before stating conclusions.

See references/connectivity-methods.md for deeper recipes (measure selection vs field spread, spectral_connectivity_epochs, PAC with surrogates, Granger caveats, and source-space connectivity).