mne-stats

MNE Statistics (grill → analyze → critic)

Statistical inference on neurophysiology data via the MNE MCP server. This is the cross-cutting skill: it consumes the outputs of mne-erp, mne-timefreq, mne-connectivity, and mne-source and decides what can actually be claimed from them. It is skeptical by design: most statistical mistakes (conflating cluster-level with point inference, an incomplete correction family, asserting normality at small n) run without any error — so the discipline is to grill before testing and critique before believing.

Companion skills: mne-mcp-guard for technical execution safety; mne-methodology-critic for Phase 3. Loaded objects (evoked / TFR / connectivity / stc arrays) persist in one MNE session.

PHASE 1 — GRILL (before testing anything)

Do not run a test 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? (condition × condition, group × group, pre × post, vs a baseline / vs zero)
• Within- or between-subject? Paired or independent? n per cell, and is there power for it?
• Confirmatory (hypothesis + ROI/window/band pre-specified) or exploratory (whole grid, corrected)?

The two questions that decide validity

• What is the statistical FAMILY — exactly which dimensions are tested? Enumerate channels × times × freqs × ROIs × conditions. The correction must cover the entire family, including exploratory tests you ran but won't headline. An undercounted family is the single most common fatal error here.
• Cluster-level or point inference? ⚠️ A cluster-based permutation test licenses a claim about a cluster (a contiguous blob in space/time/freq), not about any specific channel, time, or frequency inside it. You may NOT say "Cz at 320 ms is significant" from a cluster test. If you need point/peak inference, you need a different, pre-specified test. Decide which claim you're making before you run.