What problem does it solve? Statistical results can pass every software test while the underlying inference is wrong — a mis-specified p-value denominator, an unstated estimand, a non-exchangeable permutation null, or uncorrected multiple comparisons. This Skill audits the statistical design itself before you trust the numbers. ## Core Features & Use Cases - Nine-dimension audit: Checks denominator correctness, Monte Carlo p-value formula, FDR/multiple comparisons, bootstrap setup, estimand stability, eligibility rules, clustering/ICC, null exchangeability, and Markov order specification. - Structured verdicts: Reports PASS / CONCERN / FAIL per dimension with the specific statistic or code location and the required fix, separating machine-checkable items from human-review judgments. - Issue catalogue: Ships a reference guide covering recurring failures (pseudoreplication, circular analysis, MICE non-convergence, IPW instability, Manski bounds misuse) with identification cues and remedies. - Use Case: Before accepting a panel-study rerun that reports significant permutation p-values, run the audit to confirm the denominator is n = min(B, total_pairs), the estimand is unchanged, and Benjamini-Hochberg FDR is applied with a defined family. ## Quick Start Audit the statistical design of my permutation-based panel analysis and report PASS, CONCERN, or FAIL for each audit dimension.
