Plan and review statistical analyses with traceable estimands, data checks, diagnostics, uncertainty, sensitivity tests, and bounded claims.
A clean-looking p-value can still rest on the wrong unit, a changed denominator, repeated looks, hidden exclusions, or an observational comparison written as a causal result. This skill makes those choices visible before a decision depends on them.
It defines the population and estimand, records source provenance, checks assignment and analysis populations, matches methods to the actual sampling structure, inventories multiplicity, and separates planned work from exploratory work. Results lead with effect size and uncertainty. Diagnostics, missing-data choices, sensitivity checks, and unsupported claims stay attached to the conclusion.
Use it for experiments, observational comparisons, before-and-after studies, segmented metrics, anomaly claims, forecasts, or an existing analysis that needs a skeptical review. A bundled read-only checker validates a structured analysis record and its cross-references. It does not run the statistics, inspect raw systems, or replace a qualified statistical or domain reviewer.
Input
Review our checkout experiment. Assignment was intended to be 50/50, but the export has 41,200 control accounts and 38,050 variant accounts. The team checked purchase rate, revenue per account, refund rate, and 18 country segments every day, then stopped after one segment reached p=0.03. Decide whether we can ship. The practical threshold is +0.5 percentage point purchase rate.
Output
A confirmatory-versus-exploratory inventory; sample-ratio mismatch as a controlling diagnostic; exact population and denominator requests; repeated-look and multiplicity treatment; effect-size and interval requirements against the +0.5 point threshold; post-hoc segment limits; and a blocked or exploratory-only verdict until assignment integrity is resolved.
Provide the decision, population, unit of analysis, exposure or intervention, comparator, outcome, time horizon, practical threshold, data sources, protocol or metric definitions, planned analyses, and any existing results. The skill returns the narrowest supportable plan or review and keeps missing provenance, failed diagnostics, exploratory choices, and unsupported claims visible.
The decision, owner, population, unit, exposure or intervention, comparator, outcome, time horizon, and smallest effect that would change the action. Missing fields limit the output to clarification or a blocked plan.
Source extracts or summaries, metric definitions, queries or revisions, exclusions, assignment or sampling details, model choices, diagnostics, results, and known limitations. Raw personal data is not required.
The person or role responsible for statistical review, domain review, data approval, and the final decision. The skill does not grant approval authority.
Required only for the optional read-only JSON record checker. It uses the standard library, reads one local file, and writes findings to stdout.
SKILL.md; statistical analysis review rubric; human-readable and JSON record templates; read-only structured record checker; Agent Skills interface metadata
No reviews yet.
Input
Users who enabled Team Spaces retained at 62% after 90 days versus 38% for users who did not. Product wants to say Team Spaces causes retention. The feature is available only on paid plans and is usually enabled after the first month. We have account size, plan, tenure, seat count, activation events, and churn dates. Build a defensible analysis review.
Output
A target estimand and time-zero definition; immortal-time, plan, tenure, and selection-bias risks; post-treatment-variable warnings; an observational analysis and sensitivity plan; overlap and missingness diagnostics; supported association wording; causal wording marked unsupported; and the evidence needed for a randomized or stronger quasi-experimental claim.
Input
Return one JSON analysis record for a before-and-after support change across 12 teams. Each team contributes weekly resolution time for 10 weeks before and 8 weeks after rollout. Rollout dates differ, ticket mix changed, two teams have four missing weeks, and the draft analysis treats all 216 team-weeks as independent. Include analyses, diagnostics, sensitivities, issues, claims, and a verdict. Do not invent results that were not supplied.
Output
One parseable JSON object with stable IDs; a team- and time-aware estimand; source and missingness states; the independence error recorded as blocking; an appropriate clustered or repeated-measures analysis plan; rollout, seasonality, case-mix, and missing-data diagnostics; no fabricated estimates; explicit unsupported claims; and a blocked verdict with owned next steps.
Creator
Llorvex74
