review-data-analysis
について
このスキルは、データ分析と機械学習パイプラインの品質、正確性、再現性をレビューします。データ品質の評価、モデルの検証、前提条件のチェック、リークの検出、再現性の確認を行います。分析作業の出版前レビュー、本番環境導入前の検証、規制監査にご利用ください。
クイックインストール
Claude Code
推奨npx skills add pjt222/agent-almanac -a claude-code/plugin add https://github.com/pjt222/agent-almanacgit clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/review-data-analysisこのコマンドをClaude Codeにコピー&ペーストしてスキルをインストールします
ドキュメント
Review Data Analysis
Eval data analysis pipeline → correctness, robustness, reproducibility.
Use When
- Review colleague analysis notebook/script pre-publication
- Validate ML pipeline pre-prod deploy
- Audit analytical report for regulatory or business decision
- Assess analysis supports stated conclusions
- Second-analyst review in regulated env
In
- Required: Analysis code (scripts, notebooks, pipeline defs)
- Required: Analysis output (results, tables, figures, model metrics)
- Optional: Raw data or data dict
- Optional: Analysis plan/protocol (pre-registered or ad-hoc)
- Optional: Target audience + decision ctx
Do
Step 1: Data Quality
Review input data before eval analysis:
## Data Quality Assessment
### Completeness
- [ ] Missing data quantified (% by column and by row)
- [ ] Missing data mechanism considered (MCAR, MAR, MNAR)
- [ ] Imputation method appropriate (if used) or complete-case analysis justified
### Consistency
- [ ] Data types match expectations (dates are dates, numbers are numbers)
- [ ] Value ranges are plausible (no negative ages, future dates in historical data)
- [ ] Categorical variables have expected levels (no misspellings, consistent coding)
- [ ] Units are consistent across records
### Uniqueness
- [ ] Duplicate records identified and handled
- [ ] Primary keys are unique where expected
- [ ] Join operations produce expected row counts (no fan-out or drop)
### Timeliness
- [ ] Data vintage appropriate for the analysis question
- [ ] Temporal coverage matches the study period
- [ ] No look-ahead bias in time-series data
### Provenance
- [ ] Data source documented
- [ ] Extraction date/version recorded
- [ ] Any transformations between source and analysis input documented
→ Data quality issues documented w/ potential impact on results. If err: data not accessible for review → assess quality from code (what checks + transformations applied).
Step 2: Check Assumptions
For each statistical method/model used:
| Method | Key Assumptions | How to Check |
|---|---|---|
| Linear regression | Linearity, independence, normality of residuals, homoscedasticity | Residual plots, Q-Q plot, Durbin-Watson, Breusch-Pagan |
| Logistic regression | Independence, no multicollinearity, linear logit | VIF, Box-Tidwell, residual diagnostics |
| t-test | Independence, normality (or large n), equal variance | Shapiro-Wilk, Levene's test, visual inspection |
| ANOVA | Independence, normality, homogeneity of variance | Shapiro-Wilk per group, Levene's test |
| Chi-squared | Independence, expected frequency ≥ 5 | Expected frequency table |
| Random forest | Sufficient training data, feature relevance | OOB error, feature importance, learning curves |
| Neural network | Sufficient data, appropriate architecture, no data leakage | Validation curves, overfitting checks |
## Assumption Check Results
| Analysis Step | Method | Assumption | Checked? | Result |
|---------------|--------|------------|----------|--------|
| Primary model | Linear regression | Normality of residuals | Yes | Q-Q plot shows mild deviation — acceptable for n>100 |
| Primary model | Linear regression | Homoscedasticity | No | Not checked — recommend adding Breusch-Pagan test |
→ Every method has assumptions explicitly checked or ack'd. If err: assumptions violated → check if authors addressed (robust methods, transformations, sensitivity analysis).
Step 3: Detect Leakage
Leakage occurs when info from outside training set influences model → over-optimistic perf:
Common patterns:
- Target leakage: Feature directly encoding target (e.g. "treatment_outcome" predicting "treatment_success")
- Temporal leakage: Future info predicting past (features computed from data unavailable at prediction time)
- Train-test contamination: Preprocessing (scaling, imputation, feature select) fitted on full dataset before split
- Group leakage: Related obs (same patient, same device) split across train/test
- Feature engineering leakage: Aggregates computed across entire dataset not within training fold
## Leakage Assessment
| Check | Status | Evidence |
|-------|--------|----------|
| Target leakage | Clear | No features derived from target |
| Temporal leakage | CONCERN | Feature X uses 30-day forward average |
| Train-test contamination | Clear | StandardScaler fit on train only |
| Group leakage | CONCERN | Patient IDs not used for stratified split |
→ All common leakage patterns checked w/ clear/concern status. If err: leakage found → est impact by re-running w/o leaked feature (if possible) or flag for analyst.
Step 4: Validate Perf
Predictive models:
- Appropriate metrics for problem (not just accuracy — consider precision, recall, F1, AUC, RMSE, MAE)
- Cross-validation or holdout strategy described + appropriate
- Perf on training vs test/validation compared (overfitting check)
- Baseline comparison (naive model, random chance, prev approach)
- Confidence intervals or std errors on metrics
- Perf eval'd on relevant subgroups (fairness, edge cases)
Inferential/explanatory models:
- Model fit stats reported (R², AIC, BIC, deviance)
- Coefficients interpreted correctly (direction, magnitude, significance)
- Multicollinearity assessed (VIF < 5–10)
- Influential obs ID'd (Cook's distance, leverage)
- Model comparison if multi specifications tested
→ Validation appropriate for use case (prediction vs inference). If err: test perf suspiciously close to training → flag potential leakage.
Step 5: Reproducibility
## Reproducibility Checklist
| Item | Status | Notes |
|------|--------|-------|
| Code runs without errors | [Yes/No] | Tested on [environment description] |
| Random seeds set | [Yes/No] | Line [N] in [file] |
| Dependencies documented | [Yes/No] | requirements.txt / renv.lock present |
| Data loading reproducible | [Yes/No] | Path is [relative/absolute/URL] |
| Results match reported values | [Yes/No] | Verified: Table 1 ✓, Figure 2 ✗ (minor discrepancy) |
| Environment documented | [Yes/No] | Python 3.11 / R 4.5.0 specified |
→ Reproducibility verified by re-running (or assess from code if data unavailable). If err: results don't reproduce exactly → determine if diff w/in floating-point tolerance or indicates problem.
Step 6: Write Review
## Data Analysis Review
### Overall Assessment
[1-2 sentences: Is the analysis sound? Does it support the conclusions?]
### Data Quality
[Summary of data quality findings, impact on results]
### Methodological Concerns
1. **[Title]**: [Description, location in code/report, suggestion]
2. ...
### Strengths
1. [What was done well]
2. ...
### Reproducibility
[Tier assessment: Gold/Silver/Bronze/Opaque with justification]
### Recommendations
- [ ] [Specific action items for the analyst]
→ Review provides actionable feedback w/ specific refs to code locations. If err: time-constrained → prioritize data quality + leakage checks over style.
Check
- Data quality assessed across completeness, consistency, uniqueness, timeliness, provenance
- Statistical assumptions checked for each method
- Leakage systematically assessed
- Model perf validated w/ appropriate metrics + baselines
- Reproducibility eval'd (code runs, results match)
- Feedback specific, refs code lines or report sections
- Tone constructive + collaborative
Traps
- Review only code: Plan + conclusions matter as much as impl.
- Ignore data quality: Sophisticated models on bad data → confident wrong answers.
- Assume correctness from complexity: Random forest w/ 95% accuracy may have leakage; simple t-test may be correct.
- Not run code: If possible, execute to verify reproducibility. Reading code not sufficient.
- Miss forest for trees: Don't get lost in code style while missing fundamental analytical err.
→
review-research— broader research methodology + manuscript reviewvalidate-statistical-output— double-programming verification methodologygenerate-statistical-tables— publication-ready statistical tablesreview-software-architecture— code structure + design review
GitHub リポジトリ
関連スキル
evaluating-llms-harness
テストこのClaudeスキルは、lm-evaluation-harnessを実行し、MMLUやGSM8Kなど60以上の標準化学術タスクでLLMをベンチマークします。開発者がモデルの品質を比較し、トレーニングの進捗を追跡し、学術的な結果を報告するために設計されています。このツールはHuggingFaceやvLLMモデルを含む様々なバックエンドをサポートしています。
cloudflare-cron-triggers
テストこのスキルは、cron式を使用してWorkersをスケジュールするためのCloudflare Cron Triggersの実装に関する包括的な知識を提供します。定期的なタスクの設定、メンテナンスジョブ、自動化されたワークフローの構築を網羅し、無効なcron式やタイムゾーン問題といった一般的な課題への対処法も含みます。開発者はこれを使用して、スケジュールされたハンドラーの設定、cronトリガーのテスト、WorkflowsやGreen Computeとの連携を構成できます。
webapp-testing
テストこのClaude Skillは、Playwrightベースのツールキットを提供し、Pythonスクリプトを通じてローカルWebアプリケーションのテストを可能にします。フロントエンドの検証、UIデバッグ、スクリーンショット撮影、ログ表示を実現し、サーバーライフサイクルを管理します。ブラウザ自動化タスクにご利用いただけますが、コンテキストの汚染を避けるため、スクリプトのソースコードを読むのではなく直接実行してください。
finishing-a-development-branch
テストこのスキルは、開発者がテストの合格を確認し、構造化された統合オプションを提示することで、完成した作業を仕上げることを支援します。実装が完了した後のマージ、PR作成、ブランチの整理といったワークフローを案内します。コードが準備できてテスト済みの際に使用し、開発プロセスを体系的に完了させましょう。
