review-data-analysis

Review Data Analysis

Evaluate a data analysis pipeline for correctness, robustness, and reproducibility.

When to Use

• Reviewing a colleague's analysis notebook or script before publication
• Validating a machine learning pipeline before production deployment
• Auditing an analytical report for regulatory or business decision-making
• Assessing whether an analysis supports its stated conclusions
• Performing a second-analyst review in a regulated environment

Inputs

• Required: Analysis code (scripts, notebooks, or pipeline definitions)
• Required: Analysis output (results, tables, figures, model metrics)
• Optional: Raw data or data dictionary
• Optional: Analysis plan or protocol (pre-registered or ad-hoc)
• Optional: Target audience and decision context

Procedure

Step 1: Assess Data Quality

Review the input data before evaluating the 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

Expected: Data quality issues documented with their potential impact on results. On failure: If data is not accessible for review, assess quality from the code (what checks and transformations are applied).

Step 2: Check Assumptions

For each statistical method or 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 |

Expected: Every statistical method has its assumptions explicitly checked or acknowledged. On failure: If assumptions are violated, check whether the authors addressed this (robust methods, transformations, sensitivity analysis).

Step 3: Detect Data Leakage

Data leakage occurs when information from outside the training set influences the model, leading to over-optimistic performance:

Common leakage patterns:

• Target leakage: Feature that directly encodes the target variable (e.g., "treatment_outcome" used to predict "treatment_success")
• Temporal leakage: Future information used to predict the past (features computed from data that wouldn't be available at prediction time)
• Train-test contamination: Preprocessing (scaling, imputation, feature selection) fitted on full dataset before splitting
• Group leakage: Related observations (same patient, same device) split across train and test sets
• Feature engineering leakage: Aggregates computed across the entire dataset rather than within the 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 |

Expected: All common leakage patterns checked with clear/concern status. On failure: If leakage is found, estimate its impact by re-running without the leaked feature (if possible) or flag for the analyst to investigate.

Step 4: Validate Model Performance

For predictive models:

• Appropriate metrics for the problem (not just accuracy — consider precision, recall, F1, AUC, RMSE, MAE)
• Cross-validation or holdout strategy described and appropriate
• Performance on training vs. test/validation set compared (overfitting check)
• Baseline comparison provided (naive model, random chance, previous approach)
• Confidence intervals or standard errors on performance metrics
• Performance evaluated on relevant subgroups (fairness, edge cases)

For inferential/explanatory models:

• Model fit statistics reported (R², AIC, BIC, deviance)
• Coefficients interpreted correctly (direction, magnitude, significance)
• Multicollinearity assessed (VIF < 5–10)
• Influential observations identified (Cook's distance, leverage)
• Model comparison if multiple specifications tested

Expected: Model validation appropriate for the use case (prediction vs. inference). On failure: If test set performance is suspiciously close to training performance, flag potential leakage.

Step 5: Assess 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 |

Expected: Reproducibility verified by re-running the analysis (or assessing from code if data is unavailable). On failure: If results don't reproduce exactly, determine if differences are within floating-point tolerance or indicate a problem.

Step 6: Write the 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]

Expected: Review provides actionable feedback with specific references to code locations. On failure: If time-constrained, prioritize data quality and leakage checks over style issues.

Validation

• Data quality assessed across completeness, consistency, uniqueness, timeliness, provenance
• Statistical assumptions checked for each method used
• Data leakage systematically assessed
• Model performance validated with appropriate metrics and baselines
• Reproducibility evaluated (code runs, results match)
• Feedback is specific, referencing code lines or report sections
• Tone is constructive and collaborative

Common Pitfalls

• Reviewing only the code: The analysis plan and conclusions matter as much as the implementation.
• Ignoring data quality: Sophisticated models on bad data produce confident wrong answers.
• Assuming correctness from complexity: A random forest with 95% accuracy might have data leakage; a simple t-test might be the correct approach.
• Not running the code: If at all possible, execute the code to verify reproducibility. Reading code is not sufficient.
• Missing the forest for the trees: Don't get lost in code style issues while missing a fundamental analytical error.

Related Skills

• review-research — broader research methodology and manuscript review
• validate-statistical-output — double-programming verification methodology
• generate-statistical-tables — publication-ready statistical tables
• review-software-architecture — code structure and design review