data-analysis

Data Analysis

Generate rigorous statistical analysis code with multi-round review.

Input

• $0 — Data source (CSV, JSON, pickle, or experiment logs)
• $1 — Research goal or hypothesis to test

References

• 4-round code review prompts: ~/.claude/skills/data-analysis/references/review-prompts.md

Scripts

Statistical summary and comparison

python ~/.claude/skills/data-analysis/scripts/stat_summary.py --input results.csv --compare method --metric accuracy --output summary.json
python ~/.claude/skills/data-analysis/scripts/stat_summary.py --input results.csv

Detects data types, recommends tests, runs comparisons, outputs effect sizes and significance stars. Requires numpy, scipy.

Format p-values

Formats p-values with stars, LaTeX notation, or plain text. Stdlib-only.

Workflow

Step 1: Generate Analysis Code

Structure the code with these sections:

1. # IMPORT — pandas, numpy, scipy, statsmodels, sklearn
2. # LOAD DATA — Load from original data files
3. # DATASET PREPARATIONS — Missing values, units, exclusion criteria
4. # DESCRIPTIVE STATISTICS — Summary tables if needed
5. # PREPROCESSING — Dummy variables, normalization
6. # ANALYSIS — Statistical tests per hypothesis
7. # SAVE ADDITIONAL RESULTS — Extra results to pickle

Step 2: 4-Round Code Review

1. Round 1 — Code Flaws: Mathematical/statistical errors, wrong calculations, trivial tests
2. Round 2 — Data Handling: Missing values, units, preprocessing, test choice
3. Round 3 — Per-Table: Sensible values, measures of uncertainty, missing data
4. Round 4 — Cross-Table: Completeness, consistency, missing variables

Step 3: Produce Results

• Every nominal value must have uncertainty (CI, STD, or p-value)
• Statistical tests must be appropriate for the data type
• Results must match actual data — never hallucinate

Allowed Packages

pandas, numpy, scipy, statsmodels, sklearn, pickle

Statistical Test Selection

Rules

• Always report p-values for statistical tests
• Account for relevant confounding variables
• Use inherent package functionality (e.g., formula = "y ~ a * b" for interactions)
• Do not manually implement available statistical functions
• Access dataframes using string-based column names, not integer indices

Related Skills

• Upstream: experiment-code, experiment-design
• Downstream: table-generation, figure-generation, backward-traceability
• See also: math-reasoning