Data Analysis Pipeline

Generate, execute, and verify analysis scripts across R, Python, Stata, and Julia.

Modes

Default: Full. Detect mode from user request or ask if ambiguous.

When to Use

• "Analyse this data" / "Run EDA on this CSV" / "Estimate the model"
• "Generate results tables" / "Create publication figures"
• Any task requiring data → script → output pipeline

When NOT to Use

• Experimental design or power analysis → experiment-design
• Generating synthetic data for testing → synthetic-data
• Auditing identification strategy → causal-design
• Proofreading or compiling the paper → proofread, latex

Shared References

• Method probing questions: shared/method-probing-questions.md — ask before running any analysis
• Validation tiers: shared/validation-tiers.md — declare tier before examining results
• Escalation protocol: shared/escalation-protocol.md — escalate when methodology answers are vague
• Distribution diagnostics: shared/distribution-diagnostics.md — mandatory DV checks before model selection
• Engagement-stratified sampling: shared/engagement-stratified-sampling.md — stratify by engagement tiers for social media data
• Inter-coder reliability: shared/intercoder-reliability.md — per-category reliability for content analysis and LLM annotation

Workflow

Phase 1: Setup

1. Detect project structure: Read CLAUDE.md, check for data/, code/, paper/ directories.
2. Detect language: Check existing scripts, user preference, or ask. Read shared/multi-language-conventions.md for the chosen language's conventions.
3. Locate data: Find datasets in data/raw/ or data/processed/. Never modify data/raw/ (per data-sensitivity rule). For social-media datasets, follow shared/engagement-stratified-sampling.md when constructing analysis samples.
4. Confirm validation tier per shared/validation-tiers.md. Tier dictates claim-strength language allowed in Phase 4 outputs and how strict the locked-design gate (step 5) is enforced.
5. Check for locked design: Look for analysis plan in log/plans/, .context/project-recap.md, or MEMORY.md estimand registry. If running Estimation or Full mode and no design exists, stop and warn: "No locked research design found. Run experiment-design or causal-design first, or confirm the specification before proceeding." Use shared/method-probing-questions.md to probe gaps if the user pushes back on the gate.

Phase 2: Exploratory Data Analysis

Generate and execute an EDA script that produces:

1. Data overview: dimensions, types, missingness summary
2. Univariate distributions: histograms/density for continuous, bar charts for categorical
3. Bivariate relationships: correlation matrix, key scatterplots, cross-tabulations
4. Outlier detection: box plots, IQR-based flags
5. Distribution diagnostics (mandatory): run distribution_diagnostics() from shared/distribution-diagnostics.md on every DV and key IVs. Report skewness, zero proportion, overdispersion, and model recommendation. Flag if OLS is inappropriate.
6. Balance tables (if treatment variable identified): pre-treatment covariate balance

Output routing:

• Exploratory figures → output/figures/ (not paper/figures/)
• Summary statistics → output/tables/ as .csv
• EDA script → code/01_eda.R (or .py/.do/.jl)

EDA mode stops here.

Phase 3: Estimation

Gate check: Verify the research design is locked before proceeding. The specification (estimand, identifying assumptions, main model) must be documented. This enforces the design-before-results rule. If the user resists the gate, follow shared/escalation-protocol.md — escalate rather than accommodate. For analyses involving human or LLM coding (content analysis, annotation), require per-category reliability per shared/intercoder-reliability.md before estimation.

Generate estimation script(s) covering:

1. Main specification — as defined in the locked design
2. Robustness checks — pre-committed alternatives (different SEs, controls, subsamples)
3. Diagnostics — specification-appropriate tests (first-stage F for IV, parallel trends for DiD, bandwidth sensitivity for RDD)

Read references/estimation-recipes.md for language-specific estimation patterns.

Output:

• Estimation script → code/02_estimation.R (or equivalent)
• Coefficient estimates → output/results/ as .rds/.pkl/.dta for downstream table generation

Phase 4: Publication Output

Generate publication-ready tables and figures. Read shared/publication-output.md for format standards and references/table-formatting.md for language-specific recipes.

1. Main results table — booktabs three-line format, exported as .tex to paper/tables/
2. Robustness tables — same format, appendix naming convention
3. Publication figures — coefficient plots, event study plots, mechanism figures → paper/figures/ as PDF
4. Inline statistics — export \newcommand definitions for key numbers referenced in text

Critical rule: All numbers in .tex files must come from generated files via \input{}. Never hard-code results (per no-hardcoded-results rule). Scripts in code/, outputs in paper/ (per overleaf-separation rule).

Output script → code/03_tables_figures.R (or equivalent)

Phase 5: Save & Review

1. Verify outputs exist: Check all expected files in paper/tables/ and paper/figures/
2. Run the code-review agent on all generated scripts (auto-invoke via skill-routing mechanism)
3. Log the analysis: Record what was done, which scripts were created, which outputs were generated
4. Suggest next steps: compilation with latex, or additional analyses

Script Structure

Every generated script follows this header template:

# ============================================================
# Script: [filename]
# Purpose: [one-line description]
# Inputs: [list of input files]
# Outputs: [list of output files]
# Dependencies: [packages used]
# Author: [from git config]
# Date: [today]
# ============================================================

Cross-References