Data Analysis

Data Analysis Workflow

Run an end-to-end data analysis in R: load, explore, analyze, and produce publication-ready output.

Input: $ARGUMENTS — a dataset path (e.g., data/county_panel.csv) or a description of the analysis goal (e.g., "regress wages on education with state fixed effects using CPS data").

Constraints

• Follow R code conventions in .claude/rules/r-code-conventions.md
• Save all scripts to scripts/R/ with descriptive names
• Save all outputs (figures, tables, RDS) to output/
• Use saveRDS() for every computed object — Quarto slides may need them
• Use project theme for all figures (check for custom theme in .claude/rules/)
• Run r-reviewer on the generated script before presenting results

Workflow Phases

Phase 0: Pre-Flight Report

Before writing any analysis code, produce a Pre-Flight Report showing you read the inputs. This prevents the common failure mode where the agent hallucinates variable names or skips project conventions.

Output block (in your response to the user, before Phase 1):

## Pre-Flight Report

**Dataset:** [path]
- Variables found: [list from head()/names()]
- Rows: [count]
- Key types: [e.g., "outcome=numeric, treatment=binary, state=factor"]
- Missing-data summary: [% missing per key var]

**Project conventions read:**
- `.claude/rules/r-code-conventions.md` — [one-line summary of most relevant rule]
- `.claude/rules/content-invariants.md` — [INV-9, INV-10, INV-11, INV-12 applicable]

**Task interpretation:** [one sentence restating what the user asked for]

**Plan:** [3-5 bullet outline of the R script structure]

If any input cannot be read (missing file, unreadable format), stop and ask the user before proceeding.

Phase 1: Setup and Data Loading

1. Create R script with proper header (title, author, purpose, inputs, outputs)
2. Load required packages at top (library(), never require())
3. Set seed once at top in YYYYMMDD format (per r-code-conventions.md), e.g. set.seed(20260415) (INV-9)
4. Load and inspect the dataset

Phase 2: Exploratory Data Analysis

Generate diagnostic outputs:

• Summary statistics: summary(), missingness rates, variable types
• Distributions: Histograms for key continuous variables
• Relationships: Scatter plots, correlation matrices
• Time patterns: If panel data, plot trends over time
• Group comparisons: If treatment/control, compare pre-treatment means

Save all diagnostic figures to output/diagnostics/.

Phase 3: Main Analysis

Based on the research question:

• Regression analysis: Use fixest for panel data, lm/glm for cross-section
• Standard errors: Cluster at the appropriate level (document why)
• Multiple specifications: Start simple, progressively add controls
• Effect sizes: Report standardized effects alongside raw coefficients

Specification ledger — every run, kept or not

Every specification estimated in this phase, including the ones dropped and the ones that failed, gets one row appended to quality_reports/spec-ledger.md as it is run. It is the record a referee's "what else did you try?" deserves: it keeps the search visible rather than preventing it, and it records specifications without advising which to run.

• Append only. Never edit or delete a row; a correction is a new row. A commit that changes a committed row fails the repo-hygiene gate.
• Status is kept, dropped or failed; Why says why for anything not kept.
• Estimate is optional. On restricted data, leave it empty until the number has cleared disclosure review (confidential-data.md).
• Escape a | inside a formula as \|, or the table breaks. A specification containing a backtick (a Stata local macro such as `x') goes in a double-backtick span: `` reg y `controls' ``.

LEDGER=quality_reports/spec-ledger.md
[ -f "$LEDGER" ] || printf '%s\n' "# Specification ledger" "" \
  "Every specification estimated, kept or not, one row each, appended as it is run. Never edit a past row; a correction is a new row." "" \
  "| Date | Commit | Script:line | Outcome | Specification | Sample | Status | Why | Estimate |" \
  "|---|---|---|---|---|---|---|---|---|" > "$LEDGER"
if REV=$(git rev-parse --short HEAD 2>/dev/null); then       # -dirty = anything uncommitted outside quality_reports/, untracked files included
  [ -n "$(git status --porcelain -- . ':(exclude)quality_reports' 2>/dev/null)" ] && REV="$REV-dirty"
else
  REV="no-commit"
fi
printf '| %s | %s ' "$(date +%F)" "$REV" >> "$LEDGER"          # one printf + one heredoc per specification
cat >> "$LEDGER" <<'EOF'
| scripts/R/03_analyze.R:42 | log_wage | `feols(log_wage ~ treat + age \| id + year, cluster = ~id)` | panel 2010-2019 | kept | main specification | 0.082 (0.021) |
EOF