econometrics-check

name	econometrics-check
description	Causal inference design audit. Dispatches the Econometrician agent in standalone mode for 4-phase review (claim, design validity, inference, polish). Covers DiD, IV, RDD, Synthetic Control, and Event Studies. Use when working on empirical papers, strategy memos, or R scripts with causal estimators.
disable-model-invocation	true
argument-hint	[paper .tex, strategy memo, R script, or directory path]
allowed-tools	["Read","Grep","Glob","Write","Task"]

Econometrics Check

Run a 4-phase causal inference audit on the target file(s) by dispatching the Econometrician agent in standalone mode.

Workflow

Step 1: Parse Input

Determine target from $ARGUMENTS:

• .tex file: Review paper for identification claims, assumption statements, estimation descriptions
• strategy-memo-*.md: Review strategy memo BEFORE code is written (early design check)
• .R / .do / .py / .jl file: Review script for code-theory alignment, package usage, SE computation
• Directory (e.g., code/analysis/): Review all scripts, then synthesize
• No argument: Review output/paper/main.tex if it exists, else ask user

Step 2: Context Gathering

Before launching the reviewer:

1. Read the target file(s)
2. Read Bibliography_base.bib to check citation availability
3. Read .claude/rules/domain-profile.md for field-specific conventions
4. If reviewing scripts: also read the paper (if available) for code-theory alignment
5. If a strategy memo exists in quality_reports/: read it for design intent

Step 3: Launch Econometrician Agent

Delegate to the econometrician agent via Task tool:

Prompt: Review [file] through all 4 phases of the econometrics review protocol.
Mode: Standalone (not within pipeline — skip orchestrator routing).
Focus on: [identified causal design if known, otherwise "identify the design first"].
Context: [brief summary of what the paper/script does, from Step 2].
Save report to: quality_reports/[FILENAME_WITHOUT_EXT]_econometrics_review.md

The agent runs 4 sequential phases:

1. Claim identification — what design, estimand, treatment, control
2. Core design validity — design-specific assumptions + sanity check (sign, magnitude, dynamics)
3. Inference — SE clustering, multiple testing, code-theory alignment
4. Polish & completeness — robustness, sensitivity, citation fidelity

Early stopping: If Phase 2 finds CRITICAL issues, the agent focuses there.

Step 4: Present Summary

Present to the user:

1. Design(s) identified (DiD staggered, IV, RDD, etc.)
2. Overall assessment (SOUND / MINOR ISSUES / MAJOR ISSUES / CRITICAL ERRORS)
3. Blocking issues (CRITICAL severity — must fix)
4. Priority action list (top 3-5 fixes, ordered by importance)
5. Positive findings (what the analysis does well)

Step 5: Score

Report the Econometrician's score. In standalone mode, this is the component score (not the weighted aggregate). Reference scoring-protocol.md for how this feeds into the overall score when run within the pipeline.

---

Principles

• Design-opinionated, package-flexible. Recommend standard packages but accept and validate alternatives.
• Respect the researcher. If this is the researcher's own methodological contribution, focus on implementation, not lectures.
• Actionable output. Every issue must have a concrete fix.
• Proportional. Not every paper needs every robustness check. Flag missing checks but note when omissions are reasonable.
• Sequential phases. Never skip to robustness before verifying the core design holds.