技能 数据科学 政策估计稳健性分析

政策估计稳健性分析

v20260724
aejpol-robustness
本指南指导用户为经济政策论文构建严谨的稳健性检验程序。核心原则是捍卫政策结论的稳定性,而非仅报告系数的显著性。内容详细介绍了如何应对各种计量挑战(如模型设定、预趋势、多重检验和可忽略变量选择偏差),确保政策结论在各种压力下依然可信。
获取技能
353 次下载
概览

Robustness — Defending the Policy Estimate (aejpol-robustness)

When to trigger

  • The headline causal estimate moves across specifications, or you do not yet know if it does
  • A referee will ask "is this robust?" and you have no organized answer
  • Inference (clustering, few clusters, multiple outcomes) is not yet airtight
  • You need to show the policy conclusion, not just a coefficient, survives stress

Principle: robustness defends the policy conclusion, not the coefficient

At AEJ: Policy, robustness is judged by whether the policy takeaway is stable — if the headline estimate is the cost-per-job or the MVPF, show that number is stable, with its uncertainty, not merely that a regression coefficient stays significant. Organize the robustness program around the threats that would change the policy conclusion, and report enough that a skeptical referee can see each threat addressed.

Robustness by threat (each maps to a concrete check)

Threat to the policy conclusion Check
Functional form / controls drive the result Specification ladder; show the estimate across a coherent set, not a single lucky spec
Pre-trends / parallel-trends violation Honest-DID (Rambachan–Roth) sensitivity bounds; placebo pre-period "effects"
Estimator bias under staggered timing Re-estimate with ≥1 heterogeneity-robust DID estimator (CS / SA / BJS / dCDH)
Bandwidth / kernel (RDD) Bandwidth sweep + bias-corrected CIs; donut-RDD if heaping at the cutoff
Weak / invalid instrument Effective F; AR-robust CI; over-ID test if available
Wrong inference / few clusters Wild-cluster bootstrap; report clustering level sensitivity
Multiple outcomes / specifications Romano–Wolf / sharpened q-values; a specification curve where many specs are run
Confounding by an omitted policy/shock Controls for co-timed policies; event-study around the focal reform only
Selection on unobservables Oster (2019) δ / bounds; argue the implied selection is implausible
Sample composition / outliers Drop influential jurisdictions; winsorize; alternative sample windows

Sensitivity that is policy-specific

  • If the policy lesson depends on a welfare parameter you calibrate (discount rate, value of a statistic, recycling rule), report the lesson across a plausible range of that parameter, not one value.
  • If external validity is the policy worry, show heterogeneity by jurisdiction characteristics and discuss which settings the estimate travels to.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. AEJ: Policy evaluates programs and reforms; the design must carry a policy-relevant magnitude, not just statistical significance.

  • Many outcomes / specifications: romano_wolf (step-down FWER, accounts for cross-test correlation) or benjamini_hochberg — report the adjusted threshold.
  • OVB sensitivity: oster_delta / sensemakr — the confounder strength that would overturn the headline.
  • Inference: wild_cluster_bootstrap (few clusters), twoway_cluster / conley.
  • Re-fit off one handle: audit_result(result_id) lists the missing checks and the exact suggest_function for each — no guessing the battery.
  • Exhibits: etable / did_summary_to_latex from the handle — no retyped numbers.

Keep the decisive checks in the body and the exhaustive (now actually-run) battery in the appendix. See the executed chain in the JF execution walkthrough.

Checklist

  • The headline policy number (not just a coefficient) is shown stable across specs
  • The single most likely referee threat is pre-empted with a dedicated exhibit
  • At least one heterogeneity-robust estimator shown where staggered timing applies
  • Inference stress-tested (wild-cluster / AR / multiple-testing as relevant)
  • Selection-on-unobservables addressed (Oster bounds or equivalent)
  • Calibrated welfare parameters varied across a defended range
  • No "kitchen-sink" robustness with no narrative — each check answers a named threat

Anti-patterns

  • A robustness section that is a wall of tables with no statement of which threat each rebuts
  • Showing the coefficient is stable while the welfare/policy number is never re-derived
  • A specification curve run but only the favorable region discussed
  • Treating "still significant" as robustness while ignoring magnitude stability
  • Calibrating one welfare parameter value and never probing it

Sequencing the robustness section for a referee

Order the section so a referee meets the answer before the doubt: (1) the main heterogeneity-robust estimate and its event-study; (2) the single most likely fatal threat with its dedicated check; (3) the inference stress-tests; (4) a compact specification curve or table of remaining variants; (5) the calibrated-parameter sensitivity for the welfare number. Each subsection ends with one sentence stating that the policy conclusion is unchanged, with its band — not merely that the coefficient stays signed.

Worked vignette (illustrative)

A staggered-DID estimate of a minimum-wage change on employment is the basis for a "small disemployment cost" policy claim. A referee will doubt staggered TWFE and pre-trends. The robustness program: CS and SA estimators (estimate within 10% of TWFE, illustrative), flat pre-period leads, an honest-DID bound showing the sign survives a pre-trend twice the largest observed lead, and wild-cluster inference across 30 states. The policy claim — disemployment cost per dollar of raised earnings — is re-derived under each and reported with its band.

Output format

【Headline policy number】the quantity whose stability you defend
【Top 3 threats】ranked by how badly each would change the conclusion
【Checks per threat】[threat → check → result]
【Inference】clustering / few-cluster / multiple-testing handling
【Calibrated-parameter sensitivity】range probed + conclusion stability
【Next step】aejpol-tables-figures
信息
Category 数据科学
Name aejpol-robustness
版本 v20260724
大小 6.5KB
更新时间 2026-07-28
语言