技能 数据科学 政策评估的因果识别方法

政策评估的因果识别方法

v20260724
aejpol-identification
这份指南提供了在实证政策论文中建立可信因果关系的严格方法论标准。它详细介绍了高级计量经济学技术,如双重差分(DiD)、工具变量(IV)、回归不连续设计(RDD)和随机对照试验(RCT),帮助研究人员超越简单的相关性分析,证明观察到的效应确实源于政策干预。
获取技能
77 次下载
概览

Identification — Credible Policy Evaluation (aejpol-identification)

When to trigger

  • The causal effect of a policy rests on OLS + controls, or TWFE on staggered policy adoption
  • A reform / threshold / experiment exists but the design's assumptions are not pinned down
  • A referee questions whether the estimated effect is really caused by the policy
  • You are unsure the design clears AEJ: Policy's credible-causal-evidence bar

The AEJ: Policy identification bar

AEJ: Policy is an empirical policy journal: the effect attributed to the policy must be credibly causal, the estimand must be the policy-relevant one, and the design must survive the obvious confound that the policy was not random. The policy variation is the research design — name it explicitly (a reform date, an eligibility cutoff, a formula kink, a randomized rollout) and defend the assumption that makes it causal. Report standard errors (no significance asterisks; see aejpol-tables-figures) and make the design reproducible for the AEA Data Editor.

Design paths

Path A: DID / event study (reforms, staggered policy adoption)

  • With staggered adoption move beyond TWFE (Callaway–Sant'Anna, Sun–Abraham, Borusyak–Jaravel–Spiess, de Chaisemartin–D'Haultfœuille); report a Goodman-Bacon decomposition to show the bias TWFE would induce.
  • Show a clean event study with pre-period leads flat around zero; do not assert parallel trends, demonstrate it (and probe with Rambachan–Roth honest-DID where pre-trends are imperfect).
  • Define the policy-relevant estimand (ATT on treated jurisdictions; weight by population/exposure if the policy lesson requires it).
  • Cluster at the policy-assignment level (often state/jurisdiction); address few-cluster issues (wild-cluster bootstrap).

Path B: IV / instrumented policy exposure

  • Strong first stage; with weak instruments use Anderson–Rubin / weak-IV-robust sets and report the effective F.
  • Defend the exclusion restriction in institutions and theory, not just statistically; argue the instrument affects outcomes only through the policy channel.
  • State the LATE complier population and whether it is the policy-relevant margin.

Path C: RDD / bunching (eligibility thresholds, tax/benefit schedules)

  • RDD: McCrary / Cattaneo–Jansson–Ma density test; data-driven bandwidth; covariate smoothness at the cutoff; bias-corrected robust CIs (rdrobust).
  • Bunching at kinks/notches in tax or benefit schedules: defend the counterfactual density and the structural elasticity it implies.
  • Be explicit that the estimate is local to the threshold and argue its policy relevance.

Path D: RCT / field experiment of a program

  • Pre-registration with a pre-analysis plan; report deviations. Detailed instructions / protocol included.
  • Randomization balance; attrition (Lee bounds if differential); multiple-hypothesis adjustment; explicit estimand and a take-up / intent-to-treat vs. treatment-on-treated distinction.
  • Tie the experimental effect to the cost of the program so a welfare reading is possible.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the identification claim, don't only argue it. Full map: execution-with-mcp. AEJ: Policy evaluates programs and reforms; the design must carry a policy-relevant magnitude, not just statistical significance.

  1. detect_designrecommend → fit with as_handle=trueaudit_result to list the checks the design still owes.
  2. Staggered DiD: callaway_santanna / sun_abraham + bacon_decomposition + honest_did_from_result (the pre-trend test is low-power, Roth 2022).
  3. IV: effective_f_test + an anderson_rubin_ci (valid under weak instruments), not a 2SLS t-stat alone.
  4. RDD: rdrobust (bias-corrected) + rddensity / mccrary_test for manipulation.
  5. OVB: oster_delta / sensemakr — how strong a confounder would have to be.

Report the economic magnitude; route the full battery to the appendix; keep every number reproducible. A run end-to-end (synthetic data, real returns) is in the JF execution walkthrough. If StatsPAI/Stata are not connected, adapt the vendored resources/code/ skeleton and flag any unverified number.

Checklist

  • The policy variation is named and the identifying assumption stated in one sentence
  • Design-appropriate diagnostics shown (pre-trends / density / first-stage F / balance)
  • Modern heterogeneity-robust estimator used where TWFE would bias
  • Estimand is the policy-relevant one (right population, right weighting)
  • Inference clustered at the assignment level; few-cluster handled
  • SEs reported (no asterisks); the causal claim never exceeds what the design supports

Anti-patterns

  • TWFE on staggered policy rollout with no heterogeneity-bias discussion
  • Asserting parallel trends instead of showing flat, precisely-estimated leads
  • An exclusion restriction defended only by a significant first stage
  • An RDD estimate generalized far from the cutoff without argument
  • An RCT with no pre-registration, no attrition analysis, or no link to program cost
  • Reporting significance with asterisks instead of standard errors

Referee pushback mapped to the fix

  • "Staggered TWFE here is biased." → Re-estimate with Callaway–Sant'Anna / Sun–Abraham; show flat leads + Bacon decomposition.
  • "Pre-trends look slightly off." → Honest-DID (Rambachan–Roth) bounds; show the conclusion survives plausible violations.
  • "This is just the effect at the threshold." → State the local estimand; argue why the threshold population is policy-relevant or extrapolate cautiously.

Output format

【Design】DID / IV / RDD-bunching / RCT
【Policy variation】the reform/cutoff/rollout that identifies the effect
【Identifying assumption】one sentence + how it is defended
【Diagnostics shown】[pre-trends / density / first-stage F / balance + attrition]
【Estimand】policy-relevant population + weighting; inference/clustering
【What it does NOT identify】[...]
【Next step】aejpol-theory-model (welfare mapping) or aejpol-robustness
信息
Category 数据科学
Name aejpol-identification
版本 v20260724
大小 6.58KB
更新时间 2026-07-28
语言