技能 数据科学 学术因果识别增强工具

学术因果识别增强工具

v20260724
eer-identification
本技能旨在帮助用户增强实证经济学论文的因果识别策略。它指导用户从方法论层面(如双重差分、工具变量、局部回归和实验设计)系统性地审查和改进研究设计,确保所有关键假设、潜在偏差和因果机制得到明确论证和严格检验,以满足高水平学者的学术标准。
获取技能
138 次下载
概览

Identification Strategy (eer-identification)

When to trigger

  • A causal claim rests on OLS + controls, or TWFE on staggered timing
  • An IV's exclusion restriction or first-stage strength is contested
  • An RDD's continuity/manipulation assumptions are unexamined
  • An experiment's estimand, balance, or pre-registration is unclear
  • You are unsure the design clears EER's credibility bar for a general-interest readership

The EER identification bar

EER publishes broadly across empirical economics, so identification is judged on credibility legible to a general reader: the mapping from variation in the data to the causal object must be explicit, the key assumption stated, and the most obvious threat pre-empted. Because review is single-anonymized, the referee is often a methods expert in your exact design — modern, design-appropriate estimators and honest inference are expected. Report standard errors and confidence intervals (EER house style; do not lean on significance stars — see eer-tables-figures). Match the size of the causal claim to what the design supports.

Branch paths

Branch A: DiD / event study

  • With staggered adoption, move beyond TWFE (Callaway–Sant'Anna, Sun–Abraham, de Chaisemartin–D'Haultfœuille); a TWFE coefficient on staggered timing must be defended against heterogeneity bias.
  • Show a clean event-study with pre-treatment leads (flat, precisely estimated) and dynamic post effects.
  • Report a Goodman-Bacon decomposition when using two-way fixed effects.
  • State the parallel-trends assumption and a pre-trends / sensitivity argument (e.g., Rambachan–Roth honest DiD).

Branch B: IV

  • Strong first stage (report the first-stage F / effective F); with weak instruments use Anderson–Rubin / weak-IV-robust sets.
  • Defend the exclusion restriction in theory, institutions, and a falsification/placebo test.
  • Be explicit about the LATE / complier interpretation; do not generalize beyond it.

Branch C: RDD

  • Density/manipulation test (McCrary or Cattaneo–Jansson–Ma); covariate smoothness at the cutoff.
  • Optimal bandwidth + bias-corrected CIs (Calonico–Cattaneo–Titiunik); show sensitivity to bandwidth.
  • State the local nature of the estimate.

Branch D: Experiment / behavioral

  • Pre-registration where applicable; report deviations; include instructions / survey transcripts.
  • Randomization balance; attrition (Lee bounds if differential); multiple-hypothesis adjustment.
  • State the estimand and external-validity scope.

Clustering at the level of treatment assignment; with few clusters use wild-cluster bootstrap. Pair this skill with eer-robustness for the specification/sample battery.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. EER is a general economics field journal; the DiD/IV/RDD chain serves its applied lane.

  • detect_designrecommend → fit with as_handle=trueaudit_result.
  • Observational causal claims: staggered DiD (callaway_santanna / sun_abraham + bacon_decomposition + honest_did_from_result); IV (effective_f_test + anderson_rubin_ci); RDD (rdrobust + mccrary_test).
  • Experiments: randomization-based inference + romano_wolf for many-outcome control.
  • Sensitivity: oster_delta / sensemakr for observational claims.

Report the magnitude in interpretable units; route the full battery to the appendix. A run end-to-end (synthetic data, real returns) is in the JF execution walkthrough.

Checklist

  • Branch chosen; the variation-to-causal-object mapping stated in one sentence
  • DiD: heterogeneity-robust estimator where TWFE would bias; flat pre-trends shown
  • IV: first-stage strength reported; exclusion defended + falsification; LATE stated
  • RDD: density test + bias-corrected CI + bandwidth sensitivity
  • Experiment: pre-registered (if applicable); balance/attrition/MHT handled; estimand stated
  • Inference: SEs/CIs reported, clustering at assignment level, few-cluster fix if needed
  • Causal claim never exceeds what the design supports

Anti-patterns

  • TWFE on staggered treatment with no heterogeneity-bias discussion
  • An IV with an asserted-but-undefended exclusion restriction
  • RDD with no manipulation test and a single hand-picked bandwidth
  • An experiment with no pre-registration mention and no estimand
  • Reporting significance with asterisks instead of SEs/CIs (against EER house style)
  • Generalizing a LATE or a local RDD effect to a population it does not identify

Worked vignette (illustrative)

A migration paper uses a staggered visa-liberalization rollout. A weak version runs TWFE and reports a wage effect with stars. An EER version re-estimates with Callaway–Sant'Anna, shows flat leads and a dynamic post path, reports the effect as -1.4% local wages (s.e. 0.5, illustrative), runs Rambachan–Roth sensitivity, and states the estimand is the effect on incumbents in receiving regions — not a national average. The general-interest lesson (how labor supply shocks transmit to local wages) is named so a non-migration economist sees the point.

Output format

【Branch】DiD / IV / RDD / experiment
【Variation→object mapping】one sentence
【Key assumption】stated + the main threat pre-empted
【Design evidence】[pre-trends / first-stage F / density test / balance]
【Inference】SEs/CIs; clustering level; few-cluster fix?
【What it does NOT identify】[...]
【Next step】eer-theory-model (if a mechanism is needed) or eer-robustness
信息
Category 数据科学
Name eer-identification
版本 v20260724
大小 6.07KB
更新时间 2026-07-28
语言