SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillser-identification

er-identification

Design & creation

View repository

Updated 10-15 20:08

About this skill

Causal Identification Strategy (er-identification)

Companion code:resources/code/stata/03_did_modern.do（DID）、04_iv.do（IV）、05_rdd.do（RDD）、06_dml.do（DML）。

When to Use

  • The empirical analysis consists only of OLS + control variables.
  • DID uses TWFE but does not respond to recent criticisms of heterogeneous treatment effects.
  • IV first stage is weak / concerns about instrument endogeneity.
  • Planning to use double machine learning but unsure how to report it.

Design Priority

Preference ranking of Economic Research Journal editorial board members (strong → weak):

  1. Policy shock + DID (including staggered / continuous treatment)
  2. Regression discontinuity (clear policy threshold)
  3. Instrumental variables (strong instruments + exclusion argument)
  4. Propensity score matching + DID
  5. Synthetic control method
  6. Double machine learning / causal forest
  7. OLS + rigorous endogeneity discussion (acceptable in structural estimation / theoretical-empirical articles)

The journal explicitly opposes a "quantitative-only tendency": no matter how elegant the identification strategy, it must return to theory and China's institutional issues. Identification is a means, not a selling point.

Branch A: Staggered (multiple time point) DID — ★ most common and most easily criticized

Staggered DID cannot report only TWFE. The standard four-step process:

  1. TWFE baseline—the starting point familiar to readers (but may be biased under staggered treatment).
  2. Goodman-Bacon (2021) decomposition（bacondecomp)—show the "bad comparison / negative weight" problem.
  3. Heterogeneity-robust estimators as the main results(choose one from the table below as the main specification, with the others as robustness checks):
Estimator Paper Stata R
group-time ATT Callaway & Sant'Anna (2021) csdid did::att_gt
Interaction-weighted IW Sun & Abraham (2021) eventstudyinteract fixest::sunab
Imputation (most efficient) Borusyak, Jaravel & Spiess (2024) did_imputation didimputation
Two-stage Gardner (2022) did2s did2s
Non-binary/reversible treatment de Chaisemartin & D'Haultfœuille (2020/24) did_multiplegt_dyn DIDmultiplegtDYN 
* Callaway-Sant'Anna：gvar = 首次受处理年份，从不处理者 = 0
csdid Y X, ivar(id) time(year) gvar(gvar) method(dripw)
estat simple        // 总体 ATT
estat event         // 动态效应
  1. Event study plotTest parallel trends and dynamic effects (use the period before treatment as the baseline, 95% CI, vertical dashed line at the treatment time).
  2. Placebo: randomly assign treatment timing/units 500–1000 times, and see whether the true coefficient falls in the tail of the distribution (see er-robustness）。

Pitfalls to avoid: "only looking at the plot without testing" parallel trends; anticipation effects/early responses; using TWFE event studies as the main dynamic results under staggered treatment (they may also be biased; use CS / SA for the main results).

Branch B: IV — report modern weak instrument diagnostics

Do not report only "F>10". The standard report has four elements:

  1. First-stage Kleibergen-Paap rk Wald F(valid under heteroskedasticity/clustering, replacing Cragg-Donald F).
  2. Compare against Stock-Yogo (2005) critical values.
  3. Effective F（Montiel Olea & Pflueger 2013，weakivtest), a single endogenous variable is more reliable.
  4. Weak-instrument-robust inference: Anderson-Rubin test and confidence intervals (weakiv); under exact identification, AR is fully robust to weak instruments.
ivreg2 Y X (D = Z1 Z2), robust first   // 自动报告 KP rk F、Hansen J
weakivtest                              // 有效 F
weakiv ivreg2 Y X (D = Z1 Z2), robust   // AR / CLR / K 稳健区间

The exclusion argument should have at least three paragraphs: theory / institutions / placebo; and also report the reduced form. Benchmark against model papers(definitely in Economic Research Journal): Wang Yongqin and Dong Wen, "How Does the Rise of Robots Affect China's Labor Market?" (Issue 10, 2020)—constructs a Bartik / shift-share instrumental variable using the trend of robot penetration in U.S. industries, and argues for exogeneity and exclusion.

Branch C: RDD

rdplot     Y X, c(0) p(1) kernel(triangular)
rdrobust   Y X, c(0) p(1) kernel(triangular) bwselect(mserd)  // 报告 Robust 行
rddensity  X, c(0)                                             // CJM 操纵检验
  • The main estimate reports robust bias-corrected CI(the core contribution of Calonico-Cattaneo-Titiunik 2014), do not report only conventional CIs.
  • For the manipulation test use rddensity(Cattaneo-Jansson-Ma, which has superseded McCrary 2008).
  • Robustness: covariate continuity, ≥3 bandwidths, donut RD, placebo cutoffs.
  • Benchmarking against model papers(confirmed to be in Economic Research Journal): Liu Shenglong, Zhou Shaojie, Hu Angang, "The Compulsory Education Law and Returns to Education in Urban China: Based on a Regression Discontinuity Design" (2016, Issue 2).

Branch D: DML double machine learning

Report: model type (partially linear PLR / interactive), number of cross-fitting folds K (5 or 10), nuisance function learners (lasso / random forest / gradient boosting) + learner robustness comparison, Neyman orthogonal score standard errors (Chernozhukov et al. 2018). Stata ddml+pystacked；Python DoubleML。

Branch E: structural estimation / theoretical empirics

Are the microfoundations clear? Are the identifying assumptions explicitly listed? Do the parameter estimates provide counterfactual analysis?

Standard errors

Scenario Approach Command
Within-group correlation Cluster-robust reghdfe ..., vce(cluster id)
Two-dimensional correlation Two-way clustering vce(cluster firm year)
Too few clusters (<~40) wild cluster bootstrap boottest(see er-robustness)
Spatial correlation Conley spatial HAC acreg

Execution bridge (StatsPAI / Stata MCP)

Actually run the designand audit it, rather than just describing it. For the complete mapping see execution-with-mcp. Economic Research Journal is a top Chinese economics journal, and identification credibility is usually the binding constraint; staggered DID, weak-instrument-robust IV, RDD, and mechanism tests.

  • detect_design → recommend → use as_handle=true to fit → audit_result to list the checks still owed.
  • **Observational causality:** 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 inference + romano_wolf for multiple-outcome family-wise error rate control.
  • Sensitivity:oster_delta / sensemakr。

Report in the main texteconomic magnitude, with the full battery in the appendix; every number must be reproducible. For an end-to-end actually-run example see JF execution walkthrough. If StatsPAI/Stata is not connected, switch to resources/code/ and flag unverified numbers.

Must-check list

  • Staggered DID: Bacon decomposition + at least one heterogeneity-robust estimator + event study plot
  • Parallel trends / smoothness / weak instrument tests have all been done (not just looking at plots)
  • Placebo tests (randomize treatment timing / randomize treated units)
  • IV: KP rk F + effective F + AR robust confidence intervals + reduced form + three-part argument for exclusion
  • RDD: report Robust row CI + rddensity manipulation test + multiple bandwidths
  • Standard error clustering level is reasonable; wild bootstrap when few clusters
  • Responded to the "treated units' anticipation / early response" problem

Anti-patterns

  • TWFE + staggered but no discussion of heterogeneous treatment bias
  • IV uses "exogenous event × previous-period endogenous variable"—reviewers will ask "why does the previous period not affect the current period"
  • "We believe the policy is exogenous to firm decisions" but no evidence is given
  • RDD uses a truncated bandwidth but does not report bandwidth sensitivity
  • Treating the identification strategy as the selling point, divorced from theory and China's institutional issues

Output format

【识别策略】交叠DID / IV / RDD / DML / 结构估计 / 其他
【交叠DID 主估计量】TWFE only（需升级）/ CS / SA / BJS / dCDH
【已完成检验】[Bacon分解, 平行趋势, 安慰剂, KP F, 有效F, AR, rddensity, ...]
【缺失检验】[...]
【聚类层次】...
【下一步】er-mechanism