restud-identification
GitHub用于实证经济学论文（如REStud）的因果识别策略选择与压力测试。涵盖DID、IV、RDD、合成控制等设计，提供决策树及具体分支建议，确保识别策略严谨且符合顶刊标准。
Trigger Scenarios
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill restud-identification -g -y
SKILL.md
Frontmatter
{
"name": "restud-identification",
"description": "Use when selecting, implementing, or stress-testing the causal identification strategy for an empirical The Review of Economic Studies (REStud) manuscript — DID (incl. staggered), IV (incl. weak-IV-robust inference), RDD, synthetic control, shift-share, or RCT. Apply before writing results; for theory papers route to restud-theory-model."
}
REStud Identification (restud-identification)
When to trigger
- The empirical core is OLS + controls and a referee will flag selection-on-observables
- DID uses TWFE on staggered data with no modern estimator
- IV first-stage F is being cited as proof of instrument strength
- RDD uses a high-order polynomial or one untested bandwidth
- A prior rejection cited identification grounds and the design needs rebuilding
REStud identification bar
REStud was founded in 1933 to advance both theoretical and applied economics, and that balance is its signature: unlike a purely empirical top-5 outlet, REStud weighs design-based causal work and structural/theory-consistent identification on equal terms. So the bar is a state-of-the-art causal design, OR — for a structural/theoretical-empirical paper — a rigorous identification of the model's parameters with the identifying assumptions stated explicitly and mapped to data features. The empirical contribution can itself be a new identification strategy; in that case the design is the paper and must be defended to the hilt. Editorial assignment reflects the breadth: handling Joint Managing Editors span IO/applied econometrics (Jan De Loecker), micro theory / mechanism design (Antonio Penta), and behavioral/information economics (Jakub Steiner) — write the identification so the right one of them sees a rigorous argument. Identification is not a section to be footnoted; if the design is fragile, no prose, controls, or sample size rescues it.
Master decision tree
Is treatment plausibly random by design (RCT, lottery)?
├── Yes → analyze as an experiment; pre-register / PAP where applicable
└── No → identification comes from variation
├── Sharp threshold in a running variable → RDD (sharp / fuzzy)
├── Policy hits some units over time, not others → DID
│ ├── One treatment date → canonical 2x2 DID
│ └── Staggered adoption → Callaway-Sant'Anna / Borusyak-Jaravel-Spiess / Sun-Abraham
├── Endogenous regressor + plausibly exogenous shifter → IV
│ ├── Shifter x exposure shares → shift-share / Bartik
│ └── Single instrument, F < 50 → weak-IV-robust inference (AR)
├── One / few treated aggregate units → synthetic control
└── Structural parameter of interest → restud-theory-model (state identifying assumptions)
Branch A — Difference-in-Differences
- Staggered adoption: do not use TWFE. Use Callaway-Sant'Anna (
csdid/did), Borusyak-Jaravel-Spiess imputation, de Chaisemartin-D'Haultfœuille, or Sun-Abraham. - Report a Goodman-Bacon decomposition to expose the weight on forbidden comparisons under TWFE.
- Pre-trends: event-study plot and the formal joint test, not just the visual.
- Honest DiD (Rambachan-Roth 2023) sensitivity bounds for the post-period.
Branch B — Instrumental Variables
- The first-stage F > 10 rule is obsolete. For just-identified models report Anderson-Rubin confidence sets as primary inference; for F < 50, AR is required, not optional. Use the Olea-Pflueger effective F.
- Exclusion restriction is a story, not a test: state it in one sentence, defend with institutional narrative + a placebo where the instrument should have no effect + Conley et al. (2012) plausibly-exogenous sensitivity.
- Report the reduced form and argue the instrument's own exogeneity.
Branch C — Regression Discontinuity
- Local linear, triangular kernel; polynomials of order > 1 discouraged (Gelman-Imbens 2019).
- MSE-optimal bandwidth with robust bias-corrected CI (
rdrobust). - Diagnostics: McCrary / Cattaneo-Jansson-Ma density test; covariate balance at the cutoff; placebo cutoffs; estimate stable across ≥ 3 bandwidths; an
rdplotwith the binning stated.
Branch D — Synthetic control / shift-share
- SCM: in-time and in-space placebos; permutation / Fisher exact p-value; weight vector in the appendix. Consider generalized SCM (Xu 2017) or synthetic DiD (Arkhangelsky et al. 2021).
- Shift-share: declare the source of identification — exogenous shares (Goldsmith-Pinkham-Sorkin-Swift; report Rotemberg weights) or exogenous shocks (Borusyak-Hull-Jaravel; Adão-Kolesár-Morales). Do not hand-wave between them.
Branch E — Structural / theory-consistent
- State each identifying assumption explicitly and map it to a feature of the data.
- Provide identification arguments (which moments identify which parameters) before estimation.
- Route the model and proofs to
restud-theory-model; supply counterfactuals. - This is squarely REStud territory: the journal's canon includes structural identification of policy-relevant parameters — e.g., Mirrlees (1971), "An Exploration in the Theory of Optimum Income Taxation," REStud 38(2), where the structure is the identification. Hold your structural argument to that standard.
Execution bridge (StatsPAI / Stata MCP)
Estimate and audit the design, don't only describe it. Full map:
execution-with-mcp. REStud is top-5 general-interest economics; credible identification with modern estimators is the bar across applied fields.
detect_design→recommend→ fit withas_handle=true→audit_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_wolffor many-outcome control. - Sensitivity:
oster_delta/sensemakrfor 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
- Design-based strategy named; not "OLS with controls" as the headline
- Modern estimator used where the textbook default is biased (staggered DID, weak IV, naive RDD)
- Pre-trends / density / balance / weak-IV diagnostics reported as required by the branch
- Inference method correct (cluster-robust at the right level / AR / wild bootstrap / permutation)
- Identifying assumption stated in one sentence in the introduction
- Placebo / falsification test present
Anti-patterns
- TWFE on staggered data with no Goodman-Bacon decomposition
- First-stage F = 12 cited as evidence of instrument strength
- RDD with a 4th-order polynomial and one untested bandwidth
- IV exclusion restriction defended only by "we control for X"
- Reporting OLS-with-controls as the main spec and IV/RD as "robustness"
- Footnoting the identifying assumption instead of stating it up front
Output format
【STRATEGY】DiD | IV | RDD | SCM | shift-share | RCT | structural
【MODERN ESTIMATOR】yes / no / which
【DIAGNOSTICS REPORTED】[...]
【INFERENCE】cluster-robust / AR / wild bootstrap / permutation
【RED FLAGS】[... or "none"]
【NEXT SKILL】restud-robustness (or restud-theory-model if structural)
Version History
- 1839142 Current 2026-07-05 14:21
