qje-identification
GitHub针对QJE论文因果识别策略进行压力测试，涵盖RCT、DID、IV、RDD等方法。在制表前审查设计合理性，确保满足QJE对可信度与重大问题的要求，提供具体改进路径以避免被拒稿。
Trigger Scenarios
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill qje-identification -g -y
SKILL.md
Frontmatter
{
"name": "qje-identification",
"description": "Use when the causal identification strategy is the bottleneck for a Quarterly Journal of Economics (QJE) manuscript — RCT, staggered DID, IV, RDD, event study. Stress-tests the design to the QJE general-interest bar before tables are drafted."
}
Causal Identification (qje-identification)
When to trigger
- The empirical core is OLS + controls with an undefended causal claim
- A DID uses two-way fixed effects (TWFE) on staggered timing without modern estimators
- An IV's first stage is weak or the exclusion restriction is unargued
- An RDD lacks a McCrary density test or bandwidth-robustness
- You are unsure your design clears the QJE causal bar
The QJE identification bar
QJE rewards credible identification serving a big question, not technique for its own sake. Editorial reality shapes the bar: five Harvard-based Editors (Barro, Katz, Nunn, Shleifer, Stantcheva as of 2024) desk-screen in roughly two weeks, so the identifying variation must be legible and convincing on a first read. QJE's own canon mixes clean experiments and quasi-experiments with reduced-form ambition — e.g., Chetty, Hendren, Kline & Saez, "Where is the Land of Opportunity?" (QJE 2014), and Chetty et al. on Project STAR kindergarten effects (QJE 2011). The credibility ranking editors and referees implicitly apply (strong → weaker):
- RCT / field experiment with pre-registration and balance
- Sharp/fuzzy RDD at a clean policy threshold
- DID / event study off a credibly exogenous policy shock (with modern estimators)
- IV with a strong first stage and a defended exclusion restriction
- Matching / selection-on-observables — acceptable only as a complement, rarely as the spine
Note the QJE-specific twist: a novel, hard-to-assemble dataset answering a first-order question can carry a paper even when the design is reduced-form, because the journal prizes the question and the lesson (Akerlof's "Market for 'Lemons'", QJE 1970, is the patron saint of idea-first papers). It will not save a narrow within-field estimate.
Branch paths
Branch A: DID / event study
- Staggered adoption? You must move beyond TWFE. Use Callaway–Sant'Anna, Sun–Abraham, or de Chaisemartin–D'Haultfœuille; report a Goodman-Bacon decomposition to expose "forbidden" comparisons.
- Pre-trends: show a clean event-study plot with leads; pre-period coefficients near zero.
- Inference: cluster at the level of treatment assignment; consider wild-cluster bootstrap with few clusters.
- Placebo / permutation tests on treatment timing and on outcomes that should not move.
Branch B: IV
- First-stage F should be strong; with weak instruments use Anderson–Rubin or other weak-IV-robust confidence sets, not just t-stats.
- Exclusion restriction argued in three registers: economic theory, institutional detail, and falsification (effect absent where the channel is absent).
- Report the reduced form and the OLS alongside IV; discuss the LATE/compliers interpretation.
- Defend the instrument's own exogeneity — not just relevance.
Branch C: RDD
- McCrary / Cattaneo–Jansson–Ma density test for manipulation at the cutoff.
- Optimal bandwidth (Calonico–Cattaneo–Titiunik) plus at least three bandwidth-robustness checks; report bias-corrected CIs.
- Covariate smoothness / balance at the threshold; placebo cutoffs.
- For fuzzy RDD, report the first stage in the discontinuity.
Branch D: RCT / field experiment
- Pre-registration / pre-analysis plan referenced; report deviations.
- Randomization balance table; attrition analysis (Lee bounds if differential).
- Multiple-hypothesis adjustment across outcomes/subgroups.
- External validity discussed: what does the experimental population teach beyond itself?
Branch E: novel descriptive / measurement paper
- Is the data genuinely new and hard to assemble? (At QJE this is a real path, but only with first-order facts.)
- Are the new facts disciplined against measurement error and alternative explanations?
- Is there a clear conceptual lesson the facts deliver to all of economics, not one subfield?
Execution bridge (StatsPAI / Stata MCP)
A QJE identification claim should be estimated and audited, not just argued. Full
map: shared-resources/empirical-methods/execution-with-mcp.md. QJE applied-micro instantiation:
detect_design→recommend→ fit withas_handle=true, thenaudit_resultto enumerate the checks the design still owes.- Staggered DiD:
callaway_santanna/sun_abraham(never bare TWFE); show thebacon_decompositionweight you are correcting;honest_did_from_result(Rambachan–Roth) because the pre-trend test is low-power (Roth 2022). - IV:
effective_f_test+ ananderson_rubin_ci(valid under weak instruments), not a 2SLS t-stat alone. - RDD:
rdrobust(bias-corrected) +rddensity/mccrary_testfor manipulation. - OVB:
oster_delta/sensemakrto state how strong a confounder must be.
Report the economic magnitude the QJE body expects; route the full diagnostic
battery to the appendix, and keep every number reproducible for qje-replication-package.
If StatsPAI/Stata are not connected, adapt the vendored resources/code/ skeleton and
flag any unverified number. A run end-to-end (synthetic data, real returns) is in the
JF execution walkthrough.
Checklist
- Identifying variation named in one sentence and defended as exogenous
- Design-appropriate diagnostics done (pre-trends / density / first-stage / balance)
- Modern estimator used where TWFE would be biased (staggered DID)
- Inference matched to assignment level; few-cluster issues addressed
- Placebo / permutation / falsification tests reported
- LATE / external-validity interpretation stated explicitly
- The claim never exceeds what the design supports
Anti-patterns
- TWFE on staggered treatment with no discussion of heterogeneity bias
- IV that is "exogenous shock × lagged endogenous variable" with no exclusion argument
- "We argue treatment is as good as random" with no falsification evidence
- RDD reporting one bandwidth and hiding sensitivity
- A clean local design oversold as a global structural parameter with no broad lesson
Output format
【Design】RCT / RDD / DID / IV / descriptive / other
【Identifying variation】one sentence
【Diagnostics done】[pre-trends, density, first-stage F, balance, ...]
【Diagnostics missing】[...]
【Inference】clustering level + few-cluster handling
【Interpretation】LATE / ATT / external validity note
【Next step】qje-theory-model
Version History
- 1839142 Current 2026-07-05 14:17
