Use when the causal identification argument is the bottleneck for an American Economic Journal: Applied Economics (AEJ: Applied) manuscript — RCT, difference-in-differences/event study, regression discontinuity, IV, or shift-share. Stress-tests the data-to-causal-estimate mapping to the AEJ: Applied credibility bar before exhibits are finalized; it does not write the prose or build the package.
SkillScore73Maintained
Automatic
◆Research pack
Last verified 8/15/2026
What it changes
Passed 16 of 22 test cases with the skill, and 3 of 22 without it.
Without skill14%
With skill73%
Checks it passed that the model alone missed
The next step section contains an aeja downstream routing identifier.
The variation-to-estimand mapping section contains exactly one sentence.
The estimator + inference section recommends Callaway-Sant'Anna estimators.
and 10 more in the full report
15 fixed1 passed either way4 still failing2 broke
Pass-rate lift
+59 pp
Tokens
+27%
Agent turns
0%
Response time
-28%
Model gemini-3.6-flashMeasured 8/15/2026v116 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 15 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 15
Prompt
We conducted a randomized trial stratified by school district. We plan to estimate the treatment effect using a simple difference in means without including school district fixed effects or controlling for stratification variables. Audit our framework using the required brackets:
【Design】
【Variation-to-estimand mapping】
【Estimand】
【Identification evidence】
【Estimator + inference】
【What it does NOT identify】
【Next step】
What was checked, without and with the skill
Checked
Without skill
With skill
The variation-to-estimand mapping section contains exactly one sentence.
Missed
Passed
The estimator + inference section specifies accounting for stratification blocks.
Missed
Passed
The identification evidence section includes reporting baseline covariate balance.
Missed
Passed
case-17 · gemini-3.6-flash · 8/15/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md88 linessingle file
Identification Strategy (aeja-identification)
When to trigger
A causal claim rests on OLS + controls, or TWFE on staggered timing
An RCT's estimand, balance, or attrition handling is not pinned down
An RD's density, bandwidth, or covariate-smoothness defense is missing
An IV's first stage is weak or the exclusion restriction is asserted, not argued
You are unsure the design clears AEJ: Applied's credibility bar
The AEJ: Applied identification bar
AEJ: Applied is identification-driven applied micro: the mapping from a source of variation to the causal estimand must be explicit, defended, and falsifiable. Editors and referees here are unusually sophisticated about modern design pitfalls — staggered-DID bias, weak IV, RD manipulation, shift-share exogeneity. State the estimand, name the identifying assumption, show the diagnostic that could have failed but didn't, and keep the claim inside what the design supports. Inference must match the design (clustering at the assignment level; few-cluster corrections).
Design paths
Path A: RCT / field experiment (own data)
Estimand stated (ITT vs. LATE/TOT); randomization unit and stratification described.
Pre-registration (AEA RCT Registry / AsPredicted / OSF); report deviations from the pre-analysis plan.
Balance table on baseline covariates; attrition examined and bounded (Lee bounds if differential).
Multiple-hypothesis adjustment across outcomes/subgroups; explicit external-validity discussion.
Path B: Difference-in-differences / event study
With staggered adoption, move beyond TWFE — Callaway–Sant'Anna, Sun–Abraham, de Chaisemartin–D'Haultfœuille, or Borusyak–Jaravel–Spiess imputation.
Clean event-study with leads for pre-trends; report a Goodman-Bacon decomposition to show which 2×2s drive the estimate.
State and defend parallel trends; consider Rambachan–Roth honest-DID sensitivity to parallel-trend violations.
Path C: Regression discontinuity
Density test (McCrary / Cattaneo–Jansson–Ma) for manipulation; covariate smoothness at the cutoff.
Local-linear with data-driven bandwidth; bias-corrected, robust CIs (rdrobust); donut and bandwidth-sensitivity checks.
State whether the estimand is the local effect at the cutoff and resist extrapolation.
Path D: IV / shift-share
Strong first stage (effective F / Montiel-Olea–Pflueger); with weak instruments use Anderson–Rubin / weak-IV-robust sets.
Exclusion restriction argued from institutions/theory + falsification (reduced-form on never-takers, placebo outcomes).
Shift-share: defend exogeneity of shares or shocks (Goldsmith-Pinkham–Sorkin–Swift / Borusyak–Hull–Jaravel) and report the implied just-identified weights.
Execution bridge (StatsPAI / Stata MCP)
Estimate and audit the identification claim, don't only argue it. Full map:
Use when selecting, implementing, or stress-testing the causal identification strategy for an empirical economics manuscript — difference-in-differences (including staggered designs), instrumental variables (including weak-IV-robust inference), regression discontinuity, synthetic control, or shift-share / Bartik. Apply before writing the introduction or results.