brycewang-stanford Identification Strategy for JAE (jape-identification-strategy)
When to trigger
- Choosing or defending the empirical design for a JAE paper
- A referee questions whether the estimand is identified on your real data
- Mapping each identifying assumption to a test you can deposit in the archive
JAE is applied: identification on real data
JAE publishes applications on real data, so identification is a credible empirical strategy, not a theorem. The bar: a clearly stated estimand, explicit identifying assumptions, and diagnostics that test them, all reproducible from deposited code/data. Separate the econometric object from the substantive claim, and justify causal language with the design, not estimator branding.
Common designs and load-bearing assumptions
- Time-series / macro-econometrics (common at JAE): unit-root and cointegration handling; lag selection; shock identification (recursive, sign, external instruments); HAC inference. Show robustness to lag length and identification scheme.
- Panel / dynamic panel: strict vs. sequential exogeneity; GMM instrument validity (Hansen/Sargan, AR(2)); avoid instrument proliferation.
- IV / GMM: relevance (first-stage / effective F) and exogeneity; weak-IV-robust inference; over-identification tests.
- Quasi-experimental (DID, event study, RDD): parallel trends / continuity; modern estimators under staggered timing or heterogeneity; placebo and pre-trend tests.
- Forecasting / structural: define the loss function, information set, and stability assumptions.
Reproducibility is part of identification
Every check must be regeneratable from the programs you will deposit in the JAE Data Archive. With confidential data, the readme must describe the source and extraction well enough for others to apply for access and re-run it. A diagnostic you cannot reproduce is not a defense.
Diagnostic-to-archive map by design
For each design, JAE referees expect a named diagnostic and the script that produces it in the eventual deposit:
| Design | Load-bearing diagnostic | Archive artifact |
|---|---|---|
| VAR / shock identification | Robustness across identification schemes; lag-order sensitivity | var_ident.do + exported IRF CSVs |
| Dynamic panel GMM | Hansen J, AR(2), instrument count vs. N | gmm_diag.do + instrument-matrix log |
| IV / 2SLS | Effective first-stage F; Anderson–Rubin CI; overid test | iv_firststage.R + AR-interval table |
| Staggered DID / event study | Pre-trend test; heterogeneity-robust estimator comparison | did_pretrends.R + event-study CSV |
| RDD | Density (manipulation) test; bandwidth sensitivity curve | rd_density.R + bandwidth grid output |
| Forecasting | Out-of-sample loss comparison; stability over subsamples | oos_eval.R + rolling-window results |
A blank archive-artifact cell means the defense exists only as prose — at this journal that counts as no defense.
Worked example: an external-instrument pass-through IV (illustrative)
Estimand: the 12-month price response to a 1% exchange-rate movement. Instrument: foreign monetary-policy surprises. First pass: effective F ≈ 8 — below comfort. The JAE-grade response is not to bury it: report the 2SLS estimate 0.31 (s.e. 0.09) next to an Anderson–Rubin 95% interval [0.07, 0.61], show OLS (0.18, s.e. 0.04) for the endogeneity direction, and add a stronger-instrument subsample where F ≈ 19 and the AR interval tightens. Every column maps to one script in the deposit; the readme names which table each program rebuilds.
Where JAE referees push back on identification
- "Parallel trends asserted, not shown" → add the pre-trend event study and a heterogeneity-robust estimator; archive both.
- "Instrument count grows with T" → collapse the GMM instrument matrix, re-report Hansen J, and state the count in the table.
- "Identification scheme drives the IRFs" → show recursive vs. sign vs. external-instrument results side by side in the appendix.
- "This diagnostic is not in the package" → treat as a code bug: add the script, rerun, update the exhibit.
Identification appendix block
Estimand: [population object, one sentence]
Assumption A1: [statement] → Test: [diagnostic] → Script: [file] → Result: [pass/fail + number]
Assumption A2: ...
Confidential-data note: [access path readers can follow, if applicable]
Execution bridge (StatsPAI / Stata MCP)
Estimate and audit the identification claim, don't only argue it. Full map:
execution-with-mcp. Applied econometrics: the estimator and its diagnostics are themselves the contribution, so foreground the weak-IV / pre-trend / sensitivity tooling.
detect_design→recommend→ fit withas_handle=true→audit_resultto list the checks the design still owes.- Staggered DiD:
callaway_santanna/sun_abraham+bacon_decomposition+honest_did_from_result(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/sensemakr— how strong a confounder would have to be.
Report the economic magnitude; route the full battery to the appendix; keep every
number reproducible. A run end-to-end (synthetic data, real returns) is in the
JF execution walkthrough. If StatsPAI/Stata are not connected, adapt the
vendored resources/code/ skeleton and flag any unverified number.
Output format
【Estimand】one sentence
【Design】time-series / panel / IV-GMM / quasi-experiment / forecasting
【Assumptions】each has a test? [Y/N]
【Inference】matched to design? [Y/N]
【Reproducible】every diagnostic regeneratable? [Y/N]
【Map】each assumption → test → script → exhibit complete? [Y/N]
Supplementary resources
../../resources/external_tools.md— estimators and inference packages../../resources/official-source-map.md— Data Archive reproducibility sources