Identification — Clean Enough to Defend Short (aeri-identification)

When to trigger

• The single headline result rests on OLS + controls, or TWFE on staggered timing
• A structural parameter is estimated but it is unclear what in the data identifies it
• An experiment's estimand or assumptions are not pinned down
• You are unsure the identification is clean enough to carry a short paper

The AER: Insights identification bar

A short paper has no room to rescue a weak design with pages of robustness. The identification must be clean, transparent, and self-contained — the central exhibit and one or two sentences should make a non-specialist believe the headline number. Because AER: Insights papers are one insight at AER-level importance, the design is held to AER credibility but expressed with extreme economy: state the data-to-object mapping in one sentence, show the single most convincing diagnostic in-text, and move the rest to the Supplemental Appendix. AEA house style: report standard errors / confidence sets, not significance asterisks, and make everything reproducible for the AEA Data Editor.

Branch paths

Branch A: Empirical causal design (the most common AER: Insights paper)

• DiD / event study: with staggered adoption move beyond TWFE (Callaway–Sant'Anna, Sun–Abraham, de Chaisemartin–D'Haultfœuille); the single event-study figure with clean leads is often the paper's central exhibit.
• RDD: density test (Cattaneo–Jansson–Ma), optimal bandwidth, covariate smoothness, bias-corrected CIs; one well-made RD plot can be the whole identification.
• IV: strong first stage; weak-IV-robust inference (Anderson–Rubin) if needed; defend the exclusion restriction in one tight paragraph.
• Inference clustered at the assignment level; few-cluster fixes (wild-cluster bootstrap).

Branch B: Experiment (own data)

• Pre-registration in a recognized registry; report deviations.
• Randomization balance, attrition (Lee bounds if differential), pre-specified primary estimand, multiple-hypothesis control if more than one outcome.
• The headline is one pre-registered effect — resist reporting every arm in-text.

Branch C: Structural / parameter identification

• Name what identifies the key parameter from a specific data moment, in one sentence — a short paper cannot hide behind "the likelihood."
• Report the sensitivity of the headline parameter to the moment that moves it; Monte Carlo recovery in the appendix.
• Keep the model minimal (aeri-theory-model) — only what the single insight needs.

Branch D: New fact / measurement

• Documented construction; show the fact is not a measurement artifact with the one most threatening alternative addressed in-text, others in the appendix.

Choosing the single in-text diagnostic

You have at most five exhibits total and the identification competes with the result for that budget. Pick the one diagnostic that most directly defends the design (the event-study leads, the RD plot, the first-stage, the balance table) for in-text; everything else (placebo cuts, alternative bandwidths, all balance rows) goes to the appendix.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the identification claim, don't only argue it. Full map: execution-with-mcp. AER: Insights is a short format built around one decisive result, so the body/appendix split is even tighter — run the design cleanly the first time.

1. detect_design → recommend → fit with as_handle=true → audit_result to list the checks the design still owes.
2. Staggered DiD: callaway_santanna / sun_abraham + bacon_decomposition + honest_did_from_result (the pre-trend test is low-power, Roth 2022).
3. IV: effective_f_test + an anderson_rubin_ci (valid under weak instruments), not a 2SLS t-stat alone.
4. RDD: rdrobust (bias-corrected) + rddensity / mccrary_test for manipulation.
5. 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.

Checklist

• Branch chosen; data-to-object mapping stated in one sentence
• The single most convincing diagnostic is identified for in-text placement
• Empirical: modern estimator where TWFE would bias; design diagnostic shown
• Experiment: pre-registered; primary estimand pre-specified; balance/attrition handled
• Structural: key parameter tied to an identifying moment; sensitivity reported
• Inference as SEs / confidence sets (no asterisks); clustering level correct
• The headline claim never exceeds what the identification supports

Anti-patterns

• Relying on a wall of robustness to compensate for a weak core design (no room for it)
• TWFE on staggered treatment with no heterogeneity-bias discussion
• Reporting every experimental arm/heterogeneity split in-text instead of one estimand
• "The estimator converged" presented as identification (structural)
• Significance asterisks instead of standard errors / confidence sets

Referee pushback mapped to the fix

• "OLS with controls is not identification." → Move to a design (DiD/RDD/IV/experiment) or reframe scope; one clean design beats many controls.
• "Staggered TWFE is biased here." → Re-estimate with Callaway–Sant'Anna / Sun–Abraham; show flat leads in the central figure.
• "Which moment identifies this parameter?" → One sentence + a sensitivity number; recovery Monte Carlo in the appendix.

Output format

【Branch】empirical / experiment / structural / measurement
【Data-to-object mapping】one sentence
【In-text diagnostic】<the single most convincing exhibit>
【Inference】SEs / confidence sets (no asterisks); clustering level
【To the appendix】placebos / alt bandwidths / full balance / MC recovery
【What it does NOT identify】[…]
【Next step】aeri-robustness (or aeri-theory-model if structural)