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

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)