Aer Identification

AER Identification

Overview

In AER-track empirical economics, identification is the paper. This skill routes among canonical designs, modern defaults, and referee-facing diagnostics.

If the design is fragile, return to aer-topic-selection; writing cannot save it.

When to Use

• Designing the empirical strategy for a new project
• The current strategy is TWFE / first-stage F / naive RDD and the referee will flag it
• A prior submission was rejected on identification grounds and the design needs rebuilding
• Choosing between two candidate identification strategies for the same question

Master Decision Tree

Is treatment assignment plausibly random conditional on observables?
├── Yes, by design (RCT, lottery) → run the RCT analysis; register PAP via AEA RCT Registry
└── No → identification must come from variation
    ├── Sharp threshold in a running variable → RDD (sharp or fuzzy)
    ├── Discrete policy change in some units, not others, over time → DiD
    │     ├── Single treatment date → canonical 2×2 DiD
    │     └── Staggered adoption → Callaway-Sant'Anna or Borusyak-Jaravel-Spiess
    ├── Endogenous regressor + plausibly exogenous shifter → IV
    │     ├── Shifter × pre-existing exposure shares → shift-share / Bartik
    │     └── Single instrument → weak-IV-robust inference if F  10 rule is **obsolete**. Modern conventions:

- Just-identified models: report **Anderson-Rubin (AR) confidence sets** as primary inference; AR keeps size under weak instruments.
- For F  1 are discouraged (Gelman-Imbens 2019).
- **MSE-optimal bandwidth (Calonico-Cattaneo-Titiunik 2014)** with the robust bias-corrected confidence interval. Use `rdrobust`.
- **Donut RDD** if bunching near the cutoff is a concern.
- **Covariate adjustment** for efficiency; main result must hold without it.

### Required Diagnostics

1. McCrary (2008) / Cattaneo-Jansson-Ma (2020) density test for manipulation of the running variable
2. Balance tests on predetermined covariates at the cutoff
3. Placebo cutoffs away from the true threshold
4. Bandwidth sensitivity — show the estimate across at least three bandwidths
5. Visual RD plot using `rdplot` with the binning method explicitly stated

## Synthetic Control

### When Appropriate

- One (or few) treated units
- Long pre-treatment outcome series (≥ 10 periods)
- A large donor pool of plausibly comparable untreated units
- Aggregate intervention (policy at the country, state, city level)

### Modern Extensions

- **Generalized synthetic control (Xu 2017)** for multiple treated units
- **Augmented synthetic control (Ben-Michael, Feller, Rothstein 2021)** for bias correction
- **Synthetic DiD (Arkhangelsky et al. 2021)** combining SCM and DiD weighting

### Required Diagnostics

1. Placebo (in-time): apply SCM to pre-treatment fake intervention dates
2. Placebo (in-space): apply SCM to every donor as if it were treated; report the distribution of placebo effects
3. Permutation inference / Fisher exact p-value
4. Weight vector reported in the appendix; donors with > 10% weight discussed

## Field Experiments and RCTs

If the paper uses a field experiment:

- **Register with AEA RCT Registry** before the intervention begins. AEA journals require this prior to submission.
- **Pre-analysis plan (PAP)** posted before unblinding. Per Olken and others, keep the PAP moderate in scope — pre-specify primary outcomes and the analysis specification, leave exploratory work clearly labeled as such.
- **Power calculations** in the manuscript or appendix.
- **Multiple-hypothesis correction** if more than one primary outcome.
- **Attrition** documented and tested for differential attrition by treatment arm.

## Mechanism vs. Identification

A common confusion: **identification answers whether X causes Y; mechanism answers why.** Mechanism evidence should not weaken the identification of the main effect. Run:

- Subgroup heterogeneity (does the effect concentrate where theory predicts?)
- Mediation analysis only if the mediator is itself plausibly exogenous (rare)
- Auxiliary outcomes consistent with the proposed channel

## Red Flags for Referees

- TWFE on staggered data with no Goodman-Bacon decomposition
- First-stage F = 12 cited as evidence of instrument strength
- RDD with a polynomial of order 4
- Synthetic control with no placebo inference
- DiD with a "control group" of eventually-treated units
- IV exclusion restriction defended only by "we control for X"
- Quoting an Angrist-Pischke citation as a substitute for showing the diagnostic

## Repository Resources

When working from the repo or plugin bundle, load only the relevant resource:

- Estimator defaults, package calls, diagnostics, and citations: `docs/methods-reference.md`
- Staggered DiD implementation: `templates/stata/03_main_did.do`, `templates/r/03_main_did.R`, or `templates/python/main_did.py`
- Worked empirical examples: `examples/aer-exemplars.md` and `examples/modern-aer-exemplars.md`

Use the methods reference before prose: it fixes the estimand,
diagnostic, inference method, and citation that the manuscript must report.

## Identification Gate

Do not advance to robustness or writing until, for the chosen design, **all** are true:

- [ ] A modern estimator is used — no TWFE on staggered data, no first-stage-F-only IV, no high-order-polynomial RDD
- [ ] Every required diagnostic for the design (see the per-design lists above) is run and reported
- [ ] Inference matches the design — cluster-robust / AR / wild bootstrap / permutation, not default OLS SEs by reflex
- [ ] The identifying assumption is stated in one sentence, ready to drop into the introduction
- [ ] No item in "Red Flags for Referees" is present

### Gate Record Mini-Example

Write the gate decision before routing onward:

```text
STRATEGY: IV
FIRST STAGE: effective F = 7.8; 2SLS CI is not primary
ROBUST INFERENCE: AR 95% CI = [-0.14, 0.52]
PLACEBO: beta = 0.003 (p = 0.71)
DECISION: advance with directional headline only

Handoff

STRATEGY: 
MODERN ESTIMATOR USED: 
REQUIRED DIAGNOSTICS REPORTED: 
INFERENCE METHOD: 
WEAK-IV / TWFE / POLY-ORDER RED FLAGS: 
NEXT SKILL: aer-robustness

Anti-Patterns

• Defending an old design ("the prior literature used TWFE") when modern estimators exist
• Reporting OLS-with-controls as the main specification and IV/RD as "robustness"
• Using more than one identification strategy as if they were independent confirmations when they share identifying variation
• Footnoting the identifying assumption instead of stating it in the introduction

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: brycewang-stanford
• Source: brycewang-stanford/AER-Skills
• License: MIT

Install and usage instructions live in the source repository linked above.