jde-identification-strategy

Identification Strategy (jde-identification-strategy)

When to trigger

• The empirical core is OLS + controls with an undefended causal claim
• A DID uses two-way fixed effects (TWFE) on staggered timing without modern estimators
• An IV's first stage is weak or the exclusion restriction is unargued
• An RCT lacks a pre-analysis plan, balance, or attrition analysis
• You are unsure your design clears the JDE causal bar

The JDE identification bar

Development economics has been a leader in the credibility revolution, and JDE referees apply a demanding identification standard while respecting that field settings are messy. The implicit credibility ranking (strong → weaker):

1. RCT / field experiment with a registered pre-analysis plan, balance, and attrition handling — the modal credible design in modern development micro
2. Sharp/fuzzy RDD at a clean program or eligibility threshold (common in targeted transfer and education programs)
3. DID / event study off a credibly exogenous policy or shock, using modern estimators
4. IV with a strong first stage and a defended, institutionally-grounded exclusion restriction
5. Selection-on-observables / matching — acceptable only as a complement, rarely the spine

A theoretical paper is judged on the development relevance and rigor of its mechanism, not on a research design. A paper with novel, hard-to-assemble data answering a first-order development question can carry reduced-form evidence — but the question and the data discipline must be exceptional.

JDE also runs a permanent pre-results review / Registered Reports track: a prospective design (hypotheses, procedures, statistical analysis plan, power analysis, pilot data if applicable) can be reviewed and accepted in principle before results exist. If your design is prospective, build it to that standard — it both strengthens identification and unlocks that route (see jde-review-process).

Branch paths

RCT / field experiment (the development workhorse)

• Pre-registered PAP (AEA RCT Registry / OSF); report deviations honestly.
• Power / MDE justified at the level of randomization; clustered designs need cluster-level power.
• Balance table; attrition analysis with Lee bounds if differential.
• Inference at the unit of randomization (cluster); few clusters → wild-cluster bootstrap or randomization inference.
• Multiple-hypothesis adjustment across outcomes/subgroups (Romano–Wolf / Westfall–Young).
• External validity: what does this population and context teach beyond the site?

DID / event study

• Staggered adoption → move beyond TWFE (Callaway–Sant'Anna, Sun–Abraham, de Chaisemartin–D'Haultfœuille); report a Goodman-Bacon decomposition.
• Clean event-study plot with leads; pre-period coefficients near zero.

IV

• Strong first-stage F; with weak instruments use Anderson–Rubin / weak-IV-robust sets.
• Exclusion argued in three registers: theory, institutional detail, falsification.
• Report reduced form and OLS alongside; discuss the LATE/complier interpretation.

RDD

• McCrary / Cattaneo–Jansson–Ma density test for manipulation at the cutoff.
• Optimal bandwidth (Calonico–Cattaneo–Titiunik) plus bandwidth robustness; bias-corrected CIs.
• Covariate smoothness and placebo cutoffs.

Worked micro-example (illustrative)

Hypothetical: a national school-construction program rolled out district-by-district over six years — a quasi-experiment exploiting staggered policy timing to estimate effects on years of schooling.

• Identifying variation: within-district changes in program exposure timing, comparing cohorts young enough to benefit against older cohorts at rollout.
• The trap: plain TWFE on the staggered rollout reporting one coefficient (e.g., +0.4 years, illustrative); with heterogeneous timing, already-treated districts contaminate the comparison group.
• JDE-shaped fix: estimate with Callaway–Sant'Anna or de Chaisemartin–D'Haultfœuille, show a Goodman-Bacon decomposition, plot the event study with leads near zero, and cluster at the district (rollout) level.
• External validity line: the estimate is a LATE for cohorts near the exposure margin in this institutional setting — say what it teaches about returns to schooling without claiming a universal parameter.

Credibility pushback and the design-level answer

| Referee concern | What clears the JDE bar |

|------------------------------------------------|--------------------------------------------------------------|

| "Staggered TWFE is biased here" | Modern estimator + Bacon decomposition + event-study plot |

| "Spillovers leak treatment into controls" | Spillover/SUTVA test; bound it; defend the controls |

| "IV exclusion is asserted, not argued" | Theory + institutions + falsification, all three |

| "RDD cutoff may be manipulated" | Density test (McCrary / CJM) + covariate smoothness |

| "One site — why generalize?" | Pin the LATE/population; mechanism evidence that travels |

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the identification claim, don't only argue it. Full map:

[execution-with-mcp](../../../shared-resources/empirical-methods/execution-with-mcp.md). Development economics leans on RCTs and observational designs alike; field experiments demand the many-outcome family-wise correction (romano_wolf).

1. detect_design → recommend → fit with as_handle=true → audit_result to list

the checks the design still owes.

1. Staggered DiD: callaway_santanna / sun_abraham + bacon_decomposition +

honest_did_from_result (the pre-trend test is low-power, Roth 2022).

1. IV: effective_f_test + an anderson_rubin_ci (valid under weak instruments),

not a 2SLS t-stat alone.

1. RDD: rdrobust (bias-corrected) + rddensity / mccrary_test for manipulation.
2. 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](../../../Journal-of-Finance-Skills/resources/worked-examples/02-execution-walkthrough.md). If StatsPAI/Stata are not connected, adapt the

vendored resources/code/ skeleton and flag any unverified number.

Anti-patterns

• TWFE on staggered treatment with no discussion of heterogeneity bias
• An RCT with no PAP, no power calculation, and unexamined differential attrition
• Inference not clustered at the randomization level
• Overclaiming a single-site LATE as a universal development parameter

Output format

【Design】RCT / RDD / DID / IV / theory / descriptive
【Identifying variation】one sentence
【Diagnostics done】[PAP, balance, attrition/Lee bounds, density, first-stage F, pre-trends, ...]
【Diagnostics missing】[...]
【Inference】clustering level + few-cluster handling + MHT
【Interpretation】LATE / ATE / external validity note
【Pre-results route?】[Y/N]
【Next step】jde-data-analysis