wber-identification

Identification Strategy (wber-identification)

When to trigger

• A causal claim rests on OLS + controls, or TWFE on staggered reform timing
• An RCT's estimand, balance, attrition, or spillover handling is not pinned down
• An RD's density, bandwidth, or covariate-smoothness defense is missing
• An IV's first stage is weak or the exclusion restriction is asserted, not argued
• The design is clean but the policy interpretation (scale-up, GE, fiscal cost) is missing

The WBER identification bar

WBER demands the same identification rigor as a top applied-micro field journal — and then one more thing that siblings let slide: the mapping from the local causal estimate to a development-policy lesson must be argued, not assumed. A pristine ITT from one trial is necessary but not sufficient; the WBER referee asks "what does this imply for a finance ministry deciding whether to scale this?" So state the estimand, name the identifying assumption, show the diagnostic that could have failed, and address external validity, general equilibrium, and scaling. Inference must match the design (clustering at the assignment level; few-cluster corrections). Report standard errors and confidence intervals — not significance asterisks.

Design paths

Path A: RCT / program evaluation (own or admin data)

• Estimand stated (ITT vs. LATE/TOT); randomization unit and stratification described.
• Pre-registration where applicable (AEA RCT Registry / OSF); report deviations from the analysis plan.
• Balance table on baseline covariates; attrition examined and bounded (Lee bounds if differential) — attrition is endemic in field settings.
• Spillovers / SUTVA: in dense developing-country settings, treatment can leak to controls; design or test for it.
• Multiple-hypothesis adjustment across outcomes/subgroups; explicit external-validity and scale-up discussion.

Path B: Difference-in-differences / event study around a reform

• With staggered adoption, move beyond TWFE — Callaway–Sant'Anna, Sun–Abraham, de Chaisemartin–D'Haultfœuille, or Borusyak–Jaravel–Spiess imputation.
• Clean event-study with leads for pre-trends; report a Goodman-Bacon decomposition.
• State and defend parallel trends; consider Rambachan–Roth honest-DiD sensitivity. Developing-country reforms are often non-random in timing — argue why.

Path C: Regression discontinuity (eligibility thresholds, geographic borders)

• Density / manipulation test (McCrary / Cattaneo–Jansson–Ma); covariate smoothness at the cutoff.
• Local-linear with data-driven bandwidth; bias-corrected robust CIs (rdrobust); donut and bandwidth-sensitivity checks.
• Many development RDs are proxy-means-test or poverty-line cutoffs — guard against score manipulation by enumerators or applicants.

Path D: IV / shift-share

• Strong first stage (effective F); with weak instruments use Anderson–Rubin / weak-IV-robust sets.
• Exclusion restriction argued from institutions/theory + falsification. Be wary of tired development instruments (rainfall, distance, colonial-era variables) — argue exogeneity, do not invoke by tradition.

The external-validity layer (the WBER differentiator)

• From LATE to policy: name whose effect you estimate and how the complier population differs from the scale-up population.
• General equilibrium: a partial-equilibrium treatment effect can vanish or reverse at scale (wages, prices, congestion). Acknowledge and, where possible, bound it.
• Fiscal cost / cost-effectiveness: WBER readers weigh effects against budgets — a cost-per-outcome or benefit-cost figure makes the result actionable.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. WBER is development economics — RCTs and observational designs in low/middle-income settings; randomization inference + DiD/IV, magnitude in policy units.

• detect_design → recommend → fit with as_handle=true → audit_result.
• Observational causal claims: staggered DiD (callaway_santanna / sun_abraham + bacon_decomposition + honest_did_from_result); IV (effective_f_test + anderson_rubin_ci); RDD (rdrobust + mccrary_test).
• Experiments: randomization-based inference + romano_wolf for many-outcome control.
• Sensitivity: oster_delta / sensemakr for observational claims.

Report the magnitude in interpretable units; route the full battery to the appendix. A run end-to-end (synthetic data, real returns) is in the JF execution walkthrough.

Checklist

• Design chosen; variation-to-estimand mapping stated in one sentence
• Estimand named (ITT/LATE/ATT/local-at-cutoff) and matched to the design
• Design-appropriate diagnostic shown (balance+attrition+spillovers / pre-trends+Bacon / density+bandwidth / first-stage+exclusion)
• Modern estimator used where TWFE or 2SLS would bias
• Inference clustered at the assignment level; few-cluster corrected
• External validity, GE, and scaling/fiscal cost explicitly addressed
• SEs and CIs reported (no asterisks); claim never exceeds what the design identifies

Anti-patterns

• A clean trial with zero discussion of external validity or scale-up — reads as a project report
• TWFE on staggered reform timing with no heterogeneity-bias discussion
• A development RD with no density test at a proxy-means or poverty-line cutoff
• Rainfall/distance/colonial IVs invoked by tradition with no exclusion argument
• Reading a local LATE as the population ATE for a national scale-up
• Asterisks instead of clustered SEs / weak-IV-robust sets

Referee pushback mapped to the identification fix

• "Staggered TWFE here is biased." → Re-estimate with Callaway–Sant'Anna or Sun–Abraham; show flat event-study leads and a Goodman-Bacon decomposition.
• "Your instrument is rainfall/distance — assert exclusion at your peril." → Argue the exclusion restriction from institutions and falsify it (reduced form on never-takers, placebo outcomes); if it fails, drop the IV.
• "This RD eligibility score is manipulable by enumerators." → Report a density test, covariate smoothness, and a donut estimate; discuss the assignment mechanism.
• "Differential attrition contaminates the trial." → Show attrition by arm and Lee bounds; if spillovers are plausible, test or design for them.
• "A three-district LATE tells me nothing about a national program." → Characterize the complier population, bound the GE channel, and report a cost-per-outcome at scale.

Worked vignette (illustrative)

A paper evaluates a fee-elimination reform rolled out across districts in staggered years using TWFE; a referee flags negative weighting. The WBER fix: re-estimate with Callaway–Sant'Anna by adoption cohort, show flat pre-trend leads, report a Goodman-Bacon decomposition (say 15% of the TWFE estimate came from forbidden already-treated comparisons, illustrative). The cohort-robust ATT settles at 4.3pp enrollment (s.e. 1.1). Then the WBER-specific step: the authors note the complier districts were poorer than average, bound the GE wage effect on teachers, and report a cost-per-additional-enrollee, turning a clean ATT into a scale-up-relevant number.

Output format

【Design】RCT / DiD / RD / IV
【Variation-to-estimand mapping】one sentence
【Estimand】ITT / LATE / ATT / local-at-cutoff
【Identification evidence】[balance+attrition+spillovers / pre-trends+Bacon / density+bandwidth / first-stage+exclusion]
【Estimator + inference】modern estimator; clustering level; weak-IV/honest-DiD sensitivity if any
【External validity】complier vs. scale-up population; GE; cost-effectiveness
【What it does NOT identify】[...]
【Next step】wber-theory-model (if interpretation needs a model) or wber-robustness