jhr-identification-strategy - SKILL.md Agent Skill

name: jhr-identification-strategy description: Use when stress-testing causal identification for a Journal of Human Resources empirical-micro manuscript, including RCT, DID, RDD, IV, event studies, decompositions, policy shocks, and reconciliation with prior estimates.

Identification Strategy (jhr-identification-strategy)

When to trigger

The paper makes a causal claim
Referees could say selection, omitted variables, or policy timing drives the result
You need a reconciliation and sensitivity plan before submission

Design checks

RCT: attrition, balance, compliance, spillovers, pre-analysis plan, multiple testing, cluster level.
DID / event study: timing, controls, pre-trends, treatment heterogeneity, anticipation, modern estimators where needed.
RDD: manipulation, bandwidth, polynomial order, covariate continuity, donut/alternative bandwidths.
IV: first stage, exclusion, monotonicity, weak-IV inference, LATE population.
Decomposition/descriptive: what is descriptive, what is causal, and what policy lesson follows.

JHR-specific layer

Re-estimate or benchmark against close prior work when results differ.
Report sensitivity tests that explain which sample/specification choices matter.
Keep the public-policy interpretation tied to identified variation.

Public-policy validity check

For every causal claim, write the policy interpretation in LATE/ATT/descriptive terms:

Who is affected: treated population, complier group, cohort, locality, or institution.
What margin changes: enrollment, employment, wages, health, fertility, retirement, or another JHR outcome.
Which policy lever is credible: eligibility rule, treatment intensity, price, access, mandate, or program design.
What does not travel: populations, periods, institutions, or equilibrium responses outside the design.

If the policy sentence needs a broader population than the estimand supports, narrow the claim before submission.

Reviewer threat matrix

For each causal design, pre-write the most damaging reviewer objection and the table or paragraph that answers it:

Threat | Why plausible here | Evidence already in paper | Missing evidence | Claim to narrow

This matrix prevents overclaiming. If the missing evidence cannot be produced, the correct repair is often to narrow the estimand or policy interpretation, not to add another robustness table. JHR referees will usually accept a precise local estimate more readily than a broad policy claim unsupported by the design.

What JHR referees probe first, by design

Design	First probe	Evidence that usually settles it
Staggered DID	Pre-trends and forbidden TWFE comparisons	Event study from a heterogeneity-robust estimator; honest-bounds sensitivity on the pre-period
RDD at an eligibility cutoff	Manipulation and sorting at the threshold	Density test at the cutoff, covariate smoothness, donut estimates
IV from policy variation	First-stage strength and exclusion stories	Per-instrument first stage with effective F; reduced form shown; weak-IV-robust CI
School/charter lottery	Differential attrition and re-application	Attrition by win/loss, balance within lottery strata, bounds for missing outcomes
Shift-share / Bartik	Shock vs. share identification	State which component is exogenous; exposure-robust SEs

The cluster question cuts across all rows: inference must sit at the level where treatment was assigned, and the paper should say how many effective clusters remain after fixed effects absorb variation.

Worked cutoff walkthrough

Illustrative RD: a selective public high school admits applicants scoring at or above 70 on a composite exam; outcome is college enrollment (numbers invented to show the decision rules):

Manipulation: density test at the cutoff is insignificant (illustrative p = 0.42); retake rules documented so referees see why bunching is unlikely.
Continuity: baseline GPA and family income are smooth through 70; the one jumpy covariate (sibling enrollment) is shown and discussed, not hidden.
Sensitivity: effect of 6.5 percentage points is stable across bandwidths roughly half to double the MSE-optimal choice and in a donut dropping exact 70 scorers.
Interpretation: a LATE for marginal admits near 70 — the paper does not claim effects for clearly inframarginal admits, and says so.

Failure modes that end JHR reviews early

Event-study figure that starts at t = 0, hiding the pre-period entirely.
One pooled first stage covering several instruments with different complier populations.
Parallel-trends defense that relies on a linear state trend doing the work.
RD without any density or covariate-continuity evidence in the draft.
Policy conclusion written as if the LATE were a population average effect.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. JHR is labor/education economics — program evaluation with selection; DiD/IV/RDD and the selection objection are central.

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.

Output format

[Claim] causal / descriptive / decomposition
[Design] RCT / DID / RDD / IV / event study / other
[Main threat] ...
[Design defense] ...
[Reconciliation needed] ...
[Next step] jhr-data-analysis