Jde data analysis

Data Analysis (jde-data-analysis)

When to trigger

• The identification is settled but the estimation, inference, or heterogeneity analysis is unconvincing
• A referee would question standard errors, attrition, measurement error, or sample construction
• You need to decide how to present treatment-effect heterogeneity across subgroups
• You are unsure the analysis would survive JDE's replication scrutiny

JDE empirical norms

JDE referees are experienced with the realities of field and survey data in developing countries — clustered sampling, panel attrition, noisy self-reports, seasonality, and small effective sample sizes. Analysis that ignores these reads as naive. Hold the work to these standards:

• Inference matched to the data structure. Cluster at the level of treatment assignment or sampling (village, school, market); with few clusters use wild-cluster bootstrap or randomization inference rather than naive cluster-robust t-stats.
• Attrition and missing data. Document panel attrition, test whether it is differential by treatment, and bound effects (Lee bounds) when it is. Survey non-response and refusal patterns belong in the appendix.
• Measurement. Be explicit about how key variables (consumption, income, yields, test scores, health) were measured and constructed; address recall error, social-desirability bias, and unit/seasonal issues. Pre-specify or transparently justify index construction.
• Heterogeneity, disciplined. Development audiences care about for whom effects bind, but data-mined subgroups are penalized. Pre-specify subgroups where possible; otherwise treat heterogeneity as exploratory and adjust for multiple comparisons.
• Magnitudes in welfare terms. Report effects in policy-comparable units (share of the poverty gap, cost-effectiveness per dollar, standard deviations) — see jde-contribution-framing.

Because JDE's replication policy lets editors or referees request data, programs, and computational details at the review stage, every number in a table must be reproducible from a script the day you submit. Build the analysis to be auditable, not just presentable.

Robustness expected

• Alternative specifications, samples, and functional forms in an extensive online appendix
• Sensitivity to outliers, winsorization, and index/aggregation choices
• Placebo / falsification outcomes that should not move
• Spillover/SUTVA checks where treatment may leak across units (common in village-level interventions)

Worked analysis (illustrative)

Hypothetical: a cluster-randomized health-extension experiment, 80 villages, ~30 households each, with 12 percent endline attrition.

• Inference: cluster at the village level; with 80 clusters report wild-cluster-bootstrap p-values too (illustrative ITT = +0.14 SD on a child-health index, p = 0.03).
• Attrition: 12 percent overall but 9 vs 15 percent across arms — differential, so add Lee bounds; the effect survives the lower bound (~+0.06 SD).
• Measurement: the health index mixes recall and anthropometric items; pre-specify weights and show robustness to inverse-covariance vs simple-average aggregation.
• Heterogeneity: only the baseline-poverty split was pre-registered; report it Romano–Wolf-adjusted, demote the rest.

Empirical-credibility pushback and the fix

Referee objection	The JDE-norm response
"SEs ignore the clustered randomization"	Re-cluster at the randomization unit; wild-cluster boot / RI
"Differential attrition biases the effect"	Report attrition by arm; add Lee bounds; show the bound holds
"Spillovers violate SUTVA in your villages"	Distance-ring / treated-neighbor checks; bound the leakage

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. Development economics leans on RCTs and observational designs alike; field experiments demand the many-outcome family-wise correction (romano_wolf).

• Many outcomes / specifications: romano_wolf (step-down FWER, accounts for cross-test correlation) or benjamini_hochberg — report the adjusted threshold.
• OVB sensitivity: oster_delta / sensemakr — the confounder strength that would overturn the headline.
• Inference: wild_cluster_bootstrap (few clusters), twoway_cluster / conley.
• Re-fit off one handle: audit_result(result_id) lists the missing checks and the exact suggest_function for each — no guessing the battery.
• Exhibits: etable / did_summary_to_latex from the handle — no retyped numbers.

Keep the decisive checks in the body and the exhaustive (now actually-run) battery in the appendix. See the executed chain in the JF execution walkthrough.

Anti-patterns

• Standard errors not clustered at the design level; ignoring few-cluster bias
• Silently dropping attritors or trimming the sample without reporting it
• A wall of unadjusted subgroup interactions presented as confirmatory
• Effects reported only in raw units, leaving importance unclear
• Tables that cannot be regenerated from the submitted code

Evidence pass for Journal of Development Economics

Treat this skill as an executable review pass, not a prose hint. First lock the development constraint, identification, welfare or distribution margin, and implementation context; then judge whether the current manuscript answers the venue's real reader: development economists who expect a development mechanism, credible design, and policy-relevant external validity.

• Do the pass: Audit the research design before polishing prose: unit of analysis, comparison set, uncertainty, sensitivity, missingness, and reproducibility must be visible.
• Return a ledger: give claim / evidence / risk / manuscript location rows, so the next agent can edit rather than rediscover the issue.
• Sibling guard: compare against World Development for broader policy audience, JPubE for fiscal/public-finance mechanisms, AER/AEJ Applied for field-wide reach; if a sibling owns the contribution, recommend re-routing before polishing format.
• Stop condition: do not give submission-ready advice until the pack's resources/official-source-map.md has been checked for volatile rules and the manuscript has one concrete fix for the largest venue-specific risk.

Output format

【Estimator】+ why it fits the design
【Inference】clustering level / few-cluster method / MHT
【Attrition】rate, differential? bounds?
【Measurement】key variable construction + error handling
【Heterogeneity】pre-specified vs exploratory
【Robustness done / missing】[...]
【Reproducible from code today?】[Y/N]
【Next step】jde-tables-figures