Data Analysis (jeg-data-analysis)

When to trigger

• You are estimating cross-country, panel, historical, or regional growth models
• A theory paper includes calibration, simulation, or transition dynamics
• Results need robustness, decomposition, or sensitivity checks for JEG

Empirical growth checklist

• Define the growth outcome: level, growth rate, convergence speed, productivity, human capital, fertility, technology, institutions, or development outcome.
• Document the unit and horizon: country-year, region-decade, cohort, household, firm, or historical panel.
• Separate long-run levels from short-run growth dynamics.
• Show sample construction, merge rules, missingness, and influential observations.
• Use specifications that match the question: convergence regressions, panel FE, IV, DID/RDD around reforms, synthetic controls, or structural estimates.

Theory / calibration checklist

• State calibrated parameters, data moments, and source for each moment.
• Separate targeted from untargeted moments.
• Report transition paths and steady states clearly.
• Stress-test key elasticities, discount rates, depreciation, fertility, human capital, and technology parameters.
• Make code reproducible enough to regenerate figures and tables.

Growth-mechanism audit table

Before drafting results, create a table with:

• Mechanism: human capital, fertility, technology, institutions, trade, finance, migration, or OLG channel.
• Object: growth rate, income level, TFP, convergence speed, transition path, or welfare.
• Discipline: data moment, calibration target, theorem assumption, or identification source.
• Main sensitivity: parameter or sample choice most likely to overturn the result.
• Replication artifact: code or file that regenerates the exhibit.

If an estimate or simulation does not map to a mechanism row, it is probably not central enough for JEG.

Spatial and historical inference discipline

Comparative-development empirics at JEG are usually geocoded, which changes the inference defaults:

• Report Conley standard errors at multiple distance cutoffs (e.g., 100/250/500 km) for any gridded or regional outcome; clustered SEs at the modern administrative level are necessary but not sufficient.
• When historical units do not coincide with modern ones, cluster at the historical unit — the level at which the treatment was assigned — and document the crosswalk.
• Pre-empt the critique that persistence t-statistics can be inflated by smooth spatial trends: include flexible geographic controls (latitude-longitude polynomials or macro-region fixed effects) plus a spatial-noise placebo test.
• For very long panels, keep measurement vintages separate: reconstructed pre-1950 series, modern national accounts, and nighttime lights are not interchangeable; show the result within each vintage where feasible.

Worked vignette — auditing a comparative-development panel

Illustrative setup: 2,400 grid cells in 41 countries; outcome is log light density in 2020; regressor is distance to a historical trade hub; candidate instrument is least-cost-path placement.

• Unit/horizon: cell-level cross-section answering a long-run level question, so convergence-dynamics machinery is unnecessary; the persistence design applies.
• Inference: coefficient 0.21; country-clustered SE 0.05, Conley 250 km SE 0.08, Conley 500 km SE 0.09 — report all three; the claim survives the widest cutoff.
• Mechanism row: schooling in 1960 absorbs roughly 40% of the coefficient (illustrative), so human capital becomes a lead exhibit, not a robustness afterthought.
• Main sensitivity: dropping cells within 50 km of modern capitals moves the estimate to 0.17; capital proximity goes into the audit table as the result's weakest joint.

Estimator defaults by growth question

• Long-run level question (deep determinants, persistence) → cross-sectional or grid design + Conley inference + mechanism decomposition.
• Convergence-speed question → panel estimation alert to Nickell bias; system GMM only with instrument-count discipline and Hansen/AR(2) reporting.
• Reform-timing question → modern staggered-adoption DID estimators with pre-trend evidence, never naive TWFE.
• Theory-driven quantitative question → calibrated model with targeted and untargeted moments kept visibly separate.
• Demographic or fertility question → cohort or census microdata aggregated to the mechanism's unit; verify the transition timing is identified by the data rather than assumed by the specification.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. JEG (growth) uses cross-country and long-run panels with deep endogeneity; foreground identification and robustness to alternatives.

• Many outcomes / specifications: romano_wolf (step-down FWER) or benjamini_hochberg.
• OVB sensitivity: oster_delta / sensemakr.
• Inference: wild_cluster_bootstrap (few clusters), twoway_cluster / conley.
• Re-fit off one handle: audit_result(result_id) lists missing checks + the exact suggest_function for each.
• Exhibits: etable / did_summary_to_latex from the handle — no retyped numbers.

Decisive checks in the body, exhaustive battery in the appendix. JF execution walkthrough.

Output format

[Paper type] empirical / theory / mixed
[Data or model object] ...
[Main estimator/calibration] ...
[Robustness or sensitivity] ...
[Reproducibility gaps] ...
[Next step] jeg-tables-figures