Tables & Figures (jde-tables-figures)

When to trigger

• The main result is a nine-column table no reader can parse
• A treatment effect or pre-trend would land harder as a figure
• Heterogeneity across subgroups is buried in interaction rows
• Notes under exhibits are too thin to stand alone for a referee

The JDE exhibit bar

Development economics has moved toward figure-forward presentation, and JDE exhibits should make the headline result legible at a glance to a development economist skimming the paper. Principles:

• Lead with a figure where the design allows it. Event-study plots (DID), discontinuity plots (RDD), and treatment-effect-by-subgroup plots communicate a field result faster than a coefficient table. RCT main effects often read best as a clean coefficient plot with confidence intervals.
• One main table, deep appendix. Keep the main coefficient table tight (headline outcomes, the preferred specification, clustered SEs noted); push alternative specifications, balance, attrition, and robustness to the extensive online appendix.
• Self-contained notes. Each exhibit must state the sample, unit of observation, the level at which standard errors are clustered, the estimator, control set, and what the bars/bands represent — a referee should not have to hunt in the text.
• Welfare-relevant scaling. Where possible, scale axes and report effects in policy-comparable units (share of the poverty gap, standard deviations, cost per outcome) so magnitude is visible, not just significance.
• Geography and maps. When spatial variation drives the design (market access, program rollout, conflict), a map or spatial figure earns its place.

JDE accepts any consistent formatting style at submission (the journal's style is applied at the proof stage), so do not burn time on house-style table formatting before acceptance — spend it on clarity, correct clustering, and reproducibility. Every exhibit must be regenerable from the submitted code.

Execution bridge (StatsPAI / Stata MCP)

Generate exhibits from the fitted result, not by retyping numbers (the usual source of body-vs-appendix drift). Full map: execution-with-mcp.

• Tables: etable (multi-model columns) or did_summary_to_latex straight from the result_id — one variable definition, one set of numbers, body and appendix in sync.
• Figures: plot_from_result / enhanced_event_study_plot / event_study_table — axis units and the SE/clustering note baked in.
• Every note names the estimator + clustering (from the result's diagnostics) and states the magnitude in interpretable units.

See a full fitted-result → exhibit chain in the JF execution walkthrough.

Checklist

• Headline result presented as a figure where the design allows
• Main table limited to preferred specification + headline outcomes
• Clustering level and estimator stated in each note
• Confidence intervals / bands shown; no chartjunk (no 3D, minimal color)
• Effects scaled to welfare/policy-comparable units where possible
• Vector output (PDF/EPS) at print resolution
• Every exhibit reproducible from the master script

Exhibit-design decision grid (by JDE design type)

Worked micro-example (illustrative numbers)

Hypothetical JDE paper: a cluster-randomized after-school tutoring program across 120 villages in a low-income setting, randomized at the village level (60 treatment, 60 control), ~25 children sampled per village.

• Wrong instinct: an eight-column table reporting test scores, attendance, a sub-index, and three interactions, all with stars. A development referee cannot find the headline.
• JDE-shaped fix: one coefficient plot — ITT on the standardized test-score index = +0.18 SD (95% CI 0.06 to 0.30), control mean printed below the axis (illustrative). Attendance and the sub-index sit beside it as secondary points; the three interactions move to an appendix figure flagged "exploratory."
• Note line: "ITT; clustering at village level (60/60), N = 2,940; bars are 95% wild-cluster-bootstrap CIs. Figures illustrative."

Referee pushback the right exhibit pre-empts

• "Is 0.18 SD a lot?" → print the control mean and a policy-comparable benchmark on the figure; significance never substitutes for size at JDE.
• "Which subgroup splits were pre-registered?" → solid markers for pre-specified subgroups, hollow for exploratory, MHT-adjusted intervals shown.

Anti-patterns

• A dense table when one event-study or coefficient plot would carry the result
• Notes that omit the clustering level or sample definition
• Significance stars with no economically meaningful magnitude
• Polishing to Elsevier house style before acceptance (style is applied at proof stage)
• A heterogeneity panel that hides which splits were pre-specified versus mined

Output format

【Headline exhibit】figure type + what it shows
【Main table】outcomes + spec + clustering noted
【Appendix exhibits】[robustness, balance, attrition, ...]
【Scaling】effects in policy-comparable units? [Y/N]
【Reproducible from code?】[Y/N]
【Next step】jde-writing-style