Empirical-economics results writing

Contract

Enforces the applied-micro / Journal-of-Human-Resources reporting conventions for any sentence, abstract, or results paragraph that states an empirical estimate. Apply whenever a regression coefficient, treatment effect, or rate change is being written up for an economics audience. These rules are non-default; follow them exactly.

Rules

R1 — Name the estimand on first use

The FIRST time a numeric result is stated, name the estimand with a specific term, never the bare word "effect". Allowed terms and the design that produces each:

| Design | Estimand | |---|---| | Lottery / IV / encouragement / nudge instrument | LATE (local average treatment effect, for compliers) | | Random assignment analyzed by assignment, ignoring take-up | ITT (intent-to-treat) | | Effect among those actually treated | ATT (average treatment effect on the treated) | | RCT with full compliance / population-average target | ATE (average treatment effect) | | Regression discontinuity | LATE at the cutoff | | OLS / FE / cross-section with no exogenous variation | descriptive gap (not an estimand of a causal effect) |

R2 — Causal verbs only for designed estimates

Use causal verbs — raises, reduces, increases, lowers, induces — ONLY for designed estimates: RCT, lottery/IV, regression discontinuity, difference-in-differences. For observational / cross-sectional / panel-FE / OLS-without-exogenous-variation results, write "is associated with" and add an explicit not-causal caveat. Never write "X raises Y" for a correlation.

R3 — Percentage points vs percent

A change in a RATE, SHARE, or PROBABILITY is reported in percentage points, never "percent". A rate moving 12% → 14% is a rise of 2 percentage points (the ≈17% relative change is secondary, never the headline unit). A linear-probability-model coefficient of 0.05 is a 5 percentage point change in the probability.

R4 — Natural-unit magnitudes

Translate every coefficient into a natural unit the reader can price: dollars at the control mean, weeks, points, standard deviations. A log-point coefficient β at control mean M is about β·M dollars (e.g. 0.083 × $4,940 ≈ $410). Report the natural unit alongside (not instead-of- omitting) the raw coefficient. Never leave the headline as bare log points.

R5 — Clustering stated once, in prose

State the clustering level exactly once in the prose (e.g. "clustered on 38 sites"); after that it lives only in table notes. Do not repeat it every sentence, and do not bury it only in a note when it is the inference unit for the headline.

R6 — The five-sentence abstract

Write the abstract as exactly five sentences, in this order:

1. the policy-relevant question (open here, not with a broad background paragraph),
2. the data and the identifying variation,
3. the headline magnitude in a natural unit,
4. reconciliation with the closest prior estimate,
5. the policy interpretation with its external-validity boundary.

Never write "important implications" (or "important policy implications") unless the next words name the actual margin.

R7 — Results-paragraph order

Order a results paragraph: claim + design → magnitude for the population → reconciliation ("compares with prior] because sample/spec/design difference]") → policy implication. Do not report a coefficient and significance with no interpretation.

Worked examples (BEFORE = base default → AFTER = conforming)

R1, estimand. Lottery-admission IV, β = 0.083 on log quarterly earnings.

• BEFORE: "The effect of training on earnings is 0.083."
• AFTER: "The lottery-based LATE implies training raises quarterly earnings by 8.3 log points (≈ $410 at the control mean) for compliers."

R1, ITT. Random tutoring offer, some decline, analyzed by assignment.

• BEFORE: "The effect on test scores is 0.12 SD."
• AFTER: "The intent-to-treat (ITT) estimate is a 0.12 SD gain in test scores from being offered tutoring."

R2, association. Cross-sectional OLS of wages on union membership, β = 0.15.

• BEFORE: "Union membership raises wages by 15%."
• AFTER: "Union membership is associated with about 15% higher hourly wages; with no exogenous variation this gap is descriptive, not causal."

R2, designed. RD on a scholarship cutoff.

• BEFORE: "There is a positive relationship between eligibility and enrollment."
• AFTER: "Crossing the eligibility cutoff raises college-enrollment probability by 20 percentage points (the LATE at the cutoff)."

R3, percentage points. Teen unemployment 12% → 14% under a DiD minimum-wage study.

• BEFORE: "Unemployment rose about 17 percent."
• AFTER: "The minimum-wage increase raised the teen unemployment rate by 2 percentage points (12% to 14%)."

R3, LPM. Experimental LPM, employed indicator, β = 0.05.

• BEFORE: "Treatment raised employment 5 percent."
• AFTER: "Treatment raised the probability of employment by 5 percentage points."

R4, natural units. Cash transfer, 0.10 log monthly consumption, control mean $300.

• BEFORE: "Consumption rose 0.10 log points."
• AFTER: "The average treatment effect is about $30 per month at the $300 control mean (0.10 log points)."

R5, clustering. Earnings LATE, SE 0.024, 38 sites.

• BEFORE: "(SE 0.024)." with the clustering level never stated, or repeated each sentence.
• AFTER: "…(SE 0.024, clustered on 38 sites)." stated once; later sentences omit it.

R6, abstract. Charter-school lottery, +0.15 SD math, low-income urban.

• BEFORE: a long background paragraph, then "we find positive effects with important implications."
• AFTER (5 sentences): "Do urban charter schools raise math achievement? Using a school-assignment lottery, we compare winners and losers. Attending a charter school raises math scores by about 0.15 SD for lottery winners. This is larger than most prior observational studies, which lacked random assignment. The gains concentrate among low-income urban applicants, bounding external validity to that setting."

R7, results paragraph. Wage subsidy, 0.11 log, control mean $1,800, prior near-zero.

• BEFORE: "The coefficient (0.11, p<0.01) is positive and significant, confirming the program works."
• AFTER: "Table 3 reports the experimental ATE of the wage subsidy. Monthly earnings rise by about $200 (0.11 log points) at the $1,800 control mean. This exceeds a prior subsidy experiment's near-zero estimate because that study targeted a higher-wage population. The implied payback supports extending the subsidy to low-wage entrants."

Edge cases & exceptions

• Relative percent IS allowed as a secondary gloss. "2 percentage points (a 17% relative rise)" is fine; the percentage-point figure must lead.
• A change measured in percent of a dollar level (e.g. earnings up 8%) is genuinely "percent" — R3 governs rates/shares/probabilities, not levels.
• IV with one-sided non-compliance still yields LATE for compliers; do not upgrade it to ATE.
• DiD and RD are designed, so causal verbs are correct even though there is no literal randomization — R2's list includes them.
• Descriptive papers may state a "descriptive gap" with causal nouns avoided; "the schooling-health gradient" is fine, "schooling improves health" is not.
• Standardized effects (SD) are themselves a natural unit; you need not also convert to dollars, but still name the estimand.
• Abstract length may compress below five sentences only if a journal's live word limit forces it; default is exactly five.

Do / Don't

• DON'T let "effect" carry the headline number. DO name LATE/ATT/ITT/ATE/descriptive gap on first use.
• DON'T write "raises/causes" for OLS. DO write "is associated with" + a not-causal caveat.
• DON'T call a rate change "percent". DO call it "percentage points".
• DON'T stop at log points. DO convert to dollars/weeks/SD at the control mean.
• DON'T repeat or hide the clustering level. DO state it once in prose.
• DON'T open the abstract with broad background. DO open with the policy-relevant question.
• DON'T write "important implications". DO name the specific margin.

Common mistakes (base's wrong defaults)

• Reporting "the effect is 0.083" with no estimand named.
• Saying a rate "rose 17 percent" instead of "2 percentage points".
• Reading a linear-probability coefficient as a percent change rather than a percentage-point change.
• Asserting causation from a cross-sectional OLS or fixed-effects regression.
• Leaving the headline in log points with no dollar/SD conversion.
• Writing six- or three-sentence abstracts, or burying the question under background.
• Closing with "important policy implications" and no named margin.

Quick checklist

Estimand named? · causal verb matches the design? · rate change in percentage points? · coefficient in natural units? · clustering stated once? · abstract exactly five sentences in order? · results paragraph reconciled and priced?