brycewang-stanford Writing Style (jhr-writing-style)
When to trigger
- The manuscript is too long for the 40-page limit
- The policy lesson or identification strategy is buried
- Results paragraphs report coefficients without interpretation
Style principles
- State the empirical-micro question in the first page.
- Make the identifying variation visible before literature detail.
- Translate coefficients into policy-relevant magnitudes.
- Reconcile with prior work explicitly.
- Keep main-text tables few and move supporting material to the Online Appendix.
First-page pattern
The first page should do five jobs in order:
- Name the policy-relevant empirical-micro question.
- State the institutional or data feature that creates variation.
- Give the headline estimate in natural units.
- Say how it compares with the closest prior estimate.
- State the policy interpretation and its boundary.
Do not open with a broad policy problem for more than one paragraph. JHR readers need to see the design and reconciliation early.
Abstract compression
Use a five-sentence abstract unless the journal's live limit forces something tighter:
- Policy-relevant question.
- Data and identifying variation.
- Headline magnitude in natural units.
- Reconciliation with the closest prior estimate or mechanism.
- Policy interpretation with the external-validity boundary.
Avoid empty phrases like "important implications" unless the next words name the actual margin. A good JHR abstract lets a reader recover the estimand, setting, magnitude, and policy lesson without opening the tables.
Results paragraph pattern
Table X estimates [claim] using [design]. The preferred specification in column Y
implies [magnitude] for [population]. This compares with [prior estimate] because
[sample/specification/design difference]. The result suggests [policy implication].
Before/after: a results paragraph rewritten for JHR
Illustrative job-training example (numbers invented to show the editing moves):
BEFORE: The coefficient on training (0.083, p<0.01) is positive and highly
significant, confirming that the program works. This has important policy
implications.
AFTER: Table 4 reports the lottery-based LATE. Column 2 implies that training
raises quarterly earnings by 8.3 log points — about $410 at the control mean —
for applicants induced by the lottery (SE 0.024, clustered on 38 sites). The
estimate sits between the prior experimental literature's near-zero results
and the larger observational ones; Appendix Table A5 shows the gap closes once
we match the prior studies' shorter follow-up window. At roughly $2,900 per
trainee, the implied payback period is under two years for compliers.
The rewrite names the estimand, converts the coefficient, states the inference unit, reconciles, and prices the policy — five JHR sentences doing five jobs.
Sentence-level conventions
- Name the estimand (ATT, LATE, ITT, descriptive gap) the first time a result is stated; never let "effect" float unanchored.
- Keep percent and percentage points rigorously distinct — a frequent JHR copyedit and referee irritant in education and health papers.
- State the clustering level once in prose and thereafter only in table notes.
- Reserve causal verbs (raises, reduces, induces) for designed estimates; descriptive results get "is associated with" and an explicit caveat.
- When heterogeneity is discussed, tie each cut to a policy margin in the same sentence, not in a separate "implications" paragraph.
Trim hierarchy for the page cap
When the manuscript runs long, cut in this order: (1) literature narration not attached to a benchmark estimate; (2) robustness prose that can become one sentence plus an appendix citation; (3) secondary heterogeneity cuts; (4) institutional history beyond what the design needs. Never trim the event-study figure, the first stage, or the reconciliation paragraph — those are what JHR referees read first.
Output format
[Section] abstract / intro / data / results / conclusion
[Main edit] ...
[Magnitude wording] ...
[Reconciliation wording] ...
[Page-limit fix] ...
[Next step] jhr-submission