# DRAFT preregistration — NOT registered. For Omer's review.
## Title
Does the "advanced maternal age" label change delivery care? Regression discontinuities at maternal ages 35 and 40 in all US births, 2016–2024

## Background and novelty
- ACOG/SMFM Obstetric Care Consensus #11 (2022) calls 35 an arbitrary threshold. It recommends considering delivery at 39 weeks for age ≥40, and says evidence is insufficient for any timing change at 35–39.
- Closest causal study: JAMA Health Forum 2022 (PMC8796879), an RD at 35 in commercial claims (~50k). It found more prenatal surveillance and a 0.39-pp drop in perinatal mortality. It explicitly could NOT study induction or cesarean, because claims under-record them.
- Other work is observational age-group comparisons: Denmark (PLOS One 2019), Sweden/Norway, US natality trends (NCHS Data Brief 554), and the 35/39 RCT (NEJM 2016).
- Searches of PubMed/Google Scholar-type sources, NBER/SSRN/RePEc and ACOG/SMFM found no RD at 35 or 40 on induction or cesarean using birth certificates, and no all-payer national study.

## Hypotheses
- H1 (SOLE primary; two-sided test at alpha = 0.05 with no multiplicity split): labor induction rises discontinuously at maternal age 35 among singleton, nulliparous births at ≥37 weeks.
- H2 (SECONDARY, prespecified and always reported, but not part of the primary verdict): labor induction rises discontinuously at age 40 in the same population. [CHANGED 2026-10-14: H2 was co-primary; Omer moved it to secondary before registration.]
- Benefit question (prespecified KEY SECONDARY outcomes at both cutoffs; they do not enter the primary verdict, which rests on H1 alone) [CHANGED 2026-10-14: these were co-primary]: cesarean delivery; delivery at 39 weeks 0–6 days; NICU admission; 5-minute Apgar < 7; a maternal morbidity composite (transfusion, 3rd/4th-degree tear, ruptured uterus, unplanned hysterectomy, ICU admission).
- Secondary: stillbirth (fetal-death files) at both cutoffs; all-parity samples; delivery at 37–38 weeks (early term).

## Data
- NCHS natality public-use files 2016–2024 via the NBER CSV mirror: https://data.nber.org/nvss/natality/csv/{year}/natality{year}us.csv (2023/2024 as natality{year}us.zip).
- NCHS fetal-death files 2018–2024: https://data.nber.org/nvss/fetaldeath/csv/{year}/fetaldeath{year}us.csv.
- Checksums in SHA256SUMS_raw. US residents only (restatus ≠ 4).
- Running variable: mother's age at delivery in single years (mager). Records with imputed age (mage_impflg = 1) are excluded; about 0.01% of records.

## Sample
- Primary: singleton births, no prior live births and no prior births now dead (nulliparous), gestation 37–44 weeks (combgest), ages within the bandwidth.
- Each outcome uses its non-missing records. Missingness by age is reported as a falsification outcome.

## Estimator
- Local linear regression in single-year age cells, separately on each side of the cutoff, with a triangular kernel; cells are weighted by births.
- Estimand: right-limit minus left-limit at the cutoff, for age a ≥ c.
- Primary bandwidth: h = 3 years (ages 32–37 for 35; ages 37–42 for 40). Sensitivity h = 2, 4, 5. A local-quadratic fit at h = 5 is also reported.
- Because age is discrete, inference uses honest confidence intervals (Kolesár & Rothe 2018, bounded second derivative). Half-width = worst-case bias bound + 1.96 SE, with M = 0.1 pp per year².
  - M is fixed a priori from published NCHS 5-year age-group rates: the largest second difference across groups is about 1 pp per 25 yr², i.e. 0.04 pp/yr². 2.5× that is used for safety.
  - Sensitivity M = 0.05 and 0.2.
- Years are pooled; a year fixed-effect version is a sensitivity check. The outcome analysis is run once, after registration.

## Falsification and density
- Covariates that should not jump at 35 or 40: bachelor's degree or higher, Medicaid payer, infertility treatment, pre-pregnancy diabetes, chronic hypertension, BMI, nulliparity share, race/ethnicity. Same honest RD; reported jointly.
- Placebo cutoffs at 31, 32, 33, 37 and 38 for the primary outcome. The placebo distribution calibrates M; if placebo jumps exceed the honest CI more than expected, M is raised.
- Density: births per single year of age; the RD in log counts at 35 and 40 is reported. No manipulation is possible because maternal age isn't chosen, but heaping or misreport would show here.
- Fuzziness note: the label applies by age at the estimated delivery date, so mothers who turn 35 between the EDD and the actual delivery are misclassified. This biases toward zero.

## Subgroups (pre-specified)
Payer (Medicaid vs private), education (BA+ vs not), race/ethnicity, year period (2016–19 vs 2020–24), and ART use.

## Decision rule (primary H1)
- Smallest effect of interest (SESOI): 1.0 pp induction (about 2.3% relative to the ~43% baseline).
- Confirm: the honest 95% CI lies entirely above 0 AND the point estimate is ≥ 0.5 pp AND the covariate falsification passes (no covariate jump outside its honest CI after Holm adjustment).
- Refute: the honest 95% CI lies entirely within (−1.0, +1.0) pp, i.e. equivalence with zero at the SESOI.
- Inconclusive: otherwise.
- H2 (secondary) is classified with the same rules using SESOI 2.0 pp. That classification is reported separately and does not change the primary verdict.
- Key secondary benefit outcomes are reported with honest CIs and Holm-adjusted p-values. "No benefit" is claimed only when the CI excludes the benefit implied by the JAMA HF 2022 estimate.

## Power (from actual counts; outcomes blind at the cutoffs)
Primary sample: 2016–2024 singleton, ≥37-week births, US residents, non-imputed age (counts_age_year.csv).
- Total births by age: 33 → 1.63M, 34 → 1.48M, 35 → 1.29M, 36 → 1.08M; 39 → 511k, 40 → 364k, 41 → 246k.
- Nulliparous: 34 → 399k, 35 → 330k; 39 → 108k, 40 → 77k.

The only outcome information used was the pooled baseline for nulliparous singletons aged 30–39 (n = 3.77M): induction 37.9%, cesarean 35.7%.

Honest power (power_honest.json; M = 0.1 pp/yr²):

| Cutoff / bandwidth | SE | Bias bound | CI half-width | MDE (80%) | Power at 0.5 / 1.0 pp |
|---|---|---|---|---|---|
| 35, h = 3, induction | 0.15 | 0.20 | 0.49 | 0.61 | 52% / 100% |
| 35, h = 4 | — | — | 0.60 | 0.71 | — / 99.9% |
| 40, h = 3 | 0.30 | 0.22 | 0.80 | 1.05 | — / 75% (2 pp: 100%) |

The refute branch (equivalence within ±1 pp at 35) is reachable whenever the estimate is within ±0.5 pp.

Falsification preview (pre-treatment covariates only; falsification_covariates.csv), jump at 35:
- BA+ −0.04 pp, Medicaid −0.08 pp, infertility treatment 0.00, pre-pregnancy diabetes −0.03 pp, chronic hypertension −0.02 pp, BMI +0.015.
- Nulliparity share +0.40 pp and log births +1.3%: both are smooth-curvature artifacts of the age distribution and are within the expected honest bias. They are reported formally with honest CIs at analysis time.
- At 40, all covariate jumps are ≤ 0.6 pp.

## Deviations policy
Any change to the sample, bandwidth, M or outcomes after registration is logged in deviations.json with its reason and date. Both the preregistered and the deviated estimates are reported. Data-quality fixes such as record-layout errors are allowed and logged; specification searches are not.

## Reproducibility
Dockerfile; `parse.py` → `counts_power.py` → `analysis.py` (written after registration). Input checksums are pinned.

## Change log
- 2026-10-14 11:00 PT (before registration): the age-40 cutoff became secondary, and the multiplicity split was dropped. H1 (induction at 35) is the sole primary test at alpha 0.05. The benefit outcomes became key secondaries. Everything else is unchanged.