How the numbers behind the maps and charts are measured — and where they come from.
← Back to the mapA desert is a desert now. The map does not distinguish a county that just lost its last delivering hospital from one that never had one. If no hospital delivers babies there today, it is black. Drawing that line any finer would make the picture look better than the reality it describes.
U.S. Centers for Medicare & Medicaid Services, Provider of Services (POS) file — the federal census of Medicare/Medicaid-certified facilities. Hospitals reporting obstetric service are aggregated to their county (FIPS code). Pre-2011 snapshots are the archived POS files redistributed by the National Bureau of Economic Research (public domain, U.S. Government work).
Twenty maternity units flagged as closed/closing (confirmed) or predicted to close within 12–24 months, identified from hospital financial signals rather than press announcements. Directional, not a roster.
As of 2025, more than one in three U.S. counties — 1,076 of the 3,107 counties in the lower 48 — has no hospital that delivers babies.
Over this century the rate falls from the high hundreds (the U.S. reached 916 in the 1918 influenza pandemic) to single digits, then the U.S. turns back up. A linear axis tall enough to hold 916 would crush the entire modern era — and the ~1990 divergence — into an invisible smear at the bottom. A logarithmic axis holds a 900-to-5 collapse and a modern 4× gap on one chart, so equal vertical steps read as equal ratios. It is the standard way this series is shown.
The pre-1960 compilation is spliced to the OECD series at 1960; most countries meet cleanly (the U.S. joins at 37.4 → 37.1). The U.K. shows a small step at the seam where the source changes.
The upward turn of the U.S. line after 1990 is robust; its precise height is correction-sensitive. The death-certificate pregnancy checkbox, phased in across states 2003–2017, raised recorded U.S. counts, and no continuous official rate was published during that transition — the dashed 2007–2018 segment is interpolated between real endpoints, not observed. On raw national rates the 2020 U.S. figure is roughly 5× the peer average; corrected cross-national comparisons put it nearer 3–3.5×. Tiny-population years (Norway, Sweden) swing on a handful of deaths and are clamped at the axis floor; follow the band, not a single point.
Those tiny-population swings are noise, and a log axis magnifies them until they bury the shape of the century. The peer-nation lines only are therefore smoothed for legibility: a centred 5-year weighted moving average, computed in log space because the axis is logarithmic. The window shrinks symmetrically at each end of a run of data, so the first and last real points are drawn at their real values, and a gap wider than the window ends a run, so an average never spans missing years.
The United States line is not smoothed — it is drawn from raw annual data in all three of its segments, so the 1918 influenza peak stands at its recorded 916.4 on 1918, and every U.S. anchor (1915 = 607.9, 2007 = 12.7, 2018 = 17.4, 2020 = 23.8) is exact. Smoothing is a display transform applied to the peer band only; the underlying annual values are unchanged. Append ?smooth=1 to the figure’s URL to draw the peer nations raw as well.
Between 2016 and 2024 the recorded rate in that low-risk group rose from 3.84% to 4.50%, a 17.2% relative increase. Among babies born before 37 weeks it rose 5.0%. The increase is largest in the group least likely to need intensive care, and it is broad: the rate rose in 46 of 51 states and for every payer.
CDC WONDER, Natality 2016–2024 expanded (dataset D149), variable “NICU Admission”, crossed with gestational age, birth weight, plurality and source of payment. Public data, standard click-through only — no data-use agreement. Pulled 8 August 2026. The series starts in 2016 because the 2003 revised birth certificate, which carries the NICU item, only reached all 50 states and DC that year.
Babies recorded Yes divided by (Yes + No). Certificates where the box was left blank are excluded from the denominator rather than counted as “No”, so incomplete forms cannot move the rate. Blanks are small and stable: 0.11% of births in 2016, 0.19% in 2024.