Maternal mortality: the stall was COVID; the real story is the denominator

Status: drafted 2026-08-10 · signed off 2026-08-10 · REVISED 2026-08-10 after the data-vintage follow-up — the verdict changed.

This is the theme bank's first verified row.

⚠️ Revision note — read this first

The signed-off v1 of this sheet was built on OWID's maternal-mortality grapher, which ends at 2020 and is still on the superseded UN MMEIG (2023) release (lastUpdated: 2024-07-08, nextUpdate: 2026-09-10).

The follow-up on open question 4 found that MMEIG published a newer release — Trends in Maternal Mortality 2000–2023 (2025) — extending the series to 2023, and that it reverses three of v1's four findings. They were artefacts of a series that stopped exactly on the COVID spike.

v1 finding (data to 2020)	Status on data to 2023
Progress "stalled" after 2015 (−3.3%)	Refuted as structural — 2015→2023 is −13.6%; the flattening was COVID
US rising, a rich-country outlier	Refuted — 17 (2015) → 22 (2020) → 17 (2023); the spike reversed
High-income countries rose +13.0% as a bloc	Refuted — 2015→2023 is −9.1%; also COVID
Composition drag masks country-level progress	Confirmed and strengthened — drag grows from +5.8% to +9.0%

Standing lesson: check the release vintage before the numbers. A series ending in 2020 makes COVID look like a structural break. Had this gone to Stage 2 as signed off, the episode's central claim would have been refuted by data published a year before we recorded it.

Segment: Maternal mortality — the stall was COVID; the real story is the denominator

• Source theme(s): Batch 25 #11. Cross-refs: Batch 25 #1 (child mortality, published), Batch 20 progress-arc cluster.
• Spine role: main story
• The framing people carry: The world made huge progress on women dying in childbirth, then progress stopped around 2015 because donors lost interest — and the United States, uniquely among rich countries, is going backwards.
• Hypothesis (from the bank): "The MMR falling 2000-2015, flattening 2016-2020, with sub-Saharan Africa carrying most of the remaining burden; the US as a rich-country outlier rising; the stall coinciding with stagnating aid for reproductive health — an intervention gap, not a knowledge gap."
• Verdict: REFUTED as framed — with a stronger replacement story. The decline and the burden concentration confirm. The stall was a pandemic, not a plateau, and it has since resumed. The US reversal is gone. What survives, and strengthens, is a claim the bank never made: global progress is being partly masked by where the world's babies are born.

The numbers (every row sourced — no exceptions)

Primary source (current): UN MMEIG, Trends in Maternal Mortality 2000–2023 (2025 release), retrieved via the World Bank API, indicator SH.STA.MMRT, accessed 2026-08-10. Tier B today — see "Datoid artifact" for why it is not yet Tier A. Births: OWID grapher number-of-births-per-year (UN WPP, estimates variant), accessed 2026-08-10. Tier A.

The texture (what narrative flattens)

• The lead finding: the denominator is eating the progress. Every income group improved 2015-2023 — low-income fastest at −29.5%, lower-middle −19.2%. Yet the world improved only −13.6%, slower than any of its large components. That is arithmetically impossible unless the mix moved, and it did: the low-income share of world births rose 15.2% → 19.2%. Hold the 2015 birth-mix fixed and the world would have improved −21.6% instead of −14.5%. About a third of the achievable improvement is offset by where babies are now born. Every country can improve while the world barely moves — the honest version of "are we making progress?"
• The stall was a pandemic, not a plateau. 2015→2020 looked like a −3.3% flatline. Extended to 2023 it is −13.6%, with the 2020-21 disruption visible as a bump that reversed. Any framing built on "progress stopped in 2015" is reading COVID as structure.
• The US spike also reversed — 17 → 22 → 17. The rich-country-outlier story does not survive current data.
• Burden concentration confirmed and stark: sub-Saharan Africa at 448 against a world figure of 197, and ~70% of all maternal deaths.

Datoid artifact (Tier-A only)

• No existing artifact covers this. Global Health Priorities is malaria + measles + child mortality.
• ⚠️ Timing problem, and it is the operational finding of this exercise. Our OWID connector reaches maternal-mortality, but OWID is on the superseded release ending 2020 and does not refresh until nextUpdate: 2026-09-10. Building the artifact today from our own connector would put the refuted story on a Datoid chart — a stall that isn't one, and a US reversal that reversed back. RESOLVED 2026-08-10 — option 2 taken: the World Bank connector is built. src/lib/server/connectors/worldbank-fetch.ts + worldbank.ts, keyless, wired into poll-source.ts. scripts/seed-worldbank-maternal.ts seeds a draft maternal-health dataset with 8 SH.STA.MMRT series (World, 4 income groups, Sub-Saharan Africa, US, Australia), 1985–2023, values matching this sheet exactly. Every number in the table above is now Tier A. The options as considered were:
  1. Wait for OWID's 2026-09-10 refresh, then build entirely Tier-A.
  2. Ingest SH.STA.MMRT from the World Bank API directly ← taken. Keyless and stable, it unlocks a large class of development indicators beyond this theme (verified: crude birth rate, adult literacy, unemployment, electricity access all fetch cleanly), and it is a standing vintage hedge against OWID's lag rather than a one-theme fix.
  3. Build now from OWID and caveat the endpoint. Rejected — publishing a chart whose story we know to be an artefact is exactly what Stage 1 exists to prevent.
• RESOLVED 2026-08-10 — the composition chart is now buildable. It needs birth weights per income group, which nothing supplied: the OWID connector hardcoded cols[0] !== "World" and so could only ever ingest the global aggregate. Its externalId is now "<slug>" or "<slug>@<entity>" (bare = World, preserving all 16 existing series), mirroring the World Bank connector. Four number-of-births-per-year@<income group> series are seeded alongside the MMR set, 1950–2023. Low-income birth share reproduces exactly: 25.29M / 131.67M = 19.2% in 2023. Rejected alternative: deriving births from World Bank SP.DYN.CBRT.IN × SP.POP.TOTL — both fetch fine, but that is a computed series we have no mechanism for.
• The chart is two-axis by construction: MMR (rate, per 100k) from the World Bank against births (count, millions) from OWID. Both land on mid-year timestamps, so they share an x-axis cleanly.
• The births-composition decomposition is the chart nobody else draws and is the reason to build this here rather than link out. It needs both MMR and births on one surface.

Story justification (Stage 1 rule 5)

• Does the argument have a reveal that needs sequencing? YES → build.
• Because: the punchline is a second dataset. World falls slowly → every income group falls faster → that is arithmetically impossible → reveal the births shifting. A dashboard panel can state that; only a sequenced reveal lands it. Built as /stories/maternal-mortality.
• Building it exposed a framework gap and forced a fix: StoryStep previously controlled only the x-domain, so every required series rendered on every step — the punchline was visible in step one. activeSeriesSlugs now lets a story introduce a dataset when the argument needs it.

Open questions / where I'm guessing

These must NOT become spoken assertions.

1. US figures still carry a measurement caveat. OWID annotates every US row "Values from 2003–2017 affected by measurement change" (the phased pregnancy-checkbox rollout). The World Bank/MMEIG series is modelled and may handle this differently, but I have not verified how. Do not make precise US-magnitude claims across 2003-2017 from either source without resolving this.
2. The composition decomposition is indicative, not exact. Four income groups do not sum to the world, and MMEIG's own aggregation may differ. The actual weighted change (−14.5%) tracks World Bank's World (−13.6%) closely, which is reassuring — but say "about a third" and "most of", never a precise share, until reproduced against MMEIG's own aggregation.
3. The bank's causal claim — "stagnating aid" — remains unsourced and now looks clearly wrong as the primary mechanism: low-income countries improved fastest. Cut it, or replace it with a sourced claim.
4. Australia's numbers are small-N and rounded to integers in the World Bank series (6 → 2). At ~300k annual births an MMR of 2 is order-6 deaths. The 30-year direction is real; the precise ratio is not. Do not say "3× better than 2015".
5. I have not read the MMEIG 2025 report itself, only its headline figures and the World Bank's published series. Before Stage 2, at least skim it for methodology changes between the 2023 and 2025 releases — a revision to history (2000 moved 334.1 → 328) means the two releases are not interchangeable mid-chart.