Four and a Half Years of Extrapolation

A new commentary in Global Health Action argues that published child mortality estimates carry no visible trace of how old the evidence beneath them is — and that verbal-autopsy sample registration is the interim bridge, not civil registration yet.

verbal autopsychild mortalityCRVSmortality data

In the 2025 round of the United Nations’ child mortality estimates, 38 countries had empirical under-five data for 2024. Another 162 did not. For those, the model extrapolated from earlier years, and the global mean length of that extrapolation was 4.5 years. Both groups of countries then appeared in the same tables, to the same number of decimal places, in a report that ministries, donors and journalists treat as a measurement.

That symmetry is the subject of a commentary published on 26 October in Global Health Action. Augustus Osborne, a health-systems researcher at the Institute for Development in Freetown, Sierra Leone, argues that the estimates used to rank countries and allocate child-survival financing cannot signal their own weakness, because the evidence behind each figure is published apart from the figure itself. He calls a system in which the headline number looks identical whether or not new evidence has arrived one that “cannot fail visibly.”

What the survey stream was doing

The arithmetic behind the claim is not contested. Roughly 40% of the world’s deaths go unregistered, and in the WHO African Region only about one death in ten is registered, according to the WHO’s civil registration fact sheet. Where the register does not work, the Demographic and Health Surveys did: Osborne reports that DHS supplied 16% of the child mortality data points behind the UN World Population Prospects, half of them in sub-Saharan Africa, and were the only source in at least one year for 85 countries.

That pipeline was cut in 2025, when United States funding for the survey programme ended with rounds under way or ready to start in 23 countries. Interim philanthropic support, he writes, has restored selected services rather than the programme. The surrounding financing has moved the same way: official development assistance fell 23.1% in real terms in 2025, and development assistance for health dropped from $80.3 billion in 2021 to an expected $38.4 billion in 2025, in 2023 dollars. A companion analysis in Scientific Data by Wahl and colleagues frames the survey programme as global public infrastructure; a 2026 paper by Asamani and co-authors in PLOS Global Public Health puts 27 of 47 African countries in the highly or very highly vulnerable category for health financing.

WHO’s own fact sheet records what is at stake in the numbers themselves: an estimated 4.9 million children died before their fifth birthday in 2024, 2.3 million of them in the first 28 days, with 58% of under-five deaths in sub-Saharan Africa and 25% in Southern Asia. Only three countries — Niger, Nigeria and Somalia — still record under-five mortality above 100 per 1,000 live births.

Three positions, and one missing column

Osborne distinguishes three responses to the contraction, only one of which has been argued at length in print. The first holds that the estimation model absorbs the loss, since a Bayesian spline borrows strength across years and countries and widens its intervals where data thin. The second holds that surveys were the wrong instrument all along and that civil registration should take over. The third, the position with a published literature behind it, is that the survey stream must be restored quickly because nothing else can produce a national rate for most high-burden countries this decade. He labels the first two as his own reconstruction of stances implicit in practice.

His objection cuts across all three. For a minister or a financing panel, what matters is not only the interval width but the age of the last observation. Rankings and country briefings, he writes from experience, carry the point estimate and drop the interval — and an interval never states that Mali’s last direct data point is six years old while Kenya’s is two. A fifth of the world’s population lives in the 43 countries whose registration systems perform least well, and no plausible rate of improvement closes that gap before the next decade of child survival financing is allocated.

The record of the models being wrong

The commentary is careful to say it is not accusing the estimates of being wrong, only of being unfalsifiable in the form they circulate. It offers the validation record as support, and the record is mixed in an interesting way. When data collected after 2005 were withheld from the under-five model, about 92% of the left-out observations fell inside 90% prediction intervals — but the predictions ran roughly 11% high at the median. In the neonatal model, 92% of withheld observations again fell inside their 90% intervals, yet only 77% of post-2005 estimates built from reduced data lay inside the full-data 80% bounds. Both tests were run across years of steady decline; neither tested a decade-long gap opening while falling aid bends the trend being extrapolated.

Divergence between modelling groups has been large before. In 2011, UN IGME and IHME estimates for 2010 differed by more than 10 deaths per 1,000 live births in a fifth of countries, largely because the groups used different input data. Between the 2012 and 2013 rounds, estimates for 2011 shifted by more than 10 per 1,000 in 33 countries, ten of them because new data had arrived; new surveys in ten countries later differed from both groups’ predictions for 2015 by a median of 15% and 19% respectively.

The timing makes this hard to see as it happens. A full birth history supports single-year estimates for about ten years before the survey, so cancelling one round removes the newest evidence first; the damage accumulates over several rounds while every published number continues to look plausible.

The bridge that has already been built once

Osborne’s practical proposal is the part with a track record. Mozambique’s COMSA system, described in the American Journal of Tropical Medicine and Hygiene in 2023, draws a provincially stratified random sample of 700 clusters covering 828,663 people, with community agents reporting pregnancies, births and deaths continuously and verbal autopsy interviews — plus social autopsy using the pathway-to-survival framework — for cause of death. For 2019–2020 it produced a neonatal mortality rate of 23 per 1,000 live births (95% CI 18–28) and an under-five rate of 80 (69–91), against UN IGME’s model estimates for 2020 built without COMSA data of 29 (19–44) and 74 (48–107).

The cause-of-death side is where the method’s limitations show. Deaths under five were coded with two computer-coded algorithms — InSilicoVA and an expert algorithm — and the resulting fractions were calibrated against verbal autopsy and minimally invasive tissue sampling pairs from CHAMPS, 340 neonatal and 426 child records pooled across sites. InterVA, the model developed at Umeå University, is one of the widely used computer-coded options in this space, though COMSA did not use it. The authors state plainly that verbal autopsy carries more uncertainty than full autopsy or medical certification, and that results should be read cautiously at the individual level. Operationally, only about 70% of reported deaths had a completed verbal and social autopsy interview by September 2021, a third of the clusters in Cabo Delgado had become inaccessible, and the highest-income urban neighbourhoods were excluded from the sample frame because residents would not accept continuous surveillance.

Sierra Leone built a comparable platform on 678 census enumeration areas covering roughly 340,000 people, work that followed its 2018–20 nationally representative mortality survey published in Lancet Global Health. Both are, as Osborne notes, still largely donor-financed — which is the fragility he wants removed. His third proposal is institutional rather than technical: place sample registration inside the national statistics office on a domestic budget line, with partners co-financing on a declining schedule, so that the national series becomes something the global models must reconcile with and explain.

What kind of document this is

Evidence level matters here. This is a single-author commentary in a peer-reviewed journal, with no new data and no new analysis — an argument assembled from published validation studies, UN IGME methods papers and two country programmes, not a validation study of its own. The author discloses that he used a generative AI tool for drafting assistance, figure preparation and formatting, and states that he checked all content and sources. The responsible editor is named, no funding is reported and no competing interests are declared. The 4.9 million and 2.3 million figures are modelled estimates with uncertainty, not counts; the Mozambique rates come from two years of a young surveillance system.

What the piece does well is separate two questions that get merged in practice: whether the estimates are close to the truth, and whether anyone downstream can tell how much evidence stands behind a given figure. Verbal autopsy enters the argument not as a solution to cause-of-death accuracy but as the component that makes sample registration possible at national scale — the mechanism by which a country can produce its own mortality series while its register is still being built. That is a more modest claim than the technology’s advocates usually make, and a more durable one. The wider pattern of survey contraction and strained registration is one this digest has followed before, most recently in Burkina Faso’s conflict-affected sampling frame.

interva.net is an independent digest and is not affiliated with the InterVA project, Umeå University, or WHO.

Sources

  1. Osborne A. Estimates that cannot fail visibly: a current debate on measuring child mortality as household surveys contract. Glob Health Action 2026;19(1):2752841 (commentary)
    pmc.ncbi.nlm.nih.gov
  2. World Health Organization. Child mortality (under 5 years) — fact sheet
    who.int
  3. World Health Organization. Civil registration and vital statistics — fact sheet
    who.int
  4. Macicame I, Kante AM, Wilson E, et al. Countrywide Mortality Surveillance for Action in Mozambique: results from a national sample-based vital statistics system for mortality and cause of death. Am J Trop Med Hyg 2023;108:5-16
    pmc.ncbi.nlm.nih.gov
  5. Carshon-Marsh R, Aimone A, Ansumana R, et al. Child, maternal, and adult mortality in Sierra Leone: nationally representative mortality survey 2018-20. Lancet Glob Health 2022;10:e114-e123
    doi.org
  6. Wahl B, Menon GI, Nyhan K, et al. Global health data precarity: safeguarding the Demographic and Health Survey program as a global public good. Sci Data 2025;12:1827
    doi.org
  7. Asamani JA, Faye S, Boris Bediakon KS, et al. Measuring health financing vulnerability due to reductions in official development assistance: a conceptual framework with empirical application across 47 African countries. PLOS Glob Public Health 2026;6:e0006282
    doi.org
  8. World Health Organization. 2022 WHO verbal autopsy instrument
    who.int

← Back to the news digest