Running the maternal mortality model without the surveys
A WHO-led modelling study reruns the UN's maternal mortality estimates with the Demographic and Health Surveys removed. The global ratio falls 15% and the uncertainty around it widens 70% — a measurement of how much of the number was borrowed.
A maternal mortality ratio is a fraction with two uncertain halves. The numerator counts deaths that somebody had to notice, define and record; the denominator counts births that somebody had to register. When the UN’s inter-agency group published its 2000–2023 round, the headline was 260,000 maternal deaths in 2023 and a global ratio of 197 per 100,000 live births, with sub-Saharan Africa carrying around 70% of the total and low- and lower-middle-income countries 92%. In the settings that produce most of those deaths, neither half of the fraction is counted in the ordinary sense. Both are assembled — from civil registration where it functions, and from household surveys where it does not.
A modelling analysis published in BMC Medicine on 16 November asks what happens to that assembly when one of its main inputs goes away. Sahar Ahmed and colleagues, most of them at the World Health Organization’s Human Reproduction Programme in Geneva, with co-authors at the WHO regional offices for Africa and South-East Asia and at the Centre d’Estudis Demogràfics in Barcelona, did not forecast the future of survey funding. They did something more clinically precise: they took the model that produces the official numbers and ran it without the survey data.
The question that produces the number
The Demographic and Health Surveys began in 1984, and maternal mortality entered the programme in 1988 through sibling survival histories — the sisterhood method. A respondent is asked whether each of her sisters is alive, and, for those who died, whether the death occurred during pregnancy or within a year of its end. The module commonly reaches back seven years, which is how a rare event becomes countable in a single household visit. It also produces large standard errors, and it produces no cause of death at all.
Two definitional adjustments follow from that design, and the paper sets them out plainly. Sibling reports miss deaths very early in pregnancy, before a woman or her sister knew she was pregnant, so estimates are revised upwards by 10%. And respondents report pregnancy-related deaths — any death during pregnancy or within 42 days of its end, whatever the cause — rather than maternal deaths, so the figures are adjusted downwards by 10% in sub-Saharan Africa and 15% elsewhere. The ratio that reaches the front of a WHO report has already been through both corrections.
Three runs of one model
The engine is BMat, the Bayesian maternal mortality estimation model that combines empirical observations with covariates, regional patterns and a global trend line. Where country data are sparse, the estimate leans on the covariates and on neighbouring countries; the thinner the direct evidence, the wider the interval. The authors ran it three times with structure and priors held constant: once with everything, once with all DHS inputs removed, and once with only the surveys conducted from 2015 onward removed — the scenario that mimics a programme that simply stops.
In the full dataset, 174 of the 193 population-based surveys are DHS. They supply 21% of all country-year observations, second only to civil registration. They exist for 64 countries; in 23 of those they are the only maternal mortality data the model has. Thirty-five post-2015 surveys cover 31 countries, and 21 of those countries have no other post-2015 source at all. In sub-Saharan Africa the DHS is not a supplement but the principal input, at 80% of observations, with 18 countries relying on it exclusively; for low-income countries the share of country-years is 90%.
What the arithmetic does without them
Remove every DHS input and the global 2023 ratio drops 15%, while the 80% uncertainty interval widens by 70%. Sub-Saharan Africa falls 14% with uncertainty up 76%; Oceania falls 36%. Regions with mature civil registration — Europe and Northern America, Australia and New Zealand, Northern Africa and Western Asia — barely move, which is the study’s own internal control.
The country detail is where the result stops being abstract. In countries where DHS is the sole source, deleting it usually produces a lower estimate and a wider interval; Afghanistan, Gabon and Indonesia shift by 58–76%, and for Gabon and Indonesia the before-and-after intervals do not even overlap. Mauritania and Togo move the other way. Among countries with several non-CRVS sources, nine estimates rise when the surveys go (Mozambique and Nepal among them) and thirteen fall (Ghana, Senegal); Zimbabwe’s interval more than doubles in width. The post-2015 scenario produces a 10% global fall, a 50% fall for Afghanistan, and — in Angola — an interval up to 90% wider.
One result runs against the grain and the authors decline to smooth it over. In Eastern and South-Eastern Asia, removing DHS data produced lower estimates and narrower intervals. They state explicitly that a narrower interval is not better precision but evidence that heterogeneous observations have been removed, and that what remains is a set of inputs disagreeing less with each other because the awkward one is gone.
The 42-day edge
Nothing in this paper concerns verbal autopsy, and that is worth saying before drawing the connection. But both instruments ask a surviving relative to describe a death, and the difference between them is what happens to the answer. The sisterhood module converts it into a ratio. A verbal autopsy interview — the WHO’s 2022 instrument is the current standard — converts it into a cause, and into the pregnancy and postpartum questions that establish whether the death was pregnancy-related in the first place.
What that second instrument shows is that the 42-day boundary, which the ratio depends on, cuts through a great deal of mortality. In a BJOG analysis of verbal autopsy data from ten health and demographic surveillance sites in The Gambia, Kenya, Malawi, Tanzania, Ethiopia and South Africa, Ulla Gazeley and colleagues examined 2,114 deaths recorded within one year of the end of a pregnancy between 2000 and 2019. Of these, 1,212 occurred within 42 days and 902 — 43% — between 43 and 365 days. Causes were assigned with InterVA-5 and InSilicoVA, two of the algorithms documented on the InterVA model page, and grouped to an adapted ICD maternal-mortality classification. Beyond 42 days, HIV and TB, other infectious diseases and non-communicable conditions took significantly larger shares, with health-system failures prominent among the reported circumstances; the HIV and TB contribution was greatest in southern Africa. These are algorithm-assigned causes, not certified ones, and the paper treats them as such — but the finding is about a definitional window, not about an algorithm’s skill.
So the ratio a country reports counts deaths inside a 42-day frame, and the survey that supplies the frame may be disappearing. The mortality outside the frame is visible only where a different instrument is in routine use.
What the scenarios are not
The authors are unusually direct about the limits. These are hypothetical runs, not predictions; the differences between scenarios are attributed to the missing DHS inputs, but the model is complex enough that shifts in the influence of the remaining data and of the priors could contribute. DHS also feeds several of BMat’s covariates — fertility, skilled birth attendance, HIV mortality — and those were left in place, so the study understates the total dependence. The reference scenario assumes every included source, DHS among them, is valid and unbiased, which the paper calls an oversimplification.
Their reading of the substitutes is equally sober. Civil registration remains the standard, but progress toward it is gradual, and a near-term gap is likely. Routine health information systems can help where facility coverage and denominators hold up. Maternal and perinatal death surveillance and response produces detailed investigations of individual deaths but, on its own, is often not suited to population-level estimation. The reason the missing surveys matter is not that DHS was a good instrument. It is that nothing in the queue is ready to take over.
The companion problem on the child-mortality side — where an absent survey stream also disappears silently rather than loudly — is one this digest covered in an earlier entry on estimates that cannot fail visibly. The finding here is the maternal version, and it has the same shape: when the input is removed, the model still returns a number, with an interval around it, and the number is different. Nothing changed in how many women died.
interva.net is an independent digest. It is not affiliated with the InterVA project, Umeå University, or WHO, and nothing here is medical or policy advice. Evidence levels: the BMC Medicine paper is a peer-reviewed modelling analysis of hypothetical data-availability scenarios, not a forecast and not a validation of any estimate; the WHO fact sheet and the inter-agency report are the source of the official 2023 figures; the BJOG paper is an open-population cohort analysis of verbal autopsy records with algorithm-assigned causes and no medical certification as reference.
Sources
- Ahmed SMA, Cresswell JA, Aujla M, Danansuriya M, Aderoba AK, Ouedraogo L, Say L. The future of maternal mortality measurement: implications of reduced DHS availability. BMC Medicine 2026;24:649 (published 16 November 2026) bmcmedicine.biomedcentral.com
- WHO — Maternal mortality fact sheet (2023 estimates: 260 000 deaths, MMR 197 per 100 000 live births) who.int
- WHO, UNICEF, UNFPA, World Bank Group, UNDESA/Population Division. Trends in maternal mortality 2000 to 2023 (Geneva, 2025) who.int
- Gazeley U, Reniers G, Romero-Prieto JE, Calvert C, et al. Pregnancy-related mortality up to 1 year postpartum in sub-Saharan Africa: an analysis of verbal autopsy data from six countries. BJOG 2024;131(2):163-174 (first published online 19 July 2023) pmc.ncbi.nlm.nih.gov
- WHO — 2022 verbal autopsy instrument (V2.0, multilingual) who.int