October 2, 2026 interva.net Editorial

A Tissue Sample and a Grieving Interview, Side by Side in 8,500 Child Deaths

Seven countries, nine years, 8,500 deaths investigated with postmortem sampling: a CDC-led surveillance report shows how much of a child's death a household interview can and cannot account for.

verbal autopsychild mortalitysurveillancecause of death

Most of what the world knows about why children die in poor countries comes from an interview. A trained fieldworker sits with a mother or grandmother weeks after a death, works through a long questionnaire about symptoms, feeding, care sought and travel to a clinic, and a coding method turns the answers into a probable cause. The whole method rests on a substitution: the memory of a household stands in for a laboratory. That substitution is usually the only one available.

On 10 September 2026, a surveillance report published in MMWR Surveillance Summaries put a number on what the substitution costs. The Child Health and Mortality Prevention Surveillance network — CHAMPS, operating in Bangladesh and six African countries — investigated deaths among stillbirths, newborns and children under five using minimally invasive tissue sampling, laboratory testing, clinical and maternal records, and a verbal autopsy interview, then had multidisciplinary panels assign causes from the combined evidence. Between December 2016 and December 2024 the network identified 18,784 eligible deaths, enrolled 15,612 of them (83.1%), completed tissue sampling for 9,415 and finished a full cause-of-death determination for 8,500.

The report is explicitly a descriptive surveillance summary rather than a validation study of any coding algorithm. But because it ran interviews and postmortem specimens in parallel and published what each contributed to the final cause, it reads as a measured account of the mental model that verbal autopsy alone imposes on a death.

One cause, or several

The report’s own framing of the contrast is unusually blunt. Conventional cause-of-death attribution in low- and middle-income countries, it notes, “rely largely on interviews with family members or other lay respondents and are designed primarily to assign a single probable underlying cause on the basis of limited information,” and “generally do not identify multiple coexisting or contributory conditions.”

The panels did the opposite. Each dossier could hold roughly 120 laboratory results and thousands of discrete clinical, maternal, pathology and interview data elements. Panelists assigned an underlying cause, and where relevant one or more antecedent and an immediate cause, applying counterfactual reasoning — would this death have happened, at that moment, without this condition? Every condition placed anywhere in the causal chain counted, so the reported proportions overlap and can sum past 100%.

What that produced is a picture of deaths that were rarely about one thing. Among newborns, 44.3% had two or more conditions in the causal chain. Among infants and older children, 67.9% did — with malnutrition, infection and a specific pathogen often sitting in the same sequence.

The findings, age group by age group

Of the 8,500 deaths analysed, 3,199 (37.6%) were stillbirths, 3,230 (38.0%) were newborn deaths and 2,071 (24.4%) were deaths of infants and children up to 59 months. Deaths of the youngest dominated the sample: stillbirths and neonatal deaths together made up roughly three quarters of everything investigated.

For stillbirths, perinatal asphyxia or hypoxia appeared in the causal chain of 79.1% — more than 85% of cases in Bangladesh, Kenya and Sierra Leone — alongside congenital birth defects in 9.0% overall, a share that reached 23.7% in Ethiopia, mostly neural tube defects. Among newborn deaths the leading conditions were complications of preterm birth (39.7%), asphyxia or hypoxia (37.7%) and sepsis (36.5%); preterm complications and asphyxia tended to be the underlying cause, while sepsis more often appeared earlier or later in the chain.

For infants and older children the leading contributors were lower respiratory infection (37.4%), sepsis (36.9%), malnutrition (27.3%), malaria (22.1%) and diarrhoeal disease (17.3%). Underlying causes looked different — malnutrition 20.8%, malaria 15.5%, lower respiratory infection 9.1% — because the panels assigned the condition that started the sequence, not the one that finished it.

What the specimens named

Pathogen-level detail is where an interview has nothing to offer at all. Among the 1,234 newborn deaths with an organism in the causal chain, 85.7% involved gram-negative bacteria: Klebsiella pneumoniae in 48.3% and Acinetobacter baumannii in 25.9%, with the latter disproportionately frequent in South African sites. Among older infants and children with a pathogen implicated, the mean was two organisms per death, with Plasmodium falciparum in 26.9%. Postmortem anthropometry added a further dimension no questionnaire records: among children with measurements, 61.3% were moderately or severely underweight, 61.3% were wasted and 43.0% were stunted.

The maternal thread that interviews miss

Upstream of the child, the panels assigned a main maternal condition for 71.1% of stillbirths and 58.2% of newborn deaths, mapped onto what the report calls the ICD-PM maternal condition framework. Hypertensive disorders of pregnancy and placental complications — abruption and other separation with haemorrhage — each accounted for 16.8% of stillbirths. Where placental, cord or membrane complications were present, asphyxia or hypoxia was the underlying cause in 71.1% of those deaths.

That is precisely the causal territory verbal autopsy handles least well. A household respondent can describe a swollen face or a convulsion; the interview cannot reliably separate pre-eclampsia from placental abruption from intrapartum hypoxia, and the standard instruments are not built to reconcile a maternal condition with a fetal one. Reported through interviews, the same cluster of deaths tends to collapse into a single perinatal category.

Where the postmortem network is thin

The report is candid about what its own design leaves out, and the omissions point exactly at the population verbal autopsy serves. Because tissue sampling demanded notification and specimen collection within 24 hours of death — 72 if the body had been refrigerated — deaths in health care facilities were more likely to be investigated. They were 88.2% of the analysed sample. The authors write that community deaths, deaths with delayed notification, “deaths evaluated primarily through verbal autopsy,” and some injury deaths “might be underrepresented.” Catchment areas, they add, may not be nationally representative.

Two further caveats matter for how the findings travel. Pathogen attribution in postmortem surveillance is hard, and the report says its aggregate organism figures should sit alongside organism-specific analyses. And the headline preventability finding — of 7,558 deaths with a classification, 80.7% were judged preventable or possibly preventable — is an expert judgement about plausible changes in care, not a causal estimate of how many lives any single intervention would save.

The interview keeps its territory

What CHAMPS demonstrates is a reference standard, not a replacement. Seven countries with defined catchment areas, mortuary laboratories, trained sampling technicians, telepathology support and multidisciplinary panels is a sentinel system; verbal autopsy is what runs at national scale inside civil registration and mortality surveillance, and it is the only instrument available for deaths that never reach a facility or a mortuary. The 2022 WHO verbal autopsy instrument V2.0 remains the standard questionnaire for that work, building on the coding framework the WHO set out two decades ago, and the method’s known limitations — recall, lay reporting, a single underlying cause per death — were described in the literature long before this report. Garenne and Fauveau’s 2006 review of the “potential and limits” of verbal autopsies in the Bulletin of the World Health Organization made much the same argument from far less data.

The CHAMPS authors suggest select components of their model could be adapted for sentinel surveillance elsewhere. Read alongside the interview-based studies that carry most national cause-of-death estimates, the report does something useful without settling any algorithmic question: it shows what a single assigned cause flattens, and roughly how much.

interva.net is an independent digest of verbal autopsy and cause-of-death research. It is not affiliated with, endorsed by, or speaking for the InterVA project, Umeå University, the WHO, the CHAMPS network or the US Centers for Disease Control and Prevention.

Sources

← Back to the news digest