What the Dead Had Been Told

A decade of verbal autopsy records from Agincourt, South Africa, was analysed not for causes of death but for the diagnoses people carried while alive. Hypertension among the dead went from a quarter to about half, and the dominant cluster moved from HIV–TB to cardiometabolic disease.

verbal autopsyNCDsfield studiesdata quality

Somewhere in a standard verbal autopsy interview there is a question about whether the person who died had ever been told by a health worker that they had high blood pressure. There are near neighbours in the schedule: diabetes, tuberculosis, HIV, asthma, cancer, dementia, kidney disease. In the ordinary use of the verbal autopsy method, those answers are grist — they help an algorithm or a physician arrive at a single line, one underlying cause, and then they are put away.

A paper published in July in BMC Public Health takes the opposite approach. Its authors, from the SAMRC/Wits Rural Public Health and Health Transitions Research Unit at Agincourt, analysed eleven years of verbal autopsy records on 5,892 people who died in rural Mpumalanga, South Africa, and set the cause of death aside entirely. What they counted instead was the illness history the bereaved families had reported: what the deceased had lived with, whether or not it appeared on the death certificate.

The authors are explicit that this is an unusual use of the material. “Our focus was not on causes of death but on the disease conditions that the deceased experienced during their lifetime as reported by their caregivers,” they write, describing it as a novel application of a questionnaire designed for something else.

Eleven years, cut into three windows

Agincourt has been running demographic and verbal autopsy surveillance since 1992, with trained fieldworkers interviewing caregivers at regular household visits and WHO questionnaires in use throughout. The study window runs from 2012 to 2022 and is divided into three periods — 1,682 deaths in 2012–2015, 2,127 in 2016–2019 and 2,083 in 2020–2022 — chosen before the analysis began, for a mixture of reasons: South Africa’s move to universal HIV test-and-treat in 2016, updated national NCD screening guidance in 2019, the pandemic, and the need for at least 1,500 deaths in each period to keep the statistics stable.

Thirteen long-term conditions were carried through, plus three behaviours (alcohol, smoking, drug use). Because the site changed instruments between the two WHO questionnaire generations, the team harmonised the relevant fields from the raw interview responses rather than from derived variables, so that 2012 and 2016 answers could be compared on the same footing. Records that reported none of the thirteen conditions were dropped from the clustering step; “unknown” and blank answers were treated as missing, not as an absence of disease.

From a quarter to about half

Among the dead, reported hypertension rose from roughly 25% in 2012 to about 52% by 2021 and above 50% by 2022. Reported diabetes climbed from 10% to around 25%. Dementia, close to invisible at the start, reached about 6%. Tuberculosis went the other way — from roughly 23% to under 10% — while HIV held flat at around a fifth of the deceased, between 18% and 22%, across the whole period.

The combinations matter more than the individual lines. In 2012, 55% of the deceased had one reported chronic condition and 8% had three or more; by 2022 the single-condition share had slipped to 50% and complex multimorbidity had doubled to 15% by 2019, then levelled off through 2020–2022. Among the 54–64 age group, two or more conditions went from 25% to over 40% in the same span. Deceased women carried more: hypertension reached 85.5% among women aged 65 and over, and complex multimorbidity in women rose to 22% by 2020.

To see whether those conditions travelled together in recognisable patterns, the authors ran two clustering methods side by side — partitioning around medoids on Gower distances, and latent class analysis — and compared the resulting groupings rather than trusting either alone. Both told the same story of a handover. In 2012–2015, the largest latent class was hypertension with HIV (43.8%), with an HIV–TB class at 28.4% and a mixed infectious–cardiometabolic class at 27.8%. By 2020–2022 a cardiometabolic class dominated at 43%, with dementia displacing heart disease inside it; hypertension alone accounted for 35.5%, and the HIV–TB–hypertension class had shrunk to 21.5%. The medoid clustering put hypertension alone at 29.9% in the final period, hypertension with diabetes at 20.6%, a dementia cluster at 6.4% and cancer alone at 3.3%.

The word is “reported”

None of those figures is a measured prevalence. Each one rests on a caregiver’s recollection that the deceased had been told something by a clinician, and that recollection is shaped by how much contact the household had with the health system in the first place. The authors list the consequences themselves: misclassification is plausible for HIV, for mental illness and for early-stage NCDs that were never diagnosed; the dementia curve could reflect a real increase, better recognition by families, or subtle changes in how the questionnaire probed cognitive symptoms across instrument versions.

Two further caveats are worth holding onto. The 2020–2022 period coincided with COVID-19, when South African lockdowns disrupted routine care and the site moved part of its interviewing to the telephone — a change the authors say “may have introduced modest measurement differences for certain conditions”. And the regression models that linked cluster membership to age, sex, marital status and substance use had to leave out education, socioeconomic status and body mass index, because the standard verbal autopsy questionnaire does not collect them. Age turned out to be the only consistent predictor, and only in the final period; several of the earlier models produced odds ratios so large, or reference cells so thin, that the authors flagged them rather than interpret them.

The sample is also, by construction, a sample of the dead. These are the severe combinations that ended in a death, not a picture of what chronic illness looks like among people still living with it.

A record of the health system, as well as of the death

Read with those limits in view, the study is less a burden-of-disease estimate than a retrospective audit of who had been diagnosed and who had not. An earlier Agincourt analysis of 1993–2013 documented the site’s shift away from HIV/AIDS mortality as treatment scaled up; this one pushes the same transition down to the level of the individual diagnosis, and finds that by the early 2020s hypertension and diabetes were present in the histories of half or a quarter of the dead respectively, even in a rural setting with a heavy infectious-disease load.

That is information a health system can act on in a way that a single cause column cannot, and it is available at no extra cost from interviews already being conducted. It also sharpens an argument the field has been making for years: a verbal autopsy record contains more than the input to algorithmic cause assignment, and more than the one cause the NCD estimates for Africa are built from. What it holds is a household’s account of the care its member did and did not receive.

Evidence level: a single-site retrospective analysis of eleven years of verbal autopsy surveillance data, peer reviewed and open access, descriptive and associational. The chronic-disease variables are caregiver-reported diagnoses, not clinically verified morbidity; the authors report no validation against medical records, and their conclusions about health-system integration are their interpretation of the pattern rather than a tested intervention.

interva.net is an independent digest and observatory. It is not affiliated with, endorsed by, or speaking for the InterVA project, Umeå University, or the World Health Organization.

Sources

  1. Chironda CT, Chirwa T, Mlambo F, Mupfuti R, Ohene-Kwofie D, Njiro BJ, Bashingwa JJH, Tollman S, Kahn K, Gomez-Olive FX, Kabudula CW. Evolution of chronic diseases and multimorbidity clusters among the deceased in rural northeast South Africa (2012–2022): insights from verbal autopsy data. BMC Public Health 2026;26:3005 (published 23 July 2026; open access; peer reviewed)
    doi.org
  2. Kabudula CW, Houle B, Collinson MA, Kahn K, Gómez-Olivé FX, Clark SJ, Tollman S. Progression of the epidemiological transition in a rural South African setting: findings from population surveillance in Agincourt, 1993–2013. BMC Public Health 2017;17:424 (PMC5424387)
    doi.org
  3. World Health Organization. 2022 WHO verbal autopsy instrument (V2.0, multilingual)
    who.int

← Back to the news digest