October 23, 2026 interva.net Editorial

Every Death, Every Fifteen Weeks — and Still No Cause Column

A rural KwaZulu-Natal surveillance site counted 6,181 adult deaths over seven years and found pandemic mortality roughly 60% above expectation. The paper's own limitations section explains why that number cannot be turned into a cause.

verbal autopsyCOVID-19surveillancemortality data

Fieldworkers in the uMkhanyakude district of northern KwaZulu-Natal walk a household round about every fifteen weeks. They ask who was born, who moved, who died. Household refusal runs under 3%. Because South African law requires every death to be registered within 72 hours, the surveillance record can be cross-checked against the national death registry, and the Africa Health Research Institute, which runs the site, describes a multi-step validation process that includes verbal autopsy classification for cause-of-death determination.

All of that machinery sits under a paper published on 16 October in Population Health Metrics, and the paper uses only one column of it: the death count. Mazibuko and colleagues took 130,710 adults aged 15 and over, followed them across 652,091 person-years from January 2015 to December 2021, and worked with the 6,181 deaths recorded in that window — an overall rate of 9.5 deaths per 1,000 person-years (95% CI 9.2–9.7).

What the monthly counts did

The design is an excess-mortality model, not a case series. Monthly death counts were fitted with a negative binomial regression — chosen after a likelihood ratio test showed the counts were overdispersed — with calendar year as a linear trend, three Fourier sine-cosine pairs to absorb seasonality, and an offset for the population at risk. The model was fitted on the pre-pandemic period only, from 1 January 2015 to 4 March 2020, and then used to predict what the district should have seen between 5 March 2020 and 31 December 2021.

Two peaks came out. One ran from November 2020 into February 2021 and overlapped the Beta wave; the second fell in July to September 2021, against Delta. The analysis stops just as Omicron was starting, and the authors note only preliminary signs of an uptick at the end. Across the pandemic window as a whole, mortality ran about 60% higher than the pre-pandemic model expected — a district-level echo of the 58,627 excess deaths the South African Medical Research Council estimated for the province between May 2020 and January 2022.

The point the paper keeps returning to is that this happened in a rural population, not a dense urban one. Early in the pandemic there was a working assumption that lower population density might cushion rural districts. On this record it did not, in a district that carries one of the world’s heaviest HIV burdens alongside a growing load of hypertension and diabetes.

The comorbidity arithmetic, and one reversal

Clinical measures came from a separate source: the Vukuzazi health fairs held between 2018 and 2020, which screened roughly 18,000 adolescents and adults for HIV, tuberculosis, diabetes and hypertension using objective tests rather than self-report. Those results were linked into the surveillance cohort by identifier, and treated as fixed baseline exposures for the rest of the follow-up.

The crude rates alone are stark. Among people living with HIV the mortality rate was 13.3 per 1,000 person-years against 8.3 among HIV-negative participants; for diabetes it was 34.2 against 10.5, and for hypertension 28.5 against 7.4. What the interaction models add is that the relative excess moved. HIV carried a relative rate of 1.8 (95% CI 1.3–2.4) before the pandemic and 1.5 (1.2–2.0) during it; diabetes went from 1.6 (1.2–2.2) to 1.4 (1.1–1.9). Hypertension went the other way, with a 15% relative increase in the death rate during the pandemic compared with before (p for interaction = 0.003), and people living with both HIV and hypertension reached a relative rate of 2.18 (95% CI 1.45–3.26).

The authors are careful about what that pattern means, and offer two readings without choosing between them: mortality rose across the whole population during the waves, which would compress relative differences between groups; or access to HIV services and chronic-disease care shifted during the pandemic. Their tuberculosis finding is flagged even more cautiously — lower relative mortality during the pandemic than before it, in a group that mixes current and past TB, where survivor bias and residual confounding are live possibilities and the confidence intervals are wide.

Three years off a fifteen-year-old’s remaining life

Life expectancy at age 15 was calculated by period life tables and then decomposed by age and sex. Women began the period at 70.1 years of remaining life at 15 (95% CI 68.1–72.1); men at 61.2 (59.3–63.0). Male life expectancy climbed to 65.6 (64.1–67.1) by 2020, one of the clearest gains in the series, and then between 2020 and 2021 both sexes lost ground sharply: 3.3 years for women, 3.2 for men.

The decomposition puts the losses in different places for the two sexes. For men, the biggest contributions came from ages 15–24 and 45–54. For women, from 35–44 and from 55 and over — and women contributed more years to the total loss than men, narrowing but not closing the female survival advantage. By health condition, hypertension at ages 45–64 and overweight or obesity at any age accounted for the largest shares of lost life expectancy, with underweight at 15–34 also prominent. That last pattern is worth holding onto: non-communicable risk showing up in the life-expectancy arithmetic of young adults, in a cohort where HIV incidence among the young is high.

The part the paper cannot supply

Here is where the story turns back on verbal autopsy. The data source section states plainly that the site classifies causes of death by verbal autopsy. The limitations section states just as plainly that cause-specific analysis is not what this paper did: “the absence of cause-specific mortality data precludes direct attribution of excess deaths to COVID-19 and prevents differentiation between deaths caused by SARS-CoV-2 infection and those resulting from indirect consequences of the pandemic, such as disruptions to healthcare services.”

That is an honest boundary, and an important one. Excess mortality measures the size of a hole. It does not say what filled it. A death from untreated HIV or a missed hypertension prescription during a lockdown and a death from SARS-CoV-2 itself both land in the same monthly count — and the two imply different accounts of what went wrong, which is precisely the kind of question cause-of-death data exists to answer.

The same district has produced that kind of data. In 2024, Chu and colleagues published a harmonised analysis of verbal autopsy records from nine health and demographic surveillance sites in six countries across eastern and southern Africa, covering 1995 to 2019. uMkhanyakude was one of them, and the paper records that it had full or near-full verbal autopsy coverage of all identified deaths. Across the nine sites, 52,484 deaths were reported among 1,071,913 people; 91.4% had a verbal autopsy, and 97.1% of those were assigned a cause using InSilicoVA, the probabilistic model implemented in the openVA toolset. All-cause mortality fell over the period, in many sites driven by reductions in HIV and tuberculosis deaths; by 2010–14 the leading causes were road traffic injuries, HIV or tuberculosis, and meningitis or sepsis among adolescents, HIV or tuberculosis among adults 20–59, and neoplasms and cardiovascular disease among those 60 and over.

That series cannot be grafted onto the pandemic years — different period, different design, different question. It shows what the instrument can deliver at this site, not what it would have said about 2020 and 2021. Worth saying too that verbal autopsy assigns causes probabilistically from reported signs and symptoms, under the WHO standard instrument; it is not a laboratory result. Validation work has repeatedly shown that accuracy varies sharply by cause and setting — the country-specific misclassification rates we looked at in an earlier entry ranged from about 5% for some causes to the mid-80s for others, depending on the cause and the algorithm.

What kind of evidence this is

An observational cohort with a modelled counterfactual, published in a specialist measurement journal, externally peer reviewed. The excess figures are estimates against an expectation, not counts of COVID-19 deaths; the timing correspondence with provincial waves is offered as reassurance about the counterfactual, not as proof of attribution. Health-condition analyses are restricted to Vukuzazi participants, who attended voluntarily and may differ systematically from those who did not, and the exposures are held constant from a screening round that closed in 2020. The combined-condition subgroup estimates carry wide intervals, and the authors describe several of them as hypothesis-generating. Code and data are on GitHub under a SAPRIN-MRC repository.

None of that undermines the headline. It sharpens it. A district that can tell you, to the month, how many more people died than should have — and that in principle holds the interviews that would say of what — published the first half of the ledger and left the second for another paper. Both halves matter, and they arrive on different schedules.

Sources

← Back to the news digest