B33 — the code that stands in for HIV

South Africa registers more than nine in ten adult deaths, yet only about a quarter of the HIV-related deaths documented in medical records appear as HIV on the death certificate. A national validation study, and the verbal autopsies run alongside it, show what happens when two documents describe the same death differently.

verbal autopsyCRVSdata qualitymortality data

A death notification form is written for at least three audiences at once: the family that has to bury someone, the official who has to file the paperwork, and the coder who will later turn the form into a statistical record. What gets written in the cause-of-death box is therefore never only a clinical judgement. In South Africa, one of the entries that appears there is “retroviral disease”. It codes to B33 — a residual ICD-10 category for viral disease not classified elsewhere — and it does not say HIV. Everybody involved understands what it means.

A research letter published in AIDS last year puts a number on how much of that substitution is going on. Linking the country’s national cause-of-death validation study to the official register, Pam Groenewald and colleagues at the South African Medical Research Council, with co-authors from the US Centers for Disease Control and Prevention, found that 2,749 of 8,191 linked records — 33.6% — had HIV/AIDS assigned as the underlying cause after doctors reviewed the hospital and forensic files. The linked death notifications from Statistics South Africa recorded 811 HIV/AIDS deaths, or 9.9%. Only 763 of the 2,749 agreed. Just over a quarter of the HIV deaths that the paper trail documented reached the register as HIV deaths.

Where the deaths went instead

The letter disaggregates the remainder, and the disaggregation is the most quietly damning part of it. Pseudonyms absorbed 22.2% of those 2,749 deaths: 14.4% written as retroviral disease and coded to B33, 7.8% written as an immune disorder and coded to D89. AIDS-defining conditions accounted for another 27.8%, of which tuberculosis was the largest single share at 22%. A further 12.3% were signed out under some other immediate cause, lower respiratory infections leading that group.

None of this is a coding error in the narrow sense. The coders coded what was written. What the letter documents is a documentary practice: certifiers in a country with a well-known epidemic avoided the words, and the classification system had ready-made places for them to avoid it into. The authors point to the legacy of AIDS denialism, to confidentiality worries, and to anticipated stigma, and they note that the deterrent survived both targeted training in medical certification and a self-sealing certificate form designed to make the entry private.

The register is not empty of these deaths. It is full of them, filed under other names.

Two readings of the same sample

Alongside the record linkage sits the verbal autopsy strand, reported in Global Health Action by Monique Maqungo and many of the same authors. Twenty-seven sub-districts were sampled across all nine provinces — 10.5% of the population — and deaths registered between September 2017 and April 2018 were followed up with interviews using the WHO 2016 instrument, the version then in force; the standard has since been revised, and the current WHO instrument is the 2022 edition. Trained fieldworkers spoke to next of kin, and two physicians independently certified an underlying cause for each interview.

On that reading, HIV/AIDS accounted for 22.3% of deaths. InterVA-5, the algorithmic model developed at Umeå University, put the same cause at 18.5% — the kind of gap between physician review and computer coding that shows up repeatedly in this literature, and one reason the InterVA model page is careful to describe output as cause fractions rather than verdicts. Statistics South Africa’s own figures, set in the same comparison, stood at 4.9%. Tuberculosis barely moved between the sources, 6.9% against 6.7%, which is exactly what makes the HIV column so conspicuous: the two instruments disagree about one cause, not about the shape of mortality overall.

The 9.9% and the 4.9% are not the same number twice. The first is the share among the 8,191 records that could be linked to the register; the second is the share in the verbal autopsy comparison against published vital statistics. Different denominators, different analysis windows. Both describe the same direction of travel, and neither is a measurement of how many people actually died of HIV.

The coding room

Verbal autopsy is often described as if the interview were the whole method. The South African numbers show how much sits downstream of it. Of 5,375 completed interviews, physicians scored 81.6% as good-quality information, but judged only 66% sufficient to assign an underlying cause. The two independent reviewers picked the same cause in 56.9% of cases, converged after discussion in another 31.7%, and needed a panel in 10.0%. At the end of that process 68% of deaths carried a usable code, 8.7% were ill-defined signs and symptoms, and 16.6% were assigned a cause too imprecise to fit anywhere inside an ICD chapter.

The same paper shows what the interview recovers that the certificate loses elsewhere. Homicide made up 35.3% of injury deaths in the reviewed interviews and 15.0% in the register — with 48.1% of registered injury deaths sitting in the ill-defined unintentional residue. Narratives also allowed four-digit coding, so that HIV deaths presenting as tuberculosis could be separated out: 49.3% of the HIV/AIDS deaths had TB recorded alongside, and 61.3% of the TB deaths had HIV underneath. The three-digit register cannot see either overlap.

Reaching the door

The recruitment numbers are the part that any country considering this would have to read twice. Of 36,976 deaths registered in the sampled areas during the study period, only 26% were identified while the study was recruiting. Nine thousand seven hundred and thirty next of kin were approached; 6,358 consented; 5,375 were interviewed — 55% overall, from 77% in the Western Cape down to 30% in Gauteng. Five hundred and sixty households could not be located at all. Once a family was found, 93.4% sat down for the interview.

The bottleneck was legal as much as logistical. South Africa’s Protection of Personal Information Act prevents Home Affairs and Health from passing next-of-kin details to researchers without consent, so the study recruited through 353 funeral practitioners and 95 Home Affairs offices, asking intermediaries to make the first approach. The authors’ own conclusion is that national verbal autopsy depends on being built into the death notification process itself, not bolted onto it afterwards. Their follow-up work tests the cheaper channel: a Western Cape pilot of telephone interviews with 229 respondents found that only one of 76 questions produced significantly more non-informative answers than face-to-face administration — and that incomplete contact details, again, were the practical obstacle.

What the exercise does not settle

Two honest limits sit over the whole comparison. The first is that the HIV assignment in the record review rested partly on clinical judgement, including a criterion for “high clinical suspicion” in a young adult with no other comorbidity and no serological confirmation; and the source records themselves varied in quality, with laboratory results often incomplete. The second is that a 55% sample realisation cannot be treated as a national profile — the authors say so plainly, and present their result as a set of correction factors rather than a corrected count.

That leaves the more interesting question untouched. The certificate and the interview are two different kinds of document, produced under different pressures, and neither is the death. The certificate records what a certifier was willing to write in front of a family and a state; the interview records what a relative remembers months later, at a median interval of 9.2 months in this study, with 16.7% of interviews conducted more than a year after the death — the kind of delay whose effect on cause-of-death findings our piece on recall periods describes as a stability question rather than an accuracy one. The register records what got filed. South Africa’s contribution is to have measured all three in the same population, and to have published the gap between them instead of averaging it away.


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 AIDS letter is a research letter reporting record linkage from a national validation sample; the Global Health Action paper is a cross-sectional feasibility and quality study with no external gold standard for cause of death; the BMJ Open paper is a small mixed-methods pilot.

Sources

  1. Groenewald P, Glass T, Kallis N, Holmgren C, Nojilana B, Maqungo M, Cheyip M, Morof D, Nannan N, Bradshaw D. Under-reporting of HIV and AIDS as an underlying cause of death in South Africa. AIDS 2025;39(12):1839-1841
    pmc.ncbi.nlm.nih.gov
  2. Maqungo M, Nannan N, Nojilana B, Nichols E, Morof D, Cheyip M, Rao C, Lombard C, Price J, Kahn K, Martin LJ, Bezuidenhout F, Laubscher R, Kabudula C, Glass T, Awotiwon O, Zinyakatira N, Funani N, Joubert J, Bradshaw D, Groenewald P. Can verbal autopsies be used on a national scale? Key findings and lessons from South Africa's national cause-of-death validation study. Global Health Action 2024;17:2399413
    pmc.ncbi.nlm.nih.gov
  3. Sant Fruchtman C, Neethling I, Bradshaw D, Cobos Muñoz D, Morof D, Ngobeni S, Ngwenya X, Edwards A, Glass T, Kahn K, Herbst K, Morden E, Zinyakatira N, Groenewald P. Telephonic verbal autopsies among adults in South Africa: a feasibility and acceptability pilot study. BMJ Open 2025;15:e090708
    pmc.ncbi.nlm.nih.gov
  4. WHO — 2022 verbal autopsy instrument (V2.0, multilingual)
    who.int

← Back to the news digest