October 3, 2026 interva.net Editorial

When Only Some Deaths Get a Verified Cause: Verbal Autopsy Meets Verification Bias

A statistical framework from the University of California, Santa Cruz treats the way deaths get selected for verification as part of the model — then tests itself on 411,491 suspected COVID-19 deaths recorded in Brazil's severe respiratory illness surveillance.

verbal autopsycause of deathdata qualityCOVID-19

Nearly 40% of the world’s deaths are never registered, and in low-income countries only around 8% of the deaths that are reported have a documented cause, according to the World Health Organization. Where no physician writes a death certificate, the most reliable record of how someone died is frequently an interview: a fieldworker seated with a bereaved family, working through the WHO verbal autopsy instrument, and passing the answers on to a coding method.

Which method matters more than it sounds. The InterVA model, developed at Umeå University, carries symptom-to-cause relationships elicited from physicians; data-driven alternatives such as InSilicoVA learn those relationships from deaths whose causes were independently verified. Both rest on an input that a genuinely new pathogen takes away.

Two things an outbreak breaks

The first is knowledge. When a disease is unfamiliar, nobody can say in advance how its symptoms map onto death, and there is little labelled experience to learn from. The second is subtler, and it is the subject of a methods paper published in the Journal of the Royal Statistical Society Series A by Yu Zhu and Zehang Richard Li, both at the University of California, Santa Cruz.

During an outbreak, only a fraction of deaths are ever verified — by a laboratory test, a physician review, a postmortem examination. That fraction is not a random slice of the dead. It reflects where testing capacity existed, who could reach a hospital, and what public health authorities prioritised that month. Train a model on those verified deaths, or calibrate a surveillance system with them, and the selection process travels silently into the results. The authors draw the parallel to verification bias in diagnostic-test research, and note that no guidelines exist for when and how to account for how reference deaths were selected.

Their answer is to model that selection outright. The framework is Bayesian and hierarchical, built on latent classes nested beneath each cause of death, so that symptoms are allowed to cluster rather than be treated as independent — an assumption the paper notes is baked into the VA algorithms currently used in practice. Its target of inference is the cause-specific mortality fraction: what share of deaths in a given group, in a given month, came from the emerging cause.

Because counting deaths separately in every age-by-sex-by-month cell leaves many cells very thin, the authors add structured priors — random walks across time — that let neighbouring sub-populations borrow strength from one another. In their simulations, smoothing narrowed the credible intervals markedly in the sparsest strata while still tracking the true trend.

411,491 deaths in Brazil

The motivating dataset is Brazil’s severe acute respiratory syndrome surveillance for January to October 2021: 411,491 deaths suspected of COVID-19, each described by fourteen binary symptom indicators — fever, vomiting, loss of taste and smell among them — alongside an independently assigned underlying cause. Test results were largely missing and were excluded. Deaths were grouped into eight age bands and two sexes across ten months.

Since the true fraction in the real data cannot be known with certainty, the researchers built 100 semi-synthetic replications, resampling half the observations within each stratum while holding the real proportions fixed, and imposed a verification mechanism that depended on time, age and a random handful of symptoms. Estimation that ignored the mechanism drifted away from the truth. The hierarchical model, the authors report, recovered it.

What the paper does and does not say

This is a statistical contribution, and it should be read as one: simulation evidence plus one country’s surveillance data, not a field validation and not an accuracy contest against InterVA, InSilicoVA or physician review. The authors are careful on this point, noting that even their comparison baselines are more flexible than the VA models in routine use. Their own list of limitations is candid — the availability of PCR and antigen tests cannot be treated as missing at random, the latent-class priors are assumed independent, and today’s VA algorithms take what they call an overly simplified view of the causal structure among the collected indicators.

The wider argument in the discussion is about provenance. As long as deaths are chosen for verification under a documented design, the assumptions a model needs can be made plausible; skip that documentation and, in the authors’ words, failures to account for the data collection process can produce severely biased results even with sophisticated models. They flag the point as machine learning and artificial intelligence spread through population health research.

Verification, of course, has other forms. Where minimally invasive tissue sampling is available — as in the CHAMPS network, which pairs postmortem investigation with the WHO questionnaire across seven countries — the reference classification is a pathology result rather than a laboratory flag. That work answers different questions about why children die; the Santa Cruz paper is aimed at an urgent, cheaper question: how many people a new pathogen is killing, and in whom. The code for the method is public, and the full Brazilian dataset is available on request with the health ministry’s permission.

Sources

← Back to the news digest