The Months Between a Death and Its Interview

Two large analyses of verbal autopsy timing — 2,113 deaths in South Asia and the Philippines interviewed twice, 10,882 deaths from a South African surveillance site — measure different things, and disagree in a way that matters for anyone reading a cause-of-death table.

verbal autopsyInterVAdata qualityfield studies

A verbal autopsy produces a questionnaire, a narrative and a cause. It also produces a date, and that date is usually the least examined number in the dataset. The gap between a death and the interview about it is set in the field — by mourning customs, transport, staffing and survey calendars — and it varies more than most method sections admit. In one South African surveillance site the median interview happened seven months after the death, and the range ran from one month to four years.

The World Health Organization’s guidance, as the authors of a 2016 validation study summarise it, is to interview as soon as possible after a period of mourning, and to treat recalls beyond a year with caution. That guidance rested for years on thin evidence: the paper traces it to an unpublished 2001 doctoral thesis that found no difference in sensitivity or specificity up to three years of recall. Two large analyses have since approached the question from opposite ends, and their disagreement is more instructive than either result alone.

Two thousand deaths, questioned twice

The first study was built into a validation programme. The Population Health Metrics Research Consortium assembled deaths in Andhra Pradesh, Bohol and Manila whose cause had been established against clinical criteria, then went back to the same households for a second interview. The published dataset covers 2,113 decedents and 4,226 interviews: a first round collected a mean of 1.84 months after death, a second a mean of 17.17 months after, with the same person answering both times in 86.7% of cases. The Tariff 2.0 algorithm assigned a cause to each interview, and each assignment was scored against the known cause.

Peter Serina and colleagues reported that the probability of a correct assignment fell by 0.55% per month. The underlying odds ratio was 0.989 per month, with a confidence interval from 0.974 to 1.004. Put as a comparison rather than a slope, interviews conducted between four and twelve months after a death were, on average, 95.9% as accurate as those conducted within three months; those at a year or more were 95.6%. Both intervals are wide enough to include no decline at all, and the authors say so plainly, twice. Their recommendation is modest: three months is better, where mourning allows.

Ten thousand deaths, one interview each

The second analysis used an archive rather than an experiment. The Agincourt health and demographic surveillance site in rural north-east South Africa has followed roughly 90,000 people since 1992; by 2011 fieldworkers had completed 11,187 verbal autopsies with next of kin, and 10,882 carry a recorded interval. The median was seven months. Causes were transformed to the WHO 2012 list and interpreted with the InterVA-4 probabilistic model, which returns up to three likely causes per death plus an indeterminate residual.

Comparing cause-specific mortality fractions for interviews held within six months against those held at six to eleven months, only three WHO groups shifted beyond chance at the 99% confidence level: acute respiratory infections, diarrhoeal disease, and the catch-all category of other unspecified non-communicable disease. Within a year, the pattern held. Beyond a year, neonatal deaths broke it — an odds ratio of 1.69 (p = 0.004), rising to 2.58 once background factors were adjusted for. Lucinda Hussain-Alkhateeb and colleagues do not claim to know why.

One detail deserves more attention than it gets. The delay was not random: interviews came later, on average, for deaths of infants under one year and of people over 65. A site reporting cause fractions by age while its interview interval varies by age is mixing two effects, and only one is about who died.

What the two results actually measure

Read side by side, the studies look contradictory, and the reason is the reference standard. Serina’s analysis scores each interview against a known cause, so it can watch individual assignments decay. Agincourt has no external truth: it watches whether a population cause distribution stays put as recall lengthens. Both can be true at once, because scattered individual errors move a distribution far less than concentrated ones do.

The consequence follows from the exceptions, not the rule. If late interviews shift the neonatal, respiratory and diarrhoeal fractions, then a programme that reorganises its fieldwork — a shorter funding cycle, a delayed round, a new transport contract — can move its published cause distribution without any change in who is dying. That is a data artefact wearing the costume of a finding.

Agreement is not accuracy, and repeatability is not validity

Two older studies sharpen the warning. A 2013 analysis in Ethiopia put 408 adult interviews through InterVA-4 and through physician review: raw agreement 72.1%, kappa 0.48, with weaker agreement for interviews conducted within a year of the death, among other subgroups. No gold standard was involved, so a low kappa shows two readers diverging, not that either was wrong.

The other, from Burkina Faso and Indonesia, had 91 and 116 women of reproductive age interviewed twice and read with InterVA-M. Because the algorithm is deterministic, the same answers always give the same probabilities, which isolated the interview itself. Repeatability of the chain down to cause fractions was good, the interview material only moderately so, and it fell as recall lengthened, particularly past two years.

The number worth reporting

The questionnaire is now standardised — the WHO’s current verbal autopsy instrument dates from 2022 — but the interval before it is administered is, in practice, a programme’s own decision, and the cost of that decision is measurable at roughly half a percentage point of individual-level accuracy per extra month. Elsewhere on this site we have described a community tissue-sampling study in Bushbuckridge that deliberately deferred its verbal autopsy interviews until families had finished mourning, completing 143 of 206 by its June 2026 cut-off. That was the right call for the families involved. It also has a price, and the price falls unevenly across causes.

Three things are worth asking of any verbal autopsy table. What was the median interval, not the mean, and how wide was the range. Did the interval vary by the age of the dead. And which causes were sensitive to time in that setting — those are the rows where a change in the fieldwork calendar will eventually be read as a change in the epidemic.

Sources

  1. Serina P, Riley I, Hernandez B, Flaxman AD, Praveen D, Tallo V, Joshi R, Sanvictores D, Stewart A, Mooney MD, Murray CJ, Lopez AD. What is the optimal recall period for verbal autopsies? Validation study based on repeat interviews in three populations. Population Health Metrics 2016;14:40
    pmc.ncbi.nlm.nih.gov
  2. Hussain-Alkhateeb L, Petzold M, Collinson M, Tollman S, Kahn K, Byass P. Effects of recall time on cause-of-death findings using verbal autopsy: empirical evidence from rural South Africa. Emerging Themes in Epidemiology 2016;13:10
    pmc.ncbi.nlm.nih.gov
  3. Tadesse S. Agreement between physicians and the InterVA-4 model in assigning causes of death: the role of recall period and characteristics specific to the deceased and the respondent. Archives of Public Health 2013;71:28
    pmc.ncbi.nlm.nih.gov
  4. Byass P, D'Ambruoso L, Ouédraogo M, Qomariyah SN. Assessing the repeatability of verbal autopsy for determining cause of death: two case studies among women of reproductive age in Burkina Faso and Indonesia. Population Health Metrics 2009;7:6
    pmc.ncbi.nlm.nih.gov
  5. World Health Organization. 2022 WHO verbal autopsy instrument, version 2.0 (instrument and cause-of-death list)
    who.int

← Back to the news digest