Twelve Dollars, Three Hundred Dollars: The Two Ledgers Behind a Verbal Autopsy
Two programme costings price the same interview at $12 in rural Andhra Pradesh and at $303 in Mozambique. The gap is not waste — it is the delivery model, the denominator, and how many deaths a system manages to report in the first place.
Between October 2003 and September 2007, health workers in 45 villages of Andhra Pradesh walked to the households of the dead, sat with a relative for about an hour, and worked through a verbal autopsy questionnaire. The system ended up with a cause for 96.7% of the 5,895 deaths it recorded. Its published costing puts the first year at $16.66 per death, and each of the following three years at $12.
Fifteen years later and a continent away, a different machine did something similar at national scale. Mozambique’s sample registration system, COMSA, covered 700 randomly selected clusters and a population above 800,000, with a community agent reporting every death on a smartphone and a trained interviewer following up with a tablet. Its costing, published in 2023, puts the operating cost at $303 per verbal and social autopsy in 2019 and $339 in 2020.
Both numbers are peer-reviewed, both are geographically specific, and neither is a mistake. The distance between them is the most useful thing in the literature about what it costs to learn why people die.
The village arithmetic
The Andhra Pradesh system was run with the Byrraju Foundation, a local development organisation, in a population of 185,629. Each village had one non-physician health worker — a resident with ten years of schooling and eighteen months of primary-care training — who spent roughly one day a week on surveillance, completing about three interviews a month. Deaths were found through a network of unpaid key informants: the village headman, the panchayat, priests, cremation staff, community leaders. The interview happened within four weeks of the death and took about 60 minutes.
The first year cost INR 1,133,491, or $24,943, which is $16.66 per death and 13 cents per person covered. Physician coding was the largest single item at 33%, project management took 30%, and the health workers’ salaries 22%. Over four years the whole system cost $79,259, or about ten cents per person per year.
The most consequential line in that ledger is not a cost but an experiment. In year one, two physicians coded every questionnaire independently, with a third resolving disagreements. On 1,329 deaths, the two coders assigned the same ICD chapter 94% of the time (kappa 0.93). From year two the project dropped to a single coder at roughly $3 per questionnaire, halving the analysis bill. The price of that saving shows up in the share of deaths given only a symptom code rather than a cause: 18% in year one, 10% in years two and three, 16% in the final year.
What the ledger leaves out matters as much as what it contains. Office space, electricity and water came free from the local partner. The baseline census had already been done. Research salaries and international travel were excluded by design. And the population was already inside a health and development initiative — an unusually favourable starting point for a costing that presents itself as the price of a surveillance system.
The national arithmetic
COMSA was launched in January 2017 by Mozambique’s National Institute of Health and National Institute of Statistics with Johns Hopkins University, funded by the Gates Foundation, and began collecting in March 2018. Each of the 700 clusters held about 300 households. On the ground sat 700 community surveillance agents, 48 verbal and social autopsy interviewers with tablets, and 15 drivers.
Over 2017 to 2020 the programme spent $5,804,169, including the endline assessment. About $2 million of that — 35% of the four-year total — went on start-up, with more than 80% of the start-up money on infrastructure, technology and training. Operating costs ran just under a million dollars a year: 66% personnel, including cluster-level salaries, incentives and per diems, and 21% infrastructure. Fuel alone was about $90,000 a year and vehicle maintenance about $36,000, or roughly $6,000 and $2,400 per vehicle. All 700 project phones had to be replaced after damage or theft, at $81,000. COVID-19 added $100,000 in emergency spending and shut fieldwork down between March and July 2020.
The outputs in 2019 were 29,237 vital events reported, at $38 each, and 3,627 autopsies conducted, at $303 each. Per cluster the annual cost was $1,568; per person covered, about $1. The authors are candid about two omissions: technical assistance from Johns Hopkins was not counted, and the component activities run alongside the national platform — a CRVS pilot in Inhambane province, hospital-based surveillance in Zambezia, minimally invasive tissue sampling — were treated as out of scope. They also could not separate the cost of reporting a death from the cost of interviewing about it, because in a cascade system the first is the precondition for the second.
The denominator is the argument
There is a second COMSA paper that changes how the first one should be read. An independent retrospective assessment, matching households and events across two data sources, found that community agents reported 54% of births and 45% of deaths between March 2018 and December 2019. Rural deaths were more likely to be captured than urban ones (adjusted odds ratio 1.41), adult deaths more than child deaths (1.49), and births in the smallest clusters more than in the largest (1.45).
So the $303 divides by deaths that were reported, not deaths that occurred. A division the authors do not make, on their own figures, puts the cost per death that actually happened at something more than double the headline. The completeness assessment itself cost $650,843, two thirds of it wages — the price of knowing how much you are missing. The team’s recommendation follows from that: a sample registration system needs a periodic independent check on completeness, smaller clusters than a community worker can realistically cover, and particular attention to cities.
The COMSA authors compare their figures directly with the Andhra Pradesh study and explain the difference in scale: system size, the number of clusters, and how often field staff contact households. That is part of it. The rest is what a unit means. The Indian ledger divides by every death in a small, densely covered population where the interviewer is a resident who walks to the house. The Mozambican ledger divides by interviews actually conducted on a fraction of deaths across eleven provinces, with vehicles, fuel, per diems and a national supervisory cascade. One is the marginal cost of an interview in a place where the infrastructure already exists; the other is the average cost of a country.
What the money does not buy
Neither costing measures precision, and a third study shows why that omission is expensive to fix. Comparing three platforms that used verbal autopsy for maternal mortality — a post-census mortality survey in Mozambique, Bangladesh’s 2010 maternal mortality survey, and Zambia’s SAVVY pilot — Curtis and colleagues counted 259 maternal deaths in Mozambique, 132 in Bangladesh, and 18 in Zambia. The Zambian pilot covered 17,000 households over almost two years. Bangladesh covered 168,629 households with a three-year recall period and still found only 132 maternal deaths, enough for a national ratio of 197 per 100,000 live births but not much more. Agreement between physician coders at the first stage ranged from 61% to 79%.
Their conclusion is blunt: none of these interim platforms is suited to detecting short-term change, because the sample sizes required are prohibitive. Estimates have to be accumulated over several years to be stable, which is why the authors end where the field keeps ending — with continuous civil registration as the only durable answer.
Read alongside the costings, that reframes the number. The unit of purchase is not a cause of death; it is a cause of death with a denominator large enough to say something about a population. Those are different products at very different prices, and the second one is where budgets actually break.
The line item that moves
Both ledgers converge on the same lever: the humans who assign causes. In Andhra Pradesh physician review was the biggest single cost in year one and fell by half once dual coding was dropped, a decision justified by a kappa of 0.93. The authors went further in their discussion, arguing that automated assignment — they name the Tariff method — would cut what they called the most expensive and time-consuming part of the work, and that freeing physicians from coding also stops diverting them from clinical duties. The WHO’s 2022 instrument was built with software assignment in mind, and the InterVA model, developed at Umeå University, along with InSilicoVA and Tariff implementations in the openVA package, exists precisely to do that coding without a clinician in the loop.
What no study in this set does is price an algorithm-only pipeline. The saving is an argument in a discussion section, not a measured figure, and it would not touch the fieldwork costs that dominate the Mozambican budget anyway: the agents, the interviewers, the fuel, the supervision.
Which leaves the honest summary. A verbal autopsy costs whatever its delivery model costs, and the published figures — $12, $17, $303, $339 — are the price of four specific arrangements, in two countries, in two different price years, with different things counted and different things quietly donated. Anyone quoting a single number as the cost of verbal autopsy is quoting one project. The papers, at least, are careful to say so. Whether the ministries choosing between a sample system and a registration system can afford to be is a question the costings raise and do not answer.
Sources
- Jiwani SS, Mavie VA, Williams E, Kante AM, Amouzou A. Implementing the Countrywide Mortality Surveillance in Action in Mozambique: How Much Did It Cost? Am J Trop Med Hyg 2023;108(5 Suppl):40–46. doi:10.4269/ajtmh.22-0438 pmc.ncbi.nlm.nih.gov
- Kante AM, Mulungo A, Ibraimo M, et al. Completeness and Factors Affecting Community Workers' Reporting of Births and Deaths in the Countrywide Mortality Surveillance for Action in Mozambique. Am J Trop Med Hyg 2023;108(5 Suppl):29–39. doi:10.4269/ajtmh.22-0537 pmc.ncbi.nlm.nih.gov
- Macicame I, Kante AM, Wilson E, et al. Countrywide Mortality Surveillance for Action in Mozambique: Results from a National Sample-Based Vital Statistics System for Mortality and Cause of Death. Am J Trop Med Hyg 2023;108(5 Suppl):5–16. doi:10.4269/ajtmh.22-0367 pmc.ncbi.nlm.nih.gov
- Joshi R, Praveen D, Jan S, Raju K, Maulik P, Jha V, Lopez AD. How much does a verbal autopsy based mortality surveillance system cost in rural India? PLoS One 2015;10(5):e0126410. doi:10.1371/journal.pone.0126410 pmc.ncbi.nlm.nih.gov
- Curtis SL, Mswia RG, Weaver EH. Measuring Maternal Mortality: Three Case Studies Using Verbal Autopsy with Different Platforms. PLoS One 2015;10(8):e0135062. doi:10.1371/journal.pone.0135062 pmc.ncbi.nlm.nih.gov
- World Health Organization. 2022 WHO verbal autopsy instrument (V2.0, multilingual) who.int