October 14, 2026 interva.net Editorial

Four Numbers Read Off a Single Interview

Verbal autopsy programmes usually find out that an interview went wrong months later, once the causes have been coded. A preprint proposes indicators that can be computed on one record the day it arrives — and tested them on datasets from Tanzania, Eswatini and Nigeria.

verbal autopsydata qualityvalidationfield studies

A verbal autopsy interview is a long conversation with a family about a death: the symptoms, the order they appeared in, the care that was sought, the care that was not, the weeks or months of it. Whatever a coding algorithm later concludes rests entirely on how faithfully that conversation was written down. The method has always carried that dependency, and the field has always known it.

What it has not had is a fast way to notice when the writing-down goes wrong. Completeness and consistency checks in most programmes happen late — at or around the moment the records are pooled for cause-of-death assignment. By then the interview is months old, the respondent may be unreachable, and the practical choices have narrowed to flagging the record or throwing it away.

A preprint posted on Research Square on 25 September argues that the loop can be closed much earlier, and that the signals needed to do it can be read from a single interview record rather than from a population. The paper is a methodological evaluation, not a validation study, and it has not been peer reviewed; it is under review at a journal. Eleven authors are listed, led by Isaac Lyatuu; the co-author list includes staff from the Tanzanian and Eswatini ministries of health and from the Clinton Health Access Initiative and Katsina State Primary Health Care Agency in Nigeria, who the paper says reviewed the manuscript and confirmed the results against their own country datasets.

Four measures, all computable on one record

The framework is deliberately built on the standard WHO verbal autopsy instrument — the field names of both the 2022 edition and the 2016 version — so that nothing new has to be collected at the household.

Indicator	What it reads from the record
Informative Completeness Score	The share of yes/no questions answered definitively, rather than “don’t know” or “refused”
Respondent Reliability Score	How close the respondent was to the deceased and how long after the death the interview took place
Internal Consistency Index	Nine logical checks, including symptom durations longer than the total illness, interview dates before death dates, and pregnancy questions answered for a male decedent
Median Interview Duration	Elapsed time from interview start to end, summarised as a median

The authors are careful about what this is for. Training, questionnaire design, skip logic and range constraints at the point of entry all remain the first line of defence; the indicators are meant to catch what survives them, and to catch it while a supervisor can still act — a retraining visit, a recheck, a look at a specific interviewer’s uploads.

What the three datasets showed

The indicators were computed on records from three countries: 4,344 Tanzanian interviews collected with the 2016 instrument, 380 from Eswatini and 5,474 from Nigeria, both collected with the 2022 instrument.

On completeness, all three look disciplined. Mean informative completeness was 96.4% in Eswatini, 97.4% in Tanzania and 99.8% in Nigeria, with occasional low outliers tied to interviewers leaning on “don’t know”. Respondent reliability averaged 81.8 in Eswatini and 84.1 in Tanzania, with roughly two thirds of records in the paper’s top tier. Across the nine consistency rules, 91.8%, 92.8% and 92.0% of records respectively passed every check.

The interesting output is in the failures. Cough duration recorded as longer than the whole illness ran at 6.1% of Eswatini records (23 cases) against 1.6% in Tanzania and 0.2% in Nigeria — a gap the authors attribute to interviewer training, how respondents interpret “illness duration” versus “symptom duration”, or a data-entry convention, and say should be investigated by interviewer. In Nigeria, 323 records — 5.9% — carried pregnancy-related responses for male decedents, against 0.7% in Tanzania and none in Eswatini. That pattern is unlikely to be a coding artefact: the paper reads it as a sign that skip logic built into the form was disabled or bypassed somewhere in the collection or export pipeline, exactly the kind of silent breakage that would otherwise surface only later as implausible causes of death.

Tanzania’s flagged pattern is temporal. In 152 records, 3.5%, the interview date preceded the death date — arithmetically impossible. Those records cluster with a stretch of low completeness scores between mid-2024 and late 2025, which the authors describe as a discrete period of reduced data-entry oversight rather than an instrument-wide fault.

On duration, the medians were 27.3 minutes in Eswatini and 33.3 minutes in Tanzania, with means roughly double the medians because paused and multi-day interviews stretch the upper tail; the paper argues for reporting the median. For scale, the abridged questionnaire derived from the WHO instrument and piloted in Kolkata has been reported in the 30–35 minute range, against 45–60 minutes or more for the full tool.

What the framework does not establish

Read the limitations before the results.

None of the four indicators was compared against an independent measure of whether the interview was actually right — no re-interviews, no expert record review, no audio audits, no concordance with known causes of death. The paper therefore calls its own output an initial methodological evaluation rather than validation. The thresholds that turn a score into a “high” or “excellent” tier are expert-derived and have not been tested against whether cause assignment came out correct.

The consistency rules cover part of the questionnaire only: no paediatric or neonatal sections, and no checks specific to genuine female maternal deaths. The reliability score leans on a proxy, using the deceased’s education where the instrument does not record the respondent’s. The three datasets were not random samples and the paper does not claim they represent their countries. And the duration measure turns out to be sensitive to the data pipeline itself: the same country’s records extracted six years apart — one pull from an older ODK Aggregate server, one from its successor — carried timestamps at different precision, changing what the indicator could see. Values, the authors write, are properties of a documented data pull, not fixed properties of a programme.

Why timing is the real claim

Existing quality tooling mostly works elsewhere. The WHO data quality assurance toolkit checks completeness and internal consistency at facility or district level, in aggregate, and was not built around the structure of a verbal autopsy record. Tools for the assignment step — the InterVA model, InSilicoVA, Tariff, physician review — need a corpus before their output means much; the paper notes that InSilicoVA wants on the order of a hundred cases for a stable result. A programme with ten interviews has nothing to code and, until now, nothing to check.

That is the modest, useful core of the proposal: four arithmetic measures, no extra fieldwork, meaningful on record one. The reference implementation is published as an open-source Python package so other programmes can run and dispute them — the underlying interview data are not public, held by the ministries and agencies that shared them, listed in the paper’s data-availability statement.

What the preprint cannot yet say is whether an interview that scores badly is more likely to produce a wrong cause of death. That is the study the framework still needs. Until someone runs it, these four numbers describe the quality of a conversation as recorded, which is a real and checkable thing — and, unlike an accuracy estimate, it arrives while the family’s account is still fresh.

This digest is an independent observer: it is not affiliated with the InterVA project, Umeå University or WHO, and nothing here is advice to any health ministry.

Sources

← Back to the news digest