The register at the sub-centre has a column headed cause of death, and for the entry I was looking at, a man of fifty-four who died at home in September, it said fever. That is the whole entry. The person who wrote it was a health worker who had not been present, recording what the family reported four days later, in a form that offers no other realistic option.
That register feeds upward. Somewhere above it, the entry becomes a number in a state total, and the state total becomes a national estimate, and the national estimate becomes a sentence in a press release about the burden of a disease.
I want to state my claim carefully, because there is a version of it that is nihilistic and I do not hold that version. Most of India's published disease-burden figures are model outputs presented in the register of measurements, and the difference matters far more than the numbers are usually wrong.
Start with what exists. India's civil registration of deaths is now close to complete in the sense that the fact of death is registered for the large majority of deaths. That is a real achievement and it was not true twenty-five years ago. But medical certification of the cause of death — a physician assigning a cause according to the international classification — is available for a minority, and that minority is drawn overwhelmingly from deaths that happened in hospitals in cities. The man of fifty-four died at home in a village, and that is where most deaths in this country still happen.
So for most deaths, the cause is not measured. It is inferred, by a method called verbal autopsy: a trained interviewer asks the family a structured set of questions weeks or months later, and an algorithm or a physician panel assigns a probable cause from the answers.
I need to say clearly that verbal autopsy is not a fudge. It is a serious, validated method with three decades of work behind it, it has been evaluated against hospital records in several countries, and for broad categories — did this person die of an injury, of an infection, of a cardiovascular event — it performs considerably better than most people assume. The Million Death Study built a genuinely useful picture of Indian mortality with it, including findings on snakebite and on tobacco that changed what anyone thought they knew. Anybody who dismisses the method has not read it.
What verbal autopsy cannot do is distinguish inside a category. It will tell you, with reasonable confidence, that a death was from an acute febrile illness. It will not reliably tell you which one. And the diseases we run programmes about, and publish burdens for, and declare progress against, sit inside categories, not across them.
So the chain runs: a family's recollection, months later, of a man who had a fever; a structured interview; a probabilistic assignment; an adjustment for the fraction of deaths not registered; a scaling to the population. Every one of those steps is defensible and documented. What emerges at the end is an estimate with an uncertainty interval that is, in the published literature, often very wide.
Then it is quoted without the interval.
That is the entire substance of my complaint and it is smaller than it sounds, which is why I want to be exact about it. I am not saying the numbers are fabricated or that the modellers are careless — the papers state their intervals plainly and their methods sections are thorough. I am saying that the interval is stripped off in transmission, somewhere between the paper and the press release, and that the number then behaves in public as though it had been counted.
The other example is the one that made me stop treating this as an academic irritation.
A district I worked in reported a substantial decline in deaths from a particular condition over a four-year programme period, and the decline was used, correctly by the district's own lights, to argue for continuing the programme. I went and looked at where the numbers came from. The before figure and the after figure came from the same model, but the model's inputs had changed in between: registration coverage in that district rose over exactly those four years, and the facility-death share rose with it, because a new facility opened.
More deaths were being certified. Certified deaths are assigned differently from inferred ones. The apparent decline was consistent with a real decline, and it was also consistent with the category simply being re-sorted as the measurement improved, and nothing in the district's data could separate the two.
I raised it. The response, which was not hostile and which I have thought about a great deal since, was that the programme was good, the decline was probably real, and that undermining the number would cost the programme its funding without making anybody healthier. I could not say that was wrong. I still cannot. It is the most honest defence of a bad epistemic practice I have ever been given and it names the real cost of what I am asking for.
What I am asking for is narrow. Publish the interval, in the press release, in the same font. Label a model output as an estimate in the ministry's own communications rather than only in the underlying paper. And where a trend is being claimed, state whether registration coverage moved over the same period, which is a single line the same office already holds.
None of that makes the estimate better. It makes it honest about what it is, and it costs the price of a sentence.
The register at that sub-centre has, in the same column, six entries reading fever and two reading old age, for the year. The health worker who fills it in showed it to me without any embarrassment at all, because she was not doing anything wrong. She was writing down what the family told her, which is what the form asks for, and she is the last person in the chain who has any contact with a fact.
The district responded to this in writing and the response is worth recording. They accept that registration coverage rose over the programme period, they do not accept that this accounts for the decline, and they have offered to share the facility-level series so that somebody can try to separate the two. That is more than I expected and I have said yes.
