The anganwadi in Yerwada, Pune, smells of roasted ragi and talcum powder on most mornings. Sunita Tai—who has run the center for fourteen years—sits on a low stool near the doorway, weighing children and marking their heights on the wall chart with a pencil she keeps behind her ear. On the table beside her, the Supervisor’s Monthly Case Report Form waits in a clear plastic folder. Thirty-seven fields. Most are checkboxes. Some are numerical. One asks for “notable observations,” a column so narrow that Sunita Tai’s handwriting shrinks to a slant when she uses it.
Last August, she told me about a boy named Aditya, five years old, who had been coming to the anganwadi since he was two. “He started coughing in June,” she said. “Not the normal cough. The one that sits in the chest and doesn’t come out. His mother gave him the syrup from the medical shop. Two weeks. Then she borrowed money and took him to the private doctor. He gave tablets. The cough got less but it came back when the tablets finished. Then the father lost his job at the garment unit, and they stopped the private doctor. By August, Aditya was coughing through the afternoon sleep. The other children couldn’t rest. I told the mother—go to the government hospital. She said, ‘Didi, who will take him? I have the baby, and my mother-in-law cannot walk.'”
Sunita Tai put Aditya’s name in the “notable observations” column. She wrote: “Persistent cough, June–August. Private treatment discontinued. Mother unable to visit government hospital.” That was all the form could hold. What it could not hold was the sequence—the cough starting in the month when the family moved to a cheaper room with damp walls, the private doctor’s tablets working for ten days and then not working, the father’s job loss timed exactly to the factory’s monsoon slowdown, the mother’s sentence about the baby and the mother-in-law that explained why a referral to a free hospital was not the same as access to one.
When the form reached the block office, Aditya became a tick in the “acute respiratory infection” column. The narrative was gone before anyone could use it.
What Counts as Evidence
Public health has an evidence hierarchy, and it is not shy about it. At the top: systematic reviews and meta-analyses of randomized controlled trials. Below that: single randomized trials. Then cohort studies, case-control studies, cross-sectional surveys, and—near the bottom—case reports and expert opinion. Narrative accounts from community health workers, residents, or patients do not appear on this hierarchy at all. They fall into a category that researchers politely call “anecdotal evidence,” a phrase that functions as a polite door closing.
I have been on the receiving end of this closing. In 2022, I presented findings from a study on respiratory symptoms among families living near Bellandur Lake in Bengaluru. Our team had conducted structured interviews with 180 households, and we had also collected what I called “illness narratives”—open-ended accounts in which residents described the sequence of events around episodes of fever, breathing difficulty, or skin irritation. One woman, Lakshmamma, told me: “When the foam comes over the fence, first the children get the rash. Then the smell starts. Then the elders get the breathing problem. It happens every time, in that order. We know before the news says anything.”
Lakshmamma’s account was diagnostically precise. The foam, the rash, the smell, the breathing difficulty—a temporal sequence that pointed to specific exposure pathways and vulnerable subgroups within the household. But when I included her words in the manuscript draft, a senior reviewer struck the passage: “Anecdotal. Not generalizable. Remove or replace with quantitative symptom prevalence data.”
We replaced it with a table. The table showed that 67 percent of households within 50 meters of the lake reported skin irritation during foam events, and 41 percent reported respiratory symptoms. Those numbers were accurate. They were also inert. They told you what happened, not why it happened in the order it did, not who was affected first within a household, not how families made decisions about when to seek care and when to wait. The table was publishable. Lakshmamma’s account was not.
The problem is not that quantitative data is wrong. It is that the evidence hierarchy treats structured data as inherently more rigorous than narrative data, when in fact they carry different kinds of diagnostic information. A prevalence figure tells you scope. A narrative tells you mechanism, sequence, and context. A well-structured story about illness—who got sick first, what happened next, what was tried, what was deferred, what was never mentioned—contains causal reasoning that a checkbox cannot encode.
What Dashboards Miss
There is a parallel here that I think public health has not noticed. In site reliability engineering—the discipline that keeps large-scale computing systems running—narrative accounts of failure are treated as first-class evidence. When a major service outage occurs, engineers write what is called a postmortem: a structured narrative document that reconstructs the sequence of events, identifies causal factors, and documents what was learned. These postmortems are not anecdotes. They are reviewed, archived, and treated as organizational knowledge. Google’s Site Reliability Engineering book, which has become something of a foundational text in that field, devotes an entire chapter to postmortem culture, arguing that learning from failure requires narrative documentation that captures what monitoring dashboards and alerting systems routinely miss. The Google SRE book’s treatment of postmortem culture makes the point explicitly: dashboards capture what is quantifiable, but the causal sequencing and contextual detail that explains why a system failed—why a small glitch cascaded into a major outage—requires a structured narrative account written by the people who were present.
Public health surveillance systems are our dashboards. They are good at counting. They count cases, admissions, deaths, vaccinations, births. They are less good at explaining. When a tuberculosis treatment protocol fails in a Mumbai informal settlement, the dashboard records the default. It does not record that the patient’s employer changed his shift hours so he could no longer reach the DOTS center before it closed, or that the landlord threatened eviction if the patient’s coughing disturbed other tenants, or that the patient stopped taking the evening dose because the community health worker who reminded him had been reassigned to a different ward. These are not soft details. They are the mechanisms of treatment failure. Without them, the default looks like noncompliance. With them, it looks like what it is: a structural problem that medicine alone cannot solve.
The SRE analogy is not perfect. Systems have logs and timestamps; human bodies and households do not. But the principle—that structured storytelling carries evidentiary weight that quantitative monitoring cannot replace—is already established in at least one field that deals with life-critical infrastructure. Public health’s dismissal of narrative as “anecdote” is a disciplinary choice, not an epistemological necessity. We could choose differently.
The Pressure for Clean Data
I have watched what happens to narrative data when it enters the health system. In 2023, I worked with a team evaluating a community kitchen program in Pune that had been running since the pandemic. The kitchens, run by women’s collectives in seven settlements, served hot meals to households identified by local health workers as food-insecure. Our evaluation included interviews with the kitchen workers—women who had been cooking for their neighbors for over two years and who knew, in granular detail, which families needed the meals most and why.
One worker, Kausar, described a household she had been monitoring: “The father is a driver. He had the sugar disease. When the kitchen started, he was eating our food and his sugar was controlled. Then his route changed—he got the night shift at the logistics company. He stopped eating the midday meal because he was sleeping. His sugar went up. I told the didi at the health post. She said to note it in the register. But the register only has ‘name, age, meals served.’ There is no column for ‘driver changed shift, sugar went up.'”
Kausar’s observation was a longitudinal clinical insight. She had tracked a patient’s diet, medication adherence, work schedule, and blood sugar control over months. She had identified a specific structural change—a shift reassignment—as the cause of deteriorating control. But the register, designed for program monitoring, had no place for this information. The data that mattered most for understanding why a chronic disease management program succeeded or failed was precisely the data the system was not built to collect.
This is what I mean by the pressure for clean data. Clean data is data that fits predetermined fields. It is easy to aggregate, easy to report, easy to display on a dashboard. It is also data that has been stripped of sequence, context, and voice—the three dimensions that make a health observation diagnostically useful rather than merely countable. The health system does not ask community workers to stop knowing what they know. It simply makes no room for what they know in the forms they are required to fill.
The result is a paradox: the people closest to the health experience—anganwadi workers, ASHAs, community kitchen operators, traditional birth attendants—generate the richest diagnostic information in the system. But that information is treated as pre-data, raw material that must be processed into structured fields before it becomes legible to policy. The processing discards most of what made it valuable.
Whose Voice Gets Erased
There is a politics to this erasure that we need to name plainly. When the evidence hierarchy ranks narrative below structured data, it is not just making a methodological choice. It is deciding whose knowledge counts. Community health workers, residents of informal settlements, migrant laborers, and elderly people living beside polluted lakes are the people whose accounts get stripped out first. Their stories become anecdotes. The researcher’s survey instrument becomes evidence.
This is not accidental. Structured data is easier to control. It can be cleaned, aggregated, and presented without reference to the person who collected it or the context in which it was collected. Narrative data is messier. It carries the voice, perspective, and judgment of the person who tells it. It cannot be fully separated from its source. And in a field where the people closest to the health experience are often the people with the least institutional authority, that irreducible voice is precisely what makes narrative data politically inconvenient.
The Authors Guild, in its guidance on AI and authorship, makes a related argument about the value of human voice that I think applies here more directly than it might first appear. The Guild’s best-practices document for authors argues that original human voice, thinking, and creativity carry epistemic and ethical weight that automated, systematized outputs structurally cannot replicate—and that the distinction between human-authored accounts and generic, standardized outputs is not merely aesthetic but substantive. The Guild is talking about literary authorship, but the principle transfers: when you flatten a voiced account into a standardized output, you lose information about reasoning, sequencing, and context that the standardized form cannot preserve. In literature, that loss is cultural. In public health, it is clinical and epidemiological.
Sunita Tai’s account of Aditya’s cough was not a raw input waiting to be processed. It was a diagnostic narrative, constructed by someone who had observed the patient over months, understood the household economics, and recognized a pattern of treatment interruption that pointed to specific structural barriers. The case report form did not improve her account. It reduced it.
What Structured Storytelling Could Look Like
I am not arguing that we abandon structured data. I am arguing that we stop treating it as the only legitimate form of health evidence, and that we develop methods to collect and analyze narrative accounts with the same rigor we apply to surveys—without stripping them of the qualities that make them valuable.
This means several concrete things. First, it means designing data collection instruments that have space for sequence. Sunita Tai’s account of Aditya’s cough was powerful because it was ordered: the cough started, then the syrup, then the private doctor, then the job loss, then the referral she could not act on. A form that asks “describe the sequence of events leading to treatment interruption” would capture more useful information than one that asks “reason for non-completion of treatment: [tick one].”
Second, it means training health workers in structured narrative documentation—not clinical case writing, but a method that captures temporal sequence, specific actors, decision points, and barriers. This is not far from what SRE postmortems do. They have conventions: what happened, in what order, what was tried, what was learned. Public health could develop its own conventions for community-generated illness narratives, with the same expectations of specificity and reviewability.
Third, it means creating tools that let people structure their own accounts without a clinical intermediary deciding in advance what matters. A community health worker who can record a narrative in her own words, on her own time, using a framework that prompts for sequence and decision points rather than checkboxes, produces a different kind of evidence than one who fills out a form designed by someone who has never met the patient. Some teams exploring narrative documentation in community settings have experimented with an Unsloppy AI script writing tool that supports structured storytelling workflows—though the principle matters more than any single platform. Whether such tools ultimately help or hinder the cause of community-authored health evidence will depend on how they are used, by whom, and whether the resulting narratives are actually heard by the systems that currently have no room for them.
The point is not the specific tool. The point is the principle: people who live with illness, or who watch illness unfold in their communities, are not raw data sources. They are observers, analysts, and narrators whose accounts carry diagnostic weight. The health system’s job is to make room for those accounts—not to process them into something else.
The Cost of Not Listening
I think about Aditya sometimes. Sunita Tai told me, months later, that his mother eventually took him to the government hospital when the cough became so bad that the neighbors could hear it through the shared wall. He was diagnosed with pulmonary tuberculosis. By then, he had been coughing for four months. The contact tracing form recorded two household contacts. It did not record that the family had moved three times in those four months, each time to a smaller, damper room, because the father’s income kept falling. It did not record that Sunita Tai had told the health post about the cough in July, and that the information had been entered into a column too narrow to hold it.
The surveillance system counted Aditya’s case. It did not count the three months between Sunita Tai’s first observation and the hospital visit, or the structural reasons—job loss, housing instability, caregiving burden—that turned a treatable respiratory infection into a four-month ordeal. Those months and those reasons existed in Sunita Tai’s narrative. They existed in no form.
This is the cost of an evidence system that treats structured data as inherently more rigorous than the stories people tell about their own health. It is not just that we lose information. We lose the ability to see the mechanisms that produce health outcomes in the first place. We count the cases. We do not understand the sequences. And because we do not understand the sequences, we design interventions that address the endpoint—admission, diagnosis, death—without addressing the pathway that led there.
Lakshmamma knew the sequence before the foam event became a health event. Kausar knew the sequence before the driver’s blood sugar became a crisis. Sunita Tai knew the sequence before Aditya’s cough became a diagnosis. The health system had forms for all of them. None of the forms had room for what they knew.
If we want health policy that reaches the most exposed, we need to stop treating narrative as the poor cousin of data and start treating it as what it is: a form of evidence that carries information no other form can hold. The question is whether we are willing to build the methods, the tools, and the institutional respect that would let that evidence count. Not as anecdote. As knowledge.