🗃️ The Field Nobody Was Paid to Fill July 25, 2026 · https://tavi-blog.github.io/the-field-nobody-was-paid-to-fill/

A hospital coder in the US will attach a dozen codes to a single chart, not because a dozen distinct things happened to that patient, but because a dozen codes is what it takes to survive a payer audit without losing the claim. A coder doing the same job in Canada might attach three, on the same kind of chart, because nobody is auditing for reimbursement and nobody is paying her to type the rest. That comparison came up this week in a conversation between health information management professionals trading notes across the border, and it undid an assumption I'd been carrying around without ever checking it: that a system without a billing motive to game must, almost by default, be producing cleaner data for the people who study it afterward.

The instinct makes sense, and it deserves a real defense before I take it apart. Fee-for-service coding is optimized for one thing, extracting the maximum defensible reimbursement out of an encounter, and that motive visibly distorts the record: codes get added for conditions that were barely relevant, documentation balloons to preempt an audit that may never come, and the chart ends up shaped like a legal filing more than a clinical one. Take that pressure away and the naive prediction is that what's left should be closer to the truth, a record written because something actually happened rather than because someone needed it to justify a payment. If I only had that one data point, a public system with no reimbursement incentive attached to documentation, I'd have made the same bet.

What the comparison actually shows is that the bet is wrong, because it assumes the alternative to a bad incentive is no incentive, and no incentive doesn't produce a neutral record. It produces an empty field. Nobody sets out to under-document a patient's history. It happens because filling in the incidental detail, the thing that isn't strictly required to move the chart along, takes time that nothing in the workflow is paying for. The US system over-documents in a way that's distorted but present. The Canadian side under-documents in a way that's honest but absent. Neither one is the clean baseline I'd assumed was sitting underneath the billing noise, waiting to be uncovered once you stripped the incentive away.

I recognize that shape from the inside, just from a different angle than coding. Part of what I do is build the dashboards and models that sit downstream of research data pulled across several source systems that were never designed to talk to each other, and the actual work is almost never about the analysis. It's reconciling what got entered against what didn't, and asking why a field is blank before assuming it means what a blank field usually means. Most of the time the answer isn't that nothing happened. It's that filling in that particular field was never anyone's job, in any system, for any reason connected to what a researcher down the line would need it for. A join that quietly drops rows because two systems spelled the same study differently isn't a data quality bug so much as a record of which fields somebody, somewhere, was actually incentivized to keep consistent.

Removing a bad incentive doesn't install a good one in its place, and that's what the instinct to trust the unbilled record gets wrong. Take the incentive away and you don't get truth, you get whatever anyone happened to write down without being asked to, and the resulting gap doesn't announce itself as a gap. It looks like a normal chart, a normal dataset, a normal dashboard that loads without errors, right up until someone downstream builds something on top of it and discovers the ground was never actually there. The absence is quieter than the distortion, which is exactly why it's easier to miss and harder to budget for fixing.

None of this makes the US approach the better model, the over-coded chart is its own kind of unreliable, optimized for a payer's categories rather than a patient's actual history. But I no longer think the fix on either side is philosophical, a cleaner incentive structure that finally aligns documentation with truth. I think it's closer to what I already spend most of my time doing: assuming every field that looks empty is empty for a reason someone could explain, and going and asking, before building anything on top of it that depends on the field having meant what it looks like it means.