Ideas · Essay · September 2026 · 4 min read
Data Doesn't Decide
An essay on making the assumptions and trade-offs behind a decision explicit.
By Lina Song · Published · Updated
Consider a health system that has invested in both data systems and expert analysis. First it bought the systems that record what happens — the EHR, the warehouse, the dashboards. Then it bought the people and reports that explain what happened — analysts, consultants, board packets.
And then, on a Tuesday afternoon, the board faces the question all of that was supposedly for: should we acquire the practice, employ the physicians, or walk away? — and the decision is made from a spreadsheet, a slide, and memory.
The usual diagnosis is that something is missing: not enough data, not enough discussion, not enough documentation. I want to argue something more uncomfortable.
Even where analytical resources exist, important reasoning may remain implicit. This essay focuses on that gap, not on organizations that lack data or analytical capability.
A decision that commits capital always answers a chain of questions on its way to "yes": What are we optimizing? What were the real options? What do we believe happens under each one? What are we trading away? Which assumptions carry the answer? Where are our hard lines? What would change our mind?
These questions can be answered only implicitly—or overlooked. Making them explicit creates an opportunity to examine the reasoning. When they aren't made explicit, they don't disappear — they get answered by default, by whoever happens to be standing closest:
This is an illustrative review framework, not a survey of how all health systems operate or a list of verified software capabilities.
| The step | Where reasoning may remain implicit | Potential consequence | On the record |
|---|---|---|---|
| Objectives — what are we optimizing? | In the weights someone silently built into the spreadsheet | The institution votes on a principle it never chose | A governing principle, chosen and signed before results |
| Options | Whatever made it onto the slide | The unexamined alternative is the one you live with | Options enumerated, including "do nothing," each scored |
| Prediction — what happens if we act? | Last year's trend, extended | History breaks precisely when you intervene | Structural models that carry the intervention |
| Trade-offs | A hallway argument after the meeting | The losing value never gets priced, just outvoted | Every axis scored; the tension made explicit |
| Sensitivity | "Well, it depends" — said once, then dropped | Nobody knows which assumption carries the answer | The two or three assumptions that actually move the decision, named |
| Constraints & thresholds | A gut feeling nobody signed | Hard lines discovered after they're crossed | Guardrails stated up front, on the record |
| Dissent | Remembered selectively, later | "I never agreed to this" — unfalsifiable | Who disagreed, about what, recorded |
| The decision | A motion in the minutes | The why evaporates within a quarter | A signed recommendation under a named principle |
| Reversal | Revisited by crisis, or never | Decisions outlive the facts that justified them | Explicit conditions that reopen the decision — monitored |
| Outcome | Rarely compared to the promise | The institution never learns from its own choices | The record stays alive; assumptions checked against reality |
Read the middle column again. That is not a picture of negligence. It is a picture of competent people whose reasoning may not be fully explicit. The weighting was chosen — by whoever built the model. An argument about assumptions is not a substitute for testing how those assumptions affect the result. The reversal condition exists — as a feeling someone will have in eighteen months.
This is why "we need better data" so rarely fixes decision-making, and why one more dashboard never ends the debate. Data may be one missing piece. Another is the middle — the layer where evidence, values, trade-offs, and thresholds get assembled into a choice someone must sign.
Healthcare has spent a decade making its data AI-ready, and that work was necessary. But making data ready for analysis does not, by itself, settle which choice to make. The question a board signs its name to is "what should we do — and can we defend it?" That second question is not a reporting problem or a vibes problem. It is a discipline — decision science has spent fifty years building exactly this middle layer — and it deserves the same infrastructure we built for records and billing.
Whether a more explicit process makes decisions faster or easier to review is something to test—not a benefit to assume.
Updated October 2026. Updated to clarify the scope of the argument and distinguish a proposed framework from observed practice.