A Threshold Is a Policy Choice, Not a Fact

A score of 69.9 and a score of 70.0 may be nearly indistinguishable as measurements and completely different as decisions. One applicant advances. One alert fires. One patient enters a follow-up pathway. One transaction is blocked. The line looks objective because it is expressed as a number. The number does not make the line natural.

A threshold is a rule that maps a continuous or uncertain signal into a category or action. It can be necessary. Institutions cannot always respond with an essay; they need to admit, inspect, warn, allocate, stop or continue. But a threshold is not discovered in the same way as the underlying observation. It is selected for a purpose, under constraints, with consequences for errors on both sides.

Measurement and decision are two operations

The first operation estimates something about the world: temperature, risk, probability, concentration, performance or similarity. The estimate may include noise, sampling error, model error and disagreement about what the construct means. The second operation decides what to do with that estimate. Blending the two makes a policy judgment look like a property of nature.

Consider a screening test. The measured signal does not carry a universal label saying positive. A cutoff determines which signals are reported as positive for a particular use. FDA guidance for diagnostic devices repeatedly asks developers to explain how a cutoff was determined, validate performance near it and justify tradeoffs in sensitivity and specificity. That requirement matters because moving a cutoff changes who is missed and who is sent through unnecessary follow-up. The appropriate balance depends on the condition, downstream test, population and harm—not arithmetic alone.

The same structure appears outside medicine. A fraud score becomes a block. A moderation score becomes removal. A flood forecast becomes an evacuation trigger. A school score becomes eligibility. In each case, the measured or modeled variable and the response rule should be inspectable separately.

Why systems need lines

Rejecting certainty theater does not require rejecting decisions. Thresholds can make responsibility and response consistent. They can protect scarce capacity, create a common escalation rule and make a system testable. A smoke alarm that never commits to an alarm state is not useful.

The mistake is treating the cutoff as self-justifying. A useful threshold has a named objective, a defined population, evidence about performance, a plan for borderline cases and a review date. Without those parts, numerical form supplies the appearance of rigor while hiding the actual judgment.

Every cutoff distributes error

Lowering a detection threshold may catch more true cases while increasing false alarms. Raising it may reduce unnecessary interventions while missing more real cases. There is no general instruction to minimize all errors at once. The errors compete, and their costs are not symmetrical.

Accuracy alone cannot answer these questions. Nor can the average performance of a model if errors cluster in a subgroup. The average is not the experience, and a threshold can amplify a modest calibration difference into a large difference in treatment.

Near the line, uncertainty matters more

A result just above a threshold is not necessarily meaningfully different from one just below it. If repeat measurement, another sample or a small model change could reverse the category, the system should not present the boundary as a cliff in reality.

Good practice can include a gray zone, repeat testing, a second source, confidence intervals, manual review or a delayed decision when delay is safe. The design should match the stakes. A low-cost recommendation can tolerate more automation than a decision that restricts liberty, access, health care or livelihood.

Precision also matters. A value printed to three decimal places can invite an overly sharp cutoff even when the instrument is not accurate at that resolution. Before arguing about 69.9 versus 70.0, use the precision-versus-accuracy distinction to ask whether the measured difference is reliable at all.

A threshold belongs to a context

Performance established in one population, season, device environment or operating workflow may not transfer. Base rates change predictive value. Data drift changes score distributions. A new downstream process changes the cost of escalation. A cutoff that once balanced a system can become stale while remaining numerically identical.

That is why validation is more than calculating a line on historical data. It includes checking the line in the intended setting, examining subgroups, testing behavior near the boundary and monitoring after deployment. A threshold should travel with its scope and assumptions, not as a detached number.

Capacity is a value choice in disguise

Many operational cutoffs are partly capacity controls. A review team can inspect only so many cases. A benefit program has a budget. A hospital has limited beds. Capacity is real, but calling a capacity-derived line a scientific threshold confuses a resource constraint with evidence about need.

Name the constraint. Then ask whether a queue, lottery, prioritization score, staged review or additional capacity would be more honest. If the line exists because only one hundred cases can be handled, the public explanation should not imply that case 101 became unworthy by nature.

A seven-part threshold record

1. Signal: define what is measured or estimated, with units and known limitations.
2. Purpose: name the action the cutoff initiates and the objective it is meant to serve.
3. Selection: document who chose the line, which evidence was used and which alternatives were considered.
4. Error tradeoff: identify false-positive and false-negative consequences, including who bears them.
5. Boundary handling: define repeat measurement, gray zones, human review and appeal.
6. Scope: state the population, setting, time period and conditions for which performance was evaluated.
7. Review: assign an owner, monitoring signals and a date or trigger for reconsideration.

This record is an application of the assumption-register method: keep the invisible premises beside the rule they support. It also makes disagreement more productive. People can argue about a stated cost or objective instead of pretending that the number settled the question.

Use the same discipline personally

Personal systems also hide thresholds: the balance that triggers anxiety, the notification count that means behind, the productivity score that means good day, or the wearable metric that changes a plan. A consumer dashboard may color a number red without explaining the evidence, uncertainty or individual context behind the band.

Before adopting the alert, write the action you will take and the evidence that makes that action useful. If no meaningful response exists, the threshold may be producing vigilance rather than agency. If the consequence is medical, financial or legal, use the device or score as a prompt to consult an appropriate professional, not as a diagnosis.

Claims and boundaries

Sourced fact: FDA diagnostic-device guidance treats cutoff selection, validation near the cutoff and sensitivity/specificity tradeoffs as matters requiring evidence. Inference: the same separation between signal, cutoff and action improves many automated and institutional decisions. Judgment: high-stakes systems should provide review or appeal near consequential boundaries. Not claimed: all thresholds are arbitrary, numeric rules are illegitimate or one error tradeoff fits every domain.

Primary sources and further reading

• U.S. FDA: Assay Migration Studies for In Vitro Diagnostic Devices, for the distinction between numeric data, cutoffs and qualitative result categories.
• U.S. FDA: Respiratory Viral Panel guidance, for justifying cutoffs with sensitivity, specificity and performance near the line.
• NIST: AI Risk Management Framework, for context-sensitive risk management, measurement and governance.