Surveillance quality has no single optimum

#Question

Why can improving one surveillance attribute make the overall system worse?

#Answer in brief

Because surveillance attributes interact. CDC explicitly warns that strengthening one attribute can adversely affect another: increasing sensitivity can reduce positive predictive value; adding complexity can harm acceptability, timeliness and flexibility; improving several attributes can increase cost.^[1] ECDC similarly treats completeness, validity, timeliness and representativeness as distinct dimensions that need to be monitored in context.^[2]

There is therefore no universal "maximum-quality" surveillance system. There is a system whose compromises are more or less appropriate for its objective.

#Quality is not one axis

A common mental model imagines surveillance quality as a single vertical scale: low quality at the bottom, high quality at the top.

That model fails because the system must satisfy multiple objectives at once.

A surveillance design can be:

• sensitive but nonspecific;
• timely but provisional;
• complete among reporters but unrepresentative of the population;
• detailed but burdensome;
• stable but inflexible;
• simple but unable to answer newly important questions.

Each combination may be rational or harmful depending on what decision the system supports.

#Sensitivity and predictive value

Suppose an alert rule is widened so that almost every true outbreak generates a signal.

Sensitivity rises.

But more non-outbreak periods may also generate alerts. Positive predictive value falls. Investigators spend more time resolving false signals.

CDC uses exactly this kind of interdependence to caution against treating surveillance attributes independently.^[1]

The correct threshold depends on consequence.

For a rare event with catastrophic implications, a low predictive value may be acceptable because missing one true event is worse than investigating many false alarms.

For a high-volume routine condition with limited investigation capacity, the same threshold can paralyze response.

#Timeliness and verification

A provisional record can be made available quickly.

A final record can be made more accurate after laboratory confirmation, episode reconciliation or clinical review.

If the system waits for every uncertainty to disappear, it may become too slow for early action. If it publishes every provisional signal without qualification, decision-makers can mistake unstable information for final evidence.

The solution is often not choosing "speed" or "accuracy." It is creating explicit states:

• preliminary;
• validated;
• revised;
• final.

That preserves timeliness without hiding uncertainty.

#Detail and completeness

Every additional field has a cost.

A long reporting form may capture travel, occupation, exposures, comorbidities and detailed clinical course. Those variables can support important analyses.

But if reporters cannot complete the form during routine workload, missingness grows and notification is delayed. A shorter form may be more complete and timely but analytically poorer.

The right question for each field is not "Would this be interesting?" It is "Which decision becomes materially better if this field is collected?"

WHO's decision-oriented multisource approach reinforces this backwards design: identify the decision question, then the information required, then the sources needed.^[3]

#Completeness and representativeness

A system can achieve 99% completion of mandatory fields among participating facilities while missing entire categories of care.

ECDC notes that geographical coverage, urban-rural differences, populations not reached and absence of particular clinical settings can affect representativeness. It also notes that representativeness can change over time as legislation, infrastructure and clinical practice change.^[2]

Completeness answers: How much is present among expected records?

Representativeness asks: How well do the observed records describe the target event in the target population?

Those are different questions.

#Simplicity and flexibility

Simple systems are easier to operate, train and maintain.

But a system designed around one rigid threat can become fragile when:

• definitions change;
• a new laboratory method appears;
• a new population needs monitoring;
• an emergency requires additional variables;
• data must be combined with another source.

Flexibility generally requires abstraction, configurable rules, governance and versioning—all of which add complexity.

The goal is not minimum complexity. It is necessary complexity that remains understandable.

#Stability and innovation

A mature surveillance system may depend on old technology that operators understand well. Replacing it can improve interoperability or automation while introducing new failure modes, training needs and vendor dependencies.

Stability therefore should not be confused with technological modernity.

A system is stable when it can reliably perform its required function, especially under stress. A new platform that fails during a surge is not an improvement merely because its architecture is contemporary.

#Acceptability is a technical variable

Surveillance often depends on humans entering, validating, correcting or interpreting information.

If a reporting process is too burdensome, confusing or punitive, people adapt:

• fields are left blank;
• defaults are overused;
• reports are delayed;
• workarounds emerge;
• participation becomes nominal rather than real.

Acceptability is therefore not a "soft" consideration outside system performance. It changes the data-generating process.

#A synthetic optimization problem

Imagine a surveillance programme for an uncommon invasive infection.

The current definition has:

• high specificity;
• moderate sensitivity;
• two-day median notification;
• low investigation burden.

A proposed change adds a broad clinical criterion. Simulations suggest more true cases would be captured, but alerts would triple.

If laboratory confirmation usually arrives within 12 hours, the broader definition may create little practical benefit: the marginal gain in early detection may be small relative to investigation burden.

If laboratory confirmation takes four days and early prophylaxis of contacts matters, the same sensitivity gain may be worth the cost.

The correct answer depends on the decision window, not on sensitivity in isolation.

#A useful evaluation format

Instead of one overall rating, report each material attribute with four fields:

Attribute	Current evidence	Decision consequence	Trade-off
Sensitivity	estimated from known cases/events	missed events	more alerts reduce PVP
Timeliness	event-to-action intervals	late intervention	faster data are more provisional
Representativeness	coverage by population/setting	biased burden estimates	wider coverage may increase complexity
Completeness	field/record completion	missing stratifiers	more required fields may delay reports
Acceptability	participation and burden	under-reporting	lower burden may reduce detail
Stability	availability under load	interrupted surveillance	redundancy increases cost

This makes disagreement productive. A programme manager can decide which compromise matters rather than arguing over a synthetic score.

#Quality follows purpose

CDC's surveillance-evaluation guidance begins with purpose and objectives.^[1] WHO's 2025 multisource guidance starts from decision-making questions.^[3]

Both approaches imply the same principle:

A surveillance attribute is valuable because of what it enables, not because high values are inherently virtuous.

A system optimized for outbreak detection can look "worse" on specificity than a burden-estimation system and still be better for its purpose.

#Limitations and uncertainty

The relationships described here are conceptual and operational; the shape and magnitude of trade-offs need empirical estimation in each system. Not every attribute is in tension with every other attribute, and technical improvements can sometimes improve several dimensions simultaneously. The principal CDC framework is from 2001 and ECDC's handbook from 2014; they remain useful for the attribute model, but current disease-specific, interoperability, equity and legal requirements may impose additional constraints.^[1][2]