Case-Definition Drift¶

The surveillance failure in which the operational definition of 'a case' silently changes across time, place, or data source while counts are still reported and compared under a stable label — conflating real epidemiological change with an artefact of what the instrument now counts.

Core Idea¶

Case-definition drift is the epidemiological and public-health surveillance failure in which the operational definition of "a case" of a condition changes — across time, across reporting jurisdictions, or across data sources — while aggregate counts and trends derived under that definition continue to be reported and compared as though the definition were stable. The drift may be deliberate and formally announced (a revised ICD code, an updated CSTE case definition, a new screening guideline that catches earlier-stage disease), administrative and implicit (a billing-code split, an EHR documentation template change, a shift in clinicians' threshold for ordering a confirmatory test), or methodological (an upgrade in assay sensitivity that reclassifies subclinical cases as positive). In every variant the mechanism is the same: the operational categorisation instrument that the surveillance system uses to map events in the world to counts in a dataset has changed, while the output time-series label has not. The apparent change in disease frequency, severity, or geographic distribution thus conflates two inseparable contributions — real epidemiological change in the underlying population and artefactual change in what the measurement instrument counts as an event. The artefact component is invisible without deliberate audit of the definition history, because the surveillance system's published output gives no internal signal of the definitional shift. Within epidemiology, two worked examples define the pattern's scope: the decades-long apparent rise in autism prevalence, where successive DSM revisions (DSM-III to DSM-IV to DSM-5) progressively broadened the diagnostic perimeter while awareness and diagnostic intensity also increased, making it methodologically nontrivial to partition the observed trend between real change and definitional expansion; and the "Will Rogers phenomenon" in cancer staging, where refined staging technology moves borderline patients to a higher stage, improving the observed survival rates of both the stage they left and the stage they entered without any change in actual patient outcomes — stage-migration as a threshold-shift form of case-definition drift. The intervention vocabulary the pattern unlocks is specific to surveillance engineering: freezing the operational definition within a surveillance window, maintaining parallel old-definition and new-definition counts during a bridging period, applying retrospective re-coding where feasible, computing bridging coefficients for trend analysis that crosses a definitional change, and publishing the definition-history alongside the time-series so that downstream analysts can audit the commensurability of cross-period comparisons.

Structural Signature¶

Sig role-phrases:

the surveillance system — an apparatus mapping events in the world through an operational case definition to a count or rate
the operational case definition — the categorization instrument (criteria, ICD/CSTE/DSM codes, thresholds, forms) that decides what counts as "a case"
the stable output label — the time-series name and reporting frame that stay fixed even as the definition underneath them moves
the definitional shift — a change to the instrument across time, jurisdiction, or data source: deliberate (revised criterion), administrative (billing-code split), or methodological (assay-sensitivity upgrade)
the conflated observed trend — the published series read as real epidemiological signal plus a definitional-boundary function plus sampling noise
the invisible artefact — the boundary-shift contribution gives no internal signal in the output, so commensurability is never given and must be audited from the definition history
the both-directions sign rule — a broadened definition inflates counts (false alarm); a tightened one deflates them and can mask real worsening (false reassurance)
the bridging toolkit — freeze the definition within a window, dual old/new counting across a bridging period, retrospective re-coding, bridging coefficients, and publish the definition history

What It Is Not¶

Not necessarily a real change in disease. An apparent rise or fall may be partly an artefact of how "a case" came to be counted, not biology, behavior, or an intervention's effect. The observed series is real epidemiological signal plus a shifting definitional boundary plus noise, and crediting a causal story before auditing the definition history mistakes the instrument's move for the world's.
Not Goodhart distortion. Case-definition drift is upstream of agents gaming a measure: the operational definition itself changes, not the strategic response of people to being measured. No one need be optimizing against the metric — the categorization criteria simply moved under a stable label.
Not selection bias. Drift alters which events are counted as cases, irrespective of how the sample was drawn; it is a change in the categorization instrument, not a skew in sampling. A perfectly representative sample still yields incomparable counts across a definitional change.
Not inherently inflationary. Drift does not only manufacture spurious increases. A broadened definition inflates counts (a false alarm), but a tightened one deflates them and can mask real worsening (a false reassurance) — so the direction of the definitional shift sets the sign of the artefact, and the analyst must reason both ways.
Not visible in the published output. The artefact component gives no internal signal in the time-series; the system reports counts as commensurable whether or not its criteria changed. Comparability across a definitional change is never given by the data and must be established by deliberate audit of the definition history.
Not regression to the mean or survivorship bias. It is not a statistical artefact of extreme baselines, nor a selection pathology acting on outcomes. Case-definition drift operates on the input categorization — what gets counted as a case — which is why threshold-shift forms like the Will Rogers phenomenon fall under it as boundary-move special cases.

Scope of Application¶

Case-definition drift lives across the surveillance and trend-analysis subfields of epidemiology and public health, enumerated here by the kind of condition and definitional instrument that drifts; its reach is within that domain. The structurally identical siblings in ML annotation, psychometric rating, and manufacturing metrology are true co-instances of the parent instrument_interpretive_drift, carried there rather than by the epidemiology label.

Infectious-disease surveillance — the salient recent habitat: COVID-19 case counts under repeatedly revised CSTE definitions (symptom-only → PCR-confirmed → antigen-inclusive → reinfection rules), producing apparent step-changes that conflate real waves with definitional moves.
Notifiable-disease reporting — CSTE/CDC case-definition revisions across the reportable conditions, where the agencies publish definition-update histories precisely to support trend analysis across drift.
Psychiatric and developmental epidemiology — the multi-decade autism-prevalence debate under successive DSM revisions (DSM-III → IV → 5), the field's canonical problem of apportioning an observed rise between real change and definitional broadening.
Cancer staging and survival epidemiology — the Will Rogers phenomenon under successive AJCC editions, where refined staging migrates borderline patients and lifts the apparent survival of both stages with no real outcome change (a threshold-shift special case).
Critical-care / sepsis surveillance — the Sepsis-½/3 transitions, which partly disagree on who counts as septic and make cross-definition incidence and mortality comparison non-trivial.
Chronic-disease threshold setting — hypertension and diabetes guideline updates (JNC-8 vs. ACC/AHA 2017; ADA thresholds) that reclassify tens of millions overnight, the threshold-reset form of drift.
Mortality / cause-of-death coding — ICD-9 → ICD-10 → ICD-11 revisions producing step-discontinuities in cause-specific mortality that require bridging coefficients.

Clarity¶

Naming case-definition drift reframes an apparent change in disease as a possible change in the measuring instrument, and that single reframe is its central clarifying force. Without the label, a rising prevalence curve is read as biology, behavior, or the effect of an intervention, and the epidemiologist reaches for causal explanations of a trend that may be partly an artifact of how "a case" came to be counted. With it, the analyst is obliged to ask, before any causal story, whether the operational case definition shifted, where, and how its magnitude compares to the observed move — a discipline that has both deflated false alarms and surfaced false reassurances, as when a tightened definition masked real worsening. The concept thereby decomposes a surveillance time-series into separable contributions — genuine epidemiological signal versus a definitional-boundary function that maps events to counts — and insists they cannot be read off the published output alone, because the system gives no internal signal that its categorisation instrument changed under a stable label.

The label also sharpens the distinction between drift and its neighbors, which determines the correct remedy. It is upstream of Goodhart distortion: the definition itself changes, not agents' strategic response to being measured. It is not selection bias, since it alters which events are counted as cases irrespective of how the sample was drawn. And it subsumes threshold-shift forms — the Will Rogers phenomenon in cancer staging, where reclassifying borderline patients lifts the apparent survival of both the stage they leave and the stage they enter with no change in any patient — as a boundary-move special case rather than a separate puzzle. Localizing the breakdown to the definition, and recognizing that comparability across a definitional change is not given but must be engineered, is exactly what licenses the surveillance-specific toolkit the concept implies: freezing the definition within a window, dual old/new counting across a bridging period, retrospective re-coding, bridging coefficients, and publishing the definition history so downstream analysts can audit whether two periods are even commensurable.

Manages Complexity¶

Surveillance data is beset by a long and heterogeneous catalogue of pathologies that can each bend a reported trend: a revised ICD code, an updated CSTE or DSM case definition, a billing-code split, an EHR documentation-template change, a screening-program rollout that catches earlier-stage disease, an assay-sensitivity upgrade that reclassifies subclinical cases, a guideline that resets a diagnostic threshold and reclassifies tens of millions overnight, a registry-form revision, a shift in clinicians' test-ordering threshold. Confronting an apparent rise or fall in disease without the concept, the epidemiologist faces a sprawl of possible causes and a standing temptation to reach immediately for a biological, behavioral, or interventional story for a movement that may be partly an artifact of how "a case" came to be counted — re-litigating each surveillance series from scratch with no shared structure linking the autism-prevalence debate to the sepsis-incidence transition to the cause-of-death discontinuity at an ICD revision.

Case-definition drift collapses that catalogue into a single diagnosis by reducing any surveillance time-series to a fixed compositional decomposition: the observed series equals a real underlying epidemiological signal, plus a definitional-boundary function that maps events in the world to counts in the dataset, plus sampling noise. Every item in the heterogeneous catalogue — code revision, template change, threshold reset, assay upgrade — enters that decomposition through the same middle term, because each is just another way the operational categorisation instrument changed while the output time-series label stayed fixed. So instead of inventing a bespoke explanation for each anomaly, the analyst tracks one thing: the definition history of the series, and the magnitude of its boundary-shift contribution measured against the size of the observed move. That single regularity — instrument changed under a stable label — is what the analyst reads, and it carries a hard methodological consequence the framing makes explicit: the artefact component is invisible in the published output, so commensurability across a definitional change is never given by the data and must be established by deliberate audit of the definition history.

The decomposition then fixes a compact branch structure for both diagnosis and remedy. Diagnostically, before any causal story, the analyst asks whether the operational definition shifted, where, and how its magnitude compares to the observed change — a fork that has both deflated false alarms (an apparent rise that was definitional broadening) and surfaced false reassurances (a tightened definition masking real worsening). The framing also sorts drift from its neighbors in a way that selects the correct fix: it is upstream of Goodhart distortion (the definition changes, not agents' strategic response to being measured), distinct from selection bias (it changes which events are counted as cases irrespective of sampling), and it subsumes threshold-shift forms such as the Will Rogers phenomenon in cancer staging as a boundary-move special case rather than a separate puzzle. Once drift is identified, the remedy menu is fixed and shares one shape regardless of which catalogue item caused it: freeze the operational definition within a surveillance window, maintain parallel old-definition and new-definition counts across a bridging period, apply retrospective re-coding where feasible, compute bridging coefficients for trend analysis that crosses the change, and publish the definition history alongside the series so downstream analysts can audit commensurability. The epidemiologist thereby reasons from one tracked quantity — the definition history against the observed move — straight to whether a trend is real, which neighbor-effect it is not, and which bridging tool restores comparability, replacing a pathology-by-pathology, series-by-series investigation with a single low-dimensional decomposition and a fixed corrective toolkit.

Abstract Reasoning¶

The concept's master move is a compositional decomposition the epidemiologist runs before any causal story: read an observed surveillance series as real underlying epidemiological signal plus a definitional-boundary function that maps events to counts plus sampling noise, and ask which term moved. The reasoning is FROM "autism prevalence rose across decades" or "COVID cases fell 40% year-on-year" not to "biology changed" but to "decompose the series — how much of this move is the boundary function shifting under a stable label?" The decisive diagnostic discrimination is real change versus instrument change: the epidemiologist treats an apparent epidemiological shift as a candidate measurement-instrument shift until the definition history is audited, reasoning that the artefact component is invisible in the published output (the system gives no internal signal that its categorisation criteria changed), so commensurability across a definitional change is never given by the data and must be established by deliberate audit.

This diagnostic runs in both directions, which is the move's predictive teeth. The epidemiologist predicts that a broadened definition will inflate counts with no real change (deflating a false alarm — an apparent rise that was definitional expansion), and that a tightened definition will deflate counts and can mask real worsening (surfacing a false reassurance). So the reasoner does not assume drift always manufactures spurious increases; it reasons from the direction of the definitional shift to the sign of the artefact, and weighs that artefact's magnitude against the size of the observed move before crediting any trend.

The boundary-drawing move sorts drift from its neighbors in a way that selects the correct remedy. The epidemiologist reasons that drift is upstream of Goodhart distortion (the definition changes, not agents' strategic response to being measured), distinct from selection bias (it changes which events are counted as cases irrespective of sampling), and that it subsumes threshold-shift forms — classifying the Will Rogers phenomenon in cancer staging, where reclassifying borderline patients lifts the apparent survival of both the stage they leave and the stage they enter with no change in any patient, as a boundary-move special case rather than a separate puzzle. Each classification routes the case to a different analytic treatment.

The interventionist move follows a fixed toolkit, each tool predicted to restore comparability across a definitional change: freeze the operational definition within a surveillance window, maintain parallel old-definition and new-definition counts across a bridging period, apply retrospective re-coding where feasible, compute bridging coefficients for trend analysis that crosses the change, and publish the definition history alongside the series so downstream analysts can audit whether two periods are even commensurable. The reasoner predicts that a naive cross-period comparison spanning a definitional change is not interpretable without one of these, and restricts cross-site or cross-time comparison to windows where the operational definitions are commensurable — reasoning from one tracked quantity, the definition history against the observed move, straight to whether a trend is real, which neighbor-effect it is not, and which bridging tool the situation requires.

Knowledge Transfer¶

Within epidemiology and public-health surveillance the concept transfers as mechanism, carrying its compositional decomposition and its bridging toolkit intact across every condition and every kind of definitional change. The same real-signal-plus-boundary-function-plus-noise decomposition, the same both-directions diagnostic (broadened definitions inflate, tightened definitions deflate and can mask worsening), and the same remedy menu (freeze the definition within a window, dual old/new counting across a bridging period, retrospective re-coding, bridging coefficients, publish the definition history) carry across infectious-disease case counts under shifting COVID-19 and CSTE definitions, the multi-decade autism-prevalence debate under successive DSM revisions, cancer stage-migration under successive AJCC editions (the Will Rogers phenomenon as a threshold-shift special case), the Sepsis-½/3 transitions, hypertension and diabetes threshold resets that reclassify tens of millions overnight, and cause-of-death step-discontinuities at each ICD revision. They carry because the substrate is constant: a surveillance system mapping events through an operational case definition to a count, with a downstream comparison machinery that presents counts as commensurable across the change. Within this surveillance range it is mechanism, not analogy — operational case definition, bridging coefficient, and commensurability across a definitional change are literal everywhere in the field. (Two whole disciplines also transfer literally into this domain on the same grounds: gauge-drift statistical-process-control from metrology, and rater-drift / inter-rater-reliability methodology from psychometrics, because they describe the identical instrument-stability problem.)

Beyond epidemiology the honest verdict is shared abstract mechanism — a strong, genuine recurrence, not metaphor. The pattern that recurs is a categorical measurement instrument that maps events through an operational definition to a category, whose definition silently changes over time or across implementers while the output continues to be reported as a commensurable time-series. That mechanism really recurs as a co-instance — not a borrowed shape — in annotation drift in ML datasets (the instrument is a labeling guideline), rater drift in psychometrics, performance-review rubric drift in HR, gauge drift in manufacturing metrology, photometric drift in astronomical surveys, and regulatory-category drift in compliance. Because it is the same mechanism in each, the right unit to carry the cross-domain lesson is the general parent, which the seed names explicitly: the emergent candidate instrument_interpretive_drift (the categorical instrument producing a time-series changes while the series is reported as if it had not). "Case-definition drift" is the public-health surveillance profession's name for that broader primitive in one substrate — sibling to annotation drift, rater drift, gauge drift, and photometric drift in theirs. What stays home-bound is the surveillance accent: the operational case definition vocabulary, the CSTE / DSM / AJCC / ICD revision machinery, the CDC bridging-coefficient methodology, and the specific worked cases (autism prevalence, Will Rogers staging, sepsis transitions). A drifting ML label set has no CSTE archive; an HR rubric has no ICD revision — yet all share the instrument-interpretive-drift mechanism, which is why the lesson belongs to that parent. A caution the seed flags directly: the tempting "four-domain" sketch (cybersecurity signature-set updates, aviation occurrence-classification updates, food recall-categorisation updates, financial audit-rule updates) names parallel substrates of the broader pattern, not transfers of the case-definition-drift named concept — so the correct move there is to invoke instrument_interpretive_drift, not to stretch the epidemiology label onto signatures and audit rules. This is exactly the boundary drawn in Structural Core vs. Domain Accent: the categorical-instrument-changes-under-a-stable-label skeleton lifts to instrument_interpretive_drift and recurs as true co-instances across measurement substrates; the public-health-surveillance accent — case definitions, CSTE/ICD revisions, bridging coefficients — stays home.

Examples¶

Canonical¶

The Will Rogers phenomenon in cancer staging, named by Alvan Feinstein and colleagues (New England Journal of Medicine, 1985) after the quip "when the Okies left Oklahoma and moved to California, they raised the average intelligence level in both states," is the sharpest worked instance. When improved imaging (CT, MRI) detects small metastases that older methods missed, some patients previously called Stage I are reclassified to Stage II — without any change to their disease or treatment. A toy version makes the arithmetic explicit: suppose Stage I mean survival is 10 units and Stage II is 4, and imaging moves a borderline group whose true survival is 6 out of Stage I into Stage II. Removing 6-survivors (below Stage I's mean of 10) raises Stage I's mean; adding them (above Stage II's mean of 4) raises Stage II's mean. Both stages' survival improves, yet not one patient did better.

Mapped back: The staging criteria are the operational case definition, and the imaging upgrade is a definitional shift — a methodological reclassification of who counts as which stage. "Survival by stage" is the stable output label whose comparability silently breaks. The dual improvement with no patient benefit is the invisible artefact: the published survival figures give no internal signal that the boundary moved, so the trend is the conflated observed trend until the staging history is audited.

Applied / In Practice¶

COVID-19 surveillance was a live, high-stakes case. Across the pandemic the CSTE and CDC repeatedly revised the surveillance case definition — from symptom-and-exposure criteria, to PCR-confirmed, to including antigen-positive "probable" cases, to added reinfection rules — while public dashboards kept reporting a single "cases" line. Each revision could shift counts independent of true transmission: adding antigen positives raised counts, tightening confirmation rules lowered them. Epidemiologists therefore treated apparent step-changes as candidate definitional artefacts, compared jumps against the timing of definition updates, and public-health agencies published case-definition change logs so analysts could judge whether two periods were comparable.

Mapped back: The reporting pipeline is the surveillance system and the CSTE criteria are the operational case definition; each revision is a definitional shift under an unchanged "cases" label — the stable output label. The pandemic showed the both-directions sign rule directly: broadening (antigen inclusion) inflated counts, tightening deflated them. Publishing the change logs is exactly the bridging toolkit's "publish the definition history" step, restoring auditable commensurability across the changes.

Structural Tensions¶

T1: Naming the terms versus solving for them (a decomposition that can be unidentifiable). The master move decomposes an observed series into real epidemiological signal plus a definitional-boundary function plus noise — a clarifying insistence that the artefact term exists and must be reckoned with. But naming the terms does not mean one can partition the observed move between them, and often the two are genuinely confounded: the autism-prevalence case is, in the concept's own words, methodologically nontrivial precisely because real change, diagnostic intensity, awareness, and definitional broadening all rose together, with no clean way to apportion the trend. So the framework guarantees the artefact is present while leaving its magnitude unidentified, and a decomposition that cannot be solved can license a lazy verdict in either direction — "it's mostly definitional" or "it's mostly real" — dressed in the authority of a formula. Diagnostic: Can the definitional and real contributions here actually be separated (parallel counts, a natural experiment), or is the decomposition being invoked to assert a split the data cannot identify?

T2: Freezing for comparability versus revising for validity (the definition should sometimes change). The bridging toolkit's core instinct is to stabilize the operational definition — freeze it within a window, resist letting it move — so counts stay commensurable across time. But definitions change for good reasons: a more sensitive assay, refined staging, or broadened criteria that catch real earlier-stage disease all make the instrument more accurate. Freezing therefore protects the time-series at the cost of measurement validity, holding a surveillance system to an outdated, worse definition to keep its trend readable. Parallel old/new counting and bridging coefficients mitigate the conflict but do not dissolve it — they double the counting burden and the coefficients are themselves estimates. Longitudinal comparability and cross-sectional accuracy pull against each other, and the concept's remedies lean toward the former. Diagnostic: Is preserving comparability with the old definition worth counting by a criterion now known to be less accurate — or does the improved definition's validity outweigh the broken trend?

T3: A debunking discipline versus its own over-correction (skepticism that manufactures false reassurance). The concept installs a valuable default: audit the definition before reaching for a causal story, which has deflated real false alarms where an apparent rise was definitional broadening. But the same skeptical reflex over-fires in the opposite direction. Reflexively attributing an alarming, genuine rise to "probably drift" can wave away a true outbreak as a coding artefact, and the concept that corrects naive realism can breed a naive anti-realism just as harmful — the tightened-definition-masking-real-worsening case is exactly where over-eager drift-attribution kills people. The both-directions sign rule is the intended guard, but the standing prior "suspect the instrument first" can produce false reassurance as readily as it deflates false alarms. Diagnostic: Is the drift explanation here supported by an actual definitional change of the right size and sign, or is "it's just drift" being used to dismiss a real signal the data support?

T4: Audit-the-history remedy versus the missing archive (the worst drifts leave no record). The concept's sharp claim — the artefact is invisible in the published output, so commensurability must be established by auditing the definition history — is correct, but it presupposes that a definition history exists and is accessible. That holds for formal, announced changes (CSTE/DSM/AJCC/ICD revisions with published logs) and is exactly why those cases are tractable. It fails for the implicit, administrative, and methodological drifts the concept also names: a clinician's quietly shifting test-ordering threshold, an EHR template change, a coding-practice drift that no one logged. These are often the most insidious precisely because they are gradual and unannounced — and they are the ones with no archive to audit. The remedy tells you where to look while the record you need is frequently the one that was never kept. Diagnostic: Is there an actual documented definition history to audit here, or is the drift the implicit, unlogged kind whose absence of a record is what makes it invisible?

T5: Autonomy versus reduction (a surveillance concept or the instrument-interpretive-drift parent). Case-definition drift is a precise public-health concept with real accent — the operational-case-definition vocabulary, the CSTE/DSM/AJCC/ICD revision machinery, the CDC bridging-coefficient methodology, the worked cases (autism prevalence, Will Rogers staging, sepsis transitions) — and within epidemiology it transfers as full mechanism. But the entry is explicit that it is the surveillance profession's name for a broader primitive, instrument_interpretive_drift (a categorical instrument whose definition changes while its output is reported as commensurable), with annotation drift, rater drift, gauge drift, and photometric drift as true co-instances in other substrates. A drifting ML label set has no CSTE archive; an HR rubric has no ICD revision — yet all share the one mechanism. Diagnostic: Resolve toward instrument_interpretive_drift when the drifting instrument is a labeling guideline, rubric, gauge, or audit rule; toward the named case-definition drift only where operational case definitions, CSTE/ICD revisions, and bridging coefficients are literally in play.

Structural–Framed Character¶

Case-definition drift sits on the framed side of the spectrum — best read as framed-leaning, and if anything a touch further from structure than its methodological cousin carryover effect, because the very thing that drifts is itself a human artifact. Its human_practice_bound character is high and doubly so: "case-definition drift" is the public-health surveillance profession's name for a broader primitive appearing in one substrate, so strip away the surveillance practice — the operational case definition, the reporting pipeline, the commensurable-time-series expectation, the bridging machinery — and there is no drift left to diagnose; and the categorization instrument that moves (a case definition) is not a physical state but a human-made classificatory rule, so even the object of the concept is a practice-artifact rather than a fact of nature. Institutional_origin is pronounced for the same reason: the drifting instruments are institutional furniture — CSTE/DSM/AJCC/ICD revisions — and the remedy apparatus (bridging coefficients, definition-history logs, the CDC methodology) is drawn inside epidemiology. Evaluative_weight is mild but present: drift is framed as a surveillance failure, an artefact to correct, a defect-valence, though the underlying mechanism — an instrument changing under a stable label — is in itself neutral. And vocab_travels is low: the operational-case-definition vocabulary and the CSTE/ICD revision machinery are pinned to surveillance, and off it only the parent primitive carries.

What keeps it off the framed pole is that its diagnosis rests on a genuine substrate-neutral primitive and transfers by recognition, not import, within its domain. The core — a categorical instrument mapping events through an operational definition to counts, its definition silently shifting while the output series is reported as commensurable, the observed trend decomposing into real signal plus a boundary-shift function plus noise — is a real relational structure, recognized intact across infectious-disease counts, autism prevalence, cancer staging, sepsis transitions, and cause-of-death coding. That structural spine is what a pure verdict-label lacks.

The portable skeleton is instrument_interpretive_drift — a categorical measurement instrument whose operational definition changes over time or across implementers while its output continues to be reported as a commensurable time-series. That primitive is genuinely substrate-spanning and recurs as real co-instances (annotation drift in ML datasets, rater drift in psychometrics, gauge drift in metrology, photometric drift in astronomy), but it is precisely what case-definition drift instantiates from its parent, not what makes "case-definition drift" itself travel: the cross-domain lesson belongs to instrument-interpretive-drift, while the operational-case-definition vocabulary, the CSTE/ICD revision machinery, and the CDC bridging methodology stay home in epidemiology. Its character: a doubly practice-bound, mildly defect-framed surveillance failure whose distinctive instrument and remedy apparatus are institutional furniture, structural only in the instrument-interpretive-drift primitive it instantiates and recognizes intact across its own measurement substrates.

Structural Core vs. Domain Accent¶

This section decides why case-definition drift is a domain-specific abstraction and not a prime, and it carries the case for its domain-specificity — there is no separate section for that. The case is unusually clean because the entry's substrate-neutral skeleton is an already-named parent, and everything laid on top is surveillance-engineering machinery.

What is skeletal (could lift toward a cross-domain prime). Strip the surveillance practice and a thin relational structure survives: a categorical measurement instrument maps events through an operational definition to a count, its definition silently changes over time or across implementers, yet its output continues to be reported and compared under a stable label — so the observed series decomposes into real signal plus a boundary-shift function plus noise, and the boundary term is invisible in the output and audit-only. The pieces that travel are abstract — a categorization instrument, an operational definition that can move, a stable output series that hides the move, a both-directions sign rule (broaden inflates, tighten deflates), and the consequence that commensurability is never given but must be established. That skeleton is genuinely substrate-portable, which is exactly why it is already housed in the catalog as the emergent parent instrument_interpretive_drift, recurring as real co-instances in annotation drift, rater drift, gauge drift, and photometric drift — the parent the entry instantiates. But it is the core it shares, not what makes case-definition drift distinctive.

What is domain-bound. Everything the concept adds over that parent is public-health-surveillance furniture and none of it survives extraction intact: the operational case definition vocabulary; the specific revision machinery (CSTE/CDC case definitions, DSM diagnostic criteria, AJCC staging editions, ICD cause-of-death codes, sepsis Sepsis-½/3 transitions, hypertension/diabetes threshold resets); the CDC bridging-coefficient methodology and its dual old/new counting, retrospective re-coding, and definition-history publication; and the worked cases the field actually studies (the autism-prevalence debate, the Will Rogers phenomenon in cancer staging, the COVID case-count step-changes). These are the concrete instrument, remedy toolkit, and empirical cases of epidemiological surveillance. The decisive test: strip the case definition, the CSTE/ICD archive, and the bridging coefficients and the concept does not become a looser version of itself reaching new domains — it collapses to bare instrument_interpretive_drift, the generic "instrument changed under a stable label." A drifting ML label set has no CSTE archive; an HR rubric has no ICD revision; a manufacturing gauge has no bridging coefficient — remove the surveillance apparatus and there is no "case-definition drift" left, only the parent primitive.

Why this does not clear the prime bar. A prime is a relational structure whose vocabulary travels and whose cross-domain transfer is recognition of the same mechanism, not analogy. Case-definition drift's transfer is bimodal. Within epidemiology and public-health surveillance it travels intact — infectious-disease counts, notifiable-disease reporting, psychiatric/developmental epidemiology, cancer staging, sepsis surveillance, chronic-disease thresholds, mortality coding — because the substrate is constant (events mapped through an operational case definition to a count reported as commensurable), so the compositional decomposition, the both-directions diagnostic, and the whole bridging toolkit are the same objects, and the field's own vocabulary is literal everywhere; that is recognition, not analogy. Beyond it the mechanism genuinely recurs, but as instrument_interpretive_drift, not as case-definition drift: annotation drift, rater drift, gauge drift, and photometric drift are true co-instances of the parent in other substrates, none inheriting the CSTE/ICD machinery, and stretching the epidemiology label onto cybersecurity signature sets or audit rules is naming a parallel substrate, not transferring this concept. And when the bare structural lesson is needed cross-domain — an instrument's definition changing under a stable output label conflates real change with artefact — it is already supplied in more general form by the parent the entry instantiates: instrument_interpretive_drift. The cross-domain reach belongs to that parent; "case-definition drift," as named, carries its operational-case-definition vocabulary, its CSTE/DSM/AJCC/ICD revision machinery, and its bridging-coefficient methodology as baggage that does not and should not travel.

Relationships to Other Abstractions¶

Current abstraction Case-Definition Drift Domain-specific

Parents (1) — more general patterns this builds on

Case-Definition Drift is a kind of Instrument Interpretive Drift Prime

Case-Definition Drift is Instrument-Interpretive Drift specialized to an epidemiological case-classification instrument whose stable output label conceals changed inclusion practice.

Hierarchy paths (2) — routes to 2 parentless roots

Case-Definition Drift → Instrument Interpretive Drift → Temporal Decay and Degradation → Entropy (Thermodynamic Sense)

Show alternative path (1)

Not to Be Confused With¶

The Will Rogers phenomenon (stage migration). Not a rival but a special case subsumed under the entry — the threshold-shift form of drift, in which refined staging technology reclassifies borderline patients into a higher category and lifts the apparent survival of both the category they leave and the one they enter without any patient's outcome changing. It is one boundary-move instance of the general "operational definition shifted under a stable label"; case-definition drift also covers non-threshold shifts (whole-criterion revisions, code splits, assay upgrades). Tell: is the definitional change specifically a threshold moving units across a boundary between existing bins, lifting both bins' averages (Will Rogers), or any change at all in what counts as a case (the broader drift it instances)?
Ascertainment / detection / surveillance bias. A distinct epidemiological artefact in which more looking — expanded screening, better imaging, increased testing intensity — finds more of a disease whose true frequency is unchanged, inflating counts. The definition of "a case" can be perfectly stable throughout; what moved is the effort and reach of detection, not the categorization criterion. Case-definition drift instead moves the criterion itself. Tell: did the rule for what qualifies as a case stay fixed while surveillance simply searched harder (ascertainment bias), or did the operational definition of a case actually change (drift)? A screening rollout that both looks harder and redefines eligible disease is both effects at once — separate them.
Goodhart's law / Goodhart distortion. The failure in which a measure, once it becomes a target, is gamed by the agents being measured, degrading its meaning. Case-definition drift is strictly upstream of this: the categorization criterion changes at the surveillance-engineering level, with no strategic actor optimizing against the metric required. Tell: does the distortion require someone responding to being measured (Goodhart), or does it occur even with wholly cooperative, non-strategic reporters simply because the criteria moved under a stable label (drift)?
Annotation drift, rater drift, gauge drift, photometric drift (sibling co-instances). The structurally identical failures in other measurement substrates — a labeling guideline shifting in an ML dataset, an inter-rater standard sliding in psychometrics, a manufacturing gauge losing calibration, a survey telescope's photometric zero-point wandering. These are not the same construct as case-definition drift; they are co-instances of the same parent (instrument_interpretive_drift) in domains that have no operational case definition, no CSTE/ICD archive, and no bridging coefficient. Tell: is the drifting instrument a labeling rubric, a gauge, or a photometric standard (a sibling instance) rather than an epidemiological case definition (case-definition drift)? Same mechanism, different substrate — invoke the parent, not the epidemiology label.
Instrument-interpretive-drift (the parent it instances). The substrate-neutral primitive — a categorical measurement instrument whose operational definition changes over time or across implementers while its output keeps being reported as a commensurable time-series. Case-definition drift is the public-health-surveillance specialization of exactly this, keyed to case definitions and their revision machinery. Tell: strip away the operational case definition, the CSTE/ICD revisions, and the bridging coefficients and what remains — an instrument that changed under a stable output label — is the parent primitive, which is what carries the lesson to signature sets, audit rules, and rubrics; the epidemiology label stays home. (Treated more fully in the sections above.)

Neighborhood in Abstraction Space¶

Case-Definition Drift sits in a sparse region of the domain-specific corpus (88^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (309 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-07-12