Carryover Effect¶

The validity threat in crossover and within-subject designs where residual influence from an earlier treatment persists into a later measurement window, biasing the contrast — its magnitude set by the unit's relaxation time against the inter-treatment gap.

Core Idea¶

The carryover effect is the validity threat in crossover, repeated-measures, and within-subject experimental designs in which residual influence from an earlier treatment or condition persists into the measurement window of a later one, so that the response attributed to the later condition is contaminated by the earlier condition's lingering effect. The mechanism requires two features. First, the experimental unit — a patient, participant, animal, or system — has an internal state that the earlier treatment modifies and that does not return to baseline before the later treatment's measurement window begins: a pharmacokinetic drug concentration that has not cleared, a practised skill, muscle fatigue, sensitisation of a receptor population, an updated belief or expectation. Second, the design assumes that each measurement window samples an independent, clean realisation of the treatment-outcome relationship; when the internal state is still influenced by the prior treatment, that independence assumption is violated, and the estimated contrast between conditions is biased by whatever amount of the prior treatment's effect persists.

The magnitude of the bias depends on the relaxation time of the internal state relative to the inter-treatment gap: a drug with a 36-hour half-life whose carry-in concentration after a 24-hour washout still occupies a substantial fraction of the receptor pool contaminates the second treatment period by an amount proportional to that residual concentration. The standard design interventions are: washout periods — intervals between treatments long enough for the internal state to return to a negligible residual, commonly operationalised as five half-lives for pharmacokinetic carryover; counterbalancing — assigning different participants (or units) to different treatment orders in a Latin-square or Williams design, so that carryover residues are distributed symmetrically across treatment conditions and cancel at the contrast level; and explicit modelling — including a carryover term in the statistical model (often as a lag of the prior treatment in a mixed-effects framework) to estimate and remove its contribution. The concept applies across crossover clinical trials (where it is a major concern in ICH E9 regulatory guidance), behavioural and psychological experiments (where practice, fatigue, sensitisation, and demand-characteristics produce carryover), product-analytics A/B tests run sequentially on the same users, and sensory evaluation panels where palate carryover from one sample to the next requires cleansing protocols and order counterbalancing.

Structural Signature¶

Sig role-phrases:

the unit of analysis — a subject, participant, animal, or system carrying an internal state-variable that treatments modify (a drug concentration, a practised skill, a fatigued muscle, an updated belief)
the successive treatments — the within-subject/crossover design in which the same unit is measured under one condition then another
the independence-of-windows assumption — the design's commitment that each measurement window samples a clean, independent realisation of the treatment-outcome relationship
the earlier treatment's perturbation — the move whereby the first condition writes a residue into the unit's internal state
the incomplete relaxation — the dynamic that the state does not return to baseline before the later treatment's measurement window opens, set by the relaxation time-constant versus the inter-treatment gap
the contaminated contrast — the outcome: the later-condition response is a function of "later treatment plus residual of the earlier," biasing the estimated contrast by an amount proportional to the surviving residual
the remedy menu — the relaxation-indexed fixes: washout (a gap sized to the state, five half-lives when pharmacokinetic), counterbalancing (Latin-square/Williams order distributing the residue symmetrically so it cancels), and an explicit carryover term modelled and subtracted

What It Is Not¶

Not random noise or measurement error. Carryover is a deterministic residue inside the unit of analysis left by its own prior treatment, not stochastic scatter and not a property of the instrument. It biases the contrast in a signable direction proportional to the surviving residual, where noise would merely add variance — so the analyst looks at the carried state, not the measuring device.
Not external confounding. A confounder is an outside variable that co-varies with treatment; carryover is internal, a state the earlier condition itself wrote into the subject and failed to erase. Assignment can be perfectly correct and confounder-free, and carryover still contaminates the later window — the locus is the unit's own history, not a third variable.
Not selection bias. Selection bias concerns who lands in which condition; carryover assumes correct, even randomised, assignment and concerns within-condition contamination from a prior treatment. The threat is to the independence of measurement windows, not to the comparability of groups.
Not regression to the mean. Regression to the mean is an artefact of extreme baselines reverting on remeasurement; carryover is an artefact of prior exposure persisting forward. One arises because a starting value was unusually high or low; the other because the unit's state has not relaxed — different causes, different remedies.
Not hysteresis itself. Hysteresis — a system's state depending on its history, not just current input — is the substrate-independent primitive; carryover is the experimental-design profession's name for that primitive when it appears in the unit of a crossover or repeated-measures design and biases a contrast. A ferromagnet shows hysteresis but runs no biased estimator and has no washout or counterbalancing; the cross-domain lesson travels via hysteresis, not via "carryover."
Not neutralised by any washout whatever. A washout works only when it is long relative to the state's relaxation time (commonly five half-lives for pharmacokinetic residue); a too-short gap leaves substantial residual and a still-biased contrast. Nor is counterbalancing a guaranteed fix — it cancels the residue only when it is symmetric across orders, an assumption that must itself be checked, failing which the carryover must be modelled explicitly.

Scope of Application¶

The carryover effect lives across the within-subject and crossover application contexts of statistics and experimental-design practice; its reach is within that one methodological domain, restaging identically across substantive settings that differ only in which internal state carries the residue. The substrate-independent state-from-history primitive (ferromagnetic remanence, control-system memory, ecological legacy effects) travels under hysteresis, not under this construct.

Clinical crossover trials — the home: residual drug or metabolite from period A biasing the response in period B, codified in ICH E9 regulatory guidance and Senn's crossover-trials literature.
Pharmacology and PK/PD studies — half-life-driven washout calculation, residual receptor occupancy from a prior dose, and desensitisation hysteresis.
Behavioural and psychological experiments — practice, fatigue, sensitisation, and demand-characteristics carryover, with counterbalanced order (Latin-square, Williams designs) the standard control.
Educational research — skill, vocabulary, habit, or motivation retained from an earlier instructional method contaminating measurement of a later one.
Sequential A/B testing on the same users — learning effects, novelty fatigue, and anchoring from variant A distorting the variant-B window, recovered by randomising direction at a cost in power.
Policy and programme evaluation under sequential reforms — behavioural adaptation and institutional learning persisting from reform A into the estimate of reform B.
Sensory-evaluation and panel studies — palate residue from one sample biasing the next, cleared by cleansing protocols and order counterbalancing.

Clarity¶

Naming the carryover effect isolates a validity threat that casual analysis tends to dissolve into its neighbours. A contaminated within-subject contrast is easy to mislabel as noise, as measurement error, or as confounding by some external variable — but carryover is none of these: it is a deterministic residue inside the unit of analysis, left by the unit's own prior treatment, not a property of the instrument or of an outside cause that happens to co-vary. That precision separates it cleanly from selection bias (which concerns who lands in which condition, given here a correct assignment), from confounding (an external variable moving with treatment), and from regression to the mean (an artefact of extreme baselines rather than of prior exposure). The single word tells the methodologist where to look: not at the assignment mechanism or the measuring device, but at the state the earlier condition wrote into the subject and failed to erase.

The construct's sharper payoff is the set of questions it makes mandatory. To invoke carryover is to be forced to specify three things the design otherwise leaves implicit: which internal state-variable carries the residue (a drug concentration, a practised skill, a fatigued muscle, an updated expectation), over what timescale that state relaxes, and whether the later measurement window opens before it has relaxed. Posing the threat in those terms converts a vague worry that "earlier conditions might still be affecting things" into a quantitative comparison between a relaxation time-constant and an inter-treatment gap — which is exactly what makes the standard remedies decidable rather than ritual. A washout is no longer a fixed waiting period but a length chosen against the state's own relaxation (five half-lives, when the residue is pharmacokinetic); counterbalancing is recognised as the move that distributes the residue symmetrically across conditions so it cancels in the contrast rather than biasing it; and an explicit carryover term becomes the option of last resort when the design cannot purge the residue and the analysis must estimate and subtract it. The framework also makes legible what each remedy costs and cannot do — counterbalancing buys an unbiased contrast at the price of statistical power and assumes the residue is symmetric across orders, an assumption that itself becomes a thing to check — so the practitioner reasons about the trade-off deliberately instead of trusting that any washout at all has settled the matter.

Manages Complexity¶

Within-subject and crossover designs are threatened by a long, heterogeneous list of ways an earlier condition can poison a later one — residual drug in the bloodstream, a skill practised in the first arm, accumulated fatigue, a sensitised receptor population, an updated expectation, palate residue, novelty wearing off — each arising in a different substantive setting and each, untheorised, looking like its own separate problem demanding its own ad hoc fix. The carryover effect compresses that whole list to one diagnosis and a single reasoning move. Every item is the same structure: the unit carries an internal state that the earlier treatment modified and that has not relaxed to baseline before the later treatment's measurement window opens, violating the design's assumption that each window samples an independent realisation of the treatment-outcome relationship. So the methodologist stops enumerating substrate-specific pathologies and tracks just three things the diagnosis makes mandatory — which internal state-variable carries the residue, the timescale on which that state relaxes, and whether the later window opens before it has relaxed. The qualitative verdict reads off the comparison of two numbers: a relaxation time-constant against the inter-treatment gap. Gap long relative to relaxation, and the contrast is clean; gap short, and the contrast is biased by an amount proportional to the surviving residual. A vague worry that "earlier conditions might still be affecting things" becomes a single quantitative inequality.

That same compression makes the remedy menu decidable rather than ritual, and that is the second sprawl it tames. Because the bias is a function of relaxation-time versus gap, the three standard interventions stop being independent tricks and become readable responses to one parameter: a washout is a gap chosen against the state's own relaxation (five half-lives when the residue is pharmacokinetic), not a fixed waiting period; counterbalancing (Latin-square or Williams order) is the move that distributes the residue symmetrically across conditions so it cancels at the contrast level rather than biasing it; and an explicit carryover term in a mixed-effects model is the last-resort option of estimating and subtracting the residue when the design cannot purge it. The framework even prices each choice — counterbalancing buys an unbiased contrast at the cost of statistical power and on the assumption that the residue is symmetric across orders, itself now a checkable thing — so the practitioner reasons about the trade-off deliberately. The construct further keeps the diagnosis from collapsing into its neighbours, which is part of how it manages the space: it is a deterministic residue inside the unit, so it is not noise, not measurement error, not external confounding, not regression to the mean, and the single word points the analyst at the subject's own carried state rather than at the assignment mechanism or the instrument. A scattered catalogue of design failures across pharmacology, behavioural experiments, sequential A/B tests, and sensory panels thus reduces to one state-relaxation parameter and a three-option remedy menu indexed by it — from which the presence of bias, its magnitude, and the appropriate fix all read off.

Abstract Reasoning¶

The carryover effect licenses state-space reasoning about the unit of analysis: the subject carries an internal state at the start of each measurement window, the treatment perturbs that state, and the outcome is read before the state has relaxed. The defining diagnostic move runs from a contaminated within-subject contrast to a specific locus, against the pull to mislabel it. A biased crossover contrast is read not as noise, not as measurement error, not as external confounding, and not as regression to the mean, but as a deterministic residue inside the unit left by its own prior treatment. The surface signature is the off contrast; the inferred cause is a carried internal state. Reasoning FROM "the B-period response looks wrong" TO "look at the state the A treatment wrote into the subject and failed to erase" is what points the methodologist at the subject's carried state rather than at the assignment mechanism or the instrument, and it is the move that separates carryover from every neighbor it is casually fused with.

The quantitative heart is a predictive move comparing two timescales. Once the threat is posed as carryover, the reasoner must name which internal state-variable carries the residue (a drug concentration, a practised skill, a fatigued muscle, an updated expectation), the relaxation time-constant of that state, and whether the later window opens before it has relaxed — and then predicts the bias by comparing the relaxation time-constant against the inter-treatment gap. Gap long relative to relaxation, and the contrast is clean; gap short, and the contrast is biased by an amount proportional to the surviving residual (a 36-hour-half-life drug after a 24-hour washout still occupies a substantial fraction of the receptor pool, contaminating the second period by roughly that fraction). Reasoning FROM "relaxation time versus inter-treatment gap" TO "is there bias, and how much" converts a vague worry that earlier conditions might still matter into a single quantitative inequality with a signable magnitude.

The interventionist moves are three, each selected by and tuned against that one parameter rather than applied by rote. Lengthen the washout to a gap chosen against the state's own relaxation (five half-lives when the residue is pharmacokinetic), predicting the residual shrinks to negligibility. Counterbalance the order (Latin-square or Williams design), predicting the residue is distributed symmetrically across conditions so it cancels at the contrast level — at the cost of statistical power and on the assumption, itself now checkable, that the residue is symmetric across orders. Or, when the design cannot purge the residue, add an explicit carryover term (a lag of the prior treatment in a mixed-effects model) and estimate-and-subtract it. Reasoning FROM "the bias is a function of relaxation-versus-gap" TO "purge it with a washout sized to the state, cancel it by symmetry, or model it away" makes the remedy menu decidable, and pricing each option (counterbalancing trades power for an unbiased contrast under a symmetry assumption) lets the practitioner choose the trade-off deliberately rather than trusting that any washout at all has settled the matter.

A boundary-drawing move scopes when the diagnosis even applies and which remedy is admissible, by checking the two structural preconditions: does the unit have an internal state the earlier treatment modifies, and does that state fail to return to baseline before the later window opens? Where the inter-treatment gap already exceeds the relaxation time, there is no carryover and no remedy is needed; where the residue is asymmetric across orders, counterbalancing will not fully cancel it and explicit modelling becomes necessary. Reasoning FROM "is the state still perturbed when the window opens, and is its residue symmetric" TO "does carryover bite here, and which fix is valid" keeps the analysis from invoking carryover where independence already holds, and from trusting counterbalancing where the symmetry it relies on fails.

Knowledge Transfer¶

Within statistics and experimental-design practice the carryover effect transfers as full mechanism, and the cargo that carries is the whole apparatus: the state-relaxation diagnosis, the relaxation-time-versus-gap inequality, the three-precondition scoping, and the washout / counterbalancing / explicit-modelling remedy menu. The concept restages identically across substrates that differ only in what state carries the residue: clinical crossover trials (residual drug or metabolite biasing the next period — the home, codified in ICH E9 and Senn's crossover-trials literature), pharmacology and PK/PD studies (half-life-driven washout, residual receptor occupancy, desensitisation hysteresis), behavioural and psychological experiments (practice, fatigue, sensitisation, demand-characteristics, with counterbalanced order the standard control), educational research (skill, vocabulary, or motivation retained from an earlier method), sequential A/B tests on the same users (learning effects, novelty fatigue, anchoring), policy and programme evaluation under sequential reforms (behavioural adaptation and institutional learning persisting from reform A into reform B), and sensory-evaluation panels (palate residue cleared by cleansing protocols and order counterbalancing). Across all of these the diagnosis, the timescale comparison, and the remedy menu are the same objects — and concrete templates port directly: pharmacology's "five half-lives" washout rule, psychology's Latin-square and Williams counterbalancing designs, and mixed-effects carryover terms all travel as ready-made tools. Even within human-subject work the concept reaches across substantive settings as mechanism, since learning_curve_effects (skill from condition A persisting into B) is simply one of its most common instances. The transfer is literal because the substrate — a unit measured under successive treatments, carrying an internal state that does not relax in time — is held fixed.

Beyond the experimental-design substrate the situation is the shared-abstract-mechanism case with an unusually crisp parent. The carryover effect is the experimental-design profession's name for what happens when hysteresis — a system's state at time t depends on its history before t, not just the current input — appears in the unit of analysis of a crossover or repeated-measures design. Strip the statistics vocabulary and "carryover" reduces exactly to the hysteresis primitive: the unit's state at the start of measurement is a function of its history. That primitive genuinely recurs across domains — magnetic and mechanical hysteresis in materials science, lag and memory in control systems, alternative stable states and legacy effects in ecology — but those recurrences are hysteresis travelling, not the carryover effect. None of them inherits the design-validity machinery; a ferromagnet exhibiting remanent magnetisation is not running a "biased contrast estimator," and there is no washout period, no counterbalancing, and no independence-of-measurement-windows assumption to violate.

So the cross-domain lesson — a system carrying state from its past will contaminate a present reading unless the past has had time to relax — should be carried by hysteresis, the substrate-independent primitive of which carryover is one domain instance, not by "carryover effect," whose value as a named construct lives inside statistics and experimental design. What stays home-bound is everything that makes this entry methodological: the violated independence-of-measurement-windows assumption, the biased-and-inconsistent contrast estimator, the washout/counterbalancing/explicit-term remedy family, and the regulatory and design literature (ICH E9, Williams designs, mixed-effects carryover modelling) built around it. The honest framing is therefore: the state-from-history shape is general and travels via hysteresis; the crossover-design validity threat and its remedy menu are the domain accent that stays in statistics and data science (see Structural Core vs. Domain Accent).

Examples¶

Canonical¶

Consider a two-period, two-treatment (AB/BA) crossover trial comparing two formulations of a drug, each patient receiving both with a gap between. Suppose the drug has a 36-hour elimination half-life. If the washout between periods is only 24 hours, the residual fraction of drug still present when the second period's measurement window opens is (½)^(24/36) = (½)^0.667 ≈ 0.63 — roughly 63% of the prior dose's concentration persists, heavily contaminating the second measurement. The standard rule is to wait five half-lives: after 5 × 36 = 180 hours, the residual is (½)^5 = 1/32 ≈ 0.03, about 3%, treated as negligible. The bias in the estimated A-versus-B contrast is thus governed directly by the ratio of the washout gap to the drug's relaxation time.

Mapped back: The patient is the unit of analysis carrying a drug-concentration state; the two dosing periods are the successive treatments under the independence-of-windows assumption. Period 1's dose is the earlier treatment's perturbation; the 63% still present at 24 hours is the incomplete relaxation, biasing the contaminated contrast. Choosing five half-lives is the washout member of the remedy menu, a gap sized to the state's own relaxation.

Applied / In Practice¶

Professional sensory-evaluation panels — wine judging, coffee cupping, food-product testing — face carryover on the palate: a strongly flavoured, tannic, or spicy sample lingers and distorts perception of whatever is tasted next, so a sample rated after an intense one scores differently than it would fresh. Panels manage this with two of the design's standard tools. Between samples, tasters cleanse the palate (water, plain crackers, and timed rest intervals) — a washout sized to how long the residue lingers. And the presentation order is counterbalanced across panelists using Latin-square or Williams designs, so that each sample follows every other sample roughly equally often across the panel, distributing any residual influence symmetrically so it cancels when sample means are compared.

Mapped back: The taster is the unit of analysis carrying a palate state; successive samples are the successive treatments whose independence-of-windows assumption palate residue violates. The lingering intense flavour is the earlier treatment's perturbation and its persistence the incomplete relaxation biasing the contaminated contrast. Palate cleansing and Latin-square ordering are the washout and counterbalancing members of the remedy menu, the latter cancelling the residue by symmetry across the panel.

Structural Tensions¶

T1: Washout as purge versus the cost of a long gap. Sizing the washout to the state's relaxation — five half-lives, so 5×36 = 180 hours for a 36-hour-half-life drug — drives the residue to negligibility (≈3%) and is the cleanest fix. But the longer the gap the more the study pays for it: participant dropout rises, the underlying condition progresses, and the "stable unit" the design assumes itself drifts over the extended interval, so a washout long enough to purge one residue can introduce a different contamination (the subject is no longer the same subject). The tension is that lengthening the gap monotonically shrinks the pharmacokinetic residue while monotonically growing the practical and biological costs, and there is no gap that is both long enough to purge and short enough to keep the unit fixed. Diagnostic: Is the washout here long enough to relax the state, yet short enough that the unit has not itself changed over the interval?

T2: Counterbalancing's symmetry cancellation versus asymmetric carryover. Counterbalancing is elegant: distribute treatment orders across units so residues land symmetrically across conditions and cancel at the contrast level, no waiting required. But the cancellation is exact only when carryover is symmetric — the residue A leaves into B equals the residue B leaves into A. When carryover is asymmetric (a strong drug contaminates a weak one more than the reverse, a hard task fatigues more than an easy one), the residues do not cancel and a biased contrast survives the counterbalancing that was supposed to remove it. The tension is that the remedy which avoids the cost of a long washout silently substitutes an assumption — order-symmetry — that is often false precisely in the cases where carryover is largest, and it costs statistical power besides. Diagnostic: Is the carryover here symmetric across orders (so counterbalancing cancels it) or asymmetric (so it survives and must be modelled)?

T3: Explicit modelling as last resort versus the non-identifiability trap. When the design cannot purge the residue, adding a carryover term to estimate and subtract it looks like the general escape hatch. But in the standard two-period, two-treatment crossover the carryover effect is aliased with the treatment-by-period interaction: the same data cannot separately identify how much of the period-2 difference is treatment and how much is lingering period-1 residue. Modelling the carryover then does not cleanly remove it; it forces a choice between two explanations the design cannot distinguish. The tension is that the remedy of last resort — model it away — presupposes an identifiability the very design that needed the remedy typically lacks, so "just include a carryover term" can manufacture false precision about a quantity the data do not pin down. Diagnostic: Can this design separately identify the carryover term from the treatment-by-period interaction, or is the modelled carryover an artefact of an under-identified specification?

T4: Within-subject efficiency versus the exposure it creates. The whole appeal of crossover and repeated-measures designs is efficiency: each unit serves as its own control, between-subject variance is removed, and far fewer subjects are needed than a parallel design. Carryover is the price of exactly that structure — a between-subjects (parallel) design is immune to carryover because no unit is measured twice, and it pays for that immunity with the very between-subject variance the crossover eliminated. The design's signature advantage and its signature vulnerability are one feature: reusing the unit is what buys the power and what admits the residue. The tension is that you cannot capture within-subject efficiency without accepting carryover exposure, so the choice of design is a choice of which threat to bear. Diagnostic: Is the within-subject efficiency worth the carryover exposure it necessarily brings, or would a parallel design's immunity be cheaper than the washout, counterbalancing, and modelling this contrast will require?

T5: Carryover as contamination to purge versus as the phenomenon of interest. Framed as a validity threat, carryover is a residue to be washed out, cancelled, or modelled away — a bug in the contrast estimator. But the same persistence is often exactly what a study wants to measure: the durability of a treatment effect after dosing stops, the retention of a learned skill, the long-run adaptation to a policy, the lingering of a sensory impression. Treating every carried state as contamination to eliminate risks purging a real effect that is the actual object of inquiry. The tension is that "residue from an earlier treatment persisting into a later window" is a defect when the target is a clean momentary contrast and a signal when the target is persistence itself — and the same mechanism, the same measurement, reads as either depending on the question. Diagnostic: Is the carried state here a contaminant of the contrast you want, or is its persistence the very effect you are trying to measure?

T6: Autonomy versus reduction (an experimental-design threat or the hysteresis primitive). "Carryover effect" is the experimental-design profession's fully specified construct — the violated independence-of-windows assumption, the relaxation-versus-gap inequality, the washout/counterbalancing/explicit-term remedy menu, the ICH E9 and Williams-design literature — and within statistics it transfers intact as mechanism across clinical trials, PK/PD, behavioural experiments, A/B tests, and sensory panels. But strip that apparatus and what remains is exactly hysteresis: the unit's state at the start of measurement is a function of its history, not just the current input. That primitive recurs across materials science (magnetic remanence), control systems (lag/memory), and ecology (legacy effects) — but those are hysteresis travelling, not carryover, since none inherits a biased-contrast estimator or a washout. The tension is that the state-from-history shape is fully general while the design-validity machinery is home-bound. Diagnostic: Resolve toward hysteresis when carrying the state-from-history lesson to any system with memory; toward carryover effect when a crossover or repeated-measures contrast is biased by a unit's own prior treatment in situ.

Structural–Framed Character¶

The carryover effect sits on the framed side of the spectrum — best read as framed-leaning: a methodological construct constituted by the practice of experimental design, held just off the pole by the genuine substrate-neutral primitive it rests on. Its human_practice_bound character is high and is the load-bearing point: "carryover effect" is the experimental-design profession's name for what happens when a substrate-independent primitive appears inside the unit of a crossover or repeated-measures design and biases a contrast — strip away the design practice (the independence-of-measurement-windows assumption, the contrast estimator, the washout and counterbalancing protocols) and there is no carryover effect left, only the bare fact that a unit's state depends on its history. A ferromagnet with remanent magnetisation exhibits that fact but runs no biased estimator and has no washout, exactly because it is outside the practice that constitutes the concept. Institutional_origin points the same way: the construct is methodological furniture — the ICH E9 guidance, the Williams and Latin-square designs, the mixed-effects carryover term — drawn inside statistics, not a fact of nature. Evaluative_weight is mild but present: carryover is framed as a validity threat, a contaminant to purge, which carries a defect-valence (though the entry's T5 notes the same persistence is sometimes the very effect of interest, so the valence is question-relative rather than intrinsic). And vocab_travels is low: the design-validity machinery is pinned to experimental design, and off it only the underlying primitive carries.

What keeps it off the framed pole is that its diagnosis rests on a genuine substrate-neutral mechanism and transfers by recognition, not import, within its domain. The core — a unit carrying an internal state the earlier treatment modified, not yet relaxed when the later window opens, biasing the contrast by an amount set by the relaxation time against the inter-treatment gap — is a real relational structure, and it restages identically across clinical trials, PK/PD, behavioural experiments, sequential A/B tests, and sensory panels, the same objects recognized in each. That structural spine is what a pure judgment-label lacks.

The portable skeleton is hysteresis — a system's state at time t depending on its history before t, not just the current input. That primitive is genuinely substrate-spanning and recurs as real co-instances (magnetic remanence, control-system lag, ecological legacy effects), but it is precisely what the carryover effect instantiates from its parent, not what makes "carryover effect" itself travel: the cross-domain lesson — a system carrying state from its past contaminates a present reading unless the past has had time to relax — belongs to hysteresis, while the violated independence assumption, the biased contrast estimator, and the washout/counterbalancing/explicit-term remedy menu stay home in statistics. Its character: a practice-constituted, mildly defect-framed experimental-design validity threat whose distinctive remedy machinery is methodological furniture, structural only in the hysteresis primitive it instantiates and recognizes intact across its own substrates.

Structural Core vs. Domain Accent¶

This section decides why the carryover effect 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 prime, and everything laid on top of it is experimental-design machinery.

What is skeletal (could lift toward a cross-domain prime). Strip the experimental-design practice and a thin relational structure survives: a system's state at a given moment is a function of its history, not just its current input, so a present reading is a mixture of the present drive and an unrelaxed residue of the past — and the size of that residue is set by the system's relaxation time against the gap since the past drive. The pieces that travel are abstract — a state-carrying unit, a perturbation that writes into that state, an incomplete relaxation before the next reading, and a residue whose magnitude reads off relaxation-time-versus-gap. That skeleton is genuinely substrate-portable, which is exactly why it is already housed in the catalog as hysteresis, recurring as real co-instances in magnetic remanence, control-system lag, and ecological legacy effects — the parent prime the entry instantiates. But it is the core it shares, not what makes the carryover effect distinctive.

What is domain-bound. Everything the carryover effect adds over hysteresis is experimental-design furniture and none of it survives extraction intact: the independence-of-measurement-windows assumption that the residue violates; the biased-and-inconsistent contrast estimator the residue corrupts; the relaxation-time-versus-inter-treatment-gap inequality operationalized as "five half-lives" for pharmacokinetic washout; the washout / counterbalancing (Latin-square, Williams) / explicit-carryover-term remedy menu and its costs (power loss, the order-symmetry assumption, the carryover-vs-period-interaction aliasing); and the regulatory and design literature (ICH E9, Senn's crossover trials) built around it. These are the worked apparatus and empirical cases (crossover clinical trials, PK/PD washout, sensory-panel palate cleansing) that experimental-design practice actually runs on. The decisive test: strip the crossover design and the contrast estimator and the carryover effect does not become a looser version of itself reaching new domains — it collapses to bare hysteresis, the mere fact that a unit's state depends on its history. A ferromagnet with remanent magnetization exhibits that fact but runs no biased estimator, has no washout, no counterbalancing, and no independence assumption to violate; remove the design practice and there is no "carryover effect" left, only state-from-history.

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. The carryover effect's transfer is bimodal. Within statistics and experimental design it travels intact — clinical crossover trials, pharmacology/PK-PD, behavioural and psychological experiments, educational research, sequential A/B tests, policy evaluation under sequential reforms, sensory panels — because the substrate is held fixed (a unit measured under successive treatments carrying an internal state that does not relax in time), so the diagnosis, the timescale inequality, and the whole remedy menu are the same objects, and even concrete templates (five-half-lives washout, Williams counterbalancing, mixed-effects carryover terms) port as ready-made tools; that is recognition, not analogy. Beyond the experimental-design substrate it does not travel as itself: materials science, control systems, and ecology inherit none of the design-validity machinery, and what recurs there is hysteresis, not carryover. And when the bare structural lesson is needed cross-domain — a system carrying state from its past contaminates a present reading unless the past has had time to relax — it is already supplied in more general form by the parent the entry instantiates: hysteresis. The cross-domain reach belongs to that parent; "carryover effect," as named, carries its violated independence assumption, its biased contrast estimator, and its washout/counterbalancing/modelling remedy menu as baggage that does not and should not travel.

Relationships to Other Abstractions¶

Current abstraction Carryover Effect Domain-specific

Parents (1) — more general patterns this builds on

Carryover Effect is a kind of Hysteresis Prime

Carryover is hysteresis specialized to an experimental unit whose prior treatment state persists into a later measurement window.

Hierarchy paths (4) — routes to 4 parentless roots

Carryover Effect → Hysteresis → Path Dependence → Dependency

Show alternative paths (3)

Not to Be Confused With¶

Period effect. In the same crossover design, a systematic change in outcome tied to when a measurement is taken — first period versus second — that affects every unit regardless of which treatment it received or what it received before (equipment recalibration between periods, seasonal drift, participants relaxing into the protocol). Carryover is history-dependent: it is the residue of the specific earlier treatment the unit actually got. A period effect is order-position-dependent but treatment-history-blind. Tell: would the effect still appear if every unit had received an inert placebo in the first period? If yes it is a period effect; if it requires the earlier active treatment to have written a residue, it is carryover.
Treatment-by-period interaction. The statistical term the carryover effect is aliased with in the standard two-period, two-treatment crossover — the possibility that a treatment's effect genuinely differs by the period in which it is given. The design cannot separately identify this from lingering first-period residue, so "add a carryover term" often just forces a choice between two explanations the data cannot distinguish. Carryover is the substantive mechanism (an unrelaxed state); the interaction is the estimand it becomes confounded with. Tell: are you naming a physical residue inside the unit (carryover) or an unidentifiable parameter in the model that could equally be that residue or a real period-varying treatment effect (the interaction)?
Order effects (the psychology super-type). The umbrella term in experimental psychology for any way the sequence of conditions influences responses — practice, fatigue, boredom, sensitisation, anchoring. Carryover is the member of this family defined specifically as an earlier treatment's residue biasing the contrast in a within-subject design; order effects also include sequence-driven changes that are not residues of a particular prior treatment. Tell: order effects is the whole family of sequence-sensitivities; carryover is the specific case where a named prior treatment leaves an unrelaxed internal state — part, not whole.
Learning-curve effects. A common instance of carryover, not a rival — a skill, familiarity, or strategy acquired under condition A that persists into condition B and inflates its measured performance. It is one substrate (a practised-skill state) filling the general carryover slot; the diagnosis, timescale comparison, and remedy menu are identical. Tell: is the carried state specifically a learned competence that improves later performance (learning-curve effect), or any unrelaxed residue at all — drug concentration, fatigue, palate residue (the broader carryover effect it instances)?
Confounding. An external variable that co-varies with treatment assignment and biases the comparison across units. Carryover's locus is entirely different: it is internal to the unit and self-generated — the state the earlier condition wrote into the subject and failed to erase — so assignment can be perfectly randomised and confounder-free while carryover still contaminates the later window. Tell: is the contaminating cause a third variable outside the unit that moved with treatment (confounding), or a residue of the unit's own prior treatment carried forward in time (carryover)?
Hysteresis (the parent prime it instances). The substrate-independent primitive — a system's state at time t is a function of its history before t, not just the current input — that recurs as magnetic remanence, control-system lag, and ecological legacy effects. Carryover is the experimental-design profession's name for that primitive when it appears inside the unit of a crossover design and biases a contrast; the design-validity machinery (independence assumption, contrast estimator, washout, counterbalancing) is what carryover adds and what hysteresis lacks. Tell: strip away the biased estimator, the washout, and the counterbalancing and what remains is bare state-from-history — at which point the travelling concept is hysteresis, not carryover, which is why a ferromagnet shows the primitive but runs no washout. (Treated more fully in the sections above.)

Neighborhood in Abstraction Space¶

Carryover Effect 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