Context-Dependent Preference

Version

v4 · 2026-08-30 · History

Prime #

1340

Origin domain

Decision Science

Subdomain

choice theory → Decision Science

Aliases

Context Sensitive Preference, Menu Dependent Preference

Related primes

Preference

Core Idea

Context-dependent preference occurs when the focal options remain unchanged but their ordering changes after an unchosen comparator, menu composition, presentation, or evaluation procedure changes. Value is constructed relative to the choice context rather than residing as a fixed context-free property of each option. ^[1]

The canonical identity is narrower than the phrase's everyday use. Four commitments must hold together: a focal set of at least two alternatives present in every condition, at least two contexts differing in something other than those alternatives, a contextual element that need not itself be chosen, and a systematic change in the ordering or in the constructed value. The contrast must hold outcomes substantially fixed, so that altered information about what the options deliver is never mistaken for altered context. ^[2]

What the prime asserts, put structurally, is that preference is not the two-place relation the standard notation writes. An agent does not simply prefer x to y; it prefers x to y under some comparison set, arrangement, and response procedure, and that third argument is doing work that the two-place model silently absorbed into the options. The consequence is not that agents are erratic, but that an ordering can be stable, reproducible, and predictable while remaining unrecoverable from behavior unless the context it was indexed to is recorded alongside it.

Structural Signature

Focal set held fixed + non-focal context generator varied → systematic reordering among the fixed alternatives, reversible when the generator is neutralized.

The formula is a controlled contrast rather than a single observation, and the controls are the whole content. One condition establishes nothing; the pattern lives entirely in the difference between two conditions that were supposed to be equivalent. ^[2]

Recurring features:

• A focal set present in every condition. Two or more alternatives whose intrinsic attributes, prices, and feasible consequences are identical across the contrast, ideally down to wording.
• A context generator that is not an outcome. The varied element: a third alternative nobody need take, the size or composition of the assortment, spatial or temporal arrangement, the number of items evaluated at once, the response mode (pick one, name a price, rate, rank), or the scoring rule applied to a fixed dossier.
• Independent manipulability. The generator must be settable separately from the focal options. If moving it also moves what the focal options deliver, the options changed and the contrast has been lost.
• An ordering statistic that actually reorders. Choice share, rank, willingness to pay, or panel score, showing a reversal or a monotonicity violation — not the redistribution any added option produces trivially.
• Information invariance. Nothing in the manipulation alters the agent's beliefs about the focal options' consequences; otherwise the shift is updating.
• A neutralization counterfactual. Removing, fixing, or randomizing the generator restores the baseline ordering or materially reduces the gap, which is what separates a context effect from a coincidence of two samples. ^[2]
• Exclusion: changes to the chooser rather than the surround. Depletion, budget, mood, or satiation index the ordering to internal state, a different indexing with different remedies.

What It Is Not

The prime does not claim that agents lack preferences, or that preference is arbitrary. Context dependence is lawful: the reversals it names are directional and predicted in advance rather than idiosyncratic, and where the conditions of the original demonstrations are reproduced they recur in the predicted direction, which is why they can be engineered by designers and detected by auditors. An agent whose ordering flips unpredictably from one presentation to the next is showing noise, and noise is the null hypothesis this pattern has to beat, not an instance of it. ^[3]

It does not claim that a context effect is a defect. Calling a reversal a failure presupposes that a context-free ordering is the standard the agent should have met, and that standard is a modelling convention rather than a fact about the agent. Some menu-driven shifts are defensible on their own terms, and the prime is deliberately silent on whether any particular one is an error, a reasonable inference, or an exploitation.

It does not treat every share movement as qualifying. Adding an option must take share from somewhere, and context-independent choosers redistribute in proportion. The marker is the anomaly: an added alternative that raises an incumbent's share, or an incumbent chosen from a larger set that was rejected from a subset of it. Absent that, there is nothing beyond arithmetic. ^[4]

It makes no claim about magnitude or prevalence. Nothing in the structure says context effects are large, that they dominate stable attribute preferences, or that context-free utility approximates poorly in any given market. Demonstrating an effect under a designed contrast is far weaker than showing it governs behavior in the wild, and the two are constantly conflated.

Finally, it is not a claim about manipulation. There is no context-free presentation available to fall back on: every assortment has a composition, every list an order, every survey a response scale. Whoever builds the choice situation has already chosen a context, whether or not they were thinking about preference at all. The prime identifies a dependence; it does not identify an intent behind it.

Broad Use

The pattern surfaces wherever an evaluation is performed against a comparison set that someone — a designer, an institution, a piece of software, or the environment — gets to compose.

Consumer choice and assortment design: the asymmetric-dominance decoy, where an option dominated by one incumbent and not the other lifts its dominator's share; the compromise effect, where an extreme addition promotes whatever now sits in the middle; the similarity effect, where a near-duplicate cannibalizes its twin rather than the distant alternative. Assortment size, shelf adjacency, and the displayed price range act as generators too.

Foraging and animal choice: birds, insects, and fish reverse between two constant food sources when an inferior third source is added to the array, and reverse again when it is withdrawn. Mate choice under varying candidate pools shows the same shape. The behavior is legible without attributing deliberation to the organism. ^[5]

Voting and collective choice: an entrant who wins nothing changes which of two unchanged candidates prevails under plurality; the order of pairwise votes on an agenda determines the survivor among fixed proposals; independence conditions in aggregation theory are the formal statement of the property being violated.

Interface and choice architecture: list position, default selection, items visible per screen, simultaneous versus sequential presentation, and the response widget all move outcomes among fixed alternatives.

Algorithmic ranking and recommendation: a scorer applied over a varying candidate set inherits menu dependence mechanically — a normalized score allocates probability across whatever is present, so injecting a near-duplicate splits its twin's share without any psychology involved. Upstream retrieval therefore determines downstream ranking among items it never changed. ^[6]

Valuation and survey methodology: question order, response-scale range, and whether goods are valued alone or against each other change elicited magnitudes and their ordering.

Procurement, hiring, and grant review: the shortlist, the scoring rubric, and whether candidates are reviewed one at a time or side by side each determine which fixed dossier wins.

Clinical and legal decision-making: the treatments presented for consent, and the comparison cases available to a sentencer or damages panel, are comparison sets in the same technical sense.

Clarity

The confusion this prime dissolves is the assumption that two disagreeing measurements of one person's values must include a wrong one. Confronted with a respondent who chooses A over B and then prices B above A, analysts reach for measurement error, inattention, or incoherence, and then argue about which reading to keep. All three responses assume the quantity being measured is a single context-free ordering that one procedure got right. If the ordering carries a context index, both readings can be accurate, and the question "which is the true preference?" is malformed until a context is named. ^[7]

That reframing explains why elicitation methods refuse to converge. Choice, pricing, matching, and rating are not four noisy instruments aimed at one latent scale, so averaging them yields not less error but a number indexed to no context at all. Discrepancy across modes stops being a nuisance to smooth and becomes the measurement's most informative output.

It also separates two diagnoses that decision-makers routinely merge: the agent changed, and the situation changed. A sales decline after a catalogue redesign, a swing in poll results after a candidate withdrew, a jump in plan enrollment after a tier was added — each invites a story about shifting tastes, and in each the tastes may be untouched while the comparison set moved underneath them.

Finally, it retires the idea of a neutral presentation. Once one accepts that the ordering has a context argument, "we just showed them the options" describes a specific setting of that argument rather than an abstention from setting it, and the defense that a design was not intended to influence anyone stops being available. ^[8]

Manages Complexity

The abstraction lets an analyst stop reconstructing the chooser's interior. Nothing in the diagnosis requires a utility function, an attention model, or an account of what the comparison felt like; it requires two conditions, one fixed focal set, and a statistic. That is a substantial saving, because the internal story is the part that cannot be observed, cannot be settled by argument, and reliably turns a technical dispute into a dispute about human nature.

It also collapses a long catalogue of separately named effects into one generator with one remedy family. Decoy, compromise, similarity, range and frequency effects, assortment-size effects, order and position effects, joint-versus-separate evaluation, and response-mode reversals are not a dozen phenomena to be learned and countered one at a time; they are a dozen settings of the same third argument. Whoever controls, fixes, randomizes, or logs that argument has addressed all of them at once. ^[9]

And it ends the search for a hidden scalar. In valuation work, the effort spent triangulating toward one true willingness to pay can be redirected to a cheaper and more honest object: an ordering paired with the context that produced it. A pair is harder to put in a spreadsheet cell and much easier to defend.

What it does not let anyone stop tracking is the context itself. The compression is on the psychological side and comes at the cost of bookkeeping: if the option set, its arrangement, and the elicitation procedure are not recorded with the response, the response cannot be re-indexed later and the dataset is uninterpretable no matter how large it grows.

Abstract Reasoning

The prime licenses a diagnostic that runs on a design, not on a person.

Step one: fix the focal set and audit its invariance. List the alternatives present in both conditions and verify that attributes, prices, and feasible consequences are identical across them. Any drift here disqualifies the contrast before it starts.

Step two: name the generator and prove it non-focal. Identify precisely what varied, and check that it was neither chosen nor capable of changing what the focal options deliver. A third alternative that alters a price, unlocks a bundle, or signals a delivery date has changed the options rather than their surround. ^[2]

Step three: declare the ordering statistic and predicted direction in advance. Choice share, rank, elicited value, or score, with the expected direction stated before the data are seen. A shift found by inspection afterward describes one sample.

Step four: test for reordering rather than redistribution. Two conditions carry the weight. Regularity: no added alternative should raise any incumbent's share. Contraction consistency: an option chosen from a set should still be chosen from a subset containing it. A violation of either is the positive finding; proportional reallocation is not.

Step five: eliminate the rivals in a fixed order. Menu-as-evidence, where the added option informs beliefs about quality or the going rate; state change in the chooser; capacity limits that slow or degrade choice without reordering it; and genuine learning about the focal options. The sharpest boundary case is the last. Coming to prefer a component after reading that its defect rate is lower is evidence-responsive updating, and the tell is that the focal options are no longer informationally identical across conditions — something true of them was disclosed in one condition and not the other. ^[10]

Step six: run the neutralization counterfactual. Withdraw the decoy, standardize the response mode, restore the original shortlist. If the baseline ordering returns, the generator was load-bearing; if it does not, something else moved and the attribution fails.

Knowledge Transfer

What travels is the role structure and the whole apparatus resting on it. A focal set, a non-focal generator, an ordering statistic, and a neutralization test can be instantiated in a shopper, a foraging bee, an electorate, a hiring panel, a recommender's candidate pool, or an automated evaluator scoring outputs against a varying comparison batch. The bearer changes and the test does not, which is why an experiment designed for a supermarket aisle can be re-run almost unmodified on a ranking model. The countermeasures travel with it, since fixing, randomizing, and logging a comparison set are operations on a procedure and require no theory of minds. ^[11]

What does not travel is the mechanism. Attribute-wise comparison, the search for a justifying reason, loss aversion in attribute space, and attentional weighting are hypotheses about why a human chooser reorders, and they have no counterpart in a normalized scorer that produces the identical input-output signature through arithmetic alone. A substrate can satisfy every role by a route sharing nothing with the psychological account, so exporting the explanation along with the diagnosis converts a structural finding into a metaphor.

The normative reading travels least of all. Describing a reversal as a failure of context independence measures behavior against a standard of stable context-free utility that is native to choice theory and foreign elsewhere. In a ranking system, sensitivity to the candidate set is often the designed and desirable behavior; in an ecological setting it may be an efficient use of locally available comparisons. The structure is portable; the verdict must be re-earned in each substrate.

Examples

Formal/abstract

Let a choice function C select from subsets of three alternatives described on two attributes, quality and price. A is high on quality and expensive; B is moderate on both; D is slightly worse than B on quality and slightly more expensive than B, while being neither dominated by nor dominating A.

From the pair, C({A, B}) selects A in 60 percent of trials and B in 40. From the triple, C({A, B, D}) selects A in 35 percent, B in 55, and D in 10. Nothing about A or B was touched, and D is taken one time in ten.

Two consistency conditions fail simultaneously. Regularity fails because expanding the menu raised B's share from 0.40 to 0.55, which no fixed context-free ordering over the three alternatives can produce. Contraction consistency fails because B is the modal choice from the triple and not from a subset containing it. Withdraw D and the 60/40 split returns, which rules out a drifting sample. ^[4]

The mechanism is visible in the geometry: D supplies a comparison that B wins outright and A is not party to, so the triple contains a local victory the pair did not.

Mapped back: {A, B} is the focal set, invariant by construction. D is the generator — non-focal, since it need not be chosen, and independently manipulable, since it alters no attribute of A or B. Choice share is the ordering statistic and the reversal is the finding, distinguished from mere redistribution by the regularity violation. Deleting D is the neutralization counterfactual, and its success closes the diagnosis.

Applied/industry

A firm shortlists two vendors for a platform migration. The proposals are locked: pricing, service levels, references, and a technical annex on migration support are fixed documents that no one edits during the evaluation.

Under the first procedure, each of six panelists reads one proposal at a time and scores it against an abstract rubric before opening the next. Vendor A wins comfortably. Under the second, the same panelists see both proposals side by side in a single comparison table and score them together. Vendor B wins, by a similar margin.

The asymmetry is not mysterious once the attributes are sorted by how evaluable they are in isolation. Headline price and an uptime percentage mean something on their own; A is stronger on both. Depth of migration support means almost nothing on its own — a reviewer holding one annex has no scale for "adequate" — and becomes legible only when there is a second annex to read it against. B is much stronger there. Sequential review therefore weights the attributes that survive isolation, and joint review weights the attribute that only exists as a comparison.

A second generator operated in the same process. A third bid had been withdrawn before scoring began; while it was in the running, its unusually long implementation timeline made both surviving timelines look aggressive, and early straw polls favored A more strongly still. The controlled version is cheap: re-run both procedures with randomized panels on the same dossiers and record the pair of orderings rather than a winner.

Mapped back: the two locked proposals are the focal set, invariant to the letter. The evaluation mode is the generator, and it satisfies the non-focal requirement in an unusually clean way, since it is not an alternative at all but a property of the procedure applied to the alternatives. Panel score is the ordering statistic, and the outcome is a reversal rather than a margin change. The withdrawn third bid is a second, independent generator of the menu-composition type, which is why the case exercises both limbs of the pattern. Neutralization here is not deletion but standardization: fix one procedure, and the ordering it yields is stable — stable, and indexed.

Structural Tensions

T1 — The privileged baseline is a convention, not a discovery. Saying the ordering shifted requires treating one condition as the true reading and the other as a distortion, but every condition has a menu, an arrangement, and a response mode, so no condition is context-free. In practice the baseline is whichever setting came first or was cheapest to run, and a normative judgment enters disguised as a control. The disciplined form reports a pair of context-indexed orderings and names no winner, which is almost never what the commissioning decision-maker wanted.

T2 — Menu as evidence versus menu as constructor. An added alternative can genuinely inform: its presence reveals what the market charges, that a category exists, or what a seller thinks buyers want. Under that reading the shift is sound updating about the focal options and this prime does not apply. Under the constructive reading, value is assembled out of comparisons the menu happens to supply. The same reversal supports both, and separating them requires varying what an option could signal without varying what it compares — rarely attempted, and hard to design.

T3 — The ethical line is invisible to the diagnostic. Because context cannot be switched off, every designer selects one, and an identical decoy is unremarkable when it falls out of an inventory decision and objectionable when introduced to move share. The structural test returns the same result in both: same focal set, same generator, same reversal. Intent, the only thing under dispute in most regulatory arguments, is precisely what the pattern is built to be silent about, so the prime supplies evidence for a charge it cannot substantiate.

T4 — Cross-mode reversals need a mapping the modes do not supply. The most striking evidence for the pattern comes from comparing a choice with a price or a rating, which requires a translation between incommensurable response scales. If the translation is wrong, an apparent reversal is a units artifact rather than a fact about preference. Restricting comparison to one mode dissolves the objection and discards the strongest demonstrations, since within-mode menu effects are smaller and easier to attribute to sampling. The evidence is best exactly where it is most contestable.

T5 — Aggregate reversals are not individual reversals. Most demonstrations are between-subjects: different people face different menus, and no individual is ever observed changing their mind. A population-level violation does imply that some members are context-sensitive, but it identifies neither who nor how many, and the modal member may be perfectly consistent. Reading a share reversal as a claim about a typical chooser's constructed value is a level error, and it is the error most often committed when a laboratory result is carried into a policy argument about what people really want.

T6 — Every remedy trades one distortion for another. Standardizing the context makes responses comparable and abolishes the observed shift, but it does not recover a context-free ordering; it freezes the answer at one arbitrary setting and conceals the dependence rather than removing it. Randomizing the generator converts the bias into variance, which protects an average at the cost of inflating error and destroying the information about which contexts matter. Reporting multiple contexts preserves everything and yields no single number to act on. There is no option that both eliminates the dependence and answers the question that prompted the measurement.

Structural–Framed Character

Context Dependent Preference sits on the framed side of the structural–framed spectrum, labeled mixed-framed at an aggregate of 0.5. All five diagnostics read at half: an even balance rather than a grade one criterion decides.

The skeleton that travels is a controlled contrast. Two focal alternatives appear in both conditions with intrinsic attributes and feasible consequences held fixed; something unchosen changes — a comparator, menu composition, a presentation relation, an evaluation mode; the ordering or choice probability among the unchanged options shifts systematically; and neutralizing the context generator restores or materially reduces the shift. Consumer choice, animal foraging, voting, interface design and algorithmic ranking exhibit the same relational failure of context independence.

Human-practice-bound does the most work at its half weight: an ordering over options presupposes an evaluating system, which keeps the prime off structural ground — but foraging animals and ranking algorithms instantiate the pattern as fully as shoppers, and the entry states the abstraction can be given formally beyond human psychology, which is why it reads half, not full. Institutional origin is half: behavioral decision research is the home ground, not a requirement for the effect. Vocabulary travels at half: preference, framing, anchoring and reference point come along, which the entry works to hold apart. Evaluative weight is half: calling this a failure of context independence measures against a normative standard of stable context-free utility. Import-vs-recognize is half.

The grade puts the burden on the controls: same focal options, fixed consequences, only the context varying. Without that discipline the label describes any preference change and does no work.

Substrate Independence

Context-Dependent Preference is about as substrate-independent as a prime can be — composite 5 / 5 on the substrate-independence scale. Its content is a controlled contrast rather than a claim about minds: hold a focal set fixed, vary something outside it that nobody need choose, watch the ordering among the unchanged alternatives reverse, then return when the element is neutralized. Put that way, the prime is the failure of preference to be the two-place relation standard notation writes, which is why it lands unaltered on consumer choice, animal foraging, voting rules, interface layout, and algorithmic ranking. Abstraction, breadth and evidence line up because the abstraction is a protocol: anything yielding an ordering statistic can be run through it, and the reversals recorded across those fields are literal rather than analogical.

• Composite substrate independence — 5 / 5
• Domain breadth — 5 / 5
• Structural abstraction — 5 / 5
• Transfer evidence — 4 / 5

Relationships to Other Abstractions

Current abstraction Context-Dependent Preference Prime

Parents (1) — more general patterns this builds on

• Context-Dependent Preference is a kind of Preference Prime

  Context-Dependent Preference is Preference specialized to an ordering whose value relation is indexed by a comparison menu or evaluation procedure.

Children (3) — more specific cases that build on this

• Decoy Effect Domain-specific is a kind of Context-Dependent Preference

  Decoy Effect is the asymmetric-dominance menu species in which an unchosen inferior alternative shifts preference between two unchanged focal options.

• Subjective theory of value Domain-specific is a kind of Context-Dependent Preference

  The proposed strict upward parent is prime:context_dependent_preference.

• Joint vs. Separate Evaluation Prime is a kind of Context-Dependent Preference

  Joint versus Separate Evaluation is the evaluation-procedure species in which simultaneous comparison changes attribute evaluability and can reverse judgments.

Hierarchy path (1) — routes to 1 parentless root

• Context-Dependent Preference → Preference

Neighborhood in Abstraction Space

Context-Dependent Preference sits in a moderately populated region (59^th percentile for distinctiveness): it has near-neighbors but no dense thicket of synonyms.

Family — Representation Artifacts & Fresh Perception (11 primes)

Nearest neighbors

• Rank-Dependent Value — 0.75
• Preference — 0.71
• Temporal Inconsistency and Preference Reversals — 0.70
• Context — 0.70
• Framing — 0.70

Computed from structural-signature embeddings · 2026-09-10

Not to Be Confused With

Context-Dependent Preference must first be separated from Preference, its parent. Preference is the bare ordering an agent holds over a choice set, carrying no commitment about stability under changes to that set; a perfectly context-independent agent has preferences in the full sense. The child adds the third argument and the requirement that varying it moves the ordering. The practical test is blunt: if you cannot exhibit two contexts and one focal set held across both, you are describing the parent.

It is not Reference-Point Dependence, although both make value relational rather than absolute. Reference-point dependence evaluates an outcome by its signed change from a movable baseline state, so relocating the baseline can turn the identical outcome from a gain into a loss. Two differences are decisive. The relatum there is a state, not a set, and the effect operates on a single option with no menu at all — one salary, one portfolio, evaluated against where things stood. Here at least two focal alternatives must be present in both conditions, each option's absolute description untouched; only which of them wins changes. The two compose, since a menu can install a reference point, but the composition is not an identity.

It is not Framing or Reframing. Framing holds a situation fixed and changes its presentation or interpretive dress; reframing does the same deliberately and at the level of the frame itself. Presentation is one of this prime's listed generators, so the overlap is real. The separation is that framing works through the description of the focal object and can be demonstrated with a single option and no comparison set, whereas the effects catalogued here are producible with descriptions held word-for-word identical, by adding an alternative nobody takes or changing the order in which unchanged items are read. A study that varies wording is doing framing; one that varies the roster or the procedure is doing this.

It is not Anchoring. Anchoring requires temporal priority — a value encountered first that drags a later magnitude estimate toward it — and it operates on an estimate rather than on an ordering. An anchor need not be an option, need not be available, and need not even be relevant; an arbitrary number will serve. A decoy, by contrast, is a simultaneous member of the choice set with a full attribute vector, and it works by supplying a comparison that it loses. Anchoring moves a number; this moves which of two unchanged alternatives is selected.

It is not Salience, though salience is frequently the mechanism underneath a given case. Salience is a bottom-up relational score by which items stand out from their surround before deliberate attention is allocated, and a decoy plainly changes what stands out. But the two come apart in both directions. Attention can be redirected without reordering, leaving everyone choosing what they chose before and merely faster; and reordering can occur with no perceptual change at all, as when a scoring formula is swapped in a spreadsheet no human reads. Salience is a property of items relative to a surround; this prime is a claim about an ordering's dependence on that surround, and it is unsatisfied until the ordering actually moves.

It is not Rank-Dependent Value. Both index value to a comparison class, which makes the pair genuinely close. Rank-dependent value concerns the evaluation of an outcome by the position its holder occupies within a distribution, so a rank-preserving uniform shift leaves the positional component untouched. Its comparison class is a population of holders rather than a set anyone is choosing among, and the effect appears as a change in an absolute level's felt value. Here the comparison class is the option set itself, and the required evidence is a reversal among fixed alternatives.

It is not Temporal Inconsistency and Preference Reversals, which is the closest formal sibling, since it too delivers a reversal over an unchanged pair of options. The generators differ in kind. Delay is a property of the options' own consequences — when the payoff arrives is part of what the payoff is — so it cannot be varied while holding feasible consequences fixed, and it fails this prime's independent-manipulability requirement by construction. Reversals driven by an approaching horizon belong there; reversals driven by menu composition or elicitation procedure belong here, and a case that mixes the two has not isolated either.

Finally, it is not Revealed Preference, and the relation is not one of rivalry but of obstruction. Revealed preference is an inference rule: read an ordering off observed choices, assuming choices from different sets express one underlying relation. Context dependence is the precise condition under which that assumption is unsound, so an inferred ordering is indexed to the environments that generated it. Stated Preference and the Stated–Revealed Preference Gap address a different axis — which channel the report came through, not which comparison set it was made against.

Solution Archetypes

No catalogued solution archetypes reference this prime yet.

Notes

Contested claims

• [065] Reputable sources disagree on the unconditional form. Huber, Payne and Puto (2014) published a defence of the attraction effect specifically in answer to reported replication failures, and their defence is conditional: the effect appears when the conditions of the original studies are essentially reproduced, and named factors mitigate or amplify it. Lichters, Sarstedt and Vogt (2015) argue that the standard designs neglect background factors in ways that limit practical generalisability. Neither disputes that the effect is directional and predictable under the original conditions, which is what the softened sentence now claims.

References

[1] Tversky, Amos, and Itamar Simonson. "Context-Dependent Preferences". Management Science 39(10), 1993. States the pattern in which the ordering of fixed alternatives depends on the comparison context rather than on context-free values attached to each option. registry ↩

[2] Lichters, Marcel, Marko Sarstedt, and Bodo Vogt. "On the practical relevance of the attraction effect: A cautionary note and guidelines for context effect experiments". AMS Review 5(1), 2015. Sets out the design requirements for a valid context-effect demonstration: the focal alternatives held identical across conditions, a control condition in which the context element is absent, and inferences about the focal options prevented from confounding the manipulation. registry ↩a ↩b ↩c ↩d

[3] Huber, Joel, John W. Payne, and Christopher P. Puto. "Let's Be Honest About the Attraction Effect". Journal of Marketing Research 51(4), 2014. Replies to reported failures to replicate by holding that the effect appears when the conditions of the original studies are essentially reproduced, and names factors that mitigate and amplify it. registry ↩

[4] Huber, Joel, John W. Payne, and Christopher Puto. "Adding Asymmetrically Dominated Alternatives: Violations of Regularity and the Similarity Hypothesis". Journal of Consumer Research 9(1), 1982. Shows that adding an asymmetrically dominated alternative can increase the probability of choosing the option that dominates it, the regularity violation that separates reordering from mere redistribution. registry ↩a ↩b

[5] Shafir, Sharoni, Tom A. Waite, and Brian H. Smith. "Context-dependent violations of rational choice in honeybees (Apis mellifera) and gray jays (Perisoreus canadensis)". Behavioral Ecology and Sociobiology 51(2), 2002. Reports that the relative preference between two constant options changed in both an insect and a bird when a third, relatively unattractive option was introduced. registry ↩

[6] Debreu, Gerard. "Review of R. Duncan Luce, Individual Choice Behavior: A Theoretical Analysis". American Economic Review 50, 1960, pp. 186-188. Shows with the duplicate-recording example that a normalized choice rule mis-allocates share when a near-duplicate enters the set, roughly halving its twin's probability while leaving the distant alternative untouched. registry ↩

[7] Lichtenstein, Sarah, and Paul Slovic. "Reversals of preference between bids and choices in gambling decisions". Journal of Experimental Psychology, 1971. Finds that the bet chosen from a pair was often the one given the lower price when the same bets were bid on individually, so choice and pricing cannot both be reading a single context-free scale. registry ↩

[8] Thaler, Richard H., and Cass R. Sunstein. Nudge: Improving Decisions About Health, Wealth, and Happiness. Yale University Press, 2008. Argues that every choice environment already has an architecture, so presenting options neutrally is not among the options available to a designer. registry ↩

[9] Simonson, Itamar, and Amos Tversky. "Choice in Context: Tradeoff Contrast and Extremeness Aversion". Journal of Marketing Research 29(3), 1992. Gives a common account under which separately named context effects are settings of one underlying comparison mechanism. registry ↩

[10] Wernerfelt, Birger. "A Rational Reconstruction of the Compromise Effect: Using Market Data to Infer Utilities". Journal of Consumer Research 21(4), 1995. Argues that the composition of the offered set is itself information about the options, so a menu-driven shift can be rational inference rather than constructed value. registry ↩

[11] Trueblood, Jennifer S., Scott D. Brown, Andrew Heathcote, and Jerome R. Busemeyer. "Not Just for Consumers: Context Effects Are Fundamental to Decision Making". Psychological Science 24(6), 2013. Reproduces the similarity, attraction and compromise effects within a single perceptual-decision task, showing the same paradigm and the same test carry to a substrate with no consumer content at all. registry ↩