Contribution
The paper adds no new experiment. It collects the preference-reversal line from Slovic and Lichtenstein (1968) to Tversky, Slovic and Kahneman (1990) into one argument against procedure invariance, the principle that “the relation of preference should not depend on the description of the options (description invariance) or on the method of elicitation (procedure invariance)” (pp. 364–365). It records the economists’ attempts to remove reversals and their failure except under market discipline. It offers compatibility and prominence as the mechanisms, and concludes that “beliefs and preferences are often constructed—not merely revealed—in the elicitation process” (p. 365).
Key insight
A preference is not read off a stored order. It is computed when asked, and the computation depends on the question. Asked for a price, a person weights the dollar dimension, because it shares the response’s scale. Asked to choose, the person weights the most important dimension, because that gives a good reason. So two procedures that should agree give opposite orderings, and the disagreement is in the procedures, not in a cycle of preferences. Slovic’s baseball epigraph, from Tversky and Thaler, puts the third view: “They ain’t nothing till I call them” (p. 364).
Assumptions
- Procedure invariance as the normative standard. Preference reversal is defined as H ≻ L with C_L > C_H, where H is the high-probability bet, L the high-payoff bet and C their cash equivalents. If invariance holds, reversal implies an intransitive cycle (pp. 366–367).
- Scope. The constructive conception “applies to judgments and choices among options that are important, complex, and perhaps unfamiliar, such as gambles, jobs, careers, homes, automobiles, surgical treatments, and environments” (p. 369). It is not claimed for familiar, simple options.
- Evidence is between procedures and between groups. The studies compare responses to different procedures, from the same respondents or from separate groups (as in the highway problem, p. 368). None reported here follows a respondent’s preference across time.
Key results
The numbers are the cited studies’, as the paper reports them.
- The Las Vegas study (Lichtenstein & Slovic 1973, p. 365). P bet: 11/12 to win 12 chips, 1/12 to lose 24; bet got the higher selling price about 88% of the time, and 87% of those who chose the P bet priced the $ bet higher.
- Persistence (p. 366). Reversals survived Grether and Plott’s (1979) 13 controls, raised payoffs (Pommerehne et al. 1982), money on the desk (Reilly 1982), experimental markets and arbitrage. “Chu and Chu (1990) and Cox and Grether (1994) were finally able to eradicate preference reversals in market settings characterized by repetition, feedback, and harsh penalties for being inconsistent.”
- Diagnosis (Tversky, Slovic & Kahneman 1990, p. 367). Reversal rates of 40% to 50%; “only 10% of preference reversal patterns were intransitive, and the remaining 90% violated procedure invariance”. Overpricing of the L bet accounted for nearly two thirds of the patterns.
- Scale compatibility (Slovic, Griffin & Tversky 1990, p. 367). “the weight of a stimulus attribute in judgment or in choice is enhanced by its compatibility with the response mode.” Reversals fell from 41% with monetary bets to 24% with non-monetary outcomes.
- Riskless reversals (p. 368). L = 1,600 in 1.5 years: S chosen 74% of the time, L priced higher 75% of the time, reversals above 50%, again mainly from overpricing L.
- Process data (Schkade & Johnson 1989, p. 368). More time on payoffs in pricing than in choice among respondents who reversed, little difference among those who did not.
- Prominence (Tversky, Sattath & Slovic 1988, p. 368). Highway safety (Table 1): Program X, 500 casualties at 12M. 67% chose X. In matching, “Only 4% of the inferred choices favored Program X.” “In every case, the primary dimension was given more weight in choice than in matching.”
- Environment (Irwin et al. 1993, p. 370). Air quality preferred in choice, the computer upgrade valued higher in price: reversals in 41% of respondents.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Normatively equivalent elicitation procedures give systematically different preference orderings | strong | many cited replications, incentive and market controls, p. 366 |
| C2 | Most reversals violate procedure invariance rather than transitivity | strong | Tversky et al. (1990): 90% against 10%; Bostic et al. (1990) |
| C3 | An attribute’s weight rises with its compatibility with the response scale | moderate | Slovic et al. (1990), 41% to 24%; Schkade & Johnson (1989) |
| C4 | The prominent attribute weighs more in choice than in matching | strong | Tversky et al. (1988), several problems, 67% against 4% |
| C5 | For important, complex, unfamiliar options, preferences are constructed on the spot in elicitation | moderate | interpretation of C1–C4 with Payne, Bettman & Johnson; not tested directly |
| C6 | Reversals disappear with repetition, feedback and penalties for inconsistency | weak | two cited market studies, one sentence, p. 366 |
Concepts
- Procedure invariance. Preference does not depend on the method of elicitation (p. 364).
- Preference reversal. Choosing one option and pricing the other higher; a failure of invariance or of transitivity (pp. 366–367).
- Scale compatibility. A stimulus attribute weighs more when it is on the response’s scale, because mapping across scales costs effort and because “a response mode tends to focus attention on the compatible features of the stimulus” (p. 367).
- Prominence; strategy compatibility. The more important attribute weighs more in choice, an ordinal procedure, than in matching, a cardinal one (pp. 368–369).
- Construction strategies. “anchoring and adjustment, relying on the prominent dimension, eliminating common elements, discarding nonessential differences, adding new attributes into the problem frame in order to bolster one alternative” (p. 369).
Connections
It summarizes the line Fischhoff, Slovic and Lichtenstein’s labile values (1980) opened, which Feldman and Lynch build on (LIT-1196, read in NOTE-957). It cites Payne, Bettman and Johnson’s adaptive decision maker for construction and Shafir, Simonson and Tversky for reason-based choice.
Measurement context and the interpreter’s state (TERM-039, CLAIM-127). The paper’s evidence is of the kind TERM-039 calls the measurement context entering the response. Compatibility and prominence are both effects of the response mode on how attributes are weighted when the answer is computed; in A129’s model that is C entering f_Y. Nothing in the paper shows C entering the interpreter’s state. The studies compare procedures, and none asks whether the preference constructed under one procedure is carried into the next. The one sequential design it reports, Slovic (1975), in which participants “first matched different pairs of options … and, in a later session, chose between the matched options” (p. 368), is analysed for the discrepancy between the two tasks, not for carry-over. So the paper is a source for the route A129 draws, direct influence of context on the response, and leaves TERM-039’s open route open. Its baseball epigraph (“They ain’t nothing till I call them”, p. 364) reads as the stronger claim, but the review argues only that the response is computed under the procedure.
CLAIM-341. The claim cites the paper for preferences “constructed — not merely revealed — in the process of elicitation”. That wording is the manuscript’s abstract. The published abstract says “constructed in the process of elicitation” (p. 364), and the phrase “not merely revealed” appears in the body, of “beliefs and preferences” (p. 365). The substance holds either way. But construction here is synchronic: each elicitation computes its own answer. That is the form CLAIM-341 sets aside, not the diachronic form it holds. The paper’s support for the claim is therefore indirect: that a judgement is a construction, which is what would make it able to change the judge. The mechanisms are classical (anchoring and adjustment, attention to compatible features, reason-giving). The paper gives no evidence for or against action being part of the loop.
CLAIM-342 and QUESTION-121. Two points bear on what damps the effect. First, scope: construction is claimed for options that are “important, complex, and perhaps unfamiliar” (p. 369), and contingent valuation is “exactly the kind of complex, unfamiliar task in which one would not expect to find stable, well-articulated pref[erences]” (pp. 369–370). So familiarity is the implied moderator, as in Feldman and Lynch, without studies of it here. Second, the only conditions reported to remove reversals are “repetition, feedback, and harsh penalties for being inconsistent” in markets (p. 366). That is outcome feedback scored against a standard the respondent does not set, which fits CLAIM-342’s independent scoring. It arrives through the consequences of the respondent’s own choices, which is what CLAIM-339 would point to.
Bearing on the record
- CLAIM-341: weak, indirect support. The paper supports the synchronic form (each elicitation computes its answer), which the claim does not hold, and says nothing about a constructed preference changing later judgements. The claim’s quotation should be checked against the published wording (p. 364 against the manuscript abstract).
- TERM-039 and CLAIM-127: evidence for C’s direct influence on the response, the route in A129’s graph; no evidence for an arrow into the interpreter’s state.
- CLAIM-342: the one reported remedy, market feedback with penalties, is independent scoring delivered through the consequences of choice.
- No bearing on contextuality: the reversals are explained classically and are differences between procedures, not order effects within one respondent.
- It carries no instruction for machine-learning practice; its practical section concerns contingent valuation, informed consent and managing one’s own preferences.
Limitations
A review of the author’s own line, from an award address. It gives the cited studies’ headline numbers without design detail or variances, and the market results that eradicate reversals get one sentence. “Constructed” is not operationalized beyond failure of procedure invariance; the paper does not distinguish a preference that does not exist before elicitation from one that exists and is reported through a procedure-dependent mapping (the paper itself notes that “the labels do not imply that choice represents one’s true preference”, p. 367).
Open questions
- Is a preference constructed under one procedure carried into a later elicitation under another, so that the order of procedures matters for the same respondent? That is the sequential test TERM-039 asks for.
- Do the market conditions that remove reversals work through feedback (evidence the respondent did not select) or through the penalty on inconsistency (a scoring rule)?