Contribution
The paper proposes the selective accessibility model of anchoring and tests three of its implications. In the standard paradigm a judge first says whether a target is higher or lower than an anchor, then estimates it. On this account the comparison is solved by testing whether the target has the anchor value. That test builds a mental model of the target from knowledge consistent with the anchor, and that knowledge stays accessible when the absolute estimate is made. The abstract states it as “anchoring is a special case of semantic priming; specifically, information that is activated to solve a comparative anchoring task will subsequently be more accessible when participants make absolute judgments.” The tests borrow the logic of priming research: accessible information is used only if it is applicable, and its direction depends on whether it is representative of the target.
Key insight
If the anchor works through the knowledge used to compare, then holding the anchor number fixed and changing what that knowledge is good for should change the effect. A numeric-priming account, in which the anchor primes a candidate answer, predicts no change. So the paper varies applicability (same versus different dimension), representativeness (same versus very different object) and the need to consult knowledge at all (plausible versus implausible anchors). It treats each departure from the rule of same number, same effect as evidence that the effect runs through activated target knowledge. The paper also says why the bias persists: “Unlike in other judgmental heuristics, there exists no better, systematic way to solve the problem. Even under optimal circumstances, judges will use the most accessible evidence.”
Assumptions
- Three ideal-type strategies for the comparison, which “may operate jointly in many situations”. If the judge knows the value, the anchor should have no effect. If the anchor is implausible, categorical knowledge suffices, and anchoring “may occur in the form of a judgmental adjustment that will terminate at the nearest boundary of the range of acceptable values”. If the anchor is plausible, the judge builds a mental model, and selective accessibility applies. The model is proposed for plausible anchors: “we propose that participants construct a mental model that selectively increases the accessibility of anchor-consistent information.”
- Positive test strategy. Judges “act as if they decided whether the target possesses exactly the value of the standard”. This is borrowed from Klayman and Ha (1987) and Chapman and Johnson (1994). It is assumed, not tested.
- Use rules from Strack (1992). Accessible information affects a judgment only if it is applicable. It assimilates if representative of the target, and if not it is excluded or used as a standard, which gives contrast.
- Self-generated knowledge. “the biasing information is not provided by an environmental source but is generated by judges themselves”. The mental model is built from what judges retrieve or generate.
- Latency as an index of accessibility (Study 3). A faster absolute estimate is read as more applicable information being active.
- Designs. Students were recruited in cafeterias for a “pretest” of a general-knowledge questionnaire and paid with a chocolate bar. To make the anchors seem uninformative, they were told the values “were randomly selected”. Answers were z-scored within each question to pool across items.
Key results
- Study 1 (applicability; N = 32, Würzburg; 2 anchor × 2 dimension, between-subjects, n = 8 per cell). The comparison asked about one dimension of the Brandenburg Gate or Cologne Cathedral (height, width or length). The estimate was on the same dimension or the other one. Anchors were set at the 15th and 85th percentiles of a calibration group (n = 28). Cell means in z units were: same dimension, high 1.20 and low −0.55 (a difference of 1.75); different dimension, high −0.20 and low −0.44 (a difference of 0.24). Anchor main effect F(1, 28) = 18.23, p < .001 (partial η² ≈ .39). Anchor × dimension interaction F(1, 28) = 10.50, p < .005 (≈ .27). Dimension main effect F(1, 28) = 7.76, p < .01. This “cannot be accounted for by numeric priming”. The authors also say the reduced effect was “mainly due to the stronger effect of high anchors in the same dimension condition”. No simple-effect test for the different-dimension cells is reported, so it is not shown whether a small anchor effect remained there.
- Study 2 (representativeness; N = 32, Trier; 2 anchor × 2 object, n = 8 per cell). The comparison was about Aristotle’s year of birth or the Antarctic’s mean winter temperature. The estimate was about the same object or a very different one (Kant, Hawaii). Anchors were about 1 SD from a calibration mean (n = 151). Cell z means were: same object, high 0.22 and low −0.22; different object, high −0.34 and low 0.34. Interaction F(1, 28) = 6.30, p < .02 (≈ .18). Anchor main effect F < 1. The contrast effect is the paper’s evidence that the anchor works through activated information rather than the number. The same-object assimilation here (0.44 z) is much smaller than in Study 1, and no simple effects are reported. The authors note the pattern was “stronger for the Antarctic question”.
- Study 3 (plausibility and latency; N = 69, 67 analysed; within-subjects, 8 critical pairs after 14 practice pairs, on computer). Plausible anchors were about 1 SD from the calibration mean. Implausible ones were more than 10 SD away, for example Einstein’s first US visit in 1215 (the text says 1212) or the Elbe at 45,000 km. Anchoring: direction main effect F(1, 66) = 7.61, p < .01 (≈ .10). Cell z means were: plausible, high .04 and low −.13; implausible, high .25 and low −.17. Plausibility × direction F(1, 66) = 3.55, p < .07. “implausible anchors were at least as effective as plausible ones.” Latencies (log-transformed): the comparison was slower with plausible anchors (8.87/8.89 against 8.79/8.78) and the absolute estimate faster (8.41/8.38 against 8.86/8.73). Plausibility × response-type interaction F(1, 66) = 72.81, p < .001 (≈ .52). Plausibility effect for comparative latencies F(1, 66) = 8.46, and for absolute latencies F(1, 66) = 61.31. Within-condition correlation of comparative and absolute latencies: r = −.21, p < .05, for plausible anchors, and r = −.02 for implausible ones. The paper does not say over what units these correlations were computed.
- Against conversational accounts. Assimilation to anchors that “can hardly be interpreted as suggestive” is taken to show that “conversational influences are not necessary to account for anchoring effects”.
- On adjustment for implausible anchors. The paper does not settle it: “The latency data suggest that this might not be the case … At this point, the data are merely suggestive, and more research about this issue is needed.”
- Untested proposals. Two anchors should debias by breaking the single-hypothesis test. The positive test strategy “seems to generate information that improves the validity of comparative judgments but distorts subsequent judgments for which no specific hypothesis has been provided.”
Bearing on the record
On CLAIM-298 (whose psychological form is now CLAIM-315): the paper supports the claim’s psychological form for plausible numeric anchors, on indirect evidence. It also shows that reachability does not fix the judgment.
- What it shows: with the anchor number held fixed, the anchor’s effect depends on what the comparison makes available and whether that applies to the later judgment (Study 1). The anchor supplies no evidence about the target: participants were told the values were random. That fits the claim that a frame can change a judgment “without supplying any evidence about what is being judged”. The paper’s own phrasing is close: the biasing information “is generated by judges themselves”.
- What it does not show: accessibility is never measured. Studies 1 and 2 infer it from moderation by applicability and representativeness. Study 3 measures how fast the absolute estimate is given, not what was retrieved. The direct test that anchor-consistent knowledge becomes more accessible (lexical decision) is not here. Harris et al. 2019 (LIT-1067, NOTE-876) locate it in Mussweiler and Strack 2000a, which the record does not hold, and report that five preregistered replications of it found no accessibility effect, with a method that also failed to detect direct priming. Mussweiler and Strack 1999 (LIT-1078, Unavailable) states the model as a model; what tests it reports was not checked. So this paper is the hypothesis and three consistent moderations, not the demonstration that anchors act on retrieval. Numeric priming of a response is argued against directly (Studies 1 and 2). Conversational inference is argued against for implausible anchors. Insufficient adjustment is not ruled out. The paper assigns adjustment to implausible anchors, and those anchored at least as strongly as plausible ones in Study 3.
- Where it narrows the claim: the paper’s own model has two stages. Accessibility is “not sufficient to guarantee its use and determine the direction of its influence”. Study 2 reversed the sign when the activated knowledge was about a different object. So the record’s J = δ(Retrieve(M, q), E) is right that the evaluation stage δ matters, and the paper shows δ can turn the same retrieved content into contrast. The claim’s defeat condition imagines a choice between acting on retrieval and the case in which “framing acts only on evaluation”. This paper’s account is that the anchor acts on both, and its designs cannot separate them. The mechanism is also construction (“retrieve or generate”, “construct a mental model”), not only selecting from a fixed store, which is looser than the sparse-distributed-memory reading drawn from LIT-269.
- Effect sizes are from small samples (n = 8 per cell in Studies 1 and 2). The Study 1 interaction is large in z units, and the Study 3 anchoring effect is small (0.17 z for plausible anchors). Nothing here bears on communication beyond the comparative-question paradigm.
On QUESTION-109: one sentence of the paper is a direct, untested answer for one case. The paper says the anchor-consistent test “improves the validity of comparative judgments but distorts subsequent judgments for which no specific hypothesis has been provided”. On this account, cue-led retrieval helps the task the cue was set for and harms a later judgment that reuses what was retrieved. It also predicts when this happens. The cost arises when the anchor is plausible (so knowledge is consulted), when the retrieved knowledge is applicable to the later judgment, and when it is representative of the target, which gives assimilation rather than contrast. The paper claims there is “no better, systematic way to solve the problem”, which bears on CLAIM-184 and CLAIM-182. It also proposes, without testing, that considering two anchors should debias. None of these trade-off statements is tested in the paper. They are the authors’ reading of their model.