Anchoring Bias

The original finding

Amos Tversky and Daniel Kahneman's 1974 paper "Judgment under Uncertainty: Heuristics and Biases" (Science, vol. 185) opens with a list of three heuristics — representativeness, availability, anchoring — that they argued people deploy when estimating uncertain quantities. The anchoring section reports a small, devastating experiment.

Subjects watched a wheel of fortune marked 0 through 100 spin in front of them. The experimenter had rigged it to stop on either 10 or 65. After it stopped, the subject was asked two questions: first, whether the percentage of African nations among UN member states was higher or lower than the wheel's number; second, to estimate that percentage. The wheel was visibly random, was acknowledged by everyone in the room to be random, and was not represented as being connected to the UN in any way.

The medians were 25 percent for the group that saw 10, and 45 percent for the group that saw 65. Twenty percentage points of separation, on a question with a definite numerical answer (about 28 percent at the time), driven by a number that everyone in the room knew was generated by a casino wheel. That is anchoring bias, and the experiment has been replicated more than a hundred times with different anchors, different domains, and different subject pools. The effect is one of the most robust findings in judgment and decision-making research.

Anchor and adjust

The mechanism Tversky and Kahneman proposed was procedural. When asked to estimate a quantity for which they have no precise prior, people start from an available number — the anchor — and mentally adjust toward what they think the right answer is. The adjustment is conscious and effortful; subjects know they are adjusting. The catch is that adjustment terminates too early. People stop adjusting when their estimate reaches the edge of a plausibility range that still includes the anchor — typically when the estimate "feels good enough" rather than when it would actually pass a calibration test. The asymmetry is built into the search: each step of adjustment costs effort, and there is no internal signal that says "you've gone too far".

anchor-and-adjust:
  estimate ← anchor
  while estimate seems implausible:
      estimate ← estimate ± Δ
  return estimate                # stops at edge of plausibility, not centre

The model predicts that lower anchors will produce lower final estimates and higher anchors will produce higher final estimates, because the adjustment in each case stops as soon as the estimate is no longer rejected — and the boundary of "not rejected" is closer to the anchor than the centre of the plausibility distribution is.

Selective accessibility — the rival explanation

Fritz Strack and Thomas Mussweiler (1997) gave a second account, which has come to dominate the modern literature. They argued that anchoring works by selectively activating information consistent with the anchor. When a subject is asked "is the price of this Mercedes higher or lower than $50,000?", the comparison primes representations of expensive cars, expensive features, expensive options — and those primed concepts then dominate the subsequent absolute estimate. The mechanism is not effortful adjustment from a starting number; it is differential accessibility of information in memory.

Selective accessibility explains several phenomena that anchor-and-adjust does not. Anchors presented subliminally (too briefly to consciously read) still work, which would be impossible under a deliberate-adjustment model. Anchors of different signs (does the answer have an even or odd parity) influence estimates of unrelated quantities through which numerical concepts they prime. And the strength of the anchoring effect tracks the semantic match between anchor and target — the more "anchor-consistent" facts exist in memory, the stronger the pull. Most modern researchers see anchor-and-adjust and selective accessibility as complementary: deliberate adjustment matters when the anchor is plausible and the subject engages with it; selective accessibility matters when the anchor is fast or implausible.

The Social Security number experiment

Dan Ariely, George Loewenstein, and Drazen Prelec ran an experiment at MIT in 2003 that pushed the anchor's irrelevance to the limit. MBA students were shown six consumer items — wine, chocolates, a wireless keyboard, a textbook, and so on. For each item, they wrote down the last two digits of their Social Security number, then indicated whether they would buy the item for that dollar amount, then made a real bid in a Becker-DeGroot-Marschak auction (a procedure in which truthful bidding is the dominant strategy).

The SSN-ending was the anchor; the auction was real money. Students whose SSN ended in 80-99 bid, on average, 60 to 120 percent more for the identical items than students whose SSN ended in 00-19. The effect held across all six items. Crucially, the subjects knew their SSN had no connection to the items' value; the experimenters were explicit that the digits should be ignored. The MBAs were taking a behavioural-economics course at the time.

Item	Avg bid: SSN 00-19	Avg bid: SSN 80-99	Ratio
Cordless trackball	$8.64	$26.18	3.0×
Cordless keyboard	$16.09	$34.55	2.1×
Design book	$12.82	$20.64	1.6×
Belgian chocolates	$9.55	$20.64	2.2×
1998 Côtes du Rhône	$8.64	$27.91	3.2×
1996 Hermitage Jaboulet	$11.73	$37.55	3.2×

Three replications by Fudenberg, Levine and Maniadis (2012) and others found smaller, sometimes statistically marginal effects with the SSN paradigm specifically, but the broader claim — that real bids respond to anchors known to be irrelevant — survives a large meta-analytic literature. The phrase "coherent arbitrariness" entered the lexicon: people's preferences are internally consistent (their orderings across items are stable) but the absolute price level is anchored on whatever first number they considered.

First offers in negotiation

If the first number on the table sets the anchor, negotiation theory predicts that whoever makes the first offer captures a disproportionate share of the surplus. Adam Galinsky and Thomas Mussweiler tested this in 2001 in a mock-negotiation lab study and found exactly that pattern: when the seller opened, the final price was significantly higher than when the buyer opened, even after controlling for asymmetric information. The result has been replicated in thousands of MBA negotiation exercises and confirmed in field studies of salary negotiations, M&A deals, and litigation settlements.

The practical implications are sharp. If you have a defensible analysis of value, open with a number at the favourable edge of your plausibility range; the eventual settlement will move toward you. If the other side opens first, you have two choices. The naive one is to make a counter-offer — but a counter-offer implicitly accepts the anchor's reference frame, and the eventual settlement will be a compromise between the two anchors. The better one is to refuse to negotiate over the first offer, restate the relevant range from your own analysis, and force the conversation back to your distribution before any number takes hold. Negotiation textbooks since Fisher and Ury (1981) call this "reframing"; behavioural economics gives the empirical reason it matters.

Where anchoring shows up in the wild

• Real estate listing prices. Northcraft and Neale (1987) gave real-estate agents identical houses with different listing prices and asked them to appraise. The agents' appraisals moved with the listing price, even when the agents denied being influenced by it. Listing price is now widely treated by behavioural agents as a strategic anchor, not as a neutral information item.
• Restaurant menu pricing. The "decoy" priciest item on a menu — the $58 lobster on a $24-entrée menu — exists primarily as an anchor. Few customers order it; many customers' second-most-expensive choice shifts up. The effect is widely documented in menu-engineering work (Yang 2012; Bohlke 2018).
• Legal damage awards. Plaintiffs who request a specific number receive larger awards on average than plaintiffs who say "let the jury decide." Englich, Mussweiler and Strack (2006) found that even random anchors — sentence requests written by participants who knew the anchor was random — moved subsequent judge-issued sentences. The mock judges were experienced practitioners.
• Salary negotiations. Industry-wide salary studies show that candidates who give a number first end up within roughly 10 percent of their number on either side; candidates who refuse to give a number first and ask the employer to anchor end up either far above or far below the company's internal target. The first-offer advantage holds whenever the candidate's analysis of market value is at least as good as the employer's.
• Charitable fundraising thermometers. Showing a high target makes intermediate donations look smaller and increases average gift size. The mechanism is purely anchor-and-adjust on what counts as a "normal" amount in the context.
• Discount pricing ("$100, now $40"). The crossed-out reference price is an explicit anchor for value; consumers anchor on the original and judge the discount relative to it rather than judging the absolute price. The legal restriction on "compare-at" pricing in several jurisdictions exists because the anchor was being manipulated by retailers who had never actually charged the original price.
• Medical diagnosis. A leading hypothesis — anchoring diagnosis — biases the gathering of subsequent evidence. Croskerry's (2003) review of diagnostic error puts anchoring among the top causes of misdiagnosis, especially in time-pressured ED settings where the first plausible diagnosis displaces consideration of alternatives.

Worked example — how big is a typical anchoring effect?

Consider a stylised version of the Tversky-Kahneman experiment with two anchors, a true value, and an estimator who follows anchor-and-adjust with a constant adjustment fraction.

Let the true answer be T, the anchor be A, and the estimate be E. The simplest anchor-and-adjust model is

E = A + α (T − A)        with α ∈ [0, 1]

α = 0:  no adjustment, full anchor (E = A)
α = 1:  perfect adjustment, no anchor effect (E = T)
α ≈ 0.5: typical empirical fit in lab studies

Plug in the wheel-of-fortune numbers. T ≈ 28 %, A ∈ {10, 65}, α ≈ 0.5:

low anchor:   E = 10 + 0.5 × (28 − 10)  = 10 + 9   = 19   (observed: ~25)
high anchor:  E = 65 + 0.5 × (28 − 65)  = 65 − 18.5= 46.5 (observed: ~45)
gap:          46.5 − 19                = 27.5     (observed: ~20)

The simple model overpredicts the gap somewhat, suggesting α is a bit above 0.5 for this kind of casually-considered question; estimates of α in the literature range from 0.4 to 0.7 depending on incentives, expertise, and how engaged the subject is with the question. The qualitative pattern — a 25-point shift in the anchor produces a roughly 15-25-point shift in the estimate — is robust.

Mitigation — what actually works

Strategy	Approximate effect	Why it works
Consider-the-opposite	Reduces bias ≈ 30 %	Forces generation of anchor-inconsistent evidence; offsets selective accessibility
Independent prior estimate	Largely eliminates bias	Subject's own number occupies the anchor slot before the manipulation arrives
Generic warning ("don't be influenced")	≈ 0 % effect	Warnings without procedural alternatives change nothing
Accuracy incentives	Modest reduction	Increases adjustment effort but does not eliminate the floor
Expertise	Reduces but does not eliminate	Narrows the plausibility range; anchor still pulls within it
Group deliberation	≈ 25 % reduction	Averages individual anchors; some members anchor on rivals' numbers
Procedural counter-anchoring	Strongest in negotiations	Refuse the opponent's anchor, force re-anchoring on your number

The empirical pattern is striking: telling people not to be anchored almost never helps; giving them a procedural alternative — write down your own number first, generate three reasons the anchor is wrong, take the average of an independent group — often does. The implication for institutional design is that debiasing has to happen at the level of process, not at the level of admonition.

Variants and related biases

• Adjustment heuristic. Generalises anchoring to any case where an initial reference (a default, a previous estimate, a self-generated number) is adjusted insufficiently. Insufficient adjustment is one of Tversky and Kahneman's three named heuristics in the 1974 paper.
• Status-quo bias. Closely related: the current state acts as an anchor for evaluating alternatives, so departures from it are systematically underweighted. Samuelson and Zeckhauser (1988).
• Default effect. When a default is set, choices anchor on it; opt-out organ donation, opt-out retirement enrolment, and pre-checked privacy boxes all exploit this. The default is the anchor; switching costs are the friction that prevents full adjustment.
• Reference-dependent preferences (prospect theory). The reference point in prospect theory is an anchor: gains and losses are evaluated relative to it, and changing the reference changes which outcomes register as gains. Kahneman-Tversky (1979).
• Hedonic anchoring. Adaptation level theory and the related literature on hedonic treadmill predict that current well-being is anchored on recent experience; large positive or negative shocks decay back toward the anchor over time.
• Numerical priming. A broader phenomenon in which exposure to any number influences subsequent numerical judgments, even when the prime is not framed as a comparison anchor. Includes irrelevant-number effects in pricing, dieting, and risk assessment.

Common pitfalls

• Confusing anchoring with rational updating. A Bayesian shifts toward an anchor only when the anchor contains information about the target. Anchoring bias is the persistent shift toward anchors with no informational content. The signature is that the effect survives explicit instructions that the anchor is uninformative.
• Mistaking selection effects for anchoring. If high-anchor and low-anchor groups differ on baseline characteristics, you can get apparent anchoring from sampling, not from the manipulation. Always randomise the anchor within subjects or use a matched design.
• Believing experts are safe. Judges, doctors, real-estate agents, and finance professionals all show measurable anchoring in their own domains. Expertise narrows the plausibility range; it does not remove the anchor effect within it.
• Assuming awareness defuses the bias. Telling subjects about anchoring before the experiment reduces the effect by less than a third. The pull is largely below the threshold of introspection.
• Treating anchoring as small. The Ariely SSN result produced 2-3× bid ratios for the same goods. The Galinsky-Mussweiler negotiation result shifts settlement prices by tens of percent. In welfare terms, the anchoring effect on aggregate purchasing and bargaining decisions is large.
• Forgetting the asymmetry. Anchors that are far above plausibility produce smaller absolute pulls than anchors near the edge of plausibility — adjustment terminates earlier for the latter. Strategic anchors are set just outside the plausible range, not at extremes.