Anchoring bias and what Tversky and Kahneman's wheel experiment showed
Anchoring bias occurs when an initial numerical value influences a later estimate, even when the starting number gives little useful information about the answer. Tversky and Kahneman's experiment found different median estimates after participants considered different numbers from a wheel (Tversky and Kahneman, 1974).
The classic demonstration involved a comparison followed by an absolute estimate; later experiments explored how that sequence changes reasoning. Research on adjustment, selective accessibility, valuations and expert judgements provides competing explanations, while a recent meta-analysis documents substantial variation across experimental settings (Schley and Weingarten, 2026).
What anchoring bias is and why initial numbers matter
Anchoring bias describes a pattern in quantitative judgement: an estimate tends towards a number that has already entered the task. Tversky and Kahneman (1974) described how an initial value can come from the framing of a question or from an incomplete calculation. A numerical reference can therefore be part of a deliberate comparison, even when the number itself does not answer the question.
The defining feature of the wheel demonstration is the contrast between an arbitrary starting point and a knowledge question. Participants were asked about membership of the United Nations, but the wheel supplied no corresponding information about countries (Tversky and Kahneman, 1974). The reference preceded the estimate. The reported medians compare experimental groups.
Anchors in quantitative judgements
Numerical estimates use reference points. In the original account, a participant could begin with a suggested number and move away from it when forming an estimate (Tversky and Kahneman, 1974). Incomplete calculation could also supply an anchor (Tversky and Kahneman, 1974). Epley and Gilovich (2001) separately studied anchors generated by participants.
Furnham and Boo (2011) reviewed anchoring across different models and tasks, including work on ability, individual traits, mood and knowledge. These variables complicate claims that a single cognitive weakness explains every anchored answer. Their review also discussed everyday decision domains. Experimental conditions vary between tasks.
Anchoring and adjustment
The initial explanation involved insufficient adjustment from a starting figure. Tversky and Kahneman (1974) described estimates that begin at an initial value and are revised before the final response. An adjustment can have a direction without reaching the figure that a respondent might have produced without encountering the starting value. The mechanism requires separate testing because a numerical response alone does not reveal whether a person adjusted from a proposed starting point or retrieved knowledge consistent with that starting point (Epley and Gilovich, 2001; Mussweiler and Strack, 1999).
Epley and Gilovich (2001) later found that participants reported adjustment particularly when generating their own anchors. Their experiments separated anchors suggested by the experimenter from values generated by the respondent. Anchor origin differed between conditions. Simmons and colleagues (2010) then challenged a strict division between the psychological mechanisms associated with those two anchor types.
How Tversky and Kahneman tested anchoring with a wheel
The wheel experiment formed part of Tversky and Kahneman's account of judgement under uncertainty, published in Science in 1974, and paired a conspicuous random number with a question asking for an estimate of an otherwise unrelated quantity. Its central design made the starting number conspicuous before the participant gave a percentage estimate. The wheel was spun in the participant's presence, and different groups received different starting points for the same underlying question (Tversky and Kahneman, 1974).
The wheel and United Nations question
Participants considered the percentage of African countries in the United Nations. For each quantity, the researchers presented a number between 0 and 100 using a wheel of fortune, then asked whether the actual percentage was higher or lower (Tversky and Kahneman, 1974). Only after this comparison did participants provide their own percentage estimate. That comparison explicitly directed attention to the arbitrary value and established a high-or-low judgement before any percentage estimate was recorded (Tversky and Kahneman, 1974).
Bornmann and Ganser (2025) described the same two-stage procedure in their later research on scientific quality assessments. Their discussion used the historical experiment to explain why a comparison question can make a numerical value relevant to a subsequent estimate. In the wheel task, the value came from the experimental apparatus, and its role was procedural rather than factual. The procedure measured an absolute estimate after the comparison, whereas separate experiments by Epley and Gilovich (2001) and Mussweiler and Strack (1999) investigated how participants constructed responses.
The reported estimates
The group shown 10 gave a median estimate of 25 for the percentage of African countries in the United Nations (Tversky and Kahneman, 1974). The group shown 65 gave a median estimate of 45 (Tversky and Kahneman, 1974). These were medians within the original experimental conditions. They are medians from the groups in this experiment (Tversky and Kahneman, 1974).
Tversky and Kahneman (1974) also reported that payoffs for accuracy did not reduce the anchoring effect in their demonstration. That result concerns the tested incentive. Later work revisited incentives under different conditions, including situations where participants were more or less certain of the direction of adjustment (Simmons et al., 2010). Direction certainty mattered in that test because accuracy motivation increased movement away from both provided and self-generated anchors only when participants were certain which way to adjust (Simmons et al., 2010).
The multiplication experiment
Tversky and Kahneman (1974) also described a separate experiment in which high school students rapidly estimated a multiplication product. One group worked from an ascending expression and another from a descending expression; the order of the same factors changed the early partial products (Tversky and Kahneman, 1974). This task did not involve a wheel or a geographical knowledge question. The anchor came from calculation.
Median estimates were 512 for the ascending sequence and 2,250 for the descending sequence, although the correct product was 40,320 (Tversky and Kahneman, 1974). Both medians underestimated the answer. The contrast is useful because the anchor arose from early arithmetic operations rather than from an externally generated random number. The exercise imposed strict time limits.
How anchoring and adjustment affect estimates
A change in the final estimate is observable, but the underlying cognitive process is not read directly from the answer. These accounts differ in their proposed intermediate operations: one focuses on revising an initial number, while another concerns the knowledge made accessible by a comparative question (Epley and Gilovich, 2001; Mussweiler and Strack, 1999). Mussweiler and Strack (1999) investigated hypothesis-consistent testing, while Epley and Gilovich (2001) examined adjustment from self-generated anchors. They tested different candidate mechanisms: Epley and Gilovich examined self-generated values and acceptance movements, while Mussweiler and Strack tested hypothesis-consistent access to information during comparisons.
Insufficient adjustment
In the adjustment account, a judge begins from an initial value and changes it until an answer seems acceptable. Epley and Gilovich (2001) found verbal references to adjustment among participants who had generated the anchors themselves. In further studies, nodding in acceptance of a value was associated with answers closer to a self-generated anchor than shaking the head in rejection. The reported pattern was specific to self-generated anchors in those experiments.
Simmons, LeBoeuf and Nelson (2010) studied what happens when participants are uncertain about the direction in which they should adjust from a supplied anchor. They found that accuracy motivation widened the distance between anchor and estimate when the direction was clear, including for provided and self-generated anchors. Uncertain direction can obscure adjustment (Simmons et al., 2010).
Selective accessibility
Selective accessibility provides another account of what an explicit comparison can do. Mussweiler and Strack (1999) reported results from four studies supporting hypothesis-consistent testing and semantic priming in anchoring tasks. In their model, consideration of a proposed value makes information consistent with that value more accessible for the subsequent estimate. The initial number can therefore influence what comes to mind while the answer is being constructed.
This explanation focuses on retrieved knowledge rather than exclusively on how far a number is adjusted. It is relevant to experiments where the starting value comes from an external comparison. Furnham and Boo (2011) treated competing models as part of the literature that needed reviewing. The models describe distinct experimental processes.
Self-generated and supplied anchors
Experimenter-provided anchors are introduced by the task. Self-generated anchors arise from what a participant already knows or calculates, a distinction highlighted in the work of Epley and Gilovich (2001). The classic wheel experiment used an externally supplied number. The multiplication experiment generated an initial value through partial computation (Tversky and Kahneman, 1974).
Simmons and colleagues (2010) found adjustment under some conditions with both provided and self-generated anchors. Their results make uncertainty about direction a relevant design feature alongside the origin of the number. Research on anchoring therefore tests mechanisms by changing how a reference is introduced, what is compared and what the respondent can infer. Response scores cannot reveal every operation.
Where anchoring bias appears in everyday decisions
Anchoring experiments have tested financial judgements alongside factual estimates. Initial values can be supplied by a listing price, produced from arbitrary identifying numbers or encountered incidentally in a setting. The tasks differed in whether an anchor was a listing price, a number derived from identifying information or simply a figure encountered incidentally before an estimate (Northcraft and Neale, 1987; Ariely et al., 2003; Critcher and Gilovich, 2008). Northcraft and Neale (1987) examined property pricing, while Ariely and colleagues (2003) studied valuations of products and experiences.
Price and valuation judgements
Northcraft and Neale (1987) asked estate agents and students to tour properties and make pricing decisions. The researchers manipulated listing prices and reported results consistent with an anchoring-and-adjustment account in an information-rich setting. Because the participants inspected properties, the study extended the question beyond brief numerical quizzes. The tested decisions involved property estimates.
A separate [account of Alonso's land-price model](/research/alonso-bid-rent-house-prices-research) examines how location enters property values. An experimental listing price and a model of the factors determining land values address different questions. Northcraft and Neale (1987) tested how an initial listed figure affected people's estimates of a property's value. Property prices also reflect location.
Ariely, Loewenstein and Prelec (2003) conducted six experiments on initial valuations of products and experiences, sometimes using arbitrary numbers derived from participants' social security numbers. They reported that valuations could vary with those anchors while subsequent judgements still responded coherently to differences in quality or quantity, a combination in which the apparent ordering of preferences coexisted with numerical starting values introduced arbitrarily by the experiments. The authors called this pattern coherent arbitrariness (Ariely et al., 2003).
Expert estimates and negotiations
Expertise is a relevant question because knowledge provides information independent of the initial number. The estate-agent study deliberately included professionals and students in the property-pricing task (Northcraft and Neale, 1987). Anchoring appeared in that property-pricing setting (Northcraft and Neale, 1987). The experiment covered property valuation.
Critcher and Gilovich (2008) examined numbers that happened to appear in the surroundings of a judgement. Athlete jersey numbers influenced estimates of performance, while product model numbers influenced estimates of domestic sales proportions. Those designs differ from the wheel experiment because the participant was not necessarily asked to compare the incidental number with the target value. The 2026 synthesis also identified some smaller or absent effects for incidental anchoring, in contrast with the positive findings from the earlier athlete and product-number experiments (Schley and Weingarten, 2026).
What later anchoring experiments found
The original wheel demonstration belongs to a family of experiments that differ in the source of the numerical reference, the respondent’s knowledge of its relevance, the comparison procedure and the form of the final estimate (Tversky and Kahneman, 1974; Furnham and Boo, 2011). Jacowitz and Kahneman (1995) developed a way to compare anchored estimates against a calibration group whose independent estimates established a reference distribution before the experimental participants answered the comparison question and supplied their own quantitative assessments. Other studies altered motivation, the apparent relevance of numbers and the task domain. The recent synthesis provides an estimate across a much larger collection while also documenting heterogeneity (Schley and Weingarten, 2026).
Selected experimental findings
Study | Experimental material | Reported finding |
Tversky and Kahneman, 1974 | Wheel numbers 10 and 65 | Median estimates 25 and 45 for the UN percentage question |
Epley and Gilovich, 2001 | Self-generated and supplied anchors | Adjustment evidence concentrated in self-generated conditions |
Ariely et al., 2003 | Products and experiences in six experiments | Arbitrary initial values affected valuations |
Critcher and Gilovich, 2008 | Jersey and product model numbers | Incidental numbers affected selected estimates |
Schley and Weingarten, 2026 | 2,601 effect sizes | Pooled Hedges' g 0.825, 95% confidence interval 0.765–0.884 |
Jacowitz and Kahneman (1995) set experimental anchors at the 15th and 85th percentiles of an independent calibration distribution. The anchored estimates could then be expressed as percentiles within the independent calibration distribution rather than treated as raw figures from questions with incomparable numerical scales (Jacowitz and Kahneman, 1995).
Bornmann and Ganser (2025) applied an anchoring design to academics assessing papers they had previously cited. They detected a small statistically significant effect from random questionnaire access-code numbers, while citation-count anchors produced scarcely any effects in the reported assessments. The random number was an experimental prompt, not a measure of research quality. Their follow-up also examined what happened when participants knew the purpose of the anchor.
Replication and variation in effects
Schley and Weingarten (2026) synthesised 2,601 effect sizes, of which 1,280 compared high with low anchors. They reported Hedges' g of 0.825, with a 95% confidence interval of 0.765–0.884, in their reported synthesis (Schley and Weingarten, 2026). Heterogeneity was I² = 93.73%, so the pooled estimate represents widely varying findings rather than one uniform response. It does not identify what any particular person would answer after seeing a given number.
The same meta-analysis found evidence of reduced or null effects for incidental anchors, anchors from different dimensions and random-number variants, as well as in some conditions involving incentives or debiasing interventions (Schley and Weingarten, 2026). Earlier positive experiments on incidental numbers, such as Critcher and Gilovich (2008), remain observations about their tested situations. The apparent tension is between particular experimental demonstrations and the distribution of results across many designs.
A survey assessment depends on how respondents interpret numerical information. Related [research on self-reported participation](/research/accuracy-of-self-reported-participation-rates-research) addresses measurement differences between reported responses and independently recorded activity. Anchoring experiments instead randomise starting values before an estimate. The Bornmann and Ganser (2025) study illustrates why an apparently irrelevant number and an informative indicator should be evaluated as different experimental treatments.
What is proven about anchoring bias and what is not
The effect of initial numbers on later numerical estimates is supported by the original experiments and by later synthesis. Tversky and Kahneman (1974) reported different medians after different wheel values, while Schley and Weingarten (2026) assembled a large set of high-versus-low anchor comparisons. Those findings concern the specific quantitative judgements recorded in studies. Experimental formats produced different patterns.
Arbitrary starting numbers can shift numerical estimates, but the size of the shift and the process producing it depend on how the anchor enters the task (Tversky and Kahneman, 1974; Schley and Weingarten, 2026).
Tversky and Kahneman (1974) reported multiplication estimates shaped by the order of a partial calculation. Epley and Gilovich (2001) then tested self-generated adjustment, and Mussweiler and Strack (1999) examined selective access to anchor-consistent information. Simmons and colleagues (2010) found that a respondent's uncertainty about the direction of adjustment could change whether accuracy motivation increased movement away from an anchor. These results leave room for interacting cognitive processes rather than a single mechanism that always operates unchanged.
The latest pooled evidence also requires care. Schley and Weingarten (2026) reported considerable heterogeneity and weaker or absent effects in some forms of random and incidental anchoring. Positive findings in real estate decisions, valuations and incidental-number tasks belong to the conditions in which researchers manipulated the references (Northcraft and Neale, 1987; Ariely et al., 2003; Critcher and Gilovich, 2008). A universal change cannot be inferred.
Knowing about an anchor does not produce the same outcome in every study. Bornmann and Ganser (2025) found fewer statistically significant effects after respondents were informed that a random access code was being used as an anchor. The incentives work of Simmons and colleagues (2010) meanwhile found that direction certainty mattered for detecting adjustment. Future tests can separate deliberate comparison, incidental exposure and self-generated calculation while holding the target judgement and outcome measure constant.
Frequently asked questions
What is an example of anchoring bias?
In Tversky and Kahneman's 1974 wheel experiment, participants estimated the percentage of African countries in the United Nations after considering a random starting value. Median estimates were 25 after an anchor of 10 and 45 after an anchor of 65 (Tversky and Kahneman, 1974). The comparison reveals a difference between experimental groups.
Why does a random number influence an estimate?
A prior comparison can make information related to a number easier to retrieve, according to Mussweiler and Strack's 1999 selective-accessibility experiments. Other studies report adjustment from an initial value, sometimes depending on whether it was generated by the respondent. Both mechanisms concern how an estimate is constructed after numerical information enters the task.
Is anchoring bias the same as confirmation bias?
No. Anchoring describes a numerical estimate shifting towards a starting point. The selective-accessibility account proposes hypothesis-consistent testing during a comparison, but its outcome is a subsequently anchored numerical judgement (Mussweiler and Strack, 1999). The terminology refers to different phenomena, even though both can involve preferential attention to compatible information.
Can experts be affected by anchoring bias?
Yes, at least in some experimental tasks. Northcraft and Neale (1987) studied estate agents and students who toured properties and made pricing decisions after listing-price manipulations. Their results were consistent with anchoring in that setting. Expertise does not imply immunity, although results from estate agents cannot be generalised to every professional decision.
Does knowing about anchoring eliminate the effect?
There is no universal experimental finding that awareness eliminates anchoring. Bornmann and Ganser (2025) reported fewer significant effects when respondents knew that an access code served as an anchor. A later meta-analysis described reduced or null effects with some debiasing interventions, while finding variation across the literature (Schley and Weingarten, 2026).
Sources
- Tversky A, Kahneman D. Judgment under Uncertainty: Heuristics and Biases. Science. 1974;185(4157):1124–1131. DOI 10.1126/science.185.4157.1124.
- Furnham A, Boo HC. A literature review of the anchoring effect. The Journal of Socio-Economics. 2011;40(1):35–42. DOI 10.1016/j.socec.2010.10.008.
- Schley DR, Weingarten E. Fifty Years of Anchoring Effects: A Theoretical Reintegration and Meta-Analysis. Management Science. 2026, online 19 May. DOI 10.1287/mnsc.2023.03238.
- Epley N, Gilovich T. Putting adjustment back in the anchoring and adjustment heuristic: differential processing of self-generated and experimenter-provided anchors. Psychological Science. 2001;12(5):391–396. DOI 10.1111/1467-9280.00372.
- Mussweiler T, Strack F. Hypothesis-Consistent Testing and Semantic Priming in the Anchoring Paradigm: A Selective Accessibility Model. Journal of Experimental Social Psychology. 1999;35(2):136–164. DOI 10.1006/jesp.1998.1364.
- Simmons JP, LeBoeuf RA, Nelson LD. The effect of accuracy motivation on anchoring and adjustment: do people adjust from provided anchors? J Pers Soc Psychol. 2010;99(6):917–932. DOI 10.1037/a0021540.
- Northcraft GB, Neale MA. Experts, amateurs, and real estate: An anchoring-and-adjustment perspective on property pricing decisions. Organizational Behavior and Human Decision Processes. 1987;39(1):84–97. DOI 10.1016/0749-5978(87)90046-X.
- Ariely D, Loewenstein G, Prelec D. Coherent Arbitrariness: Stable Demand Curves Without Stable Preferences. The Quarterly Journal of Economics. 2003;118(1):73–106. DOI 10.1162/00335530360535153.
- Jacowitz KE, Kahneman D. Measures of Anchoring in Estimation Tasks. Pers Soc Psychol Bull. 1995;21(11):1161–1166. DOI 10.1177/01461672952111004.
- Critcher CR, Gilovich T. Incidental environmental anchors. Journal of Behavioral Decision Making. 2008;21(3):241–251. DOI 10.1002/bdm.586.
- Bornmann L, Ganser C. Are quality assessments in science affected by anchoring effects? Empirical results from a survey of authors assessing previously cited papers. PLOS ONE. 2025;20(4):e0320148. DOI 10.1371/journal.pone.0320148.
Related
How Accurate Are Self-Reported Recreation Participation Rates?
Why Homes Get Cheaper Further Out: Alonso’s 1964 Model
Self-Rated Health and Education: Findings from 69 Countries
The Tuskegee syphilis study and how it changed research ethics
Learned helplessness and what classic experiments revealed about behaviour
Cognitive dissonance and why we defend beliefs that conflict with facts
The 1989 Molecular Cloning Manual: Why Labs Still Cite It