The Anchoring Effect: How the First Number You Hear Pulls Your Estimates.

Anchoring bias is the tendency of a first number to pull later estimates toward it. You may want to know how strong the pull is, where it matters and what weakens it. The short answer: the pull is real and often large. Tested methods can reduce it, but do not guarantee immunity. Studies have found it again and again over five decades.^[18][31] A review of the field ranks it among the most dependable cognitive biases.^[18]

The founding experiment used a wheel of fortune.^[1] People watched it spin and stop on a number. The wheel looked random but was set to stop on either a low number or a high one. People then judged whether the share of African countries in the United Nations was higher or lower than that number, and gave their best guess. Those who saw the high number guessed higher, by 20 percentage points. The participants had no reason to think a carnival prop held useful information, yet the number got in anyway.

How strong is the pull across the whole field? A pooled analysis of five decades of studies found a large average effect, and it stayed large after a correction for results that went unpublished.^[31] But the studies differ so much that no single "typical" size exists.^[31] An open dataset of anchoring studies run in Europe, Asia, North America and South America shows that the effect is not a quirk of Western labs.^[30]

The mind does not judge a number from a blank slate. It judges it against whatever number came first, then adjusts too little, if it adjusts at all.^[3][9]

Anchoring also resists the things that should stop it. Being clever does not fully protect you: in a consumer-goods experiment, higher cognitive ability reduced anchoring but did not remove it.^[17] A warning helps only when you came up with the anchor yourself. When someone else supplies the anchor, as in a negotiation, on a price tag or in a courtroom, knowing about the bias does little to reduce it.^[14]

Anchoring is built into how the mind makes numerical estimates when it is unsure. Knowing a topic well helps: across several studies, people who knew more were anchored less.^[23] But knowledge reduces the effect; it does not remove it. The useful question is not whether anchoring affects you, but how much, and what reduces it.

One account ties anchoring to a wider habit of mind called WYSIATI, short for "what you see is all there is". Fast, automatic thinking builds a story from whatever comes to mind, then treats that story as the whole story.^[12] A high first number brings to mind facts that fit a high answer. A low one brings to mind facts that fit a low answer.

Anchoring matters most where a number, or a first label, decides something real. One records study covered more than 100,000 emergency visits by US veterans who had congestive heart failure and arrived short of breath. When the triage note mentioned the heart failure, physicians were less likely to test for a blood clot in the lung. No clear difference was found in clot diagnoses within the following month, so the drop in testing is a pattern consistent with anchoring bias.^[29] In a simulated sentencing task, junior legal professionals with some judging experience gave different sentences after a demand set by a roll of dice.^[24] In a negotiation exercise, the first offer closely predicted the final price.^[7]

Medicine shows the stakes. In a study of 100 selected internal-medicine diagnostic errors, cognitive factors contributed to about three in four cases.^[22] Even a person's own ID number can act as an anchor. In one experiment, people first said whether they would buy each product for a dollar price equal to the last two digits of their US Social Security number. Those above the sample's median digits then offered up to about twice as much as those below that median.^[11] The paper calls this coherent arbitrariness: relative preferences can be consistent even when the starting valuation is arbitrary.^[11]

That matters because anchoring is not an error of ignorance. It appears in experts as well as novices. Whether rewards and effort reduce it depends on the task and on whether people know which way the right answer lies.^[16]

For almost thirty years, researchers argued over how anchoring works. Do people adjust too little from the anchor, or does the anchor change what they recall? The current answer is both. The founding paper proposed a simple route: people start at the anchor, adjust away from it, and stop too soon.^[1] Later work showed why the adjustment falls short: people stop at the first answer that seems plausible, not at the right one.^[9]

A second route was proposed later. Suppose you are asked whether a car is worth more or less than $50,000. Your mind does not just start at that figure and adjust. It tests whether the car could be worth that much, and so it calls up facts that fit the idea.^[2] Features that support the anchor come to mind easily. Features that contradict it stay out of view. This route is called selective accessibility.

Experiments by the team that proposed this route found that facts fitting the anchor stayed easy to recall even after the comparison question was over.^[2] Another team found that sufficiently attending to an uninformative number could shift later estimates without an explicit comparison question.^[5] In that work, people who knew more were less susceptible.^[5] A third team found that an anchor makes the features it shares with the thing being judged easier to recall. This changes what is in working memory, the mind's short-term workspace. Considering different features reduced or removed the measured effect in some tasks.^[4]

The two routes, anchoring-and-adjustment and selective accessibility, need not compete. One set of experiments suggested they work under different conditions. There, adjustment drove anchoring when the anchor was self-generated, a number you supply yourself.^[3] Asked when vodka freezes, for example, you might start from the freezing point of water, which you know, and adjust downward.^[3] Externally provided anchors are numbers that come from someone else, such as a price tag, a prosecutor's request or a first offer. For these, the same experiments suggested little adjustment, with the biased memory search doing the work instead.^[3] Later experiments found adjustment from supplied anchors too, especially when people knew the direction of the answer.^[16]

The difference even shows up in the body. Nodding, a motor acceptance cue, reduced adjustment from self-generated anchors, and shaking the head increased it, which suggests that adjustment partly responds to physical signs of yes and no.^[3] Head movements had no such effect on anchors supplied by others. A related line of work studies anchors in both belief judgments (estimates of facts) and value judgments (preferences).^[10] Anchors affect both.^[4]

A broader model of comparison explains the direction of the pull. A comparison can set off similarity testing, a search for ways the thing being judged is like the anchor. Then the estimate moves toward the anchor.^[15] It can instead set off dissimilarity testing, a search for differences. Then the estimate moves away from it.^[15] People test for similarity by default, which is why anchors usually pull estimates toward them.

Brain studies add a third line of evidence. In one brain-scanning (fMRI) study, people judged another person's preferences by starting from their own. Activity in the MPFC, the medial prefrontal cortex at the front and middle of the brain, rose in step with how far they adjusted away from themselves.^[8]

Another study recorded brain waves from the scalp (EEG) while asking people how much compensation they would need to listen to unpleasant noise. After comparison with higher rather than lower amounts, theta-band power, one type of brain-wave activity, rose more while they listened. The authors interpreted this as semantic priming in this compensation task, meaning that the anchor made related ideas easier to bring to mind.^[19]

The strongest brain evidence comes from a brain-stimulation experiment. Researchers passed a weak current through the scalp, a method called transcranial direct current stimulation (tDCS), over the right dorsolateral prefrontal cortex (DLPFC), an area linked to mental control and holding back responses. People were randomly assigned to real or sham stimulation, then said what they would pay for goods. In this experiment, the type of stimulation thought to raise activity in the area (anodal) reduced anchoring, and the type thought to lower it (cathodal) increased it.^[20] One proposed account is that the area works like a gate that filters out an irrelevant anchor.

Unlike the scanning studies, this one acted on the brain directly and watched the bias shift in the predicted direction, which supports cause rather than mere link. This account draws on one experiment on what people would pay. Whether the right DLPFC acts as a control point for anchoring awaits confirmation.

That average is large. But the spread matters as much as the average. More than nine-tenths of the variation between studies reflects real differences in the anchoring effect, not chance.^[31] So no "typical" anchoring effect exists. In comparative anchoring, people first judge whether the answer is higher or lower than a number. In incidental anchoring, people simply see a number without comparing. Anchoring was weaker or absent on average in incidental-number tasks.^[31] An explicit comparison often produces stronger effects, but some studies found effects without one when people attended to the number.^[5]

A computer model offers one reason for the spread. This resource-rational model asks how a mind with limited time and mental effort should best use them.^[25] It treats anchoring as a near-best speed-accuracy tradeoff: the mind anchors more when time is short and less when mistakes are costly.^[25] Experiments by the same group found that people adjusted less when time was costly and more when errors were costly, whether they made up the anchor or were given it.^[26] On this view, anchoring is a sensible shortcut that goes wrong when the anchor is irrelevant, as it often is in real decisions. A separate anchor integration model describes and measures anchoring effects.^[21]

A recent large replication shows why single studies need care. A different team re-ran one auction-bidding anchoring experiment with far more participants. The original rise in bids of nearly a third shrank to a few per cent, small enough that it may be zero.^[32] Many effect sizes in the older literature are inflated by small samples. This re-run tested one experiment. The pooled analysis averages thousands of results, and its average stayed large after its correction for unpublished results.^[31] How strong anchoring is in a given case has to be measured, not assumed.