Anchoring: How a Number You Just Saw Pulls Your Estimate

The first question in the demo has an obvious answer. Nobody thinks Gandhi died at 9 or lived to 140, and the comparison looks like a formality. The experiments show that it is not. People who have just considered whether he was older than 9 guess a lower age than people who have just considered whether he was younger than 140, by a wide margin, although both groups rejected the anchor in an instant.

Tversky and Kahneman's 1974 account was adjustment: people start from the anchor and move toward a plausible value, and they stop too soon, so the final estimate stays on the anchor's side. That describes the result well when people are told to adjust from a given number, and on its own it has trouble with the Gandhi case, where 9 is plainly impossible rather than a candidate answer. Strack and Mussweiler proposed in 1997 that anchoring is a case of semantic priming. To answer whether Gandhi was older than 9, you retrieve what you know about him; the comparison question makes you test the hypothesis that the anchor is right, and the search for evidence turns up knowledge consistent with it, such as his youth or his long life. That knowledge is then more accessible when the absolute question arrives, and it tilts the estimate. Their experiments found that an anchor affected a later judgment only when the knowledge it activated was applicable to it, and that implausible anchors, which can be rejected on category knowledge alone, still moved estimates, a result their 2001 follow-up explained by combining the two accounts: adjustment to the edge of the plausible range, then selective accessibility.

Anchoring is among the most robust findings in judgment research. Jacowitz and Kahneman's 1995 anchoring index, over fifteen uncertain quantities, found that estimates kept on average about half the distance between the low and the high anchor. Englich, Mussweiler and Strack showed in 2006 that experienced judges handed a sentencing demand determined by dice gave longer sentences after a high throw than a low one. The 2014 Many Labs project ran four anchoring questions across 36 samples and found the effect consistently. Furnham and Boo's 2011 review found it in experts and novices, with and without incentives for accuracy, and under warnings about it.

What reduces it is scarce. Telling people about anchoring does little; considering reasons why the anchor might be wrong has reduced the effect in several studies, and strong knowledge of the answer helps. Otherwise, the defense is to form an estimate before any number is on the table.

Amos Tversky and Daniel Kahneman described anchoring in "Judgment under uncertainty: Heuristics and biases," published in Science in 1974. In the demonstration that gave the effect its name, participants watched a wheel of fortune spun to a number between 0 and 100, said whether the percentage of African countries in the United Nations was higher or lower than it, then estimated the percentage. The wheel was rigged to stop at 10 for some groups and at 65 for others. The median estimate was 25 percent in the groups that had seen 10 and 45 percent in the groups that had seen 65. The authors explained it as insufficient adjustment from the starting point.

Fritz Strack and Thomas Mussweiler published "Explaining the enigmatic anchoring effect: Mechanisms of selective accessibility" in the Journal of Personality and Social Psychology in 1997. The Gandhi question comes from the third, in which 67 students answered a series of paired questions, each a comparison with an anchor followed by an absolute estimate, with the anchors chosen to be either plausible or absurd. For Gandhi's age at death the plausible anchors were 64 and 79 and the implausible ones 9 and 140, against an actual age of 78. The paper's tables report standardized values, and its appendix gives the means for the Gandhi item: 50 after the comparison with 9, 67 after the comparison with 140. The implausible anchors moved estimates at least as much as the plausible ones, which a simple adjustment account does not explain well, since an anchor of 9 is rejected at once. It does not follow that adjustment played no part: the authors' 2001 follow-up, "Considering the impossible," proposed that people adjust from an impossible value to the boundary of the plausible range and that selective accessibility works from there.

The paper's contribution was the explanation. In the first study, an anchor given for one dimension of an object, such as the height of a gate, moved later estimates of that dimension but not of a different one, such as its width. In the second, an anchor for one object produced a contrast effect, an estimate pushed away from it, when the object judged afterward was very different. The authors concluded that anchoring works by making anchor-consistent knowledge selectively accessible, and that account has shaped much of the research since.

The rows below are illustrations, not findings from the studies cited on this page.

Knowing about anchoring does not remove it, so the counters are procedural.

• Estimate first. Before you see a list price, a first offer or a suggested amount, write down your own number from whatever you know. A number committed to in advance is a competing anchor of your own choosing.
• Consider the opposite. List reasons the anchor might be much too high or much too low. Mussweiler, Strack and Pfeiffer found in 2000 that generating anchor-inconsistent reasons reduced the effect, which fits the selective accessibility account.
• Bring your own anchor. In a negotiation, making the first offer, grounded in evidence, sets the reference point rather than reacting to one.
• Treat an absurd anchor as a warning, not as harmless. Anchors nobody believes still move estimates, so rejecting a number does not mean it has gone.

Anchoring has several neighbors that are better named separately.

• It is not the framing effect, in which the same outcome described as a gain or as a loss changes which option is preferred. Framing changes a choice between options; anchoring changes the size of a numerical judgment.
• It is not the primacy effect, the extra weight of whatever information arrives first in forming an impression or recalling a list. Primacy concerns order and memory; anchoring concerns a specific number pulling an estimate toward it, whatever its position.
• It is not conservatism bias, the tendency to revise a belief too little when new evidence arrives. Conservatism is anchoring on one's own prior belief; the anchoring effect occurs with numbers that were never believed at all.
• It is not the status quo bias, the preference for things to stay as they are. A current price can act as an anchor, but status quo bias is about choices, not estimates, and runs on loss aversion rather than accessible knowledge.