Anchoring adjustment and contamination

An experimenter spins a 'Wheel of Fortune' device as you watch, and the Wheel happens to come up pointing to (version one) the number 65 or (version two) the number 15. The experimenter then asks you whether the percentage of African countries in the United Nations is above or below this number. After you answer, the experimenter asks you your estimate of the percentage of African countries in the UN.

Tversky and Kahneman (1974) demonstrated that subjects who were first asked if the number was above or below 15, later generated substantially lower percentage estimates than subjects first asked if the percentage was above or below 65. The groups' median estimates of the percentage of African countries in the UN were 25 and 45 respectively. This, even though the subjects had watched the number being generated by an apparently random device, the Wheel of Fortune, and hence believed that the number bore no relation to the actual percentage of African countries in the UN. Payoffs for accuracy did not change the magnitude of the effect. Tversky and Kahneman hypothesized that this effect was due to anchoring and adjustment; subjects took the initial uninformative number as their starting point, or anchor, and then adjusted the number up or down until they reached an answer that sounded plausible to them; then they stopped adjusting. The result was under-adjustment from the anchor.

In the example that opens this chapter, we first asked the expert on Substance P to guess the actual value for the strength of radio signal that would detonate Substance P, and only afterward asked for confidence bounds around this value. This elicitation method leads people to adjust upward and downward from their starting estimate, until they reach values that "sound implausible" and stop adjusting. This leads to under-adjustment and too-narrow confidence bounds.

After Tversky and Kahneman's 1974 paper, research began to accumulate showing a wider and wider range of anchoring and pseudo-anchoring effects. Anchoring occurred even when the anchors represented utterly implausible answers to the question; e.g., asking subjects to estimate the year Einstein first visited the United States, after considering anchors of 1215 or 1992. These implausible anchors produced anchoring effects just as large as more plausible anchors such as 1905 or 1939. (Strack and Mussweiler 1997.) Walking down the supermarket aisle, you encounter a stack of cans of canned tomato soup, and a sign saying "Limit 12 per customer." Does this sign actually cause people to buy more cans of tomato soup? According to empirical experiment, it does. (Wansink et. al. 1998.)

Such generalized phenomena became known as contamination effects, since it turned out that almost any information could work its way into a cognitive judgment. (Chapman and Johnson 2002.) Attempted manipulations to eliminate contamination include paying subjects for correct answers (Tversky and Kahneman 1974), instructing subjects to avoid anchoring on the initial quantity (Quattrone et. al. 1981), and facing real-world problems (Wansink et. al. 1998). These manipulations did not decrease, or only slightly decreased, the magnitude of anchoring and contamination effects. Furthermore, subjects asked whether they had been influenced by the contaminating factor typically did not believe they had been influenced, when experiment showed they had been. (Wilson et. al. 1996.)

A manipulation which consistently increases contamination effects is placing the subjects in cognitively 'busy' conditions such as rehearsing a word-string while working (Gilbert et. al. 1988) or asking the subjects for quick answers (Gilbert and Osborne 1989). Gilbert et. al. (1988) attribute this effect to the extra task interfering with the ability to adjust away from the anchor; that is, less adjustment was performed in the cognitively busy condition. This decreases adjustment, hence increases the under-adjustment effect known as anchoring.

To sum up: Information that is visibly irrelevant still anchors judgments and contaminates guesses. When people start from information known to be irrelevant and adjust until they reach a plausible-sounding answer, they under-adjust. People under-adjust more severely in cognitively busy situations and other manipulations that make the problem harder. People deny they are anchored or contaminated, even when experiment shows they are. These effects are not diminished or only slightly diminished by financial incentives, explicit instruction to avoid contamination, and real-world situations.

Now consider how many media stories on Artificial Intelligence cite the Terminator movies as if they were documentaries, and how many media stories on brain-computer interfaces mention Star Trek's Borg.

If briefly presenting an anchor has a substantial effect on subjects' judgments, how much greater an effect should we expect from reading an entire book, or watching a live-action television show? In the ancestral environment, there were no moving pictures; whatever you saw with your own eyes was true. People do seem to realize, so far as conscious thoughts are concerned, that fiction is fiction. Media reports that mention Terminator do not usually treat Cameron's screenplay as a prophecy or a fixed truth. Instead the reporter seems to regard Cameron's vision as something that, having happened before, might well happen again - the movie is recalled (is available) as if it were an illustrative historical case. I call this mix of anchoring and availability the logical fallacy of generalization from fictional evidence. (A related concept is the good-story bias hypothesized in Bostrom (2001). Fictional evidence usually consists of 'good stories' in Bostrom's sense. Note that not all good stories are presented as fiction.)

Storytellers obey strict rules of narrative unrelated to reality. Dramatic logic is not logic. Aspiring writers are warned that truth is no excuse: you may not justify an unbelievable event in your fiction by citing an instance of real life. A good story is painted with bright details, illuminated by glowing metaphors; a storyteller must be concrete, as hard and precise as stone. But in forecasting, every added detail is an extra burden! Truth is hard work, and not the kind of hard work done by storytellers. We should avoid, not only being duped by fiction - failing to expend the mental effort necessary to 'unbelieve' it - but also being contaminated by fiction, letting it anchor our judgments. And we should be aware that we are not always aware of this contamination. Not uncommonly in a discussion of existential risk, the categories, choices, consequences, and strategies derive from movies, books and television shows. There are subtler defeats, but this is outright surrender.

Continue reading here: The affect heuristic

Was this article helpful?

0 0