Dilution Effect

Nisbett et al. (1981) demonstrated that coupling diagnostic with non-diagnostic evidence produced more regressive predictions than people would otherwise have made. They also established the replicability of this dilution effect and explained it by invoking the representativeness heuristic discussed in Tversky and Kahneman (1974). In this account, people judge possible futures for a target person by comparing key features of the individual with key features of the possible outcomes and predicting the outcome most similar to, or representative of, the individual. In Tversky’s (1977) framework, similarity is a positive function of the number of common features and a negative function of the number of unique features. Common features are attributes of the target person that people associate with the outcome. For instance, the image of a student who studies only 3 hours a week is strongly associated with the image of the prototypical student with a low-grade point average. Unique features are attributes of the target individual that people rarely associate with the outcome. Thus, one reduces the similarity – and, therefore, the perceived predictive link – between the indolent student and poor grades by incorporating irrelevant details into the description of the student (e.g., tennis player, dating habits, number of plants in his apartment).

Assessed against the logical standards of multiple regression, it is unreasonable to reduce one’s confidence in the potency of a valid predictor merely because that predictor is embedded in an array of irrelevant information. It is useful, however, to look at the dilution effect from a social–functionalist vantage point. Far from representing an error, dilution may constitute a rational response to the interpersonal and institutional context. The presentation of information in dilution experiments can be likened to a conversation in which subjects assume that the experimenter, their conversational partner, is following widely accepted norms of social discourse such as truthfulness and relevance (Grice, 1975). In most conversations, people refrain from making statements that are utterly irrelevant or grossly deceptive. Given that the experimenters deemed it appropriate to include an assortment of evidence in their communications, one would expect a good Bayesian to attach a high prior probability to the relevance of the evidence. The intuitive statistician and politician are thus reconciled.

This normative reinterpretation is not entirely speculative. Several studies show that manipulations of conversational norms moderate a number of experimental effects, including the underuse of base-rate information, the overuse of recently presented information, and the overestimation of the likelihood of conjunctions of events. Schwarz (1994) stated this position forcefully: Many purported demonstrations of cognitive shortcomings occur because basic assumptions that underlie all social discourse in everyday settings are “routinely violated” in studies on judgmental biases.

Amplification of dilution by accountability could be consistent with either a judgmental-heuristics or conversational-norm interpretation. To disentangle the two explanations, Tetlock et al. (1996) manipulated whether subjects believed that conversational norms held in the experimental setting. Following a procedure developed by Schwarz (1994), some subjects were told that the axioms of conversation applied: the experimenter had screened the information to ensure its relevance for the prediction task. Other subjects were told that the research team could not vouch for the usefulness of the information as it had been randomly selected from a computer database.

If dilution is a rational response to conversational norms, the effect should disappear when conversational norms have been deactivated explicitly. Moreover, accountability should cease to magnify the dilution effect when conversational norms have been deactivated. By contrast, if dilution is the product of a judgmental heuristic that involves automatic similarity-matching of causes and effects, there is less reason to expect explicit deactivation of conversational norms to moderate either the magnitude of the dilution effect or the tendency for accountability to amplify it. Consistent with a conversational-norms interpretation, dilution disappeared among accountable subjects who thought that conversational norms did not apply. Accountability also ceased to motivate more complex patterns of thinking when conversational norms had been deactivated. By contrast, dilution persisted among subjects who had been told that conversational norms do apply, as well as among subjects in a no-norm-primed control condition, who were told nothing about the relevance of the information. Contrary to the conversational-norm hypothesis, however, and consistent with a judgmental-heuristics interpretation, deactivating conversational norms was not sufficient to eliminate dilution among unaccountable subjects.

The judgmental-heuristic and social-functionalist explanations need not, of course, be mutually exclusive. Indeed, whenever accountability amplifies a given error or bias, the two classes of processes may well be mutually reinforcing. Plausible-candidate effects include not only dilution, but also the compromise, attraction, and reason-augmentation effects (Shafir et al., 1993). Consider the reason-augmentation effect, in which “enriched options,” which possess both more positive and more negative features, are more likely to be selected when decision makers are asked to choose their preferred option, but less likely to be selected when decision makers are asked to reject their less-preferred option. This phenomenon is consistent with a view of decision makers as intuitive politicians who, prior to deciding, weigh the relative justifiability of response options, but do so in a superficial manner such that when they are asked to choose an option, they count the number of good reasons for doing so, and when they are asked to reject an option, they count the number of reasons for doing that (thereby short-circuiting the need for complex compensatory comparisons).