Norm Theory

Daniel Kahneman and Dale T. Miller

Reasoning flows not only forward, from anticipation and hypothesis to confirmation or revision, but also backward, from the experience to what it reminds us of or makes us think about. Backward thinking offers an alternative account of some observations usually explained in terms of anticipation and expectation.

The account distinguishes category norms that represent knowledge of concepts and with stimulus norms that govern comparative judgments and designate experiences as surprising. In the tradition of adaptation level theory (Appley, 1971; Helson, 1964), the concept of norm is applied to events that range in complexity from single visual displays to social interactions.

The central idea of the present treatment is that norms are computed after the event rather than in advance. We sketch a supplement to the generally accepted idea that events in the stream of experience are interpreted and evaluated by consulting precomputed schemas and frames of reference. The view developed here is that each stimulus selectively recruits its own alternatives (Garner, 1962, 1970) and is interpreted in a rich context of remembered and constructed representations of what it could have been, might have been, or should have been. Thus, each event brings its own frame of reference into being. We also explore the idea that knowledge of categories (e.g., “encounters with Jim”) can be derived online by selectively evoking stored representations of discrete episodes and exemplars.

The present model assumes that a number of representations can be recruited in parallel, by either a stimulus event or an abstract probe such as a category name, and that a norm is produced by aggregating the set of recruited representations. The assumptions of distributed activation and rapid aggregation are not unique to this treatment. The present analysis relates most closely to exemplar models of concepts (Brooks, 1978, 1987; Hintzman, 1986; Hintzman & Ludlam, 1980; Jacoby & Brooks, 1984; Medin & Schaffer, 1978; Smith & Medin, 1981). We were drawn to exemplar models in large part because they provide the only satisfactory account of the norms evoked by questions about arbitrary categories, such as, “Is this person friendlier than most other people on your block?”

Exemplar models assume that several representations are evoked at once and that activation varies in degree. They do not require the representations of exemplars to be accessible to conscious and explicit retrieval, and they allow representations to be fragmentary. The present model of norms adopts all of these assumptions. In addition, we propose that events are sometimes compared to counterfactual alternatives that are constructed ad hoc rather than retrieved from past experience. These ideas extend previous work on the availability and simulation heuristics (Kahneman & Tversky, 1982; Tversky & Kahneman, 1973). . . .

An abnormal event is one that has highly available alternatives, whether retrieved or constructed; a normal event mainly evokes representations that resemble it. This treatment of normality is guided by the phenomenology of surprise rather than by formal or informal conceptions of probability. The main difference between the two notions is that probability is always construed as an aspect of anticipation, whereas surprise is the outcome of what we call backward processing – evaluation after the fact. Probability reflects expectations. Surprise (or its absence) reflects the failure or success of an attempt to make sense of an experience, rather than an evaluation of the validity of prior beliefs. In his critique of standard notions of probability, the statistician Shafer (1976) developed the related idea that events do not merely alter the strength of belief in existing possibilities, they also evoke and shape the set of relevant possibilities.