Here is a link to an ungated copy of my paper "Cognitive Economics" as it appears in the Japanese Economics Review. By special arrangement with the Japanese Economics Review, this paper is in the public domain. The presentation I gave at the Japanese Economic Review conference for this conference volume is here.
This paper is written in the same style as my more academic blog posts. So I count it as a major blog post as well as an NBER Working Paper. It just happens to be a blog post that you need to follow a link to see in full. (And sadly, like the typical blog post, despite diligent efforts, a few typos have crept through. The number and severity of typos I find will have to reach a certain critical threshold before I put the NBER staff to the work of putting together a new version. Please let me know if you find a typo)
Let me give you a bit of a preview, in the form of an outline with one or more key quotations from each section and subsection:
… research in “Cognitive Economics” has already been underway for a long time. But as a participant in this subfield, it seems to me that research in this area has been growing in recent years.
II. Defining Cognitive Economics
Cognitive Economics is defined as the economics of what is in people’s minds. In practical terms, this means that cognitive economics is characterized by its use of a distinctive kind of data. This includes data on expectations, hypothetical choices, cognitive ability, and expressed attitudes.
The name “Cognitive Economics” might initially sound as if it might be yet another synonym for Behavioral Economics … The most obvious difference is that Cognitive Economics is narrower. Behavioral Economics addresses a huge range of issues and cuts across all of the data types listed above, while Cognitive Economics focuses primarily on innovative kinds of survey data … Second, important pieces of Cognitive Economics are inspired by the internal dynamic of economics rather than by psychology.
… there is an obvious complementarity between Cognitive Economics and Behavioral Economics. Although it is possible to consider nonstandard theories of human behavior on the basis of standard data on market decisions alone, freeing up economic theory from traditional assumptions tends to increase the number of free parameters. There is a great value to additional data that can help pin down these additional free parameters.
… let me give my opinion on existing research and future directions in Cognitive Economics, organized around three themes: using data on hypothetical choices and mental contents (1) to identify individual heterogeneity, (2) to revisit welfare economics and (3) to study finite cognition.
III. Identifying Individual Heterogeneity
Heterogeneity across individuals in preferences and cognitive ability is not at all controversial. But data limitations have often forced economists to assume uniformity. Here the kind of data discussed above can do a lot to allow economists to capture some of the heterogeneity that exists.
IV. Revisiting Welfare Economics
The use of self-reported happiness to study welfare issues illustrates a key methodological issue in Cognitive Economics. Whenever a new measure is used, its relationship to standard concepts of economic theory is at issue.
It is possible, however, that happiness data could have a tight relationship to preferences even if the level of happiness does not. In particular, to explain the data, Kimball and Willis (2006) suggest that a large component of self-reported happiness depends on recent innovations in lifetime utility. Whenever people receive good news about lifetime utility, self-reported happiness temporarily spikes up; whenever people receive bad news about lifetime utility, self-reported happiness temporarily dips down. If true, this means that while it is questionable to use the level of happiness to infer preferences, the dynamics of happiness are informative about preferences and so can be used to inform welfare economics.
V. Studying Finite Cognition
Moreover, to avoid the judgment Herbert Simon’s phrase “bounded rationality” can inadvertently suggest, I will refer instead to “finite cognition.” Finite cognition means something more than just imperfect information—it means finite intelligence, imperfect information processing, and decision-making that is costly.
 Often, the inadvertent judgment suggested by “bounded rationality” is quite inappropriate. For example, if decision-making is actually costly, which is more “rational,” to choose in a way that takes into account the costliness of decision-making or to pretend that decision-making has zero cost? If one’s intelligence is actually finite, which is more rational, taking into account the limits on one’s intelligence, or pretending that one’s thinking power is unlimited? There is certainly a sense in which knowing and adjusting to one’s own limitations can often be the height of “rationality.”
finite cognition implies that even in the absence of externalities, welfare can often be improved by economic education, setting up appropriate default choices for people, or providing disinterested, credible advice. By contrast, explanations of puzzling behavior on the basis of individuals maximizing exotic preferences imply (if true) that welfare improvements must come in the standard way from addressing externalities, or in the case of inconsistent preferences, by taking sides in an internal conflict. Once puzzling behavior that is difficult to explain on the basis of standard economic theory is identified, it is hard to think of a more important question than whether people behave that way because they want to, or simply because they are confused.
A. The Reality of Finite and Scarce Cognition.
Although the inadequacies of our current tools can make it hard to study finite cognition theoretically, the claim that human intelligence is finite–and that finite intelligence matters for economic life—scarce cognition—is not really controversial.
 There are many problems that are too hard for even very high levels of intelligence. For example, one of the problems with Bayesian updating is that, strictly speaking, it involves putting a positive probability on a much greater than astronomically huge set of possibilities. Various strategies of economizing on information processing are always essential in practice. Even the existence of a utility function itself is, in a sense, a technique of economizing on information transfer and processing. If evolution could process an infinite amount of information, and the genetic code could transmit an infinite amount of information, we could be endowed with decision rules embracing essentially all contingencies instead of mere objective functions and calculation capabilities.
B. Difficulties in Studying Finite Cognition with Standard Theoretical Tools.
One key reason it is not easy using our standard theoretical tools to model finite cognition is the “infinite regress” problem emphasized by John Conlisk (1996). The infinite regress problem afflicts models that assume a cost of computation or other decision-making cost. The problem is that figuring out how much time to spend in making a decision is almost always a strictly harder decision than the original decision.
Costs to decision-making are a natural enough assumption for economists that a substantial percentage of all applied economic theory papers might include them, if it were not for the infinite regress problem. Finessing the infinite regress problem somehow is essential if economists are to develop effective theoretical tools for studying finite cognition. There are several feasible strategies for getting around the infinite regress problem—every one of which requires breaking at least one inhibition shared by many economists.
Least transgressive are models in which an agent sits down once in a long while to think very carefully about how carefully to think about decisions of a frequently encountered type.
A second strategy is to give up on modeling finite cognition directly and use models of limited information transmission capacity as a way of getting agents to make more imperfect decisions. In other words, one can accept the fact that our standard tools require constrained optimization with its implication of infinite intelligence somewhere in the model, but handicap agents in the model by giving them a “thick skull” that is very inefficient at transmitting information to the infinitely intelligent decision-maker within (that is, the perfect constrained optimizer within).
A third feasible strategy is in the spirit of what the complexity theorists call “agent-based modeling. … This type of modeling substitutes the problem of agents that have unrealistically subhuman intelligence for the problem we have been focusing on of agents that have unrealistically superhuman intelligence. Despite this lack of realism, the results can be very instructive because the failure of realism is in the opposite direction from what economists are used to.
I would like to focus on a fourth strategy for getting around the infinite regress problem–one that seems to me less commonly used: modeling economic actors as doing constrained optimization in relation to a simpler economic model than the model treated as true in the analysis. This simpler economic modeled treated as true by the agent can be called a “folk theory.“ … A folk theory should not be confused with the Folk Theorem of repeated game theory. I am talking about folk economics in the same sense as the well established ideas of “folk psychology,” “folk physics” and “folk biology.”
C. Modeling Unawareness Requires a Subjective State Space for the Economic Actor Distinct from the True State Space.
Dekel, Lipman and Rustichini (1998) argue for relaxing what they call the “real states” assumption as follows:
In standard state-space models, states play two distinct roles: they are the analyst’s descriptions of ways the world might be and they are also the agent’s descriptions of ways the world might be. If the agent is unaware of some possibility, though, ‘his’ states should be less complete than the analyst’s. In particular, the propositions the agent is unaware of should not ‘appear in’ the states he perceives.
Departures from the real states assumption would allow agents to have a different model of the economic situation in their minds than the maintained assumptions the analyst is using to model the situation of those very agents.
Note that if someone is successfully taught a more sophisticated model, this would involve an expansion in the individual’s subjective state space. If positive probabilities were accorded to the newly added states, this must necessarily involve a departure from Bayesian updating. Presumably it is also possible for people to “see the light” even without being explicitly taught. For example, the agent might be driven to entertain an expanded model if the probability of observed events conditional on the initial folk model ever appeared sufficiently low. We all recognize the practical importance of expansions in one’s subjective state space when in scientific contexts we say “Asking the right question is half the battle.”
D. Using Folk Theories to Model Finite Cognition: A Portfolio Choice Example.
Clearly, the desirable properties for a modeled folk theory are quite different from the desirable properties for a theory proposed as a good approximation of reality. A folk theory need not be logically consistent at a deep level. Indeed, in representing reality, it may be a positive virtue for a folk theory to have logical inconsistencies of a form similar to the logical inconsistencies real people might have in their views of the world. Other than (a) descriptive accuracy as a reasonable representation of how people actually view the world, for theoretical purposes the key desirable properties for a modeled folk theory are (b) providing a clear prediction for how the people holding that folk theory will behave in various circumstances and © representing clearly what the people holding the folk theory are confused about and what they do understand. In terms discussed in Richard Herrnstein (1997)–particular in the chapters with Drazen Prelec–a folk theory should at least implicitly model the accounting framework that an agent uses, in addition to the objective function. Because it need not be logically consistent at a deep level, the argument for a folk theory can involve (correct reasoning about) logical leaps and plausible, though fallacious reasoning.
In reality, I am confident that people’s thinking about portfolio choice varies from person to person with a wild profusion of different kinds of misunderstanding. In most other contexts as well—at least where there is some complexity–any model that assumes everyone’s folk theory is of the same type is likely to be false. Realizing that people don’t always have the same mental model of a situation as the economist studying that situation is the first step toward facing the motley truth about people’s folk theories.
Economic research using more and more direct data about what is in people’s minds is flourishing. But much more can be done. Fostering continued progress in this area of Cognitive Economics calls for three inputs. First, new theoretical tools for dealing with finite cognition need to be developed, and existing theoretical tools sharpened. Second, welfare economics needs to be toughened up for the rugged landscape revealed by peering into people’s minds. Third, the statement “The data are endogenous” needs to become not only an econometrician’s warning but also a motto reminding economists that new surveys can be designed and new data of many kinds can be collected to answer pressing questions.