Pragmatic Beliefs and Overcondence

Hans K. Hvide Norwegian School of Economics and Business (LOS) January 11, 2000

Abstract Several studies indicate that humans are overcondent about their own (relative) abilities. We propose a notion of pragmatic beliefs, and show through an example that this concept can shed light on why overcondence emerges. Through the example, we also shed light on the idea that that bounded rationality may arise endogenously in a game - without assuming complexity costs.

Introduction

Several studies from psychology and experimental economics indicate that humans are overcondent about their own (relative) abilities: For a variety of skills, people seem to believe that they are more skillful than they really are.1 Although rational theories of

I thank Ken Binmore, Eddie Dekel, Gaute Torsvik, Bertil Tungodden, and seminar participants for

stimulating discussions. Hans K. Hvide, LOS Center, University of Bergen, Rosenbergsgaten 39, N-5015 Bergen, NORWAY. 1 The classic example is car-driving, as in Svensson (1981). Most of the older literature is based on interview data, while the more recent ndings of overcondence are from market and experimental settings. Camerer & Lovallo (1999), reports on overcondence in entry decisions in games of skill experiments (with monetary incentives). Golec & Tamarkin (1995) uses racetrack data to suggest that what has previously been viewed as risk-loving behavior (bets on longshots), more plausibly can be seen as indications of overcondence. Finally, Asubel (1991) indicates that the credit card market is characterized by users of credit cards being overcondent about their future ability to avoid overdrawn accounts.

belief formation, like the Bayesian paradigm, can easily be made consistent with some individuals in a population being overcondent, it is hard to make such theories consistent with most, or a majority of, a population being overcondent, and across dierent settings. Hence rational theories of belief formation seem unt to explain overcondence.2 The purpose of the present paper is to propose a notion of pragmatic beliefs, and to suggest that this concept can shed light on why overcondence can occur. In short, we attempt to show with a simple game-theoretic example that if agents form beliefs pragmatically, overcondence can be the equilibrium outcome. An agent that forms beliefs pragmatically forms beliefs that are (the most) useful to him. The pragmatic explanation of beliefs should be contrasted with rational approaches to beliefs, which focus on agents that try to reect the world as it is in their beliefs (with tools like Bayes law, the law of double expectation and so forth). As will be clear, pragmatic beliefs need only have a weak relation to the world as it is. We also contrast the pragmatic explanation of beliefs to approaches assuming that beliefs are of intrinsic value to agents, as with wishful thinking. To show how the idea of pragmatic beliefs can shed light on overcondence, we consider a simple model of a job market. An agent is matched with a rm, and the rm makes a take-it or leave it wage oer. If the agent accepts the oer, the game ends, while if the agent rejects the oer, she gets an outside opportunity. The wage oer made by the rm will in equilibrium depend partly on the workers productivity in the rm, and partly on the agents beliefs about her outside opportunity, which is commonly known between the rm and the worker. The key feature of this example is that instead of taking the agents beliefs as a primitive, we let the agents beliefs about her outside opportunity be an endogenous variable. We show that in equilibrium, these beliefs will be inated compared to the true distribution of the outside opportunity, and interpret this as overcondence. A way to understand the result is that through overcondence, the agent obtains a rst mover advantage towards the rm; overcondence acts as if the privilege of being the party making the take-it or leave-it oer is shifted from the rm to the agent. Although this analogy is useful, we also consider the case when both rms and workers form beliefs
2

Lemma 1 in Hvide (1999b) provides a formal link between degree of condence and the Bayesian

paradigm.

pragmatically, and show that the same type of overcondence results carry through to this case. An immediate criticism against the idea of pragmatic beliefs is that human agents are incapable of constructing beliefs that they do not really believe in. We should therefore stress that the main justication for pragmatic beliefs is dynamic; pragmatic beliefs in a population are formed either by the selection pressure towards agents that are born pragmatists, or by agents gradually learning that a certain way of forming beliefs is more rewarding than other ways. Importantly, we make no assumptions on the agents awareness about their own overcondence. Other criticisms against the notion of pragmatic beliefs will be discussed in a separate section. This paper is related to a new school of thinking on bounded rationality, exemplied by Rubinstein (1997), and surveyed by Lipman (1996). A central feature of this school is that humans are information processors that are imperfect when compared to the standards of the rationality paradigm. For example, Fershtman & Kalai (1993) consider a setting where a manager has limited attention span, and the rm therefore concentrates in only a few markets. Rubinstein (1993) considers a setting where consumers have heterogenous ability in understanding complex pricing schemes, and where the seller exploits this fact to protably discriminate among the consumers. Importantly, these models derive the behavioral implications of certain cognitive imperfections, taking these imperfections as exogenous. In contrast, our approach illustrates how bounded rationality can emerge endogenously from features of the physical description of the game. While the paper has a clear link to the bounded rationality literature, it also has a link to a philosophical school known as Pragmatism. According to William James (1908), one of the founders of Pragmatism, beliefs are true to the extent that they breed true in their consequences. One interpretation, as emphasized e.g., by Russell (1945), is that agents should (or do) hold beliefs that have good consequences. While good is a vague concept in most disciplines, economics seems particularly well-t to include a notion of Pragmatism, since we can understand good just as high utility.3 Therefore, the present
3

A standard line of argument against the Pragmatist position is that good can be understood to

mean true predictions, so that the Pragmatist position is impeccable, but vacuous (see e.g., Scruton, 1994). Clearly this criticism does not apply in our case.

paper may be seen as a modest attempt to link Pragmatism and Economics.4 The paper is structured as follows. In Section 2, we analyze the job market example, and in Section 3 we provide a discussion of the approach to beliefs outlined in this paper. Finally, in Section 4 we conclude.

An Example

There is one risk-neutral worker and a continuum of risk-neutral rms. The workers productivity is rm specic, and given by the distribution G(z), with density g(z), where z [0, 1]. At time 1, the worker is matched with a rm. The rm, who is the only part informed of the workers marginal productivity in that rm, makes a take-it-or-leave-it wage oer w to the worker. If the worker accepts the oer, the game ends. If the worker rejects the oer, she moves on to time 2, and is matched with another rm. We assume that the distribution of wage oers at time 2 is simply given by G(..).5 Then the game ends (hence there is no recall of the time 1 oer). The decision of the time 1 rm is which wage to oer the agent, and the decision of the worker is whether to accept the oer or not. Hence the agent faces a two-stage search problem, but where the oer at time 1 is endogenous. For simplicity we assume that there is no discounting or search costs. Nothing essential in the example hinges on these assumptions. Denote the agents productivity in the time 1 rm for M P . Clearly the optimal strategy of the rm at time 1 depends on M P but also on the agents beliefs about the
4

In this footnote, we consider some other related literature. Elster (1983), chapter IV, part 4, con-

siders the benets of non-rational beliefs, but in non-strategic settings. More closely related, Kyle & Wang (1997) considers the possible benecial strategic eects from overcondence in a nancial market. However, their denition of overcondence is dierent than ours; in particular their denition does not imply any violation of Bayesian modeling, in contrast to the present paper. [An agent is overcondent in Kyle & Wang (1997) if his posterior distribution is too tight, which may simply be due to a tight prior.] Finally, it is well-known that material payos in games can be higher for agents that are less than fully rational than for fully rational agents (e.g., Gilboa & Samet, 1989), or might be less than fully rational (Kreps et al., 1986). However, these results have to our knowledge never been linked to the notion of how agents form their beliefs. 5 Hence the wages are set competitively at time 2. For example, the agent is matched with two randomly drawn identical rms at time 2. As will be clear below, nothing hinges on this assumption.

distribution of wage oers at time 2. Again, these beliefs are assumed to be common knowledge. Since the agent is risk neutral, her beliefs can be summarized by the expected wage oer at time 2, denoted w. Clearly, the agent will only accept oers above w at time 1, and hence the optimal strategy for the rm is to set, w = w if M P > w, and w < w if M P < w Start out with the case where the worker and the rm share rational beliefs about the R1 distribution of wage oers at time 2, which is simply given w = 0 zg(z)dz. In that case, w at time 1.6 It follows immediately that the expected utility of the worker under rational belief formation simply equals w. Moreover, the equilibrium is ecient since the agent only starts working in the time 1 rm if her productivity exceeds w. We now consider the case when the agent forms her beliefs pragmatically. (1)

under rational belief formation, the workers optimal strategy is to accept any oer above

2.1

One-Sided Pragmatic Beliefs

Now consider the case when the worker holds pragmatic beliefs about the probability distribution over wage oers at time 2. The question we ask is whether an agent with pragmatic beliefs will be under- or overcondent about her outside opportunity. First a denition. Denition 2.1 If w > w, the agent is overcondent, and if w < w the agent is under condent. Hence if the agent believes that the opportunities at time 2 are better than they are according to a rational interpretation of data, she is overcondent, and if she believes that the opportunities are worse than under rational interpretation of data, she is undercondent. We then have the following. Proposition 1 The pragmatic beliefs are overcondent.
6

The agent may update her beliefs over time 2 oers when receiving the time 1 oer, but for simplicity

we ignore this possibility.

Proof. We treat the belief w as a variable and derive the pragmatic belief, dened as the optimal w (w should be thought of as being determined prior to the game, and is not a decision variable). Denote the probability of the workers productivity exceeding w for P (w). Then, w := arg max{P (w)w + (1 P (w))w)
w

(2)

where given the rms optimal strategy from (1) we have that, Z 1 g(z)dz P (w) =
w

(3)

Since

P (w) = g(w), the rst order condition becomes, w g(w)[w w] + P (w) = 0 (4)

Since P (..) > 0, we have that w > w, and hence the pragmatic belief is overcondent. Proposition 1 shows that a high belief about ones market value can to some extent be self-fullling, in that high expectations about future job oers forces the time 1 rm to give a higher oer to attract the worker, than in the case where the agent has lower expectations about his future. As can easily be seen, a pragmatic agent will enjoy a higher utility than a rational agent. This is illustrated further down. There are three things to notice about the solution. First, the overcondent solution induces an inecient allocation of workers: Since it is socially optimal that the worker starts working in any time 1 rm where her marginal productivity is at least w, it is socially optimal that the worker has correct beliefs about the distribution of oers at time 2. Second, since w is bounded below 1, there is a bound to the optimal degree of overcondence. A worker that excessively overcondent will practically always reject the oer at time 1, and hence in expected terms get a payo close to w. Still, a small degree of overcondence is better than no overcondence at all (or even worse, undercondence, in which case she would accept oers below w), since her expected payo is larger than w even for w very close to 1. We illustrate these points in a gure, calculated for g(z) = 1 (which gives w = 1 ). 2

0.75

Utility

tFLLG

0.5

0.25

0 0

Underconfidence

0.5

Overconfidence

1 Self-beliefs

Figure 1 The gure depicts the expected utility of the worker as a function of her degree of condence. While undercondence is severely punished in that the worker accepts a too low oer at time 1, a moderate degree of overcondence (w =
3 ) 4

gives the agent the

highest expected utility. When the agent becomes excessively overcondent (w 1), the expected utility converges from above to the expected utility under neutral condence (w = 1 ). 2 The third thing to notice is that there is an alternative way to derive w . Suppose that the worker (instead of the rm) makes the take-it or leave-it oer at time 1, and denote this oer by w. Conditional on the oer, the optimal strategy for the rm is then to, accept the oer if M P > w reject the oer if M P < w 7 (5)

In setting the optimal take-it or leave-it oer, the agent then trades o the probability of getting the oer accepted with the size of the oer. Denote the probability of having the oer accepted by Q(..). Then the optimal take-it or leave-it oer of the worker is the solution to, w := arg max{Q(w)w + (1 Q(w))w)
w

(6)

The rst order condition becomes, Q(w) [w w] + Q(w) = 0 (7) w R1 but given the rms optimal strategy, Q(w) = w g(z)dz = P (w), and hence w = w .

Consequently, the solution to the optimal belief problem is equivalent to the solution

to the optimal take-it or leave-it oer, in terms of utility to the worker and prots to the rm. Hence we can interpret the possibility of overcondence as a way of shifting the rst mover advantage from the rm to the worker (or as giving the agent the possibility to commit to turning down any oer below w). Given this equivalence result, a natural question is then whether the whole advantage of overcondence boils down to a question about rst mover advantage. To illustrate that more than this is at issue, we consider a twist to the story, where both parties at time 1 form pragmatic beliefs.

2.2

Two-sided pragmatic beliefs

We now consider the case when the time 1 rm also constructs its beliefs pragmatically, and, as before, makes the take-it or leave-it oer. Specically, we consider a situation where the worker forms a pragmatic belief about the distribution of wage oers at time 2, while the time 1 rm constructs a pragmatic belief about the agents productivity in its own rm. We retain the assumption that both the rm and the worker are risk-neutral. For simplicity it is assumed that the agent knows her true productivity at time 1.7 The timing of the game is now that the worker and the rm rst simultaneously set their
7

This assumption makes things simpler, but qualitatively the same conclusion holds for the game

where the agent forms his pragmatic belief based on a probability distribution of productivities at time 1.

beliefs, then these beliefs are observed, and nally the same oer game as in the previous section is played. Denote the rms belief about the workers productivity in its own rm by MP. Denote the agents belief about the time 2 (expected) oer by w, as before. Let us start out with the oer subgame. For a given MP and w we get the rms optimal oer w as, w = w if MP > w w < w if MP < w Given his beliefs, the optimal strategy of the worker is to accept the oer if w = w and to reject the oer if w < w. Given the equilibrium of the oer subgame, we solve for the equilibrium in pragmatic beliefs. We then have, Proposition 2 If M P < w, both the rm and the worker have neutral condence in equilibrium. However, if M P > w, the worker is overcondent in all Nash equilibria, and her utility is higher than under neutral condence. Proof. First we derive the agents best response in terms of w, as a function of MP. First suppose that MP < w. In that case clearly any w > MP is strictly better for the agent than w < MP, since w < MP implies that the rm will set w = w, and the worker will accept the oer in spite of w > w. Hence any w > MP is a best reply when MP < w. Now suppose that MP > w. In that case, the unique best reply for the worker is to set w = MP, since that will ensure the highest possible oer at time 1 (and furthermore the oer will exceed w). Now to the best response of the rm in terms of MP, as a function of w. For any w > M P , it is clearly suboptimal for the rm to set MP > w, since the rm will then run a decit when in the next turn oering w = w. It follows that any MP < w is a best response of the rm in the case where w > M P . Now to the case w < M P . In that case it is a unique best response for the rm to set MP = w. It follows that for M P > w, the set of Nash equilibria are the points on the 45 degree line between (w,w) and (M P ,M P ), which are all characterized by the agent being overcondent about the distribution of wage oers at time 2, and the rm being undercondent about the workers productivity. On the other hand, for M P < w, the unique Nash equilibrium is the point (w, MP ). 9 (8)

Since M P > w with probability 1 , the agent is overcondent in exactly one half of 2 the cases (while in the other half of the cases, the agent has neutral condence). Thus we have shown that the overcondence result survives letting the rm as well as the agent form pragmatic beliefs, and hence the nding of overcondence is not an artifact of letting the agent form pragmatic beliefs one-sidedly.8

Discussion

With a simple example from the job market, we have shown that if beliefs are formed pragmatically, overcondence can emerge as an equilibrium phenomenon, in contrast to under rational belief formation. In the example constructed in Section 2, overcondent beliefs (that are known to the rm) force the rm to oer a higher wage than it would to an agent with the same productivity, but more realistic self-condence. This nding, we argue, can shed light on the persistent nding of overcondence from psychology and economics.9 One criticism against the notion of pragmatic beliefs is that beliefs cannot have a commitment eect on the behavior of other agents, or any other eect, because an agents beliefs are unobservable to other agents. A possible reply to the observability criticism (in the context of self-beliefs) is that self-beliefs may indeed be observable to some extent. The starting point is to view pragmatic beliefs as being the outcome of an evolutionary process, where overcondent agents gradually drive out agents with more rational conceptions of themselves. In such an evolutionary model, rms ability to distinguish true from fake
8

Notice that in contrast to the case where only the agent is allowed to form pragmatic beliefs, the

two-sided pragmatic beliefs case is characterized by an ecient outcome; the worker accepts the oer from the case only in the case where M P > w. 9 In the experiment on entry decisions performed by Camerer & Lovallo (1999) overcondence clearly has a benecial strategic eect, since such overcondence would pre-empt agents with rational beliefs from entering. However, the documented degree of overcondence seems to overshoot the optimal degree of overcondence, so the notion of pragmatic belief formation can only be a part of the picture behind the ndings of Camerer & Lovallo (1999). For the ndings of overcondence in betting behavior by Golec & Tamarkin (1995), a pragmatic explanation seems feasible (if perhaps not probable), while Asubels (1991) ndings of overcondence in the credit card market can hardly have a pragmatic explanation, as there any benecial strategic eects from overcondence in the credit card market seem far-fetched.

10

overcondence should be endogenous, since being able to make such distinctions would contribute to the tness of the rm (alternatively, agents that are able to tell lie from truth would be attractive to the rms). Hence we can imagine an arms race where rms become increasingly sophisticated at distinguishing rational agents from pragmatic agents, and the rational agents become increasingly sophisticated at imitating the appearance of pragmatic agents. While the details of such an evolutionary story are left out for future work, the point is that it is not necessarily true that fake and true overcondence cannot be distinguished (for a similar point and for more on the details of similar arms races, see Frank, 1988).10 If we admit the possibility of overcondence being observable, we could also think of pragmatic beliefs as the outcome of a process where the agent repeatedly face a similar situation (like job interviews) and records how the payo from the situation varies with her beliefs. Then, if it turns out that overcondent beliefs give a higher payo than more neutral beliefs, then no wonder that the agent adopts those overcondent beliefs. Of course, the agent herself does not have to judge her own beliefs as overcondent (even if the rm does so). Similarly, one can think of pragmatic beliefs as being the outcome of an imitation process. Agents observe that other agents with a certain type of beliefs tend to do better than others, and consequently copy their manner of belief formation.11 Now to another type of criticism, which takes the notion of pragmatic beliefs as granted, but questions whether pragmatic beliefs are necessarily overcondent. We have shown through a job market example that overcondence can emerge. However, that only gives us a partial model where overcondence emerges, while such a partial model cannot
10

Alternatively, we could have a (non-evolutionary) reputation model along the lines of Kreps et al.

(1986), where rms believe that there are some pragmatic agents in a population, and where rational agents can obtain a higher payo by imitating the behavior of the true pragmatists. However, such a story would not explain how pragmatic agents came into the picture in the rst place. 11 The reader may have observed that under any of these interpretations, pragmatic agents will at some point easily agree to disagree with other agents (rational or pragmatic). As has been argued by many authors, disagreement about events that are in principle veriable (like the ability of an agent) should vanish, due to dutch book arguments. However, risk aversion or liquidity constraints (or dispreference for betting) can imply that disagreeing agents will not necessarily bet their whole fortune (but parts of it). For example, it may be true that dierences in beliefs betweeen overcondent and rational agents can to some extent explain why there is horse betting, as pointed out by Golec & Tamarkin (1996).

11

ensure that overcondence emerges as a more general phenomenon. Let us be more specic. There are many settings where overcondence can indeed be a strategic advantage. We have looked at one example from the job market, but can also imagine other examples from all-pay auctions, mating games, Cournot competition, and so forth. However, we can also easily imagine partial models where the pragmatic beliefs would be undercondent (think e.g., of rms forming pragmatic beliefs about their future production costs under various versions of Bertrand competition). Does this argument imply that optimal, say, self-beliefs are somewhere in the middle (perhaps at neutral condence)? That would be the conclusion if we argue that degree of self-condence must necessarily be the same across situations, holding the skill xed. The opposite view, e.g., that self-condence can vary across situations, would imply that self-beliefs are consistent within situations, but can be inconsistent when viewed between situations. Whether Nature can allow humans to hold beliefs that change according to situation (or game played) is an open question. While casual empiricism suggests that human agents are capable of forming new selfbeliefs independently of new information arriving (the mood swings of adolescents is one example), it is hard to say whether such changes has anything to do with strategic aspects of stage changes.

Conclusion

Which factors shape the manner in which agents create and update their beliefs about economic variables? The traditional answer to this question, imported from statistical theory, is that economic agents pursue some notion of truth in their belief formation. This idea is clearly formulated in e.g., textbook game-theoretic approaches. The basic idea behind the present paper is that the main force behind belief formation is incentives, and not any abstract notion of truth. From this idea, we proposed the notion of pragmatic beliefs, where such beliefs are simply those that maximize the payo in a given situation. Due to strategic eects, pragmatic beliefs are not necessarily truth-driven. To indicate that the notion of pragmatic beliefs can be useful, we considered the overcondence puzzle from various experimental studies. This puzzle, that real life agents seems to be overcondent about their own abilities, can hardly be explained by any tradi12

tional, rational, notion of belief formation. Equipped with the notion of pragmatic beliefs, however, we showed that overcondence can emerge as an equilibrium phenomenon, in that the pragmatic beliefs in a simple job market example were shown to be overcondent. Since the conclusions of the example rests on several strong assumptions that are hard to verify directly, one extension of the present work could be to perform an experimental test of the notion of pragmatic beliefs.

References

Asubel, L. M. (1991). The Failure of Competition in the Credit Card Market. American Economic Review, 81, 50-81. Binmore, K. (1987a). Modeling rational players: Part I. Economics & Philosophy, 3, 179-214. Binmore, K. (1987b). Modeling rational players: Part II. Economics & Philosophy, 4, 9-55. Blume, L. & D. Easley (1992). Evolution and Market Behavior. Journal of Economic Theory, 58, 9-40. Camerer, C. & D. Lovallo (1999). Overcondence and Excess Entry: An Experimental Approach. American Economic Review, 89, 306-318. De Bondt, W.F.M. & R. H. Thaler (1994). Financial Decision Making in Markets and Firms: A Behavioral Perspective. NBER Working Paper Series #477. Elster, J. (1983). Sour Grapes. Studies in the Subversion of Rationality. Cambridge University Press. Fershtman, C. & K. Judd (1987). Equilibrium Incentives in Oligopoly. American Economic Review, 77, 927-40. Fershtman, C. & E. Kalai (1993). Complexity Considerations and Market Behavior. Rand Journal of Economics, 24, 224-35. Frank, R. (1988). Passions within reason: The strategic role of the emotions. Penguin Books. Gilboa, I. & D. Samet (1989). Bounded versus Unbounded Rationality: The Tyranny of the Weak. Games and Economic Behavior, 1, 213-21. 13

Golec, J. & M. Tamarkin (1995). Do Bettors Prefer Long Shots Because They are Risk-Lovers or are They just Overcondent? Journal of Risk and Uncertainty, 11, 51-64. Hvide, H. K. (1999a). Self-Awareness, Uncertainty, and Markets With Overcondence. In: Machina, M. & B. Munier, Beliefs, Interactions and Preferences in Decision Making, p. 159-176. Kluwer Academic Publishers. Hvide, H. K. (1999b). On the Informational Role of Education in the Allocation of Talent. Working Paper, Berglas School of Economics, Tel-Aviv University. Available at: http://www.ssrn.com/. James, W. (1907/1975). Pragmatism: A New Name for Some Old Ways of Thinking. Harvard University Press. Kahneman, D., P. Slovic & A. Tversky (1982). Judgment under Uncertainty: Heuristics and Biases. Cambridge University Press. Kreps, D., P. Milgrom, J. Roberts and R. Wilson (1986). Reputation and Imperfect Information. Journal of Economic Theory, 27, 253-79. Kyle, A. S. & F. A. Wang (1997). Speculation Duopoly with Agreement to Disagree: Can Overcondence Survive the Market Test? The Journal of Finance, 52, 2073-90. Lipman, B. (1995). Information Processing and Bounded Rationality: A Survey. Canadian Journal of Economics, 28, 42-67. Ponssard, J. P. (1979). The Value of Information in Zero-Sum Games. Managment Science. Rubinstein, A. (1993). On Price Recognition and Computational Complexity in a Monopolistic Model. Journal of Political Economy, 101, 473-84. Rubinstein, A. (1997). Modeling Bounded Rationality. MIT Press. Russel, B. (1945). A History of Western Philosophy. Simon and Schuster, New York. Scruton, R. (1994). Modern Philosophy. An Introduction and Survey. Penguin Books. Svensson, O. (1981). Are We All Less Risky and More Skillful Than Our Fellow Drivers? Acta Psychologica, 47, 143-48. Vallone, R. P. et al. (1990). Overcondent Prediction of Future Actions and Outcomes by Self and Others. Journal of Personality and Social Psychology, 58, 582-9. Waldman, M. (1994). Systematic Errors and the Theory of Natural Selection. American Economic Review, 84, 482-497. 14