Bounded rationality in organisation

Bounded Rationality in Industrial Organization1
Glenn Ellison
MIT and NBER
January 2006
1This paper supported by NSF grant SES-0219205. Tom Chang and Moshe Cohen provided
valuable research assistance.I thank Richard Blundell, Ken Hendricks, and many World Congress
attendees for their comments.

Abstract
This paper discusses the use of bounded rationality in industrial organization. There is a
long tradition of such work: literatures from various decades have discussed irrationalities
by firms and consumers. Three main approaches are found in the recent literature: rule-of-thumb
papers specify simple rules for behavior; explicit bounds papers consider agents who
maximize payoffs net of cognitive costs; the psychology and economics approach typically
cites experimental evidence to motivate utility-like framworks. Common to each recent
literature is a focus on consumer irrationalities that firms might exploit. I discuss several
new topics that have been opened up by the consideration of bounded rationality and new
perspectives that have been provided on traditional topics.

1 Introduction
In the last few years there has been a surge of interest in bounded rationality in industrial
organization. The field has attracted the attention of a remarkable collection of top young
researchers from industrial organization and behavioral economics. Much of the recent in-terest
has been spurred by developments in psychology and economics, but there is also a
longstanding tradition of such work in industrial organization. Although the field is not
yet as coherent and advanced as most fields surveyed at an Econometric Society World
Congress, the fact that diverse strands of research and diverse researchers are coming to-gether
seemed to me to make it a topic for which a discussion could be useful.
The terms “boundedly rational” and “behavioral” have been used by different groups
of economists over the years to describe different styles of work. The “bounded rationality”
in my title is interpreted broadly, and includes work from three different traditions. One
tradition I call the rule-of-thumb approach. These papers typically assume directly that
consumers or firms behave in some simple way, rather than deriving behavior as the solu-tion
to a maximization problem. A second tradition is what I call the “explicit bounded
rationality” approach. In this literature, cognition is costly, and agents adopt second-best
behaviors taking these costs into account. The third tradition examines what happens in
industrial organization settings when consumers are subject to behavioral biases identified
in the psychology and economics literature.
The primary motivation for boundedly rational analysis is common across three tradi-tions:
in many situations the “rational” model does not seem plausible or has a hard time
incorporating factors that seem important. There are also other motivations. One is that
after twenty years of dominance of the game-theoretic paradigm, boundedly rationality ex-planations
have been much less worked over. Another is that boundedly rational models
are sometimes more tractable than rational models and make it feasible to enrich one’s
analysis in other dimensions.
Although rational game-theoretic approach has dominated IO for the past twenty five
years, the idea of using bounded rationality is far from new. This World Congress coincides
with the 50th anniversary of Simon’s (1955) call to replace the rational economic man with a
1

boundedly rational counterpart. Next year is the 50th anniversary of Strotz’s (1956) seminal
work on time inconsistency. The rule-of-thumb approach is even older. The application of
these ideas to industrial organization has waxed and waned, but by now it amounts to a
substantial literature.
I begin this essay with a discussion of several old literatures. I do this both because
I think it is useful to recount what happened before 1990, and because it provides an
opportunity to explain what the different boundedly rational approaches are, what are
their strengths and weaknesses, and where they have been most successful. I think all
three methodologies are useful and that anyone interested in boundedly rational industrial
organization should be fluent with each of them.
I then turn to the more recent literature on boundedly rational IO. Here, I see the liter-ature
as having divided into two main branches. The first is growing out of the industrial
organization literature. Papers in this branch typically begin with one of two observations:
the IO literature on some topic is unsatisfying because aspects of the existing rational mod-els
aren’t compelling; or the IO literature has had little to say about some topic (perhaps
because rational modeling would have been awkward). They then propose models with
some boundedly rational elements and discuss how they add to the literature on the topic.
The second branch is growing out of the psychology and economics literature. Papers in this
branch typically start by noting that the psychology and economics literature has shown
that consumers depart from rationality in some particular way. They then explore how a
monopolist would exploit such a bias, what the effects would be under competition, etc. I
provide overviews of each of these literatures and discuss some general themes.
Readers of this survey will note that the literatures I discuss remain sparse. Most
topics in IO have little or no boundedly rational work on them. Most behavioral biases
have received little or no consideration in IO, and even when they have been discussed it
is only the most basic IO questions that have been asked. This makes it an exciting time
to work in the field.
2

2 Early History
It is impossible to identify the beginnings of bounded rationality in industrial organization.
Throughout the period when the neoclassical profit-maximizing model was being developed
there were spirited debates between proponents and opponents of “marginal revenue =
marginal cost” as a foundation for firm behavior. Hall and Hitch (1939), for example, take
direct aim at economists who “assume the general relevance of the simple analysis in terms
of marginal cost and marginal revenue” and “that production is carried out to the point
where this elasticity is equal to the ratio P
P−MC .” They report that interviews of 38 business
executives
... casts doubt on the general analysis of price and output policy in terms of
marginal cost and marginal revenue ... they are thinking in altogether different
terms; that in pricing they try to apply a rule of thumb we call ‘full cost’.”
Hall and Hitch were mostly concerned with debating the modeling of firm behavior, but
other papers from the 1920’s, 1930’s, and 1940’s were more directly focused on industrial
organization implications. Rothschild (1947) is a delightful example that in many ways is
quite like modern behavioral IO papers. It assumes that a firm’s “desire for secure profits”
is a second objective as important as the desire for maximum profits and discusses vari-ous
implications in oligopoly settings. For example, it argues that this can lead to price
rigidity, to varying contractual terms-of-trade, to price wars, to political actions, etc. He
sees the firms in his paper as practicing full cost pricing beause they are rational with a
different objective function rather than irrational, but would rather throw out rationality
than full-cost pricing if his rationalization proves incorrect: “if business behaviour were
really irrational, this would not serve as an excuse for the neglect of such behavior. Ir-rationality
would then have to become one of the premises of oligopolistic price theory.”
Analytically, of course, the paper is quite different from modern work. The conclusions are
drawn from very loose verbal arguments. Rothschild is, however, prescient in recognizing
this limitation. He explains that “A completely novel and highly ingenious general theo-retical
apparatus for such a solution of the oligopoly problem has recently been created by
3

John von Neumann and Oskar Morganstern ... Unfortunately, at the time of writing this
article I had no opportunity of obtaining a copy of this important book.”
One tends to think that there was a long era of neoclassical dominance betweeen the
death of this old literature and the birth of interest in post-rationality bounded rationality.
In this light, I find it noteworthy that less than a decade separates the publication of
Rothschild’s paper and the publication of Simon’s (1955) “A Behavioral Model of Rational
Choice.” Simon is clearly writing for an audience steeped in rationality and sells his bounded
rationality as an enhancement of the prevailing paradigm:
the task is to replace the global rationality of economic man with a kind of
rational behavior that is compatible with the access to information and the
computational capacities that are actually possessed by organisms.
Simon discusses the various challenges involved in optimization. His critiques and visions for
a new framework are compelling. The progress he makes toward this goal is less satisfying.
He mostly ends up emphasizing a “satisficing” model in which agents search for actions
until they find one that achieves a payoff that provides them with at least some aspiration
level.
Cyert and March (1956) is an early application to industrial organization. In the mod-ern
style, they start by citing interesting papers from the psychology literature to justify
their assumptions about deviations from profit-maximization.1 Among the implications of
satisficing they discuss are that firms are most likely to expand sales when profits are low
and costs high, and that dominant firms may tend to lose market share over time. The
analytic methodology, of course, is still far from modern: other than one set of diagrams
using Stigler’s dominant firm model, the analysis is completely informal. They do present
empirical data related to their hypotheses.
One aspect of this literature that is striking in contrast to the current literature is that
the focus is almost exclusively on firms’ deviations from profit-maximization rather than
on consumer irrationality. Indeed, consumers are rarely mentioned. Their presence is, of
1For example, the notion that people satisfice is supported by the experiments showing that subject who
are asked to throw darts as close to the center of a target as possible are more accurate when the target is
smaller.
4

course, implicit in the assumption that the firm is facing a specified demand curve. Perhaps
this is why consumer irrationality was not discussed. Whether the demand curve reflects
the aggregation of utility-maximizing decisions or some other consumer behavior would not
really affect the analysis.
3 Bounded Rationality at the Time of the Game-Theoretic
Revolution
The late 1970’s and early 1980’s is usually thought of as the beginning of the game-theoretic
revolution in Industrial Organization. It was also a time when several distinct “boundedly
rational” or “behavioral” approaches were being developed.
3.1 The rule-of-thumb approach
Consider the simple problem a consumer faces on every trip to the supermarket: should he
buy any of the dozens of relatively new products he has never tried or should he continue
to buy what he habitually buys. To make a rational decision the consumer must start with
a prior over the quality of the unknown goods, he must also have a prior over how quality
covaries with cost and think through what this implies about the signals inherent in the price
of the good and the amount of shelf space the store has devoted to the product. Whether
by observing other shoppers or talking to friends, the consumer will also learn something
about the popularity of the product. Correctly accounting for this signal requires a correct
prior over the set of possible social processes by which popularity has been achieved. How
long has the product been on the shelf before I noticed it? Are the previous consumers
people who have used the good before or people trying it for the first time? How many
previous customers did they talk to before trying it? What did they know about where
those customers got their information and how they used it? Recent models by Banerjee
(1992), Bikhchandani, Hirshleifer and Welch (1992), Banerjee and Fudenberg (2004), and
Bose, Orosel, Ottaviani, and Vesterlund (2006) provide remarkable analyses showing that
the rational approach can be made tractable in some such situations.
Despite my admiration for these papers, when I think about consumers’ making such
5

decisions dozens of times in the course of an hour-long shopping trip I have the same reaction
that Smallwood and Conlisk (1979) had when contemplating this problem twenty-five years
ago:
. . . a would-be optimizing consumer who took account of market popularities
would be involved in a massive game theory problem with all other consumers.
Is it really plausible that he could solve the game?
The approach of the rule-of-thumb literature is to simply posit rules of thumb that
consumers are assumed to follow. One can think of this as similar to the game-theoretic
approach, but skipping the part of the argument in which one posits utility functions and
derives the behaviors as optimizing choices. Proponents of the rule-of-thumb approach see
it as having two advantages. First, in some models it seems implausible that consumers
would do the “rational” calculations. A desiderata for choosing a rule-of-thumb is that
the explicitly assumed consumer behavior should be more believable than the behavior
implicitly assumed in the obvious rational alternatives. Second, rule-of-thumb papers tend
to provide more analysis of robustness. One reason is inherent: the analysis of rule-of-thumb
models is simpler, making it possible to analyze more variants in the same amount
of time. Another is probably sociological: rule-of-thumb authors recognize that they will
be subject to more scrutiny than their “rational” peers and must head off critiques about
the conclusions being nonrobust to changes in the ad hoc assumptions.
Smallwood and Conlisk’s (1979) model of the effect of product quality on market share
remains a beautiful example of the rule-of-thumb approach. Consider a market in which
consumers choose between K brands at t = 0, 1, 2, . . .. Suppose that a consumer of product i
in period t experiences a “breakdown” with probability bi. This could be a literal breakdown
of a durable good or just a disappointing experience with an ordinary consumable.
In this environment, Smallwood and Conlisk examine what happens when consumers
follow particular rules of thumb. Specifically, they assume each consumer continues to
use the same product until he experiences a breakdown. When a breakdown occurs he
considers switching and ends up choosing product i with probability proportional to mi(t),
where mi(t) is the market share of product i. By varying the parameter the assumption
6

can encompass a range of behaviors. For example, = 0 is purely random choice, and
= 1 models consumers who ask one randomly selected friend which brand they used and
purchase that product.
The first theorems in the paper explore the connection between behavior at the consumer
level and the efficiency of product adoption. For example, they show that when 1,
higher quality products eventually become more popular regardless of the initial conditions,
but all products have positive market shares in the long run. When 1 this is not true
and an inferior product can come to dominate the market if its initial market share is
sufficiently high. A particularly interesting feature of the model is that when = 1 the
highest quality product (or products) always dominates in the long run. In this sense, we
see that limited rationality at the individual level – recall that = 1 means copying the
decision made by one randomly chosen user without using any information about their
satisfaction – can make product adoption socially optimal in the long run. When this
happens, it makes the rule-of-thumb assumption all the more palatable: there is no strong
incentive to deviate to some more complex behavior.
The Smallwood-Conlisk paper has other features that presage future developments in
boundedly rational IO. Stylistically, it includes a quick reference to some work by Tversky to
motivate the assumption of bounded rationality. More substantively, the second half of the
paper pairs irrational consumers with strategic firms to endogenize the product qualities:
Firms, unlike consumers, will be assumed to solve optimization problems in
choosing their bi(t). A rationale is that firms are better able to compute optima
and are penalized more if they do not (through the force of competition.)
In the recent psychology and economics-motivated literature the rational firm-irrational
consumer assumption has become the norm, and the question of what firms do to exploit
irrationality is often the primary focus.
When the game-theoretic revolution swept through industrial organization, much of
rule-of-thumb literature could not withstand the onslaught. For example, Schmalensee
(1978) used a rule-of-thumb approach to analyze a model in which higher quality products
were more costly to produce and firms chose advertising levels. He argues for the rule-
7

of-thumb approach over the rational approach on the grounds that “it seems implausible
to assume that households actually compute optimal solutions to a large number of dif-ficult
game-theoretic and information-theoretic problems.” Part of his argument for this
contention is that “buyers’ optimal use of the signals transmitted by firms’s choices of
advertising levels would depend on the strategies being employed by all sellers.” Within
just a few years, the mechanics of solving such signalling problems had become so routine
that I am sure few economists presented with this latter quote would have guessed that it
was meant as a critique of signalling theory rather than as an exposition of how to solve
signalling problems.
3.2 Explicit bounded rationality
Although Simon’s initial motivation of bounded rationality leaned very heavily on limited
human capacity for computation, the theory mostly developed along lines that had little
to do with this. Agents were assumed to satisfice rather than to maximize, but there was
little attempt to formalize why this might be easier than maximizing or to provide criterion
on which to assess the feasibility of other behaviors.
One place in which computational limitations were made explicit was in team theory.2
In the canonical model of this literature, a firm is modeled as a group of agents sharing
a common objective. The firm needs to choose a vector-valued action. Which action is
optimal depends on an unknown state of nature. Each employee has some information.
The problem of choosing an optimal action is complicated by the presence of two addi-tional
costs: each agent may incur a cost of gathering information; and there may be costs
of communicating information across agents. Given these information costs, it will gener-ally
not be optimal to gather all the available information, nor to convey what has been
gathered to a single decision maker. Instead, the firm may want to decentralize decision-making
and have agents or groups of agents choose components of the vector-valued action
independently.
Making decisions according to the optimal decentralized procedure is, of course, the
2See Marschak and Radner (1972).
8

fully rational thing to do in this model if one takes all costs into account. It can thus be
seen as providing a rational analysis of why firms are organized as they are. Team theory
models can also, however, be seen as tool for looking at a range of industrial organization
problems. The idea is that team theory provides information-cost-based microfoundations
that can guide our choice of rules of thumb. The traditional analysis of monopoly and
oligopoly problems ignores information costs. Hence, what we call “rational” models are
really rule-of-thumb models that assume firms use a particular rule of thumb: behave as
if information costs were not present. If team theory gives us some general insights into
how information costs affect firm behavior, then analyzing IO problems by using rules-of-thumb
suggested by team theory should be superior to the “rational” approach as it is
conventionally applied.
Radner’s (1975) model of cost-reduction by satisficing managers connects with both
the explicit bounds and the rule-of-thumb literatures. He examines the problem of a firm
that has only limited managerial attention to devote to minimizing its costs. Analytically,
Radner takes a rule-of-thumb approach. He defines procedures for the manager to follow
and examines the consequences.
3.3 Empiricism as “behavioral economics”
The 1970s also featured a third “behavioral” approach to industrial organization that would
not normally be thought of as behavioral today. Joskow’s (1973) “Pricing Decisions of
Regulated Firms: A Behavioral Approach,” is a good example. What Joskow does in this
paper is simply to estimate a probit model. The dependent variable is an indicator for
whether a utility applied to the New York Public Service Commission for an increase in its
electric rate. He assumes that they apply for a rate hike if and only if Xi

+ i 0 and
estimates the parameter

. The right-hand side variables Xi include things like the firm’s
earnings growth.
Economists today will wonder why this would have been called “behavioral.” Today,
the probit model would be thought of simply as a method for estimating the firm’s profit
function. If the firm is rational and profits are of the form i = Xi

distributed component of profits that is unobserved by the econometrician, then the probit
model provides a consistent estimate of

. Suppose, however, that profits are not actually
linear. In the rational paradigm one assumes that firms fully understand the nonlinearities,
but in many cases it would be equally plausible to assume that firms are unaware of the
nonlinearities and follow an irrational rule-of-thumb of applying for a rate hike whenever
Xi

+i 0. In this case, the probit regression no longer estimates of profits, but it remains
a consistent estimator of behavior. Whether the behavior we estimate is profit-maximizing
or derived from a rule-of-thumb is irrelevant in many applications.3
It is because empirical work often involves directly estimating behavior (rather than
utility or profit) that one can think of empirical economics as a behavioral approach. The
empirical literature then becomes a place where “boundedly rational” industrial organiza-tion
has quietly carried on for decades. Consider, for example, my empirical work with
Judy Chevalier (1997) on risk-taking by mutual funds. We start by estimating the relation-ship
between a mutual fund’s performance in year t and how much new business it attracts
in year t + 1. The main focus, is then on how fund companies distort their investment
decisions in order to attract new business. It may be possible to provide a plausible ratio-nal
explanation for the precise form of the relationship between investment returns and the
inflow of new business, although my prior would be that it is not.4 For the main question of
interest, however, why consumers choose between mutual funds as they do simply doesn’t
matter. The initial estimates of consumer behavior tell us everything about consumers that
we need to know to think about the firm’s optimization problem.
A similar argument could in fact be made for much of empirical industrial organization.
Consider, for example, a classic “rational” paper, Porter’s (1983) study of price wars in
a railroad cartel. Porter estimates a structural econometric model in which demand is
assumed to be of the form log(Qt) = 0 − 1 log(Pt) + 2Lakest + ut and firms’ supply
decisions are of the form predicted by the Green-Porter model of collusion with imperfect
monitoring. Again, to answer his main questions about what the firms are doing, it does
not matter whether the demand curve comes from a population of consumers who are
3One notable exception is papers doing welfare analyses.
4See Lynch and Musto (2003) and Berk and Green (2004) for more on this.
10

optimizing some utility function or a population of boundedly rational consumers. By
focusing on consumer behavior as an empirically estimable object, the approach is robust.
Discussing the whole field of empirical industrial organization is obviously far beyond
what is possible in this format. For the remainder of this paper, I will treat work that is
behavioral in the sense that Joskow’s paper was behavioral as outside the scope of my talk.
3.4 Psychology and economics
Psychology and economics was also developing rapidly in the late 1970’s and early 1980’s.
Kahneman and Tversky had published their work on prospect theory in Econometrica in
1979 and many new areas were opening up. Developments in psychology and economics have
had a huge impact on the new boundedly rational IO. Many other surveys of psychology
and economics are available (including two in this volume), so I won’t survey developments
in psychology and economics that are not focused on industrial organization.
4 Developments in the theory of bounded rationality
In the 1980’s the game theoretic revolution in industrial organization seems to have been
all-consuming. One can find slight departures form rationality in quite a few papers, but
most seem to be unintentional or intentional shortcuts.
In the field of economic theory the situation was somewhat different. The 1982 vol-ume
of Econometrica is a remarkable illustration of the progress in game theory in the
early 1980’s: Rubinstein’s bargaining paper, Kreps and Wilson’s sequential equilibrium
paper, Milgrom and Weber’s auction paper, and Crawford and Sobel’s cheap talk paper
are just some of the well-known papers in that volume. By the time Tirole’s The Theory
of Industrial Organization appeared in 1988, however, theorists were already looking be-yond
the rational agent-equilibrium paradigm. In this section I will discuss a couple of the
late 1980’s developments in theory that have been important in the growth of boundedly
rational industrial organization.
11

4.1 Learning
Fudenberg-Kreps’ (1988) draft monograph “A Theory of Learning, Experimentation, and
Equilibrium in Games,” promoted a rule-of-thumb approach to the question of whether we
should expect players to play the equilibrium of a game. Again, a two-fold argument can be
made for this approach. First, a fully rational approach to the problem of how players learn
to play the equilibrium of a game is unappealing — it essentially replaces the assumption
that agents play the equilibrium of some simple game with the assumption that they don’t
know enough to do this but somehow know enough to play the equilibrium of some horribly
complicated game in which agents with different information are all simultaneously learning
from one another. Second, a boundedly rational approach can provide an enlightening
discussion of which rules would and wouldn’t lead to equilibrium play. One way to do
this is to examine a variety of behavioral assumptions. Fudenberg and Kreps advocated
another: they ask the reverse-engineering question of what assumptions about behavior are
needed to get to particular conclusions about the outcome of the learning process.
By the early 1990’s boundedly rational learning models were clearly the hot topic in
game theory. Most of this literature was directed at game theorists, but several of the
classic papers included little industrial-organization stories. For example, Kandori, Mailath,
and Rob (1993) motivated their model using the example of a group of students buying
computers in a world with network externalities.
The nicest example I know of a paper addressing a real industrial organization prob-lem
by such methods is M¨obius’s (2001) examination of the rise and fall of competition in
local telephone service in the U.S. in the early 20th century. The basic facts about tele-phone
competition are striking. ATT was a monopolist in the U.S. telephone market until
key patents expired in 1893/4. Eight years later, three thousand firms had built separate
networks not connected to ATT’s and taken approximately half of the rapidly growing
market. The entrants continued to do well for another decade or so as the market quadru-pled
in size. But a period of consolidation followed, and by the mid 1920’s competition had
essentially ended.
M¨obius provides an explanation for this pattern using a rule-of-thumb learning model.
12

In his model, consumers and businesses are connected by social networks. The network
consists of many “islands,” each containing a several consumers and one business. Every-one
interacts disproportionately with others in their island, but consumers also benefit from
interactions with businesses on other islands. The behavior rules that describe how con-sumers
and businesses decide between competing noninterconnected telephone networks are
similar to those in Kandori, Mailath, and Rob (1993), Young (1993), and Ellison (1993). In
each period most consumers take others’ actions as given and choose myopically between
having no phone or buying service from firm A or firm B to maximize their current period
utility. Firms’ decision rules are similar, but firms may also want to get two telephones so
as to be accessible to consumers on both networks. A small fraction of consumers are as-sumed
to adopt randomly in each period. Essentially, what M¨obius has done with these few
assumptions is to describe a Markov process governing telephone adoption. What remains
is just to analyze the behavior of the process. Using both analytic results and simulations
he shows how the model can account for the growth and decline of independents. Roughly
speaking, what happens in the early period is that each network grows by adding new
subscribers within the islands in which they are dominant. Later, the “duplication” effect
becomes important. Businesses in islands dominated by the minority telephone network
start to get a second phone to communicate with consumers from other islands. Once this
happens, the minority network has only a weak hold on the islands it dominates and can
easily lose a whole island in a chain reaction if a few friends randomly decide to switch to
the majority system. He presents empirical evidence consistent with this mechanism.
Thinking about the methodology, what makes the boundedly rational approach so useful
here is more related to the second advantage mentioned above. The relative tractability
of making direct assumptions about behavior allows M¨obius to enrich his model in a way
(incorporating the social network) that that would have made it too complex to analyze
had he insisted on assuming consumer rationality.
13

4.2 Computational complexity
Explicit concern for computational complexity entered game theory at about the same
time. Rubinstein (1986) and Abreu and Rubinstein (1988) discussed complexity in the
context of repeated games. In particular, they discussed the concept of strategies’ being
implementable by finite automata and used the number of states of an automaton as a
measure of its complexity. Standard repeated game equilibria can be implemented by
automata, but they note that a conundrum appears if one thinks of agents as trying both
to maximize the payoffs they receive in the repeated game and to minimize costs inherent
in constructing a more complex automata: agents will be tempted to leave off any states
that are not reached on the equilibrium path and this may render the kinds of off-path
threats necessary to support cooperation infeasible.
Fershtman and Kalai (1993) examine some IO questions using a similar complexity
notion. They consider a multimarket monopoly and single- and multimarket duopoly com-petition.
Implicit in the complexity notion is a diseconomy of multimarket operation. If
there are two possible states of demand in each market, then there are 2n possible states
for the n market combination, and a fully optimizing monopolist would therefore need a
plan with 2n contingencies. They note that monopolists may want to avoid entering some
markets to reduce complexity and discuss the complexity of dynamic collusion.
The models of Rubinstein (1993) and Piccione and Rubinstein (2003) discussed below
develop alternate approaches to complexity. Rubinstein (1998) provides a nice survey of
old and new work in this vein.
5 The New Behavioral IO: Bounded Rationality as Part of
the IO Toolbox
In the 1980’s it seems to have been extremely important that IO theory papers be fully
rational. Today, being less that fully rational is much more acceptable. Indeed, I’d guess
that at the top general interest journals an IO paper now has a better chance of being
14

accepted if it sells itself as being “behavioral” or “boundedly rational.”5 A bias in either
direction is unfortunate, but the opening of journals to boundedly rational papers is a very
good thing: the legitimization of boundedly rational modeling provides an opportunity to
explore new explanations for old phenomena and opens up whole new topics.
In this section and the one that follows I’ll discuss recent work in boundedly rational
industrial organization. The division into two sections reflects a divide I perceive in the
literature in the way in which authors approach their topics. The papers I discuss in this
section are topic-focused. Many grow out of existing rational IO literatures and look to
bounded rationality as a complementary way to think about some phenomenon. Others
discuss topics that had received little attention. In most cases, the papers have an easy time
making the case that less than fully rational behavior seems implausible and that bounded
rationality can provide additional insights.
I organize my discussion in this section by the IO topic being addressed.
5.1 Technology adoption
At least since Griliches’s (1957) work on hybrid corn and Mansfield’s (1968) work on in-dustrial
technologies there has been a recognition in industrial organization that the slow
pace of technology adoption may have significant welfare consequences. Sociologists often
explain slow s-shaped adoption patterns as resulting from heterogeneity in individuals atti-tudes
toward new technologies. Rational explanations based on long machine lifetimes are
also plausible for many industrial technologies. It is also clear, however, that information
is important – it is not easy to know what technogies really are cost-effective – and this is
an area where rationality is easy to critique. It is awkward to assume that consumers have
correct priors over the set of new products that might be invented and even more awkward
5One anecdote to this effect is that the Journal of Political Economy (which I’ve always thought of as the
ultimate proponent of rational explanations) recently explained that it was rejecting a paper not because
of anything the referees said but because the editor’s “gut feeling is that most of the time, forgetfulness,
confusion, or other types of bounded rationality play a much more crucial role.” I would hypothesize that
the fact that IO is a relatively small field makes it more likely to be subject to such whims than other
fields. The top general interest journals always have a macroeconomist coeditor to handle the many macro
submissions, but usually get by with someone outside the field handling IO submissions. People outside the
field will have a harder time selecting the papers that are most important in the field, and may be more
likely to follow popular trends.
15

to assume that they fully understand the social processes by which some products achieve
popularity. Ellison and Fudenberg (1993) builds on the Smallwood-Conlisk approach to
explore whether dispersed information about product quality will be aggregated at the
population level and come to be reflected in adoption decisions.
We identify two mechanisms by which such “social learning” can occur. In one informa-tion
is aggregated via agents paying attention to popularity. In the other it is aggregated
by geography. In both models, there are two technologies f and g. The payoffs are random
variables related by ug
t − uf
t = + t with t uniformly distributed on [−, ]. Boundedly
rational players observe ug
t −uf
t , but don’t know the full history of payoff realizations. The
popularity model looks at a homogeneous population in which there is inertia and in which
players who do switch are more likely to use more popular actions: players choose g if
ug
t − uf
t k(mt(f) − mt(g)), where k is a positive constant and mt(f) is the market share
of technology f at time t. In this model, we show that with no popularity weights (k = 0)
both technologies get positive shares in the long run. With k there can be herding on
an inferior technology. With k = full learning occurs in the long run and all consumers
eventually adopt the superior technology.
In the geographic model, players are arrayed on a line (which could be a physical
location or a position in some social network or taste space). There is no longer a single
best technology – which technology is superior depends on the location. Players observe the
technology choices of those in a neighborhood around them, and also observe the average
payoff of any technology used in their neighborhood. In this model, we show that social
learning occurs via a geographic mechanism: the dividing point between the f-adopters
and the g-adopters shifts over time and a law-of-large-numbers mechanism keeps it close to
the optimum when the observation neighborhood is small. There is a tradeoff between the
speed of adoption and its long-term efficiency.
Ellison and Fudenberg (1995) consider a model closely related to the nonspatial model
described above. There are two primary differences: there is a random idiosyncratic com-ponent
to the payoff each agent receives in each period; and we consider rules-of-thumb
in which an agent asks k randomly selected members of the population about the payoffs
16

they received in the current period and chooses the technology that provided the highest
payoff in this random sample (with the proviso that an agent never adopts a technology
that was used by no one in his sample.) We focus on the question of how the structure of
information flows affects the learning process.
Considering first the case where the two technologies are equally good, we show that
the value of k affects whether we see “herding” on one technology or “diversity” with both
technologies retaining a positive market share in the long run. Intuitively, when agents rely
on small samples, they are unlikely to hear about unpopular technologies, which makes it
more likely that these technologies will die out. We then introduce differences in average
payoffs across the technologies, and show that long-run efficient social learning will occur
for a range of values of the sample size k. When k is smaller herding on an inefficient
technology may be possible, and when k is larger both technologies may survive in the
long run. Intuitively, there is a degree of popularity-weighting implicit in the assumption
that agents don’t adopt technologies not being used in their random sample, and this can
support social learning as in the earlier model.
Spiegler (2004) introduces price-setting into a related model. There are N firms. Each
firm has a distinct (albeit equally good) technology that provides random payoffs that
are independent and identically distributed both over time and across individuals: each
technology gives a payoff of one with probability and zero with probability 1 − . There
is also an N+1st technology that provides this same payoff at no charge. (Spiegler discusses
unscientific medical practices as a potential application and refers to the firms as quacks.)
Consumers are assumed to use a rule-of-thumb that involves finding one person using
each technology, treating the payoff, vi, that person received as if it were the expected
payoff of technology i, and choosing the technology for which vi − pi is maximized. This is
similar to the decision rule of Ellison and Fudenberg (1995), with the main difference being
that each technology is sampled once rather than a popularity-influenced random number
of times. Spiegler notes that the rule-of-thumb can also be motivated as an extreme form
of Rabin’s (2002) model of “believers in the law of small numbers” or as an application of
a solution concept in Osborne and Rubinstein (1998).
17

The pricing game between the firms turns out to have an unique mixed strategy Nash
equilibrium (not unlike the equilibria one sees in search-based models of price dispersion).
In this equilibrium, expected prices are inversely related to the common product quality.
Industry profits are nonmonotone in the number of firms. Industry profits initially increase
in N because there is more likelihood that consumers’ samples will contain at least one
firm success, but decline as N becomes large because the price competition becomes more
intense.
5.2 Models of sales
The classic models of sales by Varian (1980) and Sobel (1984) are described by their au-thors
as rational models. The papers could also have been sold as boundedly rational or
behavioral: Varian’s model features some ‘high search cost’ consumers who are ignorant of
the prices each store offers and end up going to one store chosen at random; and Sobel’s
features some infinitely impatient consumers who buy immediately regardless of the poten-tial
gain from waiting for a good to go on sale. In each case, a very rough intuition for
why there are sales is that firms want to price discriminate and give discounts to the more
sophisticated shoppers.
Rubinstein (1993) and Piccione and Rubinstein (2003) develop formal approaches to
modeling cognitive abilities and use these frameworks to provide models of sales in which
the discounts-for-sophisticated-consumers intuition is more formally grounded. In Rubin-stein
(1993), cognitive complexity is captured by the order of the “perceptron” needed to
implement a strategy. Some agents can only implement very simple strategies (buying if
price is above a threshold), whereas others can implement nonmonotone strategies involv-ing
two or more cutoffs. He writes down a model in which a monopolist wants to charge
high-cognitive-ability agents a lower price in some states, and in which the monopolist can
achieve this by randomly choosing prices in a manner that makes it unprofitable for low
cognitive-ability customers to also buy in this state. Piccione and Rubinstein (2003) intro-duce
an alternate form of differential cognitive ability: they assume that agents differ in the
length m of the price history they can recall. They again consider an environment in which
18

the firm would like to charge high-cognitive-ability agents a lower price, and show how this
can be done by alternating between regular and sale prices in a manner that high-ability
agents can recognize (letting them buy only when the item is on sale) and low ability agents
cannot (forcing them to pay the time-average price).
Kahnemann, Knetsch, and Thaler (1986) had much earlier proposed that sales might
have a different irrational origin. They conduct a survey, and find that many subjects say
that it is “unfair” for firms to raise prices when demand goes up and speculate that this may
give firms an incentive to hold sales rather than reducing “regular” prices: if firms lower
regular prices when demand is low, they will be branded as unfair if they raise prices back
to normal when demand returns to normal. Rotemberg (2005) proposes a more complex
fairness-based model to account both for firms’ occasional use of sales and for the stickiness
of prices within and across non-sale periods. In his model, consumers have reciprocal-altruism
preferences and punish firms discontinously if their estimate of the firm’s altruism
crosses a threshold. The model also relies on the firms’ objective function being a concave
function of profits and on consumers feeling regret. He argues that similar assumptions can
also help explain how firms adjust prices in response to demand shocks and inflation.
5.3 Price dispersion
The IO literature developed a number of search-cost based explanations for price discrimi-nation
in the early 1980s.6 These models are compelling for many real-world applications,
but Baye and Morgan (2004) note that dispersion seems to occur even in enviroments where
it is less plausible that substantial search costs exist: it is common for products listed for
sale on Internet price search engines; and it even occurs in laboratory experiments in which
a computer plays the buyer’s role to perfectly recreate a Bertrand duopoly.7
Baye and Morgan propose that one reason why this occurs may be that the equilibrium
of the Bertrand competition game is somewhat nonrobust to departures from rationality.
Suppose that firms are -optimizers rather than being fully rational, i.e. they may pick
any action that earns profits that are within of the maximum possible profit. One might
6Baye and Morgan (2004) provide an excellent survey.
7See Dufwenberg and Gneezy (2000).
19

think at first that the Bertrand model is fairly robust to such a change because in any
pure strategy -equilibrium industry profits would be at most . Baye and Morgan note,
however, the Bertrand game also has mixed strategy -equilibria with much higher profits.
For example, for any x 2 (0, m), suppose firms use the mixed strategy with CDF
Fx(p) =
8
:
0 if p p
1 − x/pD(p) if p 2 [p, pm)
1 if p pm
where p is such that pD(p) = x. In the Bertrand duopoly game with zero costs, the
maximum gain from deviating from this profile is x2/2m. If we set this to , we find that
p
a firm can earn at least
2m − in an -equilibrium. This is a strikingly large number
for reasonable values of . For example, if is one percent of the monopoly profit, the
aggregate industry-profits in and -equilibrium can be approximately 26% of the monopoly
profit.8
5.4 Obfuscation
One of the most basic results on the economics of information disclosure, is that firms will
disclose all relevant information to consumers if the information is costless to disclose and
disclosures are credible (either because the facts are evident or because truth-in-advertising
laws provide a guarantee)(Grossman 1981; Milgrom 1981). The simple intuition is that the
firm with the highest possible quality will always want to disclose its information to increase
consumers’ willingness to pay. Given that the best-news firm is disclosing its information,
the firm with the next best news will also gain from separating from any firms that are
pooling on nondisclosure, and so on.
Ellison and Ellison (2005) note that there are big differences between this theoretical
prediction and what we see in many real-world environments. For example, mattress man-ufacturers
put different model names on products sold through different stores and provide
sufficiently few technical specifications so as to make it very difficult to compare prices
across stores. Credit cards are another good example: it is also hard to imagine that the
8Note, however, that the play is -suboptimal only in an ex ante sense. When the realization of the
p
) inefficient in an ex post sense.
randomization calls for setting p = pm, play is O(
20

complex fee schedules in small print on the back of credit card offers could not be made
simpler. Our empirical work examines a group of small retailers selling computer parts over
the Internet via the PriceWatch search engine. The most basic observation of the paper is
that many of the firms have clearly adopted pricing practices that make it time-consuming
and difficult for consumers to understand what exactly a firm is offering and at what price.
For example, products are described incompletely and consumers may have to go through
many pages to learn all of nonstandard aspects of the offer: the restocking fee that will
apply if the product is returned, how much extra the consumer will need to pay for a
warranty, etc. A second interesting empirical observation is that retailers appear to obtain
substantial markups over marginal cost even though there is easy entry, minimal product
differentiation, and minimal inherent search costs.
The combination of these two facts – there appears to be a great deal of obfuscation;
and obfuscation appears to affect markups – makes obfuscation a natural topic for IO
economists to study. The emphasis of the current IO literature on product differentiation
as the source of markups may well be fully justified empirically, but another reason for the
absence of any mention of obfuscation in Tirole (1988) may be that it would have been an
awkward topic to discuss in the rational paradigm.
The most straightforward way to think about obfuscation using standard IO tools would
be to regard it as increasing search costs in a model with costly search like Stahl (1989).
One would want, however, to extend the standard models to allow the search costs to vary
by firm (instead of just across consumers). Even then, we would have a fairly black-box
model. A nicer bounded rationality augmentation would be to derive search costs as a
special case of costly computation.
More direct bounded rationality approaches are also appealing. Gabaix and Laibson
(2004) suggest a very simple formalization: one could regard obfuscation as increasing
the variance of the random evaluation error i in a model in which consumers have noisy
estimates of the utility they will receive from consuming a product: they think they will
get utility i +i from consuming product i when they actually get utility i. Such a model
is formally equivalent (from the firm’s perspective) to a model in which firms can invest in
21

product differentiation. Firms will invest in obfuscation just as they invest in differention
to raise markups. The equilibrium markups for a given error distribution can be derived
as in Perloff and Salop (1985).9 Gabaix and Laibson derive new results on the asymptotic
rate at which markups decline as the number of firms increases, and emphasize that for
some error distributions adding firms to the market will drive down prices very slowly.
Spiegler (2006) discusses another rule-of-thumb model. In his model, products inher-ently
have a large number of dimensions. Boundedly rational consumers evaluate products
on one randomly chosen dimension and buy the product that scores most highly on this
dimension. In this model, consumers would evaluate the products correctly if products were
designed to be equally good on all dimensions. Spiegler shows that this will not happen,
however. Essentially, firms randomize across dimensions making the product very good on
some dimensions and not so good on others. He thinks of this cross-dimensional variation
as intentional obfuscation. The comparative statics of the model may help us understand
why there is more obfuscation in some markets than in others: obfuscation increases when
there are more firms in the market, and when the outside option is more attractive.
5.5 Add-on pricing
In many industries, it is common to sell high-priced add-ons. Hotels charge high prices
for dry cleaning, telephone calls, minibar items, and restaurant meals. Credit cards have
high late-payment fees. Upgrading to a larger hard drive and adding more memory adds
substantially to the advertised price of a Dell computer.
Ellison (2005) notes that we can think of high add-on prices in a couple ways. One is
that we may simply be seeing the outcome of a standard multi-good price discrimination
model. Suppose that consumers who are more price-sensitive when choosing between firms
(“cheapskates”) also have a lower willingness to pay for quality-improving add-on.10 Then,
9Perloff and Salop is an exception to my earlier remark about departures from rationality in 1980’s IO
papers being unintended or unwanted. They have a brief section in the back of their paper noting that
the ’s in their model could capture “ ‘spurious’ as well as actual product differentiation”. They mention
Chamberlin and Galbraith as among those who had previously discussed spurious differentiation and point
to blind taste tests as providing empirical support for its importance.
10I would argue that this assumption is the most natural one to make. Note, however, that it does
differ from what is assumed in most papers discussed in Armstrong (2006) and Stole (2005). The standard
assumption has been that brand preferences and quality preferences are independent.
22

in a standard imperfect competition/price discrimination model in which firms simulta-neously
announce prices for low- and high-quality goods and all consumers are perfectly
informed and rational, firms will charge higher markups on higher quality goods.
A second way to think about high prices for add-ons is as in Lal and Matutes (1994). If
add-on prices are not observed by consumers when choosing between firms, then add-ons
will always be priced at the ex post monopoly price.11 In such a model, of course, the profits
earned on the add-on may be competed away in the form of lower base-good price. Lal and
Matutes’s “loss leader” model is an example in which the price cut on base good exactly
offsets the profits earned on the add-on. The second main observation of Ellison (2005)
is that this no longer holds in models with cheapskates. In such an environment, firms
that rely on sales of add-ons for a large part of their profits face a severe adverse selection
problem: price cuts disproportionately attract cheapskates who don’t buy add-ons. This
adverse selection problem discourages price-cutting. As a result, the joint adoption of
add-on practices raises equlibrium profits.
An interesting question raised by the profitability of the joint adoption of add-on pricing
in this second conceptualization is whether it is individually rational for firms to adopt add-on
pricing. Typically, consumers would be able to buy a more efficient bundle if add-ons
were priced closer to cost. If firms could commit to setting such prices and costlessly inform
consumers via advertising, then firms in a rational-consumer model would typically want
to do this.12 Hence, advertising costs, tacit collusion, or some other modification is needed
to account for why add-on pricing is practiced.13
Gabaix and Laibson (2005) develop a boundedly rational explanation for why add-on
prices often are not advertised. Proceeding along lines suggested in Ellison (2005), they
develop a model with two types of consumers: rational consumers and irrational consumers.
11Lal and Matutes explanation is a version of Diamond’s (1971) classic argument about search costs
leading to monopoly pricing. It can also be thought of as a version of Klemperer’s (1987) switching cost
model: one can think of the add-on being sold at a later date, with the inability of consumers to see add-on
prices being equivalent to assuming the firm can’t commit to future prices.
12See, for example, Shapiro (1995).
13Note that this only applies to the second model of why add-ons are expensive, not to the first. The
price discrimination model is a standard pricing game where add-on prices are high when optimally set and
observed by all consumers.
23

In such an environment, the profits earned serving rational consumers are increased by
advertising a higher base-good price and a lower add-on price. Profits obtained by serving
irrational consumers are reduced, however. They describe this reason why this occurs as
the “curse of debiasing”. Irrational consumers are assumed to be unaware of the add-on’s
existence unless they see an advertisement listing a price for it. Once informed, they realize
that they can take steps to avoid purchasing the add-on. This has the effect of reducing
their willingness to pay for the advertising firm’s product. If the fraction of irrational
consumers is sufficiently high, then the losses on them can outweigh the gains on the
rational consumers.
Papers on oligopoly behavior-based price discrimination, e.g. Chen (1997) and Fuden-berg
and Tirole (2000), consider a similar problem. They also, however, allow consumers
to buy the add-on from the firm that did not supply them with the base good (although
this is inefficient). In models without commitment, add-on prices are determined by the
second-period oligopoly competition. As Armstrong (2006) notes, this would make the ef-fect
of naive consumers very different. The fact that consumers don’t think about second
period prices until the second period won’t make second period prices any higher, and would
actually lower first-period prices by eliminating the effect that makes first-period behavior
less sensitive to price differences. See Fudenberg and Villas-Boas (2005) for an insightful
discussion of the mechanics of these models.
The paper of Spiegler (2006) I mentioned earlier can also be thought of as providing
an explanation for add-on prices. One can think of the differences in product quality on
different dimensions as representing both pure quality differences and differences in add-on
fees charged in various circumstances.
5.6 Performance standards
How well an individual or firm has performed (or will perform) typically has a multidimen-sional
answer. Performance assessments, however, often ultimately result in a zero or one
decision. Should a student be admitted to this college? Should a student be given an A or
a B? Should a job applicant be hired? Should a paper be published? Should a professor
24

be granted tenure? These decisions establish performance standards. Two questions about
standards are of primary interest: how will multiple dimensions be weighted and how high
will the overall hurdle be?
In some applications, one could try to develop fully rational models of performance
standards, e.g. a firm should hire a worker if and only if the firm’s profits are higher with the
worker than without. In many other applications, no clear objective function is available,
e.g. are colleges supposed to admit the students who will perform best in classes, those who
will add most to campus life, those who will benefit most from the education, those who
will go on to accomplish great things, or those who will eventually donate the most money
to the college? Even when there is a clear objective, I would argue that the fully rational
model is usually implausible because judges will typically lack the understanding necessary
to maximize the objective, e.g. a firm will not know how a student’s GPA in math classes
and her experience at the school newspaper determine her incremental contribution to the
firm’s profits. As a result, the way that many such decisions are made is to compare the
current candidate with past candidates, and to use relative performance to judge whether
to admit the candidate.
Sobel (2000, 2001) uses a rule-of-thumb approach to explore how various factors affect
whether standards tend to rise or fall. In these papers, candidates are assumed to choose
a level of costly effort and to direct this effort to achieve a multidimensional vector of
accomplishments. In the two-dimensional case, for example, the candidate’s achievements
are described by a pair (q, r). Candidates are then judged relative to the pool of recent
successful candidates.
In Sobel (2000) the way the judging is done is that a single judge aggregates the multiple
dimensions using a function v(q, r), and deems the tth candidate successful if v(qt, rt) zt.
It is assumed that the standard zt is set so that a fraction of the previously successful
candidates would meet the standard. When the function v is held fixed and there is no
randomness in the achievement process, this model predicts a bunching of performance.
Starting from any initial condition, candidates who are able to do so exert exactly the
effort necessary to achieve the 1 − th percentile level of performance, and eventually the
25

entire pool of recent sucessful candidates has achieved exactly this performance level. The
most interesting results concern what happens when weighting functions change over time,
e.g. shifting from v to v0 and back to v again. Such shifts cause standards to decline.
Intuitively, when standards shift from v to v0, the performance level of the earlier candidates
(measured in terms of v0) can be achieved at lower cost by tailoring effort to v0. Hence,
equilibrium effort declines. When the standards shift back to v, candidates can again
achieve a performance that grades as highly as the performance of the candidates who
had worked toward v0 with less effort. Comparing the first and third cohorts of agents,
performance will have unambiguously declined.
Sobel (2001) considers institutions involving voting by a panel of judges. Each judge
has a different v function. As in the earlier paper, judges vote to acceept a candidate if
the candidate’s performance is in the top of recent successful candidates. What happens
in this model depends both on the voting rules and the dimensionality of the performance
space. When votes from only a small fraction of the judges are sufficient for admission,
standards decline. When near unanimity is required, standards increase over time. In
intermediate cases the results are more nuanced.
Ellison (2002) focuses on the weighting of different dimensions rather than the overall
level of the standard. It is motivated by changes in the standards of academic journals
over the last thirty years. There are several clear trends in economics (and other fields)
over that period: papers have been getting longer; they have longer introductions, more
references and discuss more extensions; authors are required to carry out more extensive
revisions before papers are published. I argue that all of these changes can be thought of
as reflecting a shift in quality standards. Think of an academic paper as having a two-dimensional
quality (q, r), where q reflects the importance of a paper’s main contributions
and r reflects other dimensions that can be improved with incremental effort, e.g. improving
the exposition, generalizing and extending results, and performing robustness tests. The
observed trends can be thought of as caused by or reflecting an increased weight being
placed on r-quality, i.e. if articles are judged acceptable if v(q, r) = q+(1−)r z, then
we may be seeing a decrease in .
26

One way to account for such a trend would be to use the comparative statics of a rational
model in which is chosen to optimize some social welfare function. One can compute how
the distribution of paper qualities is affected by in equilibrium, and one could argue that
the distribution produced by a lower is now preferable because of some exogenous change
in the profession. Such changes are hard to identify, however, and it is also hard to find
any evidence that changes in standards were intentional. This motivates the search for
explanations in which the shift is an unconscious byproduct of rule-of-thumb behavior.
My model focuses on the behavior of referees. Referees are assumed to try to faithfully
apply the profession’s standards (rather than imposing their own preferences over quality
dimensions). They do not, however, inherently know what values of (, z) to apply. They
try to learn these by seeing what parameters best fit the data they get in the form of seeing
decisions made on papers they refereed, seeing papers in journals, and seeing decisions on
their own papers. In each period authors work to maximize their chance of acceptance
given their current understanding of the norm, referees apply the standard they currently
believe in, editors accept the fraction of papers that referees rate most highly, and beliefs
are updated given the acceptance decisions.
When referees are unbiased, this model has a continuum of equlibria. Given any weight
, there is a corresponding achievement level z() that yields an equilibrium. I then
introduce a small behavioral perturbation, combining the rule-of-thumb and psychology-and-
economics approaches. Players are assumed to be overconfident about the quality of
their own work and think that it is higher in quality than it truly is. This destabilizes
the former equilibria: players will be puzzled about why it is that their papers are being
rejected. The first effect that this would have on players is to make them think that z must
be higher than they had thought. This hypothesis, however, also cannot fully explain the
data: it cannot account for why marginal papers by others are being accepted. The result
of players’ continuing struggles to reconcile inherently contradictory data is a slow, gradual
evolution of beliefs about the social norm in the direction of decreasing . At the most
abstract level, the idea of the paper is that one way to explain a long trend is to perturb
a model with a continuum of equilibria. A slight perturbation of such a model can have a
27

unique equilibrium and the disequilibrium dynamics can feature a slow evolution along the
near equilibrium set.
6 The New Behavioral IO: Biases from Psychology and Eco-nomics
The recent surge of interest in bounded rationality in industrial organization comes on the
heels of a bigger and slightly less recent surge of interest in psychology and economics. This
literature has by now documented a plethora of ways in which real consumers depart from
the rational self-interested ideal: they discount in a nonexponential manner, exhibit loss
aversion; care about fairness; have self-serving biases; fail to update in a fully Bayesian
manner, etc. More importantly, it has developed a number of simple models that can be
adopted as representations of agents subject to such biases. Exactly how portable such
models are is subject to debate, but at least in principle one can construct the behavioral-bias
counterpart of a given rational model by replacing the utility maximization assumption
with the assumptions of one’s favorite representation of consumers subject to this behavioral
bias.
The initial papers in this branch of the behavioral IO literature have tended to focus
on how firms will choose prices and product characteristics to exploit behavioral biases and
whether competition will eliminate the exploitation. Combining the IO and psychology and
economics literatures, however, naturally gives many more than just one paper topic per
bias – we can get a whole matrix of paper topics. Think of the set of behavioral biases as the
column headings, and put all of the standard models in IO as the row headings: how will a
monopolist price, how will a monopolist selling durable goods price; how will a monopolist
price discriminate; how will oligopolists selling differentiated goods set prices; how will
some action be distorted to deter or accomodate entry, etc. It takes little knowledge or
imagination to come up with literally thousands of paper topics: Tirole’s (1988) text has
hundreds of IO models, each of which could be combined with dozens of behavioral-bias
models.
28

Hyperbolic Loss Self- Imperfect
Discounting Aversion Fairness Serving Bayesian
Monopoly pricing DellaVigna Heidhues-
Malmendier Koszegi Rotemberg
Price discrimination
Durable goods
Static oligopoly
Dynamic oligopoly
Entry deterrence
Innovation
Will this lead to thousands of behavioral IO papers in the next few years? I am pretty
sure (and hope) the answer is no. The problem is that most combinations will not produce
observations that are sufficiently deep or interesting to warrant their being published. For
example, Hossain and Morgan (2006) conduct field experiments on eBay and find that
auctioning goods with a high (but not too high) shipping charge raises more revenue than
using an equivalent minimum bid and making shipping free. It’s a striking fact and they’ve
written a very nice psychology and economics paper around it that discusses how the
results can be explained by mental accounts with loss aversion (or other explanations). The
companion how-should-firms-price paper, in contrast, is too obvious to write: if consumers
behave this way then firms should use high shipping charges.
I would argue, nonetheless, that most boxes in this matrix are worth exploring. Without
doing so, it is hard to know whether they will yield interesting alternate explanations
for previously noted phenomena, provide potential explanations for facts about particular
industries that are hard to square with existing models, or help us think about when
consumer irrationality does and doesnt matter.
In this section I’ll discuss two behavioral IO papers that build on the psychology and
economics literature. Each belongs to the monopoly pricing and bias X row of my matrix.
I will then discuss some general lessons.
29

6.1 Monopoly pricing with hyperbolic discounting
Della Vigna and Malmendier’s (2004, 2005) work on selling goods with delayed benefits (or
delayed costs) to time-inconsistent consumers nicely exhibits the potential of behavioral IO.
Their motivating example is pricing at health clubs: they think of consumers as incurring
a short-run disutility when visiting a club, and enjoying a delayed reward in the form of
better health.
The model uses a theoretically appealing (and practically relevant) degree of generality
in pricing. Consumers are initially offered a two-part tariff with an upfront payment of
L and an additional per visit charge of p. If consumers accept this offer, they learn the
disutility d that they will incur if they visit the club, and then decide whether to visit
(which costs p and gives a delayed benefit b.) Note that the decision to accept the two-part
tariff is made under symmetric information, so in a rational model the health club would
extract all consumer surplus via the fixed fee.
Consumers are assumed to have (

, ) quasi-hyperbolic discounting preferences: from
the perspective of period 0, payoffs in period t are discounted by

t. They consider both
naive hyperbolic consumers who don’t realize that they have a commitment problem, and
sophisticated hyperbolic consumers who are fully aware of their time inconsistency.
In this model, one can think of two reasons why a health club will want to distort p
away from marginal cost. First, sophisticated rational consumers would like to commit
themselves to go to the health club more often. The health club can help them to do this
by setting p below cost. Second, naive rational consumers will overestimate the number of
times that they will go to the club. Reducing p and increasing L widens the gap between
the surplus that the consumer expects to receive from accepting the contract and what the
consumer actually receives. Hence, distorting the contract in this way makes the contract
more appealing to these consumers. The two effects go in the same direction, so regardless
of what type of hyperbolic consumers we have, we reach the same conclusion: visits will be
priced below marginal cost.
Is this distortion bad? In the case of sophisticated consumers it is not. Essentially,
one can view the health club as selling a commitment device that allows the consumer to
30

overcome his or her self-control problem and achieve the first-best. In the case of naive
hyperbolic consumers things are less clear cut. The low per-visit price does get consumers
to visit the gym more often than they would with marginal cost pricing. Consumers’
misperception of the value they will get from the contract, however, leads to their receiving
negative total surplus.
One notable limitation of Della Vigna and Malmendier’s work is that consumers are
assumed to be ex ante homogeneous. This makes the paper less complete as an applied
model – it can’t explain why real-world health clubs offer a menu of contracts – and also
eliminates any possible insights that might come from analyzing conflicts between rent
extraction and surplus maximization.
Della Vigna and Malmendier’s (2005) empirical study lends credence to the view that
time-inconsistency is an important factor in health-club pricing. The health clubs they
surveyed offer contracts in which the per-visit charge is set below marginal cost, and most
consumers take them. In fact, they take these two-part tariffs even though the clubs also
offer contracts with a higher per-visit charge that would be cheaper for most consumers.
Clearly, people either must value the commitment or must overestimate their future usage.
Survey data suggest that overestimation is at least part of the story.
6.2 Monopoly pricing with loss aversion
Heidhues and Koszegi (2004) explore pricing to consumers who experience loss aversion.14
The formal specification of consumer behavior follows Koszegi-Rabin (2005). Consumers
have reference-dependent utility functions that depend not only on the number of units
x of the good they consume and on the amount of money m they have left over, but
also on reference levels rx and rm which are their ex ante expectations: ut(x,m; rx, rm) =
xvt + m + μ(x − rx) + μ(m − rm). Loss aversion comes in via the function μ, which is
assumed to be have a kink at the origin. Note that consumers feel a loss both if they are
unable to buy a product they expected to buy and if they pay more than they expected to
pay.
14Schlag (2004) is another paper examining pricing to boundedly rational consumers. His consumers
follow a regret-minimization strategy.
31

Heidhues and Koszegi consider a world in which demand and cost are both random and
time-varying. In the base model, firms are assumed to be able to commit to a distribution
of prices. Some interesting effects arise because this commitment allows the firm to manage
consumer expectations. One observation is that prices will be sticky and firms may set a
constant price (or choose prices from a discrete set) even when cost shocks are continuously
distributed. The disutility that loss-averse consumers feel when they pay more than they
were expecting is greater than the utility they derive from paying symmetrically less than
they were expecting. This provides an incentive to keep prices constant. A related observa-tion
is that markups are countercyclical. Another fundamental conclusion is that there is
substantial scope for indeterminacy: as in Koszegi-Rabin (2005) the purchase decision con-tingent
on the price is not necessarily unique because the consumer has a higher valuation
when he expects to buy, and even apart from this there can be multiple pricing equilibria
when the firm lacks commitment power.
6.3 Some thoughts
In this section I discuss some basic principles relating to how firms “distort” pricing to
exploit behavioral biases and some challenges that the psychology and economics motivated
literature will face going forward.
6.3.1 Will monopolists exploit biases?
Consider a standard quality-selection problem as in Spence (1975). Suppose that there is
a population of consumers with unit demands. A consumer of type is willing to pay at
most v(s; ) to obtain one unit of a good of quality s. Assume is uniformly distributed
on [0, 1] and that @v/@ 0 so that if q units are sold then the consumers will be those
with 2 [0, q].
Suppose that a monopolist that must choose a single quality level s for its product.
Assume that the good is produced under constant returns to scale, with a quality s good
having a unit production cost c(s). The monopolist’s problem is
max
q,s,p
q(p − c(s))
32

s.t. v(s; ) − p 0 for all 2 [0, q]
Conditional on the quantity sold being q and the quality being s, the highest price that can
be charged is v(s; q), so this reduces to
max
q,s
q(v(s; q) − c(s))
The first-order condition for the monopolist’s quality choice sm is
@c
@s
(sm) = @v
@s
(sm, qm).
The marginal cost of providing higher quality is equal to the marginal benefit for the
marginal consumer.
The social planner’s problem would be
max
q,s
Z q
=0
v(s; )d − qc(s).
The first order condition for the first-best quality choice sFB is
@c
@s
(sFB) =
1
qFB
Z qFB
=0
@v
@s
(sFB, )d.
The marginal cost of providing higher quality is equal to the marginal benefit for the average
consumer.
Spence emphasized that monopolists typically will not provide optimal quality: the
marginal and average consumer can have different valuations for quality; and the pool of
customers served by the monopolist differs from the pool the social planner would want to
serve. One case in which the monopolist’s quality choice is optimal, however, is when the
population is homogeneous. In this case, both the monopolist and the social planner will
serve all consumers and there is no difference between the marginal and average consumer.
Hence, the monopolist provides optimal quality.
The quality-selection model is relevant to the question of how firms respond to behav-ioral
biases because one can regard many decisions a monopolist makes as quality choices.
For example, in Della Vigna and Malmendier’s health-club application, the per-visit charge
a health club imposes is a product characteristic that determines the quality of a club
membership. A contract with a lower per-visit fee is a higher quality product.
33

The optimal quality result of Spence’s model implies that a monopolist facing an ex
ante homogenous unit-demand customer population will choose to sell the product s that
maximizes v(s)−c(s). When I developed the model I said v(s; ) was the willingness to pay
of the type consumer. In a rational model this is the amount of money that offsets the
utility the consumer receives from consuming the product. Spences result applies equally
well in irrational models, however. One just has to keep in mind that the appropriate v(s)
is the willingness to pay for a quality s good at the moment at which the consumer makes
the purchase decision.
In Della Vigna and Malmendier’s sophisticated hyperbolic model, just like in the rational
model, the willingness to pay of the time-zero agent is the expected utility that the time-zero
consumer receives (in equilibrium) if she signs up for the health club. Hence, the outcome
is that the health club chooses the contract that is optimal for the time-zero agent. In a
naive hyperbolic model, willingness to pay is the time-zero agent’s (incorrect) forecast of his
equilibrium utility. Hence, the contract is also designed to increase the difference between
the forecast utility and the true utility.
The same basic intuition will apply to other behavioral biases: monopolists will distort
product characteristics along whatever dimensions increase irrational consumers’ willingness-to-
pay. What these distortions are under alternate behavioral specifications, of course, still
needs to be worked out. What makes Della Vigna and Malmendier (2004) successful is not
that it shows that a monopolist will exploit boundedly rational consumers, but rather that
the exploitation is a compelling explanation for something we see in the world (low per visit
fees). A challenge for future authors is to find other applications where exploitation takes
an interesting form.
6.3.2 Does competition eliminate exploitation?
Another question that behavioral economists have raised is whether competition eliminates
the exploitation of behavioral biases. The question can be asked either in the imperfect
competition models that are common in industrial organization or in a perfect competition
framework.
34

In the homogeneous population case, the answer is obviously no. Competitive firms
will behave just like monopolists and make product-design choices s so as to maximize
v(s) − c(s). The only difference is that the surplus will be returned to consumers in the
form of lower fixed fees.
In heterogeneous population models, the precise choices of monopolists and competitive
firms will differ. How this will work depends on whether firms are or are not assumed to
be able to offer multiple quality levels.
For some applications one would want to suppose that only a single quality level will
be produced. The monopolist will design its product to appeal to its marginal consumer.
Firms engaged in imperfect competition a la Hotelling will design their products to appeal
to the consumer who is on the margin between the two firms. Although the monopoly and
competitive product designs will be different because the marginal consumers are different,
one imagines that qualitative conclusions about the direction in which product designs are
distorted will usually be similar.
For other applications it will be better to assume that firms can introduce multiple
products at different quality levels and use them to price discriminate. In a monopoly model,
the type with the highest willingness to pay is typically sold a product of the “optimal”
quality, which here means maximizing v(s; ) − c(s). Other types will be sold products of
quality lower than that which maximizes v(s; ) − c(s). Whether this adds to or offsets
the distortion that comes from exploiting the difference between v and utility will depend
on the application. In competitive price discrimination models, the pattern of quality
distortions depends on the joint distribution of preferences for quality and preferences for
firm 1 versus firm 2. Hence, comparisons with the monopoly case will depend on how
preference heterogeneity is specified.
7 Concluding Remarks
In this essay I’ve tried to give some perspective on the burgeoning field of boundedly rational
industrial organization. It’s a hard literature to summarize because one can depart from
rationality in so many ways. The irrational actors can be the firms or the consumers.
35

Several approaches to modeling irrationality can be taken: one can specify rules of thumb,
behavior can be derived as the maximizer of something other than utility/profits; or one can
explicitly introduce cognitive bounds. I’ve stressed two advantages of boundedly rational
approaches: boundedly rational behavior seems more realistic in some applications; and the
tractability of boundedly rational models can sometimes allow researchers to incorporate
additional features into a model. For the next few years at least there is probably also a third
important advantage: after twenty-five years of focusing on rational models, the questions
rational models are best suited to address have been much more thoroughly explored than
questions best addressed in boundedly rational models.
I’ve noted that the current literature on boundedly rational industrial organizaiton
seems to be proceeding along two different branches. One easy way to give advice for the
literature is to suggest that each branch adopt attractive features of the other. In many
papers in the IO branch, the particular form of the departures from rationality is motivated
by little more than the author’s intuition that the departure is plausible. More reliance
on empirical and experimental evidence could be a significant improvement. A relative
weakness of the psychology and economics branch is that the papers are less informed by
the existing IO literature on the topic and don’t live up to existing IO norms in terms of
incorporating consumer heterogeneity, imperfect competition, price discrimination, etc.
Which of the branches of boundedly rational IO is likely to be most successful in the
coming years? Let me say first that I think in the long run the two branches will come
together. Until they do, I think that having both is valuable. My answer to the question of
which will be more successful in the next few years depends on the meaning of successful. If
successful means advancing the IO literature I think the answer is the first branch. Delving
into the IO literature and focusing on topics where the existing explanations are lacking
is the most efficient way to improve a literature. If success is measured by publications or
citations, I think the answer is the second branch. I noted earlier that an attractive feature
of the psychology and economics literature is that the models developed in papers like
Laibson (1997), Fehr and Schmidt (1999), and Eyster and Rabin (2005) are mechanically
very much like rationality and can be grafted into just about any model. This creates many
36

Bounded rationality in organisation

More Related Content

What's hot (14)

Viewers also liked (19)

Similar to Bounded rationality in organisation (20)

Recently uploaded (20)

Bounded rationality in organisation