4. Modeling Human Actors
It is not possible yet to point to a single theory of human behavior that has
been successfully formulated and tested in a variety of settings.
—Elinor Ostrom
In this chapter, we address a question that lies at the core of this book: How
do we model people? In many of the models that follow, people will be the
fundamental unit of analysis. We will construct models of people who vote,
cooperate, start uprisings, participate in fads, invest in retirement accounts,
and become addicted to drugs. Within each model, we will have to make
assumptions about people. What are their objectives? Do they only care
about themselves or are they altruistic? What are their potential actions?
How do they choose what action to take, or do they not even have a choice?
We could make ad hoc assumptions for each model. But to do so would
invite confusion and miss an opportunity. We would be left with an
idiosyncratic set of constructions. Each new model would require new
thinking about how people act. The resulting heterogeneity would limit our
ability to think across and combine models. We could not be effective
many-model thinkers.
The approach we follow stresses coherence along with variety. We will
model people as either rule-based actors or rational actors. Within the set
of rule-based actors, we consider those who act based on simple fixed rules
and those who act based on adaptive rules. Someone acting based on an
adaptive rule can change her behavior based upon information or past
success or because she watches others. As we shall discuss, no bright lines
distinguish these cases from one another; an adaptive rule can sometimes
be interpreted as a fixed rule, and sometimes rational actions will take the
form of simple rules.
How we choose to model people will depend on the context and on our
goals. Are we predicting or explaining? Are we evaluating policy actions?
Are we trying to design an institution? Or are we exploring? In low-stakes
environments, such as modeling what color coat people buy or whether

they stand for an ovation after a show, we will most often assume that
people apply fixed rules. When people decide whether or not to cooperate
in a venture or to trust another person, we will assume that people learn and
adapt. Finally, in high-stakes environments, we will assume that informed,
sophisticated people make optimal choices.
Before describing our approach in more detail, we first address some
common misconceptions. Many people first encounter formal models of
social phenomena in introductory economics courses. Those models often
rely on a rudimentary rational-actor model in which everyone is selfinterested and capable of optimizing. The model may also assume everyone
has the same preferences and income level. Economists then solve for
equilibria within these models, enabling them to evaluate the effects of
shocks to a market or policy changes. These models, though based on
incorrect assumptions, are useful. They help economists to communicate
and students to understand.
Based on this experience, many people infer that formal modeling requires
a narrow, unrealistic view of human nature, in which people are selfinterested and never make mistakes. That is not the case. In fact, not even
economists think that to be true. The frontiers of economics consist of
models with imperfectly informed, heterogeneous actors who adapt in
response to what they learn, and who sometimes, though not always, care
about the payoffs to others. The extent to which people exhibit otherregarding preferences also depends on the situation; when donating to
charity or volunteering, for example, a person may be more other-regarding
than when buying a house.
Nevertheless, the unfortunate impression persists that modeling assumes
selfish, unrealistically rational people. We must disabuse ourselves of that
view. As an analogy, if you only wade a few steps into the ocean, you
might infer that it is shallow. As you swim out farther, you begin to sense
the depth. Here, we start near the shore. At times we venture further out
and show how models can accommodate other-focused, boundedly rational
people.
Whatever assumptions we do make, we cannot escape their implications.
We are tied to the mast of logical coherence. We cannot manufacture

implications. If we assume strong social influences in consumer choices,
our model will produce a handful of products with large market shares. If
we assume people obtain information through networks, then people who
fill structural holes will possess power.
The remainder of the chapter begins with an overview of some challenges
of modeling people: we are diverse, socially influenced, mistake-prone,
purposive, adaptive, and possessed of agency. We cannot include all of
these characteristics in a single model without creating a complicated mess,
so we must pick and choose. If heterogeneity matters little, perhaps we
assume identical agents. If the problem is simple or the people
sophisticated, perhaps we assume people do not make mistakes.
We next describe the rational-actor model and discuss its theoretical
underpinnings and the justifications for its use despite its descriptive
inaccuracy. We conclude that whether the rational-actor model functions as
a gold standard, a straw man, or somewhere in between depends on our
model’s purpose. The rational actor will be less successful at predicting
human behavior than as a tool for communicating, evaluating actions, and
designing policies.
We then show how we can add psychological biases and altruistic
preferences onto the standard rational-actor model. The choice of whether
to include a bias or a concern for others rests again on what we are
studying. Some human biases such as loss aversion and presentist bias—
caring more about delays today than in the future—may be necessary to
include in some instances. For example, those assumptions may be
important for models of retirement savings or riots. The assumptions may
be less important for models of driving behavior or disease transmission.
In the fourth section, we describe rule-based behavior. This category of
models has the advantage of being both flexible—any behavior we can
write down as a rule is fair game—and tractable. We need only encode that
behavior in a computer program, an agent-based model, and watch what
unfolds. That freedom comes with responsibility. As we can choose any
behavioral rule, we must guard against ad hoc assumptions. In some cases,
behavioral rules can be justified as optimal behavior given an objective
function, though this will not always be the case.

The chapter concludes with a reconsideration of the value of rationality as a
benchmark behavior. Even if people do not optimize, they do adapt to
changing circumstances and new knowledge. That observation produces a
conundrum of sorts. If we design an institution or policy based on the
assumption that people have a bias or that they act in ways not in their selfinterest, we run the risk that people will change their behaviors. People may
be fooled once, but they may be harder to fool two or three times. We need
not conclude that rationality is the only plausible assumption, but the logic
does argue for rationality as a relevant benchmark. Logic also supports
considering simple rules of behavior as well, as a lower bound on
rationality. And, in modeling any given situation, we might apply any
number of adaptive and psychological rules as a way for us to explore the
giant space in between those extremes.

The Challenge of Modeling People
Modeling people presents a challenge because while models require lowdimensional representations, people defy simple characterizations. People
are diverse, we are socially influenced, we are error-prone, we are
purposive, and we learn. In addition, people possess agency—we have the
capacity to act.
By way of contrast, physical objects, such as carbon atoms and billiard
balls, exhibit none of these six properties. Carbon atoms lack diversity
(though they can occupy heterogeneous positions within compounds, such
as in propane). Carbon atoms never violate the laws of physics nor do the
lead purposive lives. They do not change their behaviors based on past
experiences. They lack agency; they do not decide to lead uprisings or
switch careers. Hence the oft-repeated quip by social scientists: how
difficult physics would be if electrons could think. Physics would be even
harder if electrons could write models.
We can start with the problems created by diversity. People differ in our
preferences, in our capacity to act, in the social networks we form, in our
levels of altruism, and in the level of cognitive attention we allocate to
actions. Modeling would be easier if everyone were the same. Sometimes
we rely on statistical logic and assume that the behavioral diversity cancels.

For example, we might construct a model that predicts charitable donations
as a function of income. For a given income level and tax rate, some people
may be more altruistic than we assume and others may be less. If the
deviations from the model average out (and in Chapter 5 we cover models
of distributions that explain why they might), then our model may be
accurate. This canceling out of diversity will not occur unless actions are
independent. When behavior is socially influenced, extreme actions can
create spillovers. This occurs when political activists energize voters. We
will encounter this effect of diversity when we model riots.
Whether or not mistakes cancel in the aggregate depends on the context.
Errors that result from a lack of cognitive attachment may be random and
independent. Errors that arise from cognitive biases may be systematic and
correlated. For example, people may overweight recent events and recall
narratives better than statistics. A shared bias like this will not cancel out.
The next challenge relates to what people desire. A central challenge in
writing models of people will be making an accurate assessment of their
goals and objectives. Some people desire wealth and fame. Others want to
contribute to the betterment of their communities and the world. In the
rational-actor model, we represent a person’s payoff directly in the form of
a function. In rule-based models, purposes are more implicit. A behavioral
rule in which people seek to live in an integrated neighborhood but move
out of a neighborhood if the percentage of people who share their racial
identity falls below 10% embeds certain beliefs about what people desire.
The final challenge to modeling people results from the fact that people
have agency: the ability to take action, to change what we do, and to learn.
That said, in some contexts, people may be better characterized as creatures
of habit. Actions may be outside our control. Few people choose to be
addicted to opioids or to be poor. Yet people take actions that produce those
outcomes.
Often, when people take actions that produce bad outcomes, they adapt
their behavior. We can capture this by including learning in our models.
How people learn varies by context. When learning how many hours they
need to study for an exam in order to get a good grade, or how many times
a week they need to exercise, people may learn based on individual

experiences and introspection. When learning what grocery store to visit or
whether to contribute to a charity, people may learn through observing
others. In Chapter 26, we show how in non-strategic contexts, learning
generally works. People learn the best action. We also show that in strategic
contexts, which we model as games, all bets are off. Neither individual nor
social learning necessarily produces good outcomes.
Each of these six characteristics are potential model features. If we include
a feature, we must decide how much of it to include. How diverse do we
make our actors? How much social influence do we include? Do people
learn from others? How do we define objectives? How much agency do
people possess? We may possess less agency than we believe. Jonathan
Haidt describes our lack of agency with his metaphor of the rider and the
elephant. “The image I came up with for myself, as I marveled at my
weakness, was that I was a rider on the back of an elephant. I’m holding the
reins in my hands, and by pulling one way or the other I can tell the
elephant to turn, to stop, or to go. I can direct things, but only when the
elephant doesn’t have desires of his own. When the elephant really wants to
do something, I’m no match for him.”1 Sometimes we do ride the elephant.
Sometimes we do not. No single approach to modeling humans will be
appropriate in all settings, so we model humans in a variety of ways.

The Rational-Actor Model
The rational-actor model assumes that people make optimal choices given
a payoff or utility function. These actions can be decisions, where the
payoff depends only on the individual’s own action, or they can take place
within a game, where payoffs depend on what others do. In a game with
simultaneous choices or with incomplete information, the rational-actor
model also specifies beliefs about what the other actors will do.

Rational-Actor Model
An individual’s preferences are represented by a mathematical utility or
payoff function defined over a set of possible actions. The individual

chooses the action that maximizes the function’s value. In a game, that
choice may require beliefs about the actions of other players.
As an example, we construct a primitive rational-actor model of an
individual’s decision for how much income to allocate to housing. The
model characterizes her utility as a function of housing and all other
consumption, with the latter including food, clothing, and entertainment
(see box). The model assumes a price for housing and a price for all other
goods. The model is far from realistic. It treats all housing the same. And, it
lumps all other goods into one category called consumption and prices
them all the same. We can set those inaccuracies aside for the moment, as
the purpose of the model is to explain the proportion of income spent on
housing.

A Rational-Actor Model of Consumption
Assumption: An individual’s utility from general consumption, C, and
housing, H, can be written as follows:
image
Result: A utility-maximizing individual (a rational actor) spends exactly
one-third of her income on housing.2
In the model, the proportion of income a person spends on housing does not
depend on the price of housing or on income. Both results are reasonable
approximations of the data.3 Other than people at the extremes of the
income distribution, most people spend about a third of their income on
housing. The finding has policy implications: if housing prices fall by 10%,
people will buy 10% more housing. The finding also provides a
justification for assuming identical agents. If people spend a fixed
percentage of income on housing, total spending on housing depends only
on average income.
Using a utility function makes our models analyzable, testable, and
tractable. We can estimate the functions with data, we can derive optimal
actions, and we can ask “what if” questions by changing parameter values.

In assuming a utility function, we imply a coherence to preferences that
may not exist. For preferences to be representable by a utility function, they
must satisfy certain axioms. Theorems that prove the existence of utility
functions assume a set of alternatives along with a preference ordering.
Imagine that we can list all possible bundles of goods a person might buy.
A preference ordering ranks these bundles from most to least favored. A
person might prefer coffee with milk to tea with lemon; if so, she ranks the
bundle {coffee, milk} above the bundle {tea, lemon}.
A utility function represents preferences if it assigns a higher value to
bundle A than bundle B if and only if the preference ordering ranks A above
B. For preferences to be consistent with a utility function they must satisfy
completeness, transitivity, independence, and continuity. Completeness
requires that the preference ordering is defined over all pairs of alternatives.
Transitivity rules out preference cycles. If someone prefers bundle A to
bundle B and bundle B to bundle C, she must also prefer A to C. In other
words, if a person prefers apples to bananas and bananas to cheese, then
she must also prefer apples to cheese. This condition rules out inconsistent
preferences.
Independence requires that people evaluate the outcomes in a lottery
separately. A lottery is a probability distribution over alternatives, such as a
60% probability of A and a 40% probability of B. Preferences satisfy
independence if when A ranks above B, then in any lottery that includes B
as an outcome, the person prefers an alternative lottery in which we replace
B with A. Independence rules out strong risk aversion. A risk-averse person
might rank a trip to New Orleans over a trip to Disney World but prefers
knowing for certain that he will be going to Disney World over entering a
lottery that sends him to New Orleans with probability image and to
Disney World otherwise. The final condition, continuity, requires that if a
person ranks A above B and B above C, then there exists a lottery in which
she gets A with probability p and C with probability (1− p) that she likes
exactly as much as B. This assumption also rules out strong preferences for
certain outcomes.4
The assumptions of independence and transitivity, which people violate, on
top of the dubious claim that people optimize leads many to question the

widespread use of the rational-actor model, particularly by economists. Yet
there exist good reasons to assume rationality. First, people may act “as if”
they optimize. They may apply rules that produce nearly optimal behavior.
When people play pool, catch a Frisbee, or drive a car, they do not write
down mathematical equations. The mathematics required to time a leap to
catch a Frisbee would overwhelm almost anyone. Yet people catch
Frisbees. So, by the way, do dogs. Thus, both people and dogs act as if we
solve a difficult optimization problem.
This same logic extends to high-dimensional problems. An analysis of the
actions of Harold Zurcher, the superintendent of maintenance for the
Metropolitan Bus Company in Madison, Wisconsin, found that he made
near optimal decisions about when and whether to replace bus engines.5
Though Zurcher did not write down any mathematics, he relied on
heuristics. Those heuristics, informed by experience, meant that he acted
(almost) as if he were a rational actor.
Second, even if people do make mistakes, in cases where a situation is
repeated, our capacity to learn should push us toward optimal actions.
Third, in cases where the stakes are large, people should put in the time and
energy to make near-optimal choices. People may overpay 30% for coffee
or AAA batteries, but they do not overpay 30% for cars or houses. The
claim that learning and higher stakes increase rationality has ample
empirical and experimental support.6
A fourth reason for adopting the rational-actor model, paradoxically, is that
it simplifies the analysis. Most utility functions will have a unique optimal
action. A person can behave suboptimally in thousands of ways. Saying
that people do not optimize opens an enormous box of possibilities. If we
assume people make choices to maintain their identities or to enforce
cultural norms, we may lack a single clear answer. Rational choice is not
realistic, but realism comes at the cost of messiness. An answer, even if it is
known to be wrong, can be more useful then having no answer at all, as it
allows us to bring the model to data and to work through the effects of
changes in variables.7

Arguments for Rational Choice
“As if”: Intelligent rule-based behavior may be indistinguishable from
optimal or near-optimal behavior.
Learning: In situations that are repeated, people should approach optimal
behavior.
Large stakes: On important decisions, people gather information and think
slowly.
Uniqueness: Optimal behavior is often unique, making the model testable.
Consistency: Optimal behavior creates a consistent model. If people learn
the model, they will not change their behavior.
Benchmark: Optimal behavior provides a benchmark as an upper bound
on people’s cognitive abilities.
Fifth, the rational-actor assumption guarantees internal consistency. If a
model assumes suboptimal behavior and the model is in the public domain,
the model can be learned. People can change their behavior. They might not
optimize, but any assumption other than optimality is subject to the
criticism that it is not consistent. We return to this point in the discussion at
the end of the chapter.
Last, and some would argue most important, rationality can function as a
benchmark.8 When designing a policy, making a prediction, or choosing an
action, we should consider what would happen if people had rational
preferences and optimized. That exercise may point to flaws in our
thinking. We should also be open to the possibility that the exercise will
lead us to conclude that the rational-actor model does not apply and that we
should privilege other models instead. To this list we might add a seventh
reason: many-model thinking. If people apply many models, they are less
likely to make mistakes.

Psychological Biases

The rational-actor model has been challenged by psychologists,
economists, and neuroscientists, who note that it does not match up with
how humans behave. Empirical findings from laboratory and natural
experiments show that people suffer a variety of biases, including a status
quo bias. We ignore base rates when making probability calculations, we
attach too much significance to sure things, and we are loss-averse.
As researchers begin to link behavior and beliefs to processes within the
brain, evidence of hardwired biases becomes more compelling. For
example, neuroeconomics uses brain imaging studies to study economically
relevant behaviors such as attitudes toward risk, levels of confidence, and
responses to information.9 Kahneman argues that what we know so far
supports making a distinction between two types of thinking: quick,
intuitive rules (fast thinking) and deliberate contemplation (slow thinking).
Fast thinking is more likely to be subject to the aforementioned biases.10 In
the long run, we may be able to infer some behavioral patterns from brain
structures, but we should keep in mind that the brain has tremendous
plasticity. It is capable of overcoming biases by thinking slowly.
Further, we should be cautious about accepting as universal any finding
documented in just a handful of studies. Many psychological findings have
not proven robust. A recent study failed to replicate half of one hundred
findings published in leading psychology journals.11 Furthermore,
replicability need not imply universality. Subject pools for many studies
lack economic and cultural diversity.12 We might expect that more diverse
subject pools would produce fewer behavioral regularities, providing even
greater reason to avoid generalizations about behavior.
Last, in attempting to make more realistic models, we must keep
tractability in mind. More realistic models may require more sophisticated
mathematics.13 None of these concerns is so persuasive to suggest
abandoning models with psychologically realistic behaviors, but
collectively they imply that we proceed with caution and emphasize welldocumented behavioral regularities.
Two deviations that have been replicated many times are loss aversion and
hyperbolic discounting. Loss aversion states that people are risk-averse

over gains and risk-loving over losses. Kahneman and Tversky refer to this
general theory of behavior as prospect theory.14 Loss aversion does not at
first appear irrational, but it implies that people choose different actions
when an identical scenario is presented as a potential loss as opposed to a
potential gain.
For example, people prefer winning $400 for certain rather than entering a
lottery with an even chance of winning $1,000. Yet they will enter a lottery
with an even chance of losing $1,000 rather than lose $600 for certain. This
same inconsistency extends to nonmonetary domains. Doctors given
choices framed as gains are risk-averse. When choices are presented as
losses, doctors take more risks.15

Prospect Theory: Example
Gain Framing: You have two options:
Option A: Win $400 for certain.
Option B: Win $1,000 if a fair coin comes up heads, and $0 if tails.
Loss Framing: You are given $1,000 and have two options:
Option

image: Lose $600 for certain.

Option
tails.

image: Lose $0 if a fair coin comes up heads, and lose $1,000 if

A and image: are equivalent as are B and image. According to prospect
theory, more people choose A and image.
Hyperbolic discounting implies stronger discounting of the immediate
future. Standard economic models assume exponential discounting, a
constant rate at which people discount the future. A person with an annual
discount rate of 10% values $1,000 next year as equal to $900 today. She
would discount next year’s money by 10% for every year into the future.
Evidence shows that most people do not discount the future at a fixed

discount rate. Instead, they suffer from an immediacy bias: they discount
the near future far more than the later future.16 For example, if you ask
people whether they would prefer $9,500 twenty years from now or
$10,000 twenty years and one day from now, almost everyone will wait one
more day for the extra $500. If you ask those same people if they would
prefer $9,500 today or $10,000 tomorrow, many will take the $9,500 now.
This is an example of immediacy bias.17
That bias produces time-inconsistent behavior. One year from now, most
people prefer waiting one more day and taking the $10,000. Such
preferences are not logically consistent. Hyperbolic discounting has been
put forward as a reason why people run up credit card debts, eat unhealthy
foods, have unprotected sexual relations, and fail to save for retirement.
In summary, depending on how we will use our model, we may choose to
assume loss aversion and hyperbolic discounting given that these
assumptions appear to better match behavior for most people. The main
reason not to do so would be if they complicate the model without
qualitatively changing what we find, or if by assuming hyperbolic
discounting our model produces unrealistic behavior.

Rule-Based Models
We now turn to rule-based models.18 While optimization-based models
assume an underlying utility or payoff function that people maximize, rulebased models assume specific behaviors. A rule-based model might assume
that in an auction, a person will bid 10% less than her true value for an
item, or that a person will copy a friend’s action if that friend consistently
receives higher payoffs. Many people equate optimization-based models
with mathematics and rule-based models with computation. The distinction
between optimization-based models and rule-based models is not as clean
as might be thought. Think back to our model of housing consumption.
Optimal behavior took the form of a simple rule: spend one-third of income
on housing. The key difference between the two approaches lies in their
foundational assumptions. In an optimization-based model, preferences or

payoffs are fundamental. In a rule-based model, the behavior is
fundamental.
Behavioral rules can be fixed or adapt. A fixed rule applies the same
algorithm at all times. Just as rational-choice models provide an upper
bound on people’s cognitive abilities, fixed-rule models provide a lower
bound. A common fixed rule in markets, the zero intelligence rule, accepts
any offer that produces a higher payoff. It never takes a stupid (i.e., utilityreducing) action. Suppose we want to gauge the efficiency of a one-sided
market design in which sellers post take-it-or-leave-it offers for some good.
A seller following a zero intelligence rule would randomly pick a price
above her value. A buyer would purchase any good with a price below her
value. When we encode those behaviors in a computer model, we find that
in markets zero-intelligence traders produce nearly efficient outcomes.
Thus, exchange markets do not need rational buyers and sellers to function
well.19
An adaptive rule switches among a set of behaviors, evolves new
behaviors, or copies the behaviors of others. It takes these actions in order
to improve the payoff. Therefore, unlike fixed rules, adaptive rules require
a utility or payoff function. Advocates of this approach argue that within
any situation people tend toward simple and effective rules, and that if this
is what people do, we should model them in that way.20 Though rule-based
models make no explicit assumption about rationality, adaptive-rule models
exhibit ecological rationality—better rules come to predominate.21
To explain how adaptive-rule models operate, we describe the El Farol
model of self-organized coordination.22 El Farol is a nightclub in Santa Fe,
New Mexico, that features dancing every Tuesday night. Each week, a
population of 100 potential dancers decide whether to go dance at El Farol
or stay home. All 100 people like to dance, but they do not want to go if the
club is too crowded. The model assumes a stark form of preferences. A
person earns a payoff of zero from staying home, a payoff of 1 from
attending if 60 or fewer people attend, and a payoff of -1 from attending
when more than 60 people attend.

If we construct a fixed-rule model, anything might result. For example, if
we assigned everyone the rule “go the first week; if more than 60 people
attend, do not go the next week; and then go the following week,” the El
Farol would have 100 attendees the first week, zero attendees the second
week, and then 100 attendees the third week. The El Farol model creates
adaptive rules by endowing each person with an ensemble of rules. Each
rule tells the individual whether or not to go to El Farol. The rules take
several forms. Some are fixed rules: go every other week. Others are based
on trends in the number of people who attended El Farol in recent weeks.
One rule might predict that the number of people that show up this week
will be the same as last week. If fewer than 60 people attended last week,
that rule would tell the person to go this week.
An adaptive behavioral rule model might assign a score to each rule equal
to the percentage of weeks for which it gave correct advice. Each
individual might then use the model in her ensemble with the highest score.
The best rule will vary over the course of the weeks. Simulations of this
type of model find that if individuals possess a large ensemble of rules,
then approximately 60 people attend each week: coordination emerges
without any central planner. In other words, the system of adaptive rules
self-organizes into nearly efficient outcomes.

El Farol Model: Adaptive Rules
Each of 100 individuals decides independently whether or not to go to El
Farol every week for a year. An individual who goes to El Farol earns a
payoff of 1 if 60 or fewer people attend and a payoff of -1 otherwise. An
individual who does not go to El Farol earns a payoff of zero.
Each individual has an ensemble of rules to decide whether to attend. These
rules can be fixed or contingent on recent past attendances. Each week,
each individual follows the rule in his ensemble that, if followed, would
have produced the highest payoff in the past.
We can interpret behavior within adaptive-rule models, like the El Farol
model, within the micro-macro loop (see figure 4.1). At the micro-level, a
set of individuals take actions (denoted by the ai’s) according to rules.

These rules create macro-level phenomena (denoted by Macro1 and
Macro2), as represented by the upward arrows. In the El Farol problem, the
macro-level phenomena are the sequences of past attendances. The
downward arrows represent how these macro-level phenomena feed back
into the behaviors of the individuals. In the El Farol model, each person
may be applying a different rule. If the rules people apply produce a
crowded El Farol four weeks in a row, then rules that tell people to attend
less often will produce higher payoffs. As people switch to those rules,
fewer people will attend. The micro-level rules produce a macro-level
phenomenon (over-attendance) that feeds back to the micro-level rules.
image
Figure 4.1: The Micro-Macro Loop

Cognitive Closure, a Big Question, and Many
Models
The micro-macro loop elucidates a central tension as to how smart to make
our agents. Should people infer all consequences of their actions? The loop
also hints at a larger question that we encounter throughout the book as to
what class of outcome a model produces: Does it go to equilibrium,
produce randomness, create a cycle, or generate a complex series of
outcomes?
We start with the question of how smart to make our agents. Suppose that
we believe that individuals possess only modest cognitive abilities, so we
build a model with zero intelligence agents. Their actions aggregate to
produce aggregate macro-level phenomena. If the macro level produces
efficient or nearly efficient outcomes, as we noted was the case with a onesided market of buyers and sellers, then we may be justified in our
assumption. An easy-to-follow fixed rule produces good outcomes. People
would have little incentive to expend effort developing more sophisticated
rules.

The tension arises when our model produces inefficient or even lousy
macro-level outcomes. Such could be the case in the El Farol model, where
a common fixed rule could lead to a cycle in which El Farol was
overcrowded with dancers one week and empty the next. Confronted with
an inefficient outcome, we might think that people would adapt. They
might experiment. They might think through the logic of the situation to
formulate a new action. If we follow that logic to its extreme and assume a
low cost of thinking, then we find ourselves advocating the rational-actor
model. Any person not behaving optimally could do better. While that is
true, people also have to be able to formulate that better action.
This leads to a big question: What class of outcomes does the model
produce? We have four options: equilibrium, cycles, randomness, or
complexity. The class of outcome will matter for deciding how seriously
we take the argument that people should learn their way to equilibrium.
First, if the model produces randomness at the macro level, the individuals
probably cannot learn anything. Our model is fine. A similar logic applies
to models that produce complex patterns. In these cases, we would assume
that people continue to adapt new rules, but we would not necessarily
assume that they can choose optimally. To the contrary, the complexity of
the macro-level phenomena makes optimal responses implausible. People
would be more likely, as in the El Farol model, to confront complexity with
an ensemble of simple rules.
The models that produce cycles or equilibria create a stationary
environment. We therefore might expect that people can learn—that no one
would continually take a suboptimal action. As an example, suppose we
have a traffic model in which everyone chooses a route to work using a
fixed rule. In our model, the traffic system settles into an equilibrium. In
that equilibrium, one of the individuals, Layne, spends 75 minutes each
morning traveling from Calabasas to downtown Los Angeles. Given the
equilibrium, if Layne took side streets through Topanga Canyon, her trip
would take only 45 minutes. Given the value of an extra 30 minutes per day
and the frequency with which people in Los Angeles talk about traffic,
Layne would likely find the shorter route. She has no shortage of methods
for finding it. She might use a route recommender, talk to a neighbor, or
experiment.

Thus, if our model produces an equilibrium (or a simple cycle) and that
equilibrium is not consistent with optimizing behavior, then our model
suffers a logical flaw. If people have a better action available to them, they
should figure it out. They should learn. Notice that we need not assume
optimal behavior in order to reach the equilibrium. People could follow
simple rules and produce an equilibrium in which no one can benefit by
changing her action. At that equilibrium, it would look “as if” people are
optimizing, because they are. Again, that logic need not apply for complex
or random outcomes. If traffic patterns in Los Angeles produce a complex
sequence of traffic slowdowns and jams, we have little reason to believe
that Layne selects an optimal route each day. She almost surely cannot.
If adaptive rules that can adopt any action produce an equilibrium, then the
equilibrium must be consistent with behavior by optimizing agents. If those
same adaptive rules produce complexity, the agents’ behavior need not be
optimal. We can restate this idea as follows: optimal behavior may be an
unrealistic assumption, particularly in complex situations. On the other
hand, if a system produces a stable outcome in which a person has better
actions, she will probably figure out a better action to take.
An extension of this logic applies to policy interventions. Suppose that we
use data to estimate people’s behavioral rule—say, the likelihood a person
shows up at a hospital’s emergency room during lunch hour for minor
health issues. If we assume a fixed rule, we might enlarge the size of our
facility so that people do not have to wait. If people continue to follow that
fixed rule, we have a new equilibrium with short midday wait times.
However, with new, shorter wait times, people who had not been going to
the emergency room for sprained ankles or chest colds may now decide to
go. That equilibrium relies on people choosing suboptimal actions, such as
not going to an emergency room even though they would not have to wait.
If people learn, we cannot rely on past data to predict outcomes under a
policy change. This insight, known as the Lucas critique, is a variant of
Campbell’s law, which states that people respond to any measure or
standard in ways that render it less effective.23

The Lucas Critique

Changes in a policy or the environment likely produce behavioral responses
by those affected. Models estimated with data on past human behaviors will
therefore not be accurate. Models must take into account the fact that
people respond to policy and environmental changes.
As should be clear at this point, there exists no best solution for how to
model people. How rational we make them or how adaptive we make their
rules depends on the circumstances. We should exercise our best judgment
in each situation. Given the uncertainties, we should err on the side of more
models rather than fewer.
Even if we are predisposed to dismiss rational-choice models as unrealistic,
we must recognize their tractability, their capacity to reveal the directional
forces of incentives, and their value as a benchmark. Simple rule-based
behaviors, such as zero intelligence, are also unrealistic. Though wrong,
they can be of use. They are easy to analyze and can reveal how much
intelligence matters in a given setting.
Human behavior occurs within the extremes of zero intelligence and full
rationality, so it makes sense to construct models in which individuals adapt
using rules. Those rules should take into account the fact that people vary
in their cognitive attachment and capabilities within a domain. We should
therefore expect behavioral diversity. We might also expect some
consistency within groups. This too can be included in models.24
In sum, given the complexities involved in modeling humans, we have
abundant reasons to apply multiple diverse models. We may not be able to
predict exactly what people will do, but we may be able to identify the set
of possibilities. If we can, we have benefited from constructing models
because we know what could happen.
We conclude with a plea for humility and empathy. In constructing models
of people, a modeler must be humble. Given the challenges of diversity,
social influence, cognitive errors, purpose, and adaptation, our models will
inevitably be wrong, which is why we take a many-model approach.
Austere models of behavior fit some situations well and allow us to focus
on other aspects of the environment. Richer behavioral models will be more
appropriate when we have better data. We must maintain modest

expectations. People are diverse, purposive, adaptive, biased, and socially
influenced, and we possess a degree of agency. How can we not expect any
single model of human behavior to be wrong? It must be. Our aim is to
construct many models that as an ensemble will be useful.