Social Interactions - University of Chicagohome.uchicago.edu › sdurlauf › includes › pdf ›...
Transcript of Social Interactions - University of Chicagohome.uchicago.edu › sdurlauf › includes › pdf ›...
Social Interactions
Steven N. Durlauf1 and Yannis M. Ioannides2
1Department of Economics, University of Wisconsin,
Madison, Wisconsin 53706; email: [email protected]
2Department of Economics, Tufts University, Medford,
Massachusetts 02155; email: [email protected]
Annu. Rev. Econ. 2010. 2:451–78
First published online as a Review in Advance on
April 6, 2010
The Annual Review of Economics is online at
econ.annualreviews.org
This article’s doi:
10.1146/annurev.economics.050708.143312
Copyright © 2010 by Annual Reviews.
All rights reserved
1941-1383/10/0904-0451$20.00
Key Words
neighborhood effects, statistical mechanics, social interactions
empirics, endogenous social effects, estimation and identification,
social networks
Abstract
The study of social interactions has enriched both the domain of
inquiry of economists and the way economists conceptualize indi-
vidual decision making. The review aims to introduce the classes of
models that accommodate estimation of social interactions and to
examine the key areas where significant advances have been made
in the identification of social effects. It surveys linear and nonlinear
models and their applications, including results regarding partial
identification. The review also examines conceptual and methodo-
logical links with the spatial econometrics and the social networks
literatures. It considers empirical applications in the context of our
methodological overview.
451
Click here for quick links to Annual Reviews content online, including:
• Other articles in this volume• Top cited articles• Top downloaded articles• Our comprehensive search
FurtherANNUALREVIEWS
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
“Most people are other people. Their thoughts are some one else’s opinions,
their lives a mimicry, their passions a quotation.”
Oscar Wilde, de Profundis
1. INTRODUCTION
Within economics, the study of social interactions has expanded the domain of inquiry to
incorporate many ideas that are traditionally associated with sociology. Social interactions
analysis also extends the methodological individualism of economics in new directions
through its focus on the feedbacks between individual behaviors and aggregate outcomes.
By social interactions, we refer to interdependences among individuals in which the prefer-
ences, beliefs, and constraints faced by one person are directly influenced by the character-
istics and choices of others. We emphasize the word directly as these interactions do not
occur because of interdependences due to prices, as occurs in an Arrow-Debreu world.
In fact, social interactions often represent externalities. Canonical social interactions
examples include conformity effects, which occur when the utility from a given behavior
increases when others make the same choice, and social networks effects in which
information diffuses via direct contacts. Social interactions have been used to help explain
phenomena ranging from cigarette smoking to the persistence of ghettos and inner-city
poverty.1
Although social interactions models take sociological ideas seriously, they fully pre-
serve the purposeful, choice-based formulation of individual decision making that is the
hallmark of modern economics. These models simply expand the domain of factors
that determine individual decisions. Furthermore, it is also no surprise that a precursor
to the modern literature is Becker (1974). Moreover, social interactions models have
close parallels in game theory, especially quantal response equilibria (McKelvey &
Palfrey 1995).
Although still relatively new, the social interactions literature is sufficiently large and
wide-ranging that a complete survey is beyond the scope of an article of this length (for
related surveys in economics, see Manski 2000, Durlauf 2004, Ioannides & Loury 2004).
We therefore focus on providing a template for understanding the broader literature.
We outline a general theoretical framework for social interactions and then describe the
econometric implementation of social interactions ideas, focusing on identification issues.
We also describe some links between the social networks and spatial econometrics litera-
tures and social interactions.
Section 2 outlines a general discrete choice model of social interactions that illustrates
many of the interesting theoretical properties that social effects have been shown to gener-
ate. Section 3 describes the baseline econometric models that have been used in the study of
social interactions and focuses on what we call classical identification problems involving
the disentanglement of endogenous and exogenous social effects. Some prominent empiri-
cal studies are discussed. Section 4 describes contemporary identification problems, which
include, in particular, environments with self-selection and unobserved group variables
being present. Empirical studies that address these problems, including those that use
Social interactions:
direct
interdependences inpreferences,
constraints, and beliefs
of individuals, whichimpose a social
structure on individual
decisions
1Many of the ideas concerning persistent poverty in Wilson (2009) have been formalized and studied in the social
interactions literature.
452 Durlauf � Ioannides
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
experimental data, are reviewed. Section 5 discusses spatial and network approaches to
social interactions. Section 6 concludes.
2. A DISCRETE CHOICE MODEL OF SOCIAL INTERACTIONS
A basic model of social interactions, one that captures many of the interesting implications
of integrating social factors into individual behavior, can be developed in the context of
decisions over a discrete set of choices. We generalize the multinomial choice model of
Brock & Durlauf (2002, 2007) following Durlauf (1997) and Ioannides (2006). We model
a population of I individuals, each of whom chooses oi among unordered alternatives from
the choice set S¼ {1; . . . ; L}. Each agent i is associated with a group g(i), which is the subset
of the population whose behaviors and characteristics enter as direct arguments in i’s
decision problem. Our goal is to describe the population’s behavior in light of these
group-level effects.
For i, each choice ‘,‘ 2 S, produces utility Vi,‘, which is the sum of three distinct
components. The first is hi,‘, private deterministic utility. It is private in that it does not
exhibit direct dependence on the choices of others and is deterministic as it is known to the
modeler. In econometric work, this is operationalized by assuming hi,‘ is known up to some
set of estimable parameters. The second component is deterministic social utility and
captures dependence of the overall utility of a given choice on others’ choices. If individual
i chooses ‘ and j chooses s, then i’s utility is Ji,j,‘,s . We assume that agents do not observe
the choices of others, but rather form beliefs on the choices: pej;sji denotes the probability
i assigns to choice s on the part of j. The social utility components Ji; j;‘;spej;sji are additive
across j for a given ‘. A third component is a random individual utility term, ei,‘, assumed
to be independent across i and ‘. Therefore,
Vi;‘ ¼ hi;‘ þXLs¼1
XI
j6¼i
Ji; j;‘;s pej;sji þ ei;‘: ð1Þ
This utility function allows for an explicit characterization of the equilibrium choice
probabilities once the probability distributions for the random utility terms are specified.
We assume that the ei,‘’s are distributed according to the multinomial logit model with
dispersion parameter B and that expectations are rational (i.e., self-consistency of beliefs).
Thus
Probðoi ¼ ‘Þ ¼ pi;‘ ¼exp
�B�hi;‘ þ
PLs¼1
PIj6¼i
Ji; j;‘;spj;s
��
PL‘0¼1
exp
�B�hi;‘0 þ
PLs¼1
PIj 6¼i
Ji; j;‘0;s pj;s
�� ; ‘ 2 S; i ¼ 1; . . . ; I: ð2Þ
Higher B implies lower variance. At one extreme, B ¼ 0 implies all outcomes are equally
likely because the private random utility density is so diffused that the maximum of
the random utility shocks controls the choice. At the other extreme, B ¼ 1 means that
choices are deterministic as the private random utility terms are all equal to zero with
probability 1. Under the Brouwer fixed-point theorem, at least one such fixed point exists
for each of individual i’s choice probabilities, given by Equation 2.
It is common in the social interactions literature to simplify the interaction structure by
restricting social utility so that an individual only cares about the fraction of the population
www.annualreviews.org � Social Interactions 453
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
making the same choice he does. This renders the agent indifferent as to who makes the
choices. Under this simplification, the object of interest is not the matrix with elements pi,‘,
the individual choice probabilities, but rather the aggregate choice probabilities,
p‘ ¼ I�1P
ipi;‘. This leads to
pe‘ ¼ p‘ ¼ I�1XI
i¼1
exp�B�hi;‘ þ Ji;‘p‘
�PLs¼1
exp�Bðhi;s þ Ji;sps
� ; ‘ 2 S; ð3Þ
where Ji,‘ is the social utility weight i assigns to the share among the population making
the choice ‘.
A leading case in the theoretical literature focuses on binary choices with the decision
space S ¼ {1,�1}. These models typically treat social utility as a function of the expected
choice of others; that is, i’s expectation of j’s decision is mei; j ¼ pej;1ji � pej;�1ji .
2 Under self-
consistency of beliefs, mei; j ¼ mj, where me
i; j ¼ Eðoj jFiÞ, the mathematical expectation of
oj, conditional on i’s information set. Let m denote the I vector with the mj’s as elements.
Recalling the hyperbolic tangent function, tanhðxÞ � expðxÞ � expð�xÞexpðxÞ þ expð�xÞ,
mi ¼ tanh ½Bhi þ BJim�; i ¼ 1; . . . ; I; ð4Þwhere hi ¼ hi,1 � hi,�1, and Ji ¼ (Ji1; . . . ; Jij; . . . ; JiI) is an I vector of interaction coeffi-
cients Jij.
To understand the properties of the binary choice model, we consider the case in which
all heterogeneity across agents results from random utility; i.e., we assume that hi,‘ and Ji,‘are constant across agents. Under our preference assumptions, we can focus on symmetric
equilibria, i.e., equilibria in which all the expected values are the same. This means that the
expected value of the average choice of others will be the overall population mean m:
m ¼ tanhðBhþ BJmÞ: ð5ÞThe properties of this special case are straightforward to describe. If BJ > 1, and h ¼ 0,
then the function tanhðBhþ BJmÞ is centered at m ¼ 0, and Equation 5 has three roots:
a positive one (upper), m�þ; zero (middle); and a negative one (lower), m�
�, where
m�þ ¼ jm�
� j. If h 6¼ 0 and J > 0, then there exists a threshold H�, which depends on Band J, such that if Bh < H�, Equation 5 has a unique root, which agrees with h in sign.
If Bh > H�, then Equation 5 has three roots: one with the same sign as h and two with
the opposite sign. If J < 0, then there is a unique equilibrium that agrees with h in sign.
The relationship between the number of equilibria and the parameters J, h, and B can
be given some intuition. Holding B constant, it is not surprising that multiple equilibria
emerge when the strength of conformity effects, measured by J, is large relative to the
strength of private incentives, measured by h. The role of the measure of random utility
dispersion B is less obvious. Higher Bmeans lower heterogeneity. The degree of heterogeneity
in turn determines how private and social incentives interact to produce equilibria. With
large heterogeneity, then relatively large fractions of the population will experience draws
such that either ei;‘ � ei;‘
0 or ei;‘
0 � ei;‘ is large. This means in turn that a relatively high
fraction will have their decisions controlled by their idiosyncratic payoffs. But this means
that a relatively small percentage of the population remains that can engage in self-consistent
2This may be seen a special case of Equation 1, where Ji; j;‘;1 ¼ Jij; Ji; j;‘;�1 ¼ �Jij, so thatPj6¼i
Ji; j;‘;spj;sji ¼PjJijmj:
454 Durlauf � Ioannides
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
bunching because of social utility effects. In other words, if enough agents make choices
driven by symmetrically distributed payoff differences, this implicitly delimits the magnitude
of the social utility terms because it delimits the possible support ofm.
Brock & Durlauf (2002, 2007) show that these properties extend to the multinomial
choice model, when hi;‘ ¼ h; 8i; ‘. If BJL
> 1, then at least three equilibria exist, whereas ifBJL
< 1, the equilibrium aggregate choice probabilities are unique. The dependence of the
threshold for multiplicity on the number of choices L occurs because, when the ei,‘’s areindependent across i, the probability that the idiosyncratic draw for one of the possible
choices dominates the agent’s decision is increasing in the number of choices.
Similar properties hold for alternative interaction structures. Blume (1993) conceptual-
izes agents as located on a two-dimensional integer lattice. He assumes that agents interact
locally, which amounts to replacing the social utility component in Equation 1 with
JP
jj�ij¼1Efojg. It is straightforward to see that the mean choice level for the population is
characterized by Equation 5. This example illustrates a general principle about discrete
choice models of social interactions, namely that their qualitative properties are often
independent of the specifics of the interaction structure. This property, called universality
in the physics literature, helps justify the analysis of particular interaction structures
because their properties are not tied to their detailed specification.
Universality provides a segue between social interactions models and statistical mechan-
ics models in physics, in which the property was originally conceptualized. Statistical
mechanics models study the macroscopic properties of a large number of interacting
objects (e.g., atoms). The Brock-Durlauf model is mathematically equivalent to a variant
of the Curie-Weiss model of magnetism, whereas the Blume model is a variant of the Ising
model of magnetism. Another example of the application of statistical mechanics models
to economics is Topa (2001), who employs contact processes to study local interactions in
unemployment. Despite these examples, we believe the mathematics of statistical mechan-
ics has yet to be fully exploited by economists.
One must be careful in translating statistical mechanics models into social science
contexts. One reason is that statistical mechanics models take as primitives the conditional
probabilities that relate objects in a system, whereas socioeconomic contexts require that
these conditional probabilities derive from individual decision problems. The importance
of this distinction is demonstrated by Horst & Scheinkman (2006). In a social interactions
model in which agents are arranged on integer lattices, in which the individual payoff
structure is defined in terms of values for Ji,j,‘,s, the authors show that the equilibrium
conditional probabilities that relate the choices may have a different pattern of zero and
nonzero values, when agents observe one another’s choices. Therefore, the conditional
probability structure for a given statistical mechanics model does not always reveal the
payoff interdependences needed to generate it. Another reason is that exact versus approx-
imate statistical mechanics models can have substantively different interpretations in social
science contexts. Brock & Durlauf (2001a) show that the exact solution to the Curie-Weiss
model without approximation corresponds to a social planner’s problem, whereas the
mean field approximation represents an exact (to social scientists) noncooperative model
of binary choice.
Discrete choice models do provide a tight connection between theory and econometrics,
although this link has not usually been exploited in empirical work. Our canonical social
interactions model treats Ji,j,‘,s as a primitive. One can identify models that produce
www.annualreviews.org � Social Interactions 455
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
different types of social interactions as equilibrium outcomes (e.g., Bernheim 1994,
Streufert 2000, Akerlof & Kranton 2002). This type of work, however, has not been
directly integrated into the abstract theoretical structures we outline above.
3. ECONOMETRIC MODELS OF SOCIAL INTERACTIONS: CLASSICALIDENTIFICATION ISSUES
In this section we consider the problem of identifying social interactions in various
empirical contexts. This leads to variations of the classical identification problem of
simultaneous equations systems that reflect the specific structure of social interactions
models. The statistical models allow us to study both the identifiability of a role for
some type of social influence on behavior and the identifiability of types of social
influences (for estimation issues, see Krauth 2006, Aradillas-Lopez 2007, Conley &
Topa 2007, de Paula 2009).
The distinction between types of social interactions did not arise in our baseline theo-
retical model of social interactions because effects other than those associated with the Jij’s
are subsumed in the private deterministic utility term, as we were considering individuals
within a common group. Manski (1993) first emphasized the difficulties of distinguishing
types of social interactions and developed an important dichotomy. Endogenous effects
capture social interactions that directly occur between one agent’s decisions and those of
others. An essential feature of this type of interaction is that the choices are simultaneously
determined. A canonical example is peer effects in classroom effort; one student works
hard based on whether others do so (or are expected to do so). Of course, one can equally
well imagine that individuals are affected by the personal characteristics of others. The
educational level of the parents of classmates matters if these levels help form the aspira-
tions of the student (and affects students’ net utility from effort): These are contextual
effects.
In discussing the econometrics of social interactions, we assume that the unobserved
heterogeneity is independent across agents and unpredictable given individual and group
characteristics. We relax these assumptions in the next section.
Our analysis of identification focuses on the two main classes of econometric models:
linear-in-means models and discrete choice models. It will be evident that the connections
between the linear-in-means models and economic theory are tenuous; Brock & Durlauf
(2001c) show that one can reverse engineer a linear-in-means model to represent an agent’s
optimal behavior. Discrete choice models, of course, provide a tighter link between theory
and econometrics, although the empirical literature has typically eschewed fully exploring
these links.
Our discussion of identification considers individual observations across groups, g(i).
Outcomes for the linear-in-means model are denoted as yi, whereas oi denotes outcomes
for discrete choice models; xi denotes individual-level characteristics, and zg(i) denote
group-level characteristics.
3.1. Linear Models of Social Interactions
The linear-in-means model is a standard regression model:
yi ¼ a0 þ axi þ yzgðiÞ þ bmgðiÞ þ ei; ð6Þ
456 Durlauf � Ioannides
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
where mg(i) is the expected average choice in g(i), xi is an r vector of individual effects, zg(i)is an rg vector of contextual effects, and ei is unobserved individual heterogeneity. Applying
an expectations operator to both sides of Equation 6 provides an expression for mg(i) in
terms of observables that produces a reduced form for yi:
yi ¼ a01� b
þ axi þ b1� b
axgðiÞ þ y1� b
zgðiÞ þ ei: ð7Þ
Variants of Equation 6 and especially Equation 7 have been used in many contexts to
study social interactions. Datcher (1982) deserves credit for pioneering the empirical
approach along the lines of Equation 7. Durlauf (2004) provides an overview of the
many studies that have estimated regressions of this type and surveys the kinds of
variables that have been purported to represent evidence of social interactions. Promi-
nent examples include the following. Brooks-Gunn et al. (1993) relate IQ and behav-
ioral problems at 36 months of age, high school dropout rates, and nonmarital fertility.
Aizer & Currie (2004) and Bertrand et al. (2000) argue that individual use of social
services/public assistance is affected by the usage rates of others with whom one inter-
acts, and Weinberg et al. (2004) and Bayer et al. (2008) argue that individual market
outcomes are influenced contemporaneously by the labor or market status of neighbors
(see also Ioannides & Loury 2004 for a review of the literature). Corcoran et al. (1992)
claim that individual labor market outcomes are influenced by growing up in a poor
neighborhood, and Burke et al. (2003) contend that physicians’ following each others’
practices produces geographical variations in medical care practices. Gaviria & Raphael
(2001) explore endogenous peer effects in drinking and smoking among teens. An espe-
cially interesting example of this type of regression is Mas & Moretti (2009), who
employ a data set that measures supermarket employee productivity in 10-min intervals;
their data set is also impressive because the set of peers for a given worker regularly
changes due to differences in shift composition and because the spatial orientation of
workers in a store is known. This allows for analyses of such questions as whether
frequent interactions induce stronger social effects and whether physical proximity to
others matters.
One topic that has received particular attention concerns the effects of school and
classroom peers. For example, Hoxby (2000) examines peer effects in the classroom, and
Hanushek et al. (2003) find that growth in academic achievement is associated with the
average achievement of school peers. Controlling for peers’ achievement, Hoxby &
Weingarth (2006) conclude that peers’ race, ethnicity, income, and parental education have
no or very weak effects, and Henry & Rickman (2007) study peer academic achievement
and learning for preschoolers. Lavy et al. (2008) find that the proportion of low-achieving
peers has a negative effect on the performance of regular students.
Numerous studies have appeared that have failed to find evidence of social interactions.
Among recent studies, Oreopoulos (2003) fails to find evidence that differences in neigh-
borhood quality affect longer-run labor market outcomes, which contrasts with findings
such as Corcoran et al. Employing methods similar to Mas &Moretti, Guryan et al. (2007)
find little evidence of workplace effects in the context of golf tournaments. We do not
attempt to adjudicate the disparate results in the extant literature, which may result from
factors ranging from different socioeconomic contexts to different model specifications,
but merely observe that, taken at face value, the empirical evidence for social interactions is
not decisive.
www.annualreviews.org � Social Interactions 457
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
The regression literature on social interactions suffers from serious measurement prob-
lems. One difficulty with empirical studies based on Equation 7 (and, for that matter, its
nonlinear analogs we discuss below) is that economic theory does not dictate the appropri-
ate empirical measures of contextual variables that a researcher ought to use. As a result,
one finds Bertrand et al. (2000) using the product of welfare usage and own-ethnic group
intensity to explain individual welfare usage, whereas Aizer & Currie (2004) use the
utilization rate of an individual’s language group to measure social effects on public
prenatal-care utilization. Similarly, the empirical literature does not typically consider
how social variables should interact with individual decisions. If the reason why utilization
of social services depends on the usage of others is because of information transmission, as
argued by Bertrand et al., then it is unclear why the percentage of users is the appropriate
variable, as opposed to some nonlinear transformation, as presumably one only needs one
neighbor to provide the information. While considering this type of problem in studying
social effects in marriage markets, Drewianka (2003) argues that a higher marriage rate in
a community may reduce the propensity of unmarried people to marry as a higher rate
hampers search.
Cooley (2009) analyzes this measurement problem from a different vantage point,
namely the discrepancy between observable measures and causal ones. She argues that,
from the vantage point of theory, causal sources of peer effects between students in a
classroom involve two unobservables, endogenous effort and contextual ability, whereas
work on classroom peer effects typically uses observable classroom achievement as the
measure of outcomes and various ad hoc measures of student characteristics for contex-
tual efforts. Cooley demonstrates that absence of attention to the link between
unobserved causal factors and observable proxies can render reduced-form regressions
uninterpretable.
A distinct measurement problem arises because theory does not provide guidance as to
the appropriate measure of groups. Hence one finds the use of zip codes in Corcoran et al.
(1992) and census tracts and blocks in Weinberg et al. (2004) to determine neighborhoods.
Akerlof (1997) argues that social interactions are best understood as occurring in a social
space that may have many dimensions; this follows naturally when one considers the
overlapping effects of factors such as physical proximity, ethnicity, gender, and education
on the ways in which individuals interact. Conley & Topa (2002) are unusual in seeking
to identify the appropriate axes to situate the actors in social space, finding that ethnicity is
of particular importance in defining social interactions for labor market outcomes in
Chicago.
3.2. The Reflection Problem
A fundamental identification issue for the linear-in-means model arises from the fact
that the model is a variant of a simultaneous equations system in which the expected
choices of one agent appear as arguments in the equations describing the choices of
others. The possibility that this interdependence could cause identification to fail was
first recognized in a fundamental paper by Manski (1993), who dubbed the failure of
identification the reflection problem. To understand the problem, assume that the
within-group averages of the individual characteristics map one to one to the contextual
effects, xg(i) ¼ zg(i), so that both are r vectors. In this case, the reduced form for out-
comes (Equation 7) becomes
458 Durlauf � Ioannides
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
yi ¼ a01� b
þ axi þ baþ y1� b
xgðiÞ þ ei: ð8Þ
In this case, there are 2r þ 1 coefficients corresponding to 2r þ 2 structural parameters.
Identification thus fails. In contrast, a necessary condition for identification is that there
exists at least one individual characteristic whose group-level average is not a contextual
effect, as shown originally in Brock & Durlauf (2001c); this corresponds to the classical
row condition for identification in simultaneous equations systems. Durlauf & Tanaka
(2008) give a sufficient condition for identification, which corresponds to the rank condi-
tion in simultaneous equations. These conditions, of course, require prior knowledge on
the part of the researcher to produce exclusion restrictions and therefore fall prey to the
problem that social interactions models are typically open-ended,3 which means that
the presence of particular group-level determinants of behavior is logically independent of
the presence or absence of various individual-level determinants.
Once one moves away from a linear cross-section structure, the reflection problem may
not arise, even if there is a one-to-one correspondence between individual and contextual
effects. For example, non-linear-in-means variants generically cannot exhibit it. Brock &
Durlauf (2001c) show that the reflection problem does not occur for
yi ¼ a0 þ axi þ yzgðiÞ þ bm�mgðiÞ
�þ ei; ð9Þ
for @2m�mgðiÞ
�=@m2
gðiÞ 6¼ 0, relative to the space of twice differentiable functions, outside of
nongeneric cases. The intuition is straightforward; the reflection problem requires linear
dependence between group outcomes and certain group-level aggregates, which is ruled
out by the nonlinearity in Equation 9. Furthermore, the reflection problem does not arise in
linear models with dynamic forms of interactions, e.g.,
yit ¼ a0 þ axi;t þ yzgðiÞ;t þ bmgðiÞ;t�1 þ ei;t: ð10ÞThese examples in which the reflection problem fails to arise should be understood with
caution. The reflection problem occurs when there is linear dependence between regres-
sors; multicollinearity can still be a problem in models such as Equations 9 and 10, even
if linear independence holds, so that parameter estimates are highly imprecise. Further-
more, the examples given above in which the problem does not arise rely on parametric
modeling assumptions. For cross-section data, Manski (1993) provides a nonidenti-
fication result for nonparametric contexts that parallels the reflection problem in linear
models.
3.3. Variance-Based Approaches
The fact that endogenous social interactions help amplify differences in the average group
behavior has been used as another route for identification. Glaeser et al. (2003) introduce
the concept of a social multiplier to measure this effect. The social multiplier approach can
also be useful in delivering a range of estimates for the endogenous social effect when
individual data are hard to obtain, as in the case of crime data. Glaeser et al. (1996)
motivate their study of crime and social interactions by the extraordinary variation in the
3Brock & Durlauf (2001b) introduce the idea of model open-endedness in the context of economic growth models,
but the idea is implicit in earlier critiques of econometric practice such as Sims (1980).
www.annualreviews.org � Social Interactions 459
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
incidence of crime across U.S. metropolitan areas over and above differences in fundamen-
tals. They use the idea that the convergence of sample means is slowed down by social
interactions as a basis for assessing whether cross-city differences in crime rates are large,
once one has controlled for some city-wide characteristics. This approach only requires
city-level data. Formally, Glaeser et al. (1996) work with a statistic, the covariance of
per-capita outcomes across agents, a measure of social interactions, that is derived as a
function of the proportion of agents in a city who are either law-abiders or criminals. This
covariance generates overdispersion in cross-city crime rates relative to the absence of
social interactions. They show that the estimated index is highest for petty crimes but
declines for more serious crimes to become negligible for the most serious ones. Across
cities, the implied level of interactions is roughly constant. Because the statistic is related to
the expected size of the interacting group of criminals, their results suggest that more
serious crimes are associated with smaller group interactions.
Glaeser et al. (2003) extend the idea of variance comparisons to formalize the idea of a
social multiplier to measure how endogenous social interactions amplify private incentives.
They compute the social multiplier by comparing the coefficient vector produced by a
within-group regression across individuals,
yig ¼ kþ pxi þ �ig; ð11Þwith the coefficient produced by a regression across group-level averages,
�yg ¼ k0 þ p0�xg þ ��g: ð12Þ
Social multipliers are defined as the ratios of the coefficients in p0 to their counterparts
in p.4 This approach, of course, requires the existence of at least one individual effect, xi.
The social multiplier approach has also proven useful in uncovering b, even when no
individual or contextual effects are present. We return to this approach in the next section.5
3.4. Discrete Choice Models of Social Interactions
Discrete choice models of social interactions have been applied to a wide range of contexts,
although the model is used more frequently than its linear-in-mean analog. Early examples
include Crane (1991) and Evans et al. (1992) on teenage behaviors; more recent examples
range from cigarette smoking (Krauth 2005, Nakajima 2007) to medical practices (Burke
et al. 2003). As in the case of linear models, one finds conflicting results in the literature.
For example, Crane (focusing on residential neighborhoods) finds strong social effects,
whereas Evans et al. (focusing on schools) do not; similarly, Nakajima concludes that peer
effects play a larger role in smoking than does Krauth.
The econometrics of discrete choice models with social interactions is studied by
Brock & Durlauf (2001a,c, 2006, 2007). These papers show that the reflection problem
does not, under relatively weak assumptions, apply in multinomial choice models when the
analyst knows the distribution of the unobserved error terms; the standard assumption in
4For example, suppose that the true data-generating process is yig ¼ a0 þ axi þP
j6¼iyij þ �ig, where a is scalar. Then,one can show that p0
p ¼ 1ð1�bÞð1þsbÞ, where s ¼ Varð�xgÞ
VarðxiÞ :5Although the literature on the social multiplier has so far rested on linear models in static settings and emphasizes
measuring the strength of social interactions, the concept may be extended to dynamic and nonlinear settings
(Brock & Durlauf 2001a), to complete versus incomplete information (Bisin et al. 2006), and to more complex
interaction topologies (Ioannides 2006).
460 Durlauf � Ioannides
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
theoretical models that the errors are logit is inessential. Intuitively, identification holds so
long as the individual and contextual effects exhibit sufficient variation so as to imply that
choice probabilities are nonlinearly related to them. Further, Brock & Durlauf show that
for the binary choice case identification holds even if the error distribution is unknown,
following Manski (1988).
A number of variations of our baseline discrete choice model have been studied.
Soetevent & Kooreman (2007) use Brock & Durlauf (2001a) as a building block with an
inessential conceptual difference. Soetevent & Kooreman’s model involves interactions in
relatively small groups of given sizes in which choices of other individuals are assumed to
be fully observed, and therefore an individual’s payoff depends on the actual choice of
others in her group, as opposed to expected choices as in Brock & Durlauf. Like Brock &
Durlauf, Soetevent & Kooreman focus on pure Nash equilibria with binary outcomes,
oi ¼ 1,�1, and estimate the model in effect as a system of simultaneous equations by
means of simulation methods.6
Their empirical application examines the individual behavior of high school teenagers
in almost 500 school classes from 70 different schools using data from the National School
Youth Survey of the Netherlands. In their baseline model, endogenous social interactions
effects are strong for behavior closely related to school (truancy), somewhat weaker for
behavior partly related to school (smoking, cell-phone ownership, and moped ownership),
and absent for behavior far away from school (asking parents’ permission for purchases).
Intragender interactions are generally much stronger than cross-gender interactions. When
school-specific fixed effects are allowed, social interactions effects are insignificant, with
the exception of intragender interactions for truancy. This work is also significant because
its simulation-based estimation method allows the authors to account for the potential
multiplicity of noncooperative Nash equilibria and the identification problems it poses.
Another interesting variation of the discrete choice model is developed in Aradillas-
Lopez (2008, 2009), who considers the question of identification of social effects when
agents’ expectations are not necessarily correct. Aradillas-Lopez (2008) focuses on the
empirical implications of social interactions when beliefs are only required to be consistent
with the iterated elimination of dominant strategies. Probability bounds are established for
social interactions parameters that depend on the number of iterations. Aradillas-Lopez
(2009) extends this reasoning to consider environments in which agents’ behavior is
required to obey a criterion called weakly consistent equilibrium. This imposes that the
choices of other agents always lie in the support over which an agent forms beliefs about
these behaviors. These approaches provide a link between the social interactions literature
and developments in behavioral economics.
In terms of empirical work, the discrete choice model has been extended to consider
dynamic contexts and in turn has been applied to duration and optimal stopping problems.
Sirakaya (2006) uses a national U.S. sample to identify the risk factors for recidivism
among female, male, black, white, and Hispanic felony probationers. She assumes that
the individual hazard function for time to recidivism depends on various individual and
neighborhood characteristics, as well as social interactions among probationers. Because a
probationer is required to live in the jurisdiction that passed the probation sentence, the
neighborhood of probationer i, g(i), is assumed to be the jurisdiction. Sirakaya’s results
6With many agents, the two formulations are equivalent, but the finite-agent case raises issues of identification that
are common in models of systems of simultaneous outcomes (see Tamer 2003).
www.annualreviews.org � Social Interactions 461
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
point to social interactions as one of the most significant factors affecting recidivism within
all gender, ethnicity, and race groups, with unobserved neighborhood-level heterogeneity
being negligible.
A number of other studies are closely related to Sirakaya’s methodological frame-
work. Irwin & Bockstael (2002) argue that the development of adjacent parcels of land
in exurbia may confer social benefits, if proximity is desired, but also costs, if conges-
tion and environmental degradation arise; their empirical work finds the second effect is
dominant. They estimate net social interactions effects in exurban land-use development
that are negative. A second example is Costa & Kahn (2007), who use a longitudinal
data set of Union Army soldiers and a cross-sectional data set of the population of
Andersonville to examine the role of social networks in ensuring the survival of Union
Army soldiers in captivity. They estimate hazard functions for individual survival prob-
abilities as functions of the number of friends, individual characteristics, and camp
conditions and use the cross-sectional data to estimate a probit model of the probability
of survival as a function of the number of friends and of demographic characteristics.
They find that friends had a statistically significant positive effect on survival probabil-
ities and that the closer the ties were between friends as measured by such identifiers as
ethnicity, kinship, and the same hometown, the bigger the impact of friends on survival
probabilities was. De Paula (2009) uses the same data and finds evidence of bunching in
desertions that is consistent with social interactions. His analysis is based on a sophisti-
cated optimal stopping problem with endogenous social interactions whose econometric
properties are carefully examined.
4. IDENTIFICATION: CONTEMPORARY PERSPECTIVES
In our discussion of social interactions, a best-case assumption on the error structure is
assumed, namely that individual unobservables are exchangeable across individuals and
groups. Exchangeability in turn is operationalized as independence. Much of modern
econometric practice, of course, has developed in a systematic effort to avoid such assump-
tions (see Heckman 2001). In this section, we explore the implications of alternative error
structures, due to self-selection and group-level unobservables.
4.1. Identification of Social Interactions with Self-Selection toGroups and Sorting
For many contexts, economic agents choose their group memberships, and these choices
are influenced by social interactions effects. Endogenous group formation naturally affects
the distribution of the unobserved heterogeneity within groups, which depends on both
endogenous and contextual effects, as well as individual specific characteristics, i.e., in
Equation 6, Efei jxi; zgðiÞ;mgðiÞ; i 2 gðiÞg 6¼ 0, the standard self-selection problem. Because
self-selection into a group means that each member of the group has been influenced by a
common set of social factors, this leads to dependence in the conditional expectations
across individuals, which Manski (1993) terms correlated effects.
The social interactions literature has followed the broader econometric literature’s
strategies to address self-selection. A number of studies have employed instrumental vari-
ables methods (see Evans et al. 1992). Instrumental variable approaches are problematic
due to theory open-endedness, which is discussed above. Social interactions models do not
462 Durlauf � Ioannides
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
naturally generate valid instruments; rather instruments are produced via intuition. This
contrasts with contexts such as Euler equation estimation, in which the theory, which
implies combinations of forecast errors, itself generates instruments for the analyst.
A second strategy follows the control function approach pioneered by Heckman (1979).
From this vantage point, self-selection is addressed by accounting for the behaviors that
generate it. Following Heckman, suppose that evaluation of the attractiveness of a group g
may be expressed in terms of an unobservable latent quality variable Q�i;gðiÞ. Individual i
evaluates group g by means of observable attributes Wi,g(i), which enter with weights z andunobservable component #i,g(i),
Q�i;gðiÞ ¼ zWi;gðiÞ þ #i;gðiÞ; ð13Þ
and chooses the group with the highest Q�i;gðiÞ. Random shocks ei, from Equation 6 and
#i,g(i) above, are assumed to have zero means, conditional on (orthogonal to) regressors
[xi, zg(i),Wi,g(i)], across the population. Under various parametric and semiparametric
assumptions on the joint distribution of�ei; #i;gðiÞ
�; d�z;Wi;gðiÞ;Wi;�gðiÞ
�, an object that con-
verges to a value proportional to Efei j xi; zgðiÞ; i 2 gðiÞg can be constructed and used as a
regressor in Equation 6, now corrected for selection bias:
yi ¼ a0 þ axi þ yzgðiÞ þ bmgðiÞ þ kd�z;Wi;gðiÞ;Wi;�gðiÞ
�þ xi: ð14Þxi is now uncorrelated with the other regressors. Note that the selection correction term
depends on the attributes of all groups in the opportunity set, not only the one that is
chosen (see Brock & Durlauf 2001c, 2006 for further discussion).
Equation 14 integrates information about the group selection process into the behav-
ioral equation for outcomes and in a way that provides additional information on social
interactions. The additional regressor d�z;Wi;gðiÞ;Wi;�gðiÞ
�affects the possibility of the
reflection problem’s arising. To see why, consider two possibilities. First, suppose that
groups are chosen according to the individuals’ characteristics xi and the group’s charac-
teristics zg(i). The additional regressor will then take the form d�z; xi; zgðiÞ; z�gðiÞ
�, which is
an individual specific random variable (its group-level average does not appear in the
equation). Second, suppose that groups are chosen according to expected average behavior
in the group. In this case d will functionally depend on mg(i). Because d is almost always a
nonlinear function, its introduction transforms the linear-in-means model to a nonlinear
one. Both these changes are sufficient to produce identification, even if the model were not
identified under random assignment. The ability of self-selection to facilitate identification
is not surprising. Group choice naturally contains information about the determinants of
the choice. What is perhaps surprising is that randomized group assignment would work to
the detriment of identification.
Ioannides & Zabel (2008) is the first empirical study to use self-selection to facilitate
identification of social interactions. Studying housing demand, Ioannides & Zabel merge
the neighborhood cluster subsample with confidential information from the National
American Housing Survey. Their estimation of the demand equation for housing structure
allows for contextual effects at the level of the immediate neighborhood and at the census
tract of residence, and for endogenous social effects, defined as the mean housing demand
among immediate neighbors. Their estimation results confirm that the endogenous neigh-
borhood effect is significant and stronger after controlling for the endogeneity of neighbor-
hood choice.
www.annualreviews.org � Social Interactions 463
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
Self-selection also suggests that evidence of social interactions may derive from the
prices for group membership. For example, social effects of residential neighborhoods will
be reflected in housing prices, via standard hedonic price arguments. Not surprisingly, in
view of Brock & Durlauf’s (2001c) results on the ability of self-selection to facilitate
identification, Nesheim (2002) shows that preferences and the technology of social inter-
actions, measured as children’s schooling as a function of the mean parental education in
the neighborhood and of parental characteristics, may be identified. Ioannides (2008)
shows how certain functional forms produce a closed-form solution for the neighborhood
hedonic price in terms of the mean parental education, which should facilitate empirical
implementation.7
A related strategy for integrating group choice with social interactions involves estima-
tion of equilibrium models of neighborhood formation in which social interactions play a
role in locational decisions. Bayer et al. (2007) employ a discrete choice model to estimate
preferences over neighborhood composition with respect to housing choices. Calabrese
et al. (2006) estimate a sophisticated structural model that explicitly addresses the political
economy of local educational expenditure in which neighborhood composition is shown to
affect neighborhood choice.
One general feature of the various approaches to sorting and social interactions is that
they assume that once groups are formed, agents in a common group experience similar
interactions. Weinberg (2007) makes an original argument that interactions may occur at a
lower level of aggregation than the level at which groups are formed. For example, within a
classroom, friendships may occur among students with similar characteristics. Weinberg
provides empirical evidence that this type of phenomena occurs. He further makes the
argument that the capacity of intragroup sorting into subgroups will depend on group size,
as larger groups will have more opportunities for individuals to find similar types with
which to form bonds. Social interactions in social networks, which we discuss further
below, carry this notion further.
4.2. Group-Level Unobservables
The second basic problem facing econometric analyses of social interactions is the pres-
ence of unobserved group effects. In a classroom, for example, differences in teacher
quality will simultaneously affect all students in a classroom. From the perspective of our
basic social interactions models, these may be understood as replacing the initial error
assumptions in Equation 6 with ag(i) þ ei, where ag(i) is a common group-specific shock.
Of course, one can imagine more elaborate changes in the error structure; these have not
been much studied.
In our view, unobserved group effects represent the most difficult hurdle to the con-
struction of persuasive evidence of social interactions because, unlike self-selection, there is
typically no economic reasoning to facilitate modeling the influences. There is thus no
natural solution to group effects analogous to the joint modeling of group formation and
behaviors within groups. The methods that exist for overcoming their presence are thus
essentially statistical.
7Bayer et al. (2007) report some nonstructural hedonic regressions of housing prices on neighborhood characteristics
[see also Bayer & Ross (2009) who use neighborhood prices to construct a control function to proxy for unobserved
neighborhood characteristics].
464 Durlauf � Ioannides
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
One approach uses panels to difference out the group effects; Hoxby (2000) is a good
example. Intuitively, variation in zg(i) over time induces variation in mg(i) and therefore
produces identification in the differenced model. For Hoxby’s model, zg(i) is the percentage
of a student’s own ethnic group in a classroom. Brock & Durlauf (2007) show that this
argument generalizes to panels with discrete data, using ideas in Chamberlain (1984). This
justifies the approach taken by Arcidiacono & Nicholson (2005) in evaluating the effects
of peers on specialty choice in medical school. Differencing, of course, assumes that ag(i) isitself not time varying. Furthermore, the approach treats the individual fixed effects
as nuisance parameters. Arcidiacono et al. (2007) consider a linear model in which the
spillover effects operate through the fixed effects of a student’s peers, making fixed effects
for one agent influence other agents. They provide conditions under which these effects can
be exploited in a panel to facilitate identification of social interactions.
Differencing becomes problematic for models with nonadditive error structures. Cooley
(2008) shows how social interactions may be identified for a class of models with nonad-
ditive group effects, exploiting results in Imbens & Newey (2009). Cooley applies the
strategy to school data in North Carolina, finding strong evidence of intraethnic group
social interactions in achievement. Cooley’s approach does require the use of instru-
mental variables, which, of course, constitutes a second general strategy for dealing with
unobserved group effects, albeit one that we argue above can be hard to justify because of
theory open-endedness. Given that unobserved group effects are undertheorized, open-
endedness is especially hard to overcome for this case. However, as the careful discussion
in Cooley shows, it is not impossible given a specific context.
4.2.1. Variance-based approaches to uncovering social interactions with unobserved
group effects. An alternative strategy for uncovering social interactions in the presence
of unobserved group effects is based on the variance-covariance structure of outcomes.
Graham (2008) extends earlier variance comparisons such as Glaeser et al. (1996), who do
not allow for unobserved group effects. Following Graham, we rule out individual contex-
tual effects and assume that behaviors depend on the realized rather than expected choice
of others, producing the model
yi;gðiÞ ¼ kþ b�ygðiÞ þ ei;gðiÞ ¼ kþ bmgðiÞ þ b½�ygðiÞ �mgðiÞ� þ ei;gðiÞ: ð15Þ
Obviously, relative to Equation 15, no instruments exist formg(i). However, Graham shows
that identification can still be achieved. He does this by employing an approach that echoes
the classical literature on identification in simultaneous equations models, in which the
structure of the variance-covariance matrix of the reduced-form system augments identifi-
cation via exclusion restrictions. The key to Graham’s method is the assumption that
unobserved group effects are random rather than fixed, which imposes structure on the
variance-covariance matrix of outcomes. The random-effects assumption can be unappeal-
ing in contexts where groups are endogenously formed.
To see how a random-effects assumption on group unobservables facilitates identifica-
tion, decompose the residuals in Equation 15 as
ei;gðiÞ ¼ agðiÞ þ �i;gðiÞ; ð16Þwhere ag(i) is a group-level unobservable, and �i,g(i) is an individual-level unobservable;
each has mean zero and constant variance across i and g, s2a and s2� , respectively; we assume
www.annualreviews.org � Social Interactions 465
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
the two variables are uncorrelated with each other. Let the variance-covariance matrix of
the ei,g(i)’s be Vn. These assumptions imply that the variance of average outcomes in groups
of size n equals
Var½�yg� ¼ In � bn1n
� ��1
Vn In � bn1n
� ��1
: ð17Þ
This expression indicates how the parameters b; s2a ; and s2� can be recovered from size-
specific group variances; e.g., calculating Equation 17 for different n’s reveals these param-
eters. Graham (2008) and Durlauf & Tanaka (2008) discuss ways to modify Graham’s
baseline assumptions and preserve identification. An especially valuable study is Davezies
et al. (2006), which provides a systematic discussion of the use of the variance-covariance
matrix of residuals for the linear-in-means model to facilitate identification; it also pro-
vides results for binary choice models. A nice feature of their work is the clarification of
when homoskedasticity assumptions are needed.
Graham applies his methods to data from Project STAR, a class-size reduction experi-
ment in Tennessee. In that experiment, kindergarten students and teachers were randomly
assigned to large and small classrooms. Random assignment of teachers ensures that the
distribution of their characteristics is similar across the two types of classrooms, and
Equation 17 may be exploited. Graham (2008) reports differences in peer-group quality
that constitute evidence of social interactions.
4.2.2. Variance components and social interactions. The use of variance information has
been exploited in a different way by Solon et al. (2000). Their analysis focuses on the case
in which one cannot observe either family-specific or group-specific effects. Solon et al.
propose a variance decomposition of individual outcomes to bound the contribution of
group effects. One can formulate their analysis in terms of a variance components model
(Searle et al. 2006):
yi;f;n ¼ mf þ nn þ of;n þ ei;f;n; ð18Þwhere mf denotes a family effect, nn denotes a group effect, of,n denotes an interactive
effect between family and group, and ei,f,n denotes an idiosyncratic effect. Assuming
that these components are orthogonal (and such a decomposition always exists), one
can calculate the variance contribution of nn to the overall variance of yi,f,n. Solon et al.
do not use this decomposition directly; instead they use sibling versus neighbor covari-
ances to bound the variance of nn, finding little evidence that neighborhoods affect
education.
One limitation of this approach is that it reduces the vector of social interactions to a
scalar so that one cannot tell which social factors matter. In fact, it is possible for different
social factors to cancel each other out. Another limitation is that the presence of of,n can
make interpretations problematic, although Solon et al. argue that one can assume the term
is positive due to sorting across neighborhoods. For these reasons, Oreopoulos (2003) is
interesting. This paper uses data from Toronto that allow one to trace the adult labor
market outcomes of children who grew up in different public housing projects; these pro-
jects differ in terms of neighborhood characteristics such as crime. Random assignment
across public housing projects eliminates correlations between individual and neigh-
borhood characteristics that cloud the variance decomposition (Equation 18). That said,
466 Durlauf � Ioannides
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
Oreopoulos also finds that social interactions fail to play a large role in explaining differ-
ences in adult economic outcomes.
4.2.3. Partial identification with nonlinear models. A final approach to unobserved group
effects is due to Brock & Durlauf (2007), who develop partial identification results for
binary choice models under what they argue are weak assumptions about the distribution
of the unobserved group effects. This work exploits a fundamental difference between
endogenous effects and unobserved group effects: Only endogenous effects can produce
multiple equilibria. Hence, if evidence of multiple equilibria can be generated, it represents
evidence of endogenous social interactions.
Brock & Durlauf operationalize this idea in several ways. First, they consider the use of
pattern reversals in group-level outcomes. A pattern reversal occurs when the rank order of
average outcomes between two groups is the reverse of what one would predict given the
observed individual and contextual effects for the groups. Brock & Durlauf demonstrate
that, under various shape restrictions on the probability density of the unobservables,
pattern reversals can only occur because of multiple equilibria and hence endogenous
social interactions. For example, if the distribution of unobservables shifts monotonically
in observables, then pattern reversals cannot occur without social interactions. A second
approach involves the bimodality of linear combinations of contextual effects. Here
Brock & Durlauf show that, conditioning on a given average outcome, the cross-section
distribution of certain linear combinations of contextual variables must be unimodal.
This last result corrects a misconception that multiple equilibria induce multimodality of
average outcomes conditional on contextual effects, whereas multiple equilibria do induce
a mixture density of outcomes; not all mixtures are multimodal.
Brock & Durlauf (2010) extend this perspective to the study of social effects in technol-
ogy adoption. In this analysis, individuals are motivated to adopt a new technology by
both individual observable and unobservable characteristics (which may be correlated
across individuals) as well as by the percentage of the population that has adopted, which
defines the social interactions effect. They provide conditions under which a higher adop-
tion rate by individuals whose observable fundamentals predict a lower adoption rate than
another group can only occur because of social effects. This pattern reversal does not occur
because of multiple equilibria but rather because of the potential for social effects to lead
agents with different private incentives to bunch and adopt simultaneously. As such, the
analysis complements the more structural one of de Paula (2009), as well as a subtle
investigation by Young (2009) on the observational differences between learning and
conformity models in dynamic environments.
Although partial identification arguments of the Brock-Durlauf type have not been used
empirically, one does find analyses that may be interpreted in this way. For example, Card
et al. (2008) evaluate Schelling’s classic tipping model of segregation by a search for
discontinuities in the relationship between changes in neighborhood racial composition
and initial racial composition. We conjecture that this achieves partial identification of the
effect of neighborhood composition on individual utility so long as one imposes a continu-
ity assumption on the density of unobservables, which is what Brock & Durlauf (2010) call
a shape assumption.
For many forms of social interactions, endogenous group formation can produce assor-
tative matching. One intuition follows from the Becker (1973) marriage model and can be
generalized in a number of ways (Durlauf & Seshadri 2003): Social interactions can
www.annualreviews.org � Social Interactions 467
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
represent complementarities and mean that stratification is efficient. One example for this
is human capital complementarity across workers. In other contexts, such as neighbor-
hoods, social interactions can produce incentives to stratify regardless of whether the
stratification is efficient; Becker & Murphy (2000) give a nice exposition of this property
for residential neighborhoods. These theoretical results suggest that another strategy for
uncovering social interactions could involve the evaluation of degrees of stratification
across groups. Although there does exist a literature on the identification of complemen-
tarities via stratification patterns (see Siow 2009 for a recent example), this method has yet
to be used to uncover social interactions per se. The closest example of this strategy is Card
et al. (2008), discussed above, as it uses changes in the degree of segregation as the basis for
inferences. It should be noted that methods for estimating neighborhood choices as exem-
plified in Epple & Sieg (1999) are explicitly designed to explain the intragroup distribu-
tions of characteristics, so our suggestion should be understood as complementary to this
structural style of work.
Although we focus on partial identification in social interactions contexts, there is a
broader literature on partial identification in games, much of which is driven by multiple
equilibrium issues. Tamer (2003) is a seminal contribution.
4.3. Experiments
As is true for much of economics, concerns over the assumptions needed to achieve
identification with observational data have led to substantial interest in experiments and
quasi-experiments as ways to uncover social interactions. The experimental approach has
seen little direct work on econometric issues, relying on the existing treatment effect
methods to study social interactions. One exception is Hirano & Hahn (2009), who study
design issues for social interactions experiments. Another is Graham et al. (2008), who
formulate measures of changes in group composition on group outcomes that generalize
treatment effect calculations to explicitly consider social interactions.
At one end of the scale, interest in the effects of residential neighborhood effects led the
Department of Housing and Urban Development in 1994 to implement the Moving to
Opportunity (MTO) demonstration in Baltimore, Boston, Chicago, Los Angeles, and New
York. The program provides housing vouchers to a randomly selected group of families
from among residents of high-poverty public housing projects. Within this subsidized
group, families in turn were randomly allocated between two subgroups: one that received
unrestricted vouchers and another that received vouchers that could only be used in census
tracts with poverty rates below 10% (these users are termed the experimental group).
Members of the experimental group also received relocation counseling.
The data from the MTO program have been used by a number of researchers to
evaluate social interactions. Kling et al. (2007) is the state-of-the-art study both in terms
of methodology and coverage across all study sites; earlier analyses include Ludwig et al.
(2001) and Kling et al. (2005). Following Kling et al. (2007), social interactions are
inferred when evidence is found that movement to a low-poverty neighborhood affects
socioeconomic outcomes. In doing this, the authors calculate both intent to treat and
treatment on the treated effects, corresponding to the effects of receipt of the voucher and
use of the voucher, respectively.
In what is the best overview of the MTO program, Kling et al. (2007) find mixed
evidence on the effects of movement to lower-poverty communities, even though the
468 Durlauf � Ioannides
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
housing voucher lottery caused otherwise similar groups of families to reside in very
different neighborhoods. For adults, their analysis revealed little evidence that movement
to low-poverty neighborhoods affected economic self-sufficiency although some improve-
ments occurred in mental health. In terms of the children, they find beneficial effects for
teenage girls with respect to education, risky behavior, and physical health. Alternatively,
for teenage boys, relocation to lower-poverty neighborhoods was associated with deterio-
ration of a number of socioeconomic outcomes, in particular with respect to criminal
behavior. The evidence that effects of housing vouchers appear to accrue from changes in
neighborhood characteristics rather than from moves per se suggests that interventions
that substantially improve distressed neighborhoods could have effects at least as large as
those observed from moving to lower-poverty neighborhoods.
Although the MTO program is a unique evidence source, there are a number of limita-
tions to the use of existing findings of program effects to either draw policy implications or
even to infer the empirical salience of social interactions. One general problem is that only
one-quarter of eligible families signed up for MTO, which also suggests that decisions are
influenced by unobservables. These statistics also suggest that the respective populations
are not representative. In terms of policy, one criticism is due to Sobel (2006), who argues
that different respondents’ outcomes are correlated, which is known as interference among
respondents. Using the MTO demonstration as a concrete context, Sobel provides a causal
framework for policy evaluation when interference is present. He characterizes the proper-
ties of the usual estimators of treatment effects, which are unbiased and/or consistent in
randomized studies without interference. When interference is present, the difference
between a treatment group mean and a control group mean does not estimate an average
treatment effect, but rather the difference between two effects defined on two distinct
subpopulations. This result means that a researcher who fails to recognize it could infer
that a treatment is beneficial when in fact it is not. Kling et al. (2007) reject the argument of
interference by pointing to the fact that 55% of household heads who signed up for MTO
had no friends and 65% had no family in their baseline neighborhoods. However, to the
extent that interference did not occur, then it is unclear how to extrapolate the program to
larger policy interventions. In particular, Durlauf (2004) argues that moving large numbers
of poor families to more affluent communities will induce general equilibrium effects in
terms of the location decisions of other families and the ability of schools in these neighbor-
hoods to provide needed services, for example. One can additionally imagine that the
commitment of affluent families to public schools would be strained by a massive influx
of poor families into their communities.
In terms of the question of the empirical salience of social interactions, MTO studies
suffer from the lack of attention to mapping between treatment parameters and the under-
lying social interactions. A range of neighborhood characteristics generate social interac-
tions effects; the role model effects of affluent neighbors are quite distinct from the support
mechanisms that are generated by proximity to friends and relatives. MTO families who
moved to low-poverty neighborhoods did not simply move from worse to better neighbor-
hoods; rather they traded off one vector of neighborhood attributes for another. This also
renders the interpretation of positive evidence of voucher effects problematic. One exam-
ple is discussed by Kling et al. (2007): the reduction in asthma among children who move
to low-poverty neighborhoods. This could result from stress reduction; to the extent this is
generated by neighborhood characteristics, this is a social interactions effect. Another
reason for lower asthma rates, however, could be reduced exposure to vermin infestations,
www.annualreviews.org � Social Interactions 469
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
induced by changes in housing quality, in turn induced by the requirement that the
vouchers were used in low-poverty neighborhoods, which is not a social interactions effect.
A second class of experimental studies has been conducted in laboratories, such as
Falk & Ichino (2006). Their experiment involves workers who are assigned in pairs to
stuff envelopes, with a control group being provided by subjects working alone in a room.
These authors find that standard deviations of output are significantly smaller within pairs
than between pairs, and average output per person is greater when subjects work in pairs.
They also predict theoretically that peer effects are asymmetric; low-productivity workers
are more sensitive to the behavior of high-productivity workers as peers, and they find
indirect evidence that this holds. Although the findings of the effects of working in pairs
versus working singly are quite clean, they do not reveal why the effects are occurring.
A third class of studies involves natural experiments, in which the assignment of indi-
viduals to groups has occurred in a way that facilitates identification of social effects.
Sacerdote (2001) exploits the effects of randomly assigned freshman-year room- and
dormmates at Dartmouth College. Sacerdote estimates Equation 6 with an individual’s
grade point average as a dependent variable, as a function of an individual’s own academic
ability and social habits, and the academic ability and grade point average of roommates,
finding that peers have an impact on each others’ grade point average and on decisions to
join social groups such as fraternities. He does not, however, find residential peer effects in
decisions such as choice of college major. Interestingly, peer effects are smaller the more
directly a decision is related to labor market activities.
A final class of studies follows the spirit of natural experiments but does not rely on
identifying contexts where random assignment holds, but rather on other data features.
Young & Burke (2001) use differences in the terms of tenant/landowner agriculture con-
tracts across Illinois to evaluate the importance of social norms on contract terms. They
find that contract terms tend to focus on simple crop-sharing rules that cannot be
explained by standard contract theory.
Although we emphasize some limitations of experimental studies, this does not imply
either that the studies are uninformative or that their utility is lower than observational
studies. Rather, we seek to combat the presumption that experimental studies trump
observational ones and that randomization should be regarded as the gold standard for
empirical work.
5. SOCIAL INTERACTIONS IN SOCIAL NETWORKS
Our discussion of empirical work above focuses on social interactions models in which
group influences are generated by relatively crude aggregates, normally the average char-
acteristics or behaviors of others. In contrast, social networks models provide a primary
focus on the microstructure of interactions emphasizing heterogeneity in these interactions
across individual pairs. Jackson (2008) reviews the new theories. Here are examples of
social interactions analyses that have been enriched by network thinking.
5.1. Models
One reason social networks matter in social interactions analysis is because they facilitate
identification by breaking the reflection problem. This was originally recognized in Cohen-
Cole (2006) and is systematically explored in Bramoulle et al. (2009). Using Y to represent
470 Durlauf � Ioannides
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
the I-column vector of individual actions, we may generalize the linear-in-means model
(Equation 6) to allow for an arbitrary set of pairwise interactions. In a linear setting
(Equation 6), this amounts to replacing bmg(i) with bJY. Following Bramoulle et al., we
employ the former and further assume that Jik ¼ 1, if i is influenced by k, and Jik ¼ 0,
otherwise. In addition, the endogenous effect is defined as bJY. In other words, each
pairwise interaction has an endogenous interaction weight b or a weight zero, and every
contextual effect has a weight y or zero, depending on whether the agents are directly
linked. Recalling that i is a vector of ones, the vector counterpart of Equation 6 is thus
Y ¼ a0iþXaþ JXyþ bJYþ e; ð19Þwhere a and y are r-column vectors of parameters, and X is an I � r matrix of individual
effects. This system of equations retains a single parameter b for the endogenous social
effect but allows for heterogeneous interactions among individuals.
If jbj < 1, then I � bJ is invertible, and provided that each individual is socially con-
nected with at least one other individual, Equation 19 yields the reduced form:
Y ¼ 1
1� ba0iþXaþ
X1k¼0
bkJkþ1Xðabþ yÞ þX1k¼0
bkJke: ð20Þ
Bramoulle et al. (2009, proposition 1) use Equation 20 to obtain the unique solution for
EfJY j Xg and establish that identification of social effects in Equation 19 requires that
both ay þ b 6¼ 0 and matrices I, J, and J2 be linearly independent. Moreover, when no
individual is isolated, the matrices I, J, and J2 are linearly dependent if and only if
EfJY j Xg is perfectly collinear with regressors (i,X, JX). Otherwise, the network structure
is rich enough to make the variables (i, X, JX, J2X, J3X; . . . ;) serve as appropriate instru-
ments that allow endogenous and exogenous effects to be identified. Bramoulle et al. also
obtain identification results for the model with correlated effects, as in Equation 21 below.
Why does this condition allow for identification? Intuitively, accounting for network
structure means that each individual i’s behavior is associated with a set EfYgðiÞg ¼ JEfYgof variables that need to be instrumented. Consider two other agents, j and k, such that
Jij ¼ 0 (i and j are not socially connected) is imposed a priori. When Jik ¼ 0 has not been
imposed a priori, one can use the individual characteristics for j to instrument for EfYgðkÞgso long as j does directly affect k. In other words, by setting certain coefficients equal to
zero, network structure produces exclusion restrictions that are analogous to the rank and
order condition in simultaneous equations analysis.
There exists a close relationship between social interactions and spatial econometrics
models that was only recently recognized by researchers.8 Lee (2007), Lee et al. (2010),
and Lin (2008) estimate a version of the social interactions problem with a network
structure, which is specified in terms of groups g ¼ 1; . . . ; G:
Yg ¼ bJgYg þXaþ JgZyþ a0gig þ eg; ð21Þwhere Z is an ng � r matrix of contextual effects, Yg is the column ng vector of outcomes
for each member of group g, ng is the group’s size, Jg is group g’s social weights matrix, ig isa column ng vector of ones, eg is a column ng vector of shocks, and a0g is a group-specific
8The canonical version of the linear social interactions model, in vector form (Equation 19), with y ¼ 0, coincides
with the classic Cliff-Ord model of spatial autocorrelation with one spatial lag, JY, but no spatial error structure (see
Cliff & Ord 1981).
www.annualreviews.org � Social Interactions 471
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
fixed effect. The network structure in Equation 21 generalizes the standard spatial auto-
correlation (SAR) model. The shocks eg’s may have a general variance-covariance matrix.
However, the spatial econometrics literature has typically assumed that
eg ¼ rJg;eeg þ Rg; ð22Þwhere Jg,e is a matrix of coefficients that corresponds to network structure in the errors
for group g, which may or may not be the same as Jg (see Lee et al. 2010), and Rg denotesan ng column of independently and identically distributed shocks, which allows for infor-
mation from the covariance matrix of errors to assist identification.
A comparison of Equation 19 and Equations 21 and 22 indicates how network and
spatial models are interrelated. Each assumes prior knowledge of the members of each
agent’s group. Each assumes that interactions within a group are homogeneous; i.e., the same
weight is assigned to each member of the group. The key differences lie in the introduction of
an error structure corresponding to the network structure in the spatial econometrics case.
5.2. Empirics of Social Interactions in Social Networks
As emphasized above, network models require prior knowledge of the network structure.
Relatively few data sets have this feature. For this reason, economists have been especially
interested in the Add Health data set. Add Health, the National Longitudinal Study of
Adolescent Health (http://www.cpc.unc.edu/addhealth), is a U.S.-based study designed to
facilitate exploration of the causes of health-related behaviors of adolescents in grades 7
through 12, the structure of adolescent friendship networks, and individual outcomes in
young adulthood. It seeks to examine how social contexts (families, friends, peers, schools,
neighborhoods, and communities) influence adolescents’ health and risk-taking behaviors.
It has become a major resource for social networks analysis. In the remainder of this
section we describe some examples of interesting studies using this data set.
Fryer & Torelli (2006) use academic achievement as a definition of “acting white” by
nonwhite students to examine its relationship to popularity. They measure popularity in terms
of a network-specific spectral popularity measure, which identifies popularity of the members
of a group with a measure of the intensity of the social connections among the members of
that group. Specifically, this study can be seen as a simplified version of Equation 19, in
which, instead of the full system of simultaneous equations, achievement as an outcome is
related to a popularity statistic based on J in Equation 19. They demonstrate that there are
large racial differences in the relationship between popularity and academic achievement.
Calvo-Armengol et al. (2009) estimate individual school performance as a function of
the topology of their friendship networks, defined as embedded in different schools, while
controlling for individual characteristics, such as family background controls, residential
group variables, contextual effects, and network component fixed effects. Specifically, they
estimate Equation 21 imposing b ¼ 0 and assuming a given adjacency matrix for the social
network J, that is, Ji,j ¼ 1, if agents i and J are directly connected, and Ji,j ¼ 0, otherwise.
The error structure consists of network component-specific fixed effects plus a vector of
shocks e that satisfy
e ¼ aNg þ fJgeþ R; ð23Þwhere Jg is the group g–specific adjacency matrix, and Ng is a vector whose i element
is the number of direct links for each person in group g(i). Because the error structure
472 Durlauf � Ioannides
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
represents individual outcomes that are not explained for by individual characteristics,
X, and contextual effects, JX, it proxies for correlated effects and peer effects. Their
estimation of the stochastic structure of Equation 21 with Equation 23 thus subsumes
endogenous social effects into the estimation of (a, f) above. Their results suggest that
a one-standard-deviation increase in their centrality index translates into approximately
7% of a standard deviation in education outcomes. Interestingly, these authors’ esti-
mates of the effect on a school performance index of the number of each individual’s
friends and of the peer effect, (a, f), change little if social interactions are allowed to
be directional.
6. CONCLUSIONS
Scholarly interest in social interactions has rapidly expanded in many areas of economics
and has led to numerous methodological and empirical advances. For theorists, key chal-
lenges include the development of analytically tractable frameworks that allow for the
analysis for social interactions contexts in ways that accommodate increasingly rich forms
of intertemporal as well as cross-section heterogeneity.
In parallel, we see a need for a more constructive role for computer simulations when
analytical tractability is impossible. For theorists as well as econometricians, a primary
challenge is the complete integration of the measurement of social interactions with the
joint processes of group formation and subsequent behaviors within groups. The social
interactions literature has paid inadequate attention to markets both in the sense of model-
ing the mechanisms by which groups are formed and in understanding how social influ-
ences manifest themselves in equilibria. For example, socially determined effort in a firm is
manifested in wages.
Social forces, whether manifested through neighborhoods, schools, communities, or
firms, impact much of what constitutes modern economic life. We therefore conjec-
ture that social interactions become an increasingly important area of research.
SUMMARY POINTS
1. The integration of social interactions into theoretical economic models deserves
attention.
2. The identification of social interactions relies on features of different types of
data.
3. Relationships between social interactions models and spatial and network models
have only recently attracted serious scholarly attention.
FUTURE ISSUES
1. The integration of richer interactions structures into social interactions models
will benefit from endogenizing links.
2. The integration of endogenous network formation and the identification of social
interactions deserve attention in future research.
www.annualreviews.org � Social Interactions 473
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
3. The use of price data to infer social interactions effects continues to be an
important topic.
4. The use of transition versus steady-state data to infer social interactions effects
should attract attention.
DISCLOSURE STATEMENT
The authors are not aware of any affiliations, memberships, funding, or financial holdings
that might be perceived as affecting the objectivity of this review.
ACKNOWLEDGMENTS
Durlauf’s work was supported by the National Science Foundation grant SES-0518274,
the University of Wisconsin Graduate School, and the Laurits R. Christiansen Chair in
Economics. The authors thank Yann Broumoulle, Jane Cooley, Andres Aradillas-Lopez,
and Adriaan Soetevent for insightful comments and Nonarit Bisonyabut and Xiangrong Yu
for excellent research assistance. Durlauf thanks Buz Brock for years of collaboration and
conversation on social interactions that have profoundly affected his thinking.
LITERATURE CITED
Aizer A, Currie J. 2004. Networks or neighborhoods? Interpreting correlations in the use of publicly-
funded maternity care in California. J. Public Econ. 88:2573–85
Akerlof G. 1997. Social distance and social decisions. Econometrica 65(5):1005–27
Akerlof G, Kranton R. 2002. Identity and schooling: some lessons for the economics of education.
J. Econ. Lit. 40(4):1167–201
Aradillas-Lopez A. 2007. Semiparametric efficiency in models with rational expectations and strategic
interactions. Unpublished manuscript, Princeton Univ.
Aradillas-Lopez A. 2008. Using rationalizable bounds for conditional choice probabilities as control
variables. Unpublished manuscript, Univ. Wisconsin, Madison
Aradillas-Lopez A. 2009.Nonparametric probability bounds in a class of simultaneous discrete choice
models. Unpublished manuscript, Univ. Wisconsin, Madison
Arcidiacono P, Foster G, Goodpaster N, Kinsler J. 2007. Estimating spillovers using panel data, with
an application to the classroom. Unpublished manuscript, Duke Univ.
Arcidiacono P, Nicholson S. 2005. Peer effects in medical school. J. Public Econ. 89:327–50
Bayer P, Ferreira FV, McMillan R. 2007. A unified framework for measuring preferences over schools
and neighborhoods. J. Polit. Econ. 115(4):588–638
Bayer P, Ross S. 2009. Identifying individual and group effects in the presence of sorting: a neighbor-
hood effects approach. Unpublished manuscript, Duke Univ., Univ. Connecticut
Bayer P, Ross S, Topa G. 2008. Place of work and place of residence: informal hiring networks and
labor market outcomes. J. Polit. Econ. 116(6):1150–96
Becker G. 1973. A theory of marriage, part I. J. Polit. Econ. 81(4):813–46
Becker G. 1974. A theory of social interactions. J. Polit. Econ. 82(6):1063–93
Becker G, Murphy K. 2000. Social Economics. Cambridge, MA: Harvard Univ. Press
Bernheim D. 1994. A theory of conformity. J. Polit. Econ. 102(5):841–77
474 Durlauf � Ioannides
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
Bertrand M, Luttmer E, Mullainathan S. 2000. Network effects and welfare cultures. Q. J. Econ.
115(3):1019–55
Bisin A, Ozgur O, Horst U. 2006. Rational expectations equilibria of economies with local interac-
tions. J. Econ. Theory. 127:74–116
Blume L. 1993. The statistical mechanics of strategic interaction. Games Econ. Behav. 5:387–424
Bramoulle Y, Djebbari H, Fortin B. 2009. Identification of peer effects through social networks.
J. Econom. 150(1):41–55
Brock W, Durlauf S. 2001a. Discrete choice with social interactions. Rev. Econ. Stud. 68:235–60
Brock W, Durlauf S. 2001b. Growth empirics and reality. World Bank Econ. Rev. 15(2):229–72
Brock W, Durlauf S. 2001c. Interaction-based models. In Handbook of Econometrics, Vol. 5,
ed. J Heckman, E Leamer, pp. 3297–380. Amsterdam: North-Holland
Brock W, Durlauf S. 2002. A multinomial choice model with neighborhood effects. Am. Econ. Rev.
92:298–303
BrockW, Durlauf S. 2006. Multinomial choice with social interactions. In The Economy as an Evolving
Complex System III, ed. L Blume, S Durlauf, pp. 175–206. New York: Oxford Univ. Press
Brock W, Durlauf S. 2007. Identification of binary choice models with social interactions. J. Econom.
140:52–75
Brock W, Durlauf S. 2010. Social interactions and adoption curves. J. Eur. Econ. Assoc. 8:232–51
Brooks-Gunn J, Duncan G, Klebanov P, Sealand N. 1993. Do neighborhoods affect child and adoles-
cent development? Am. J. Sociol. 99(2):353–95
Burke M, Fournier G, Prasad K. 2003. Physician social networks and geographical variation in
medical care. Work. Pap. 33, Brookings Inst. Center Social Econ. Dyn.
Calabrese S, Epple D, Romer T, Sieg H. 2006. Local public good provision: voting, peer effects, and
mobility. J. Public Econ. 90:959–81
Calvo-Armengol A, Patacchini E, Zenou Y. 2009. Peer effects and social networks in education. Rev.
Econ. Stud. 76(4):1239–67
Card D, Mas A, Rothstein J. 2008. Tipping and the dynamics of segregation. Q. J. Econ.
123(1):177–218
Chamberlain G. 1984. Panel data. In Handbook of Econometrics, Vol. 2, ed. Z Griliches,
D McFadden, pp. 1247–318. Amsterdam: North-Holland
Cliff A, Ord J. 1981. Spatial Processes, Models and Applications. London: Pion
Cohen-Cole E. 2006. Multiple groups identification in the linear-in-means model. Econ. Lett.
92(2):157–62
Conley T, Topa G. 2002. Socio-economic distance and spatial patterns in unemployment. J. Appl.
Econ. 17(4):303–27
Conley T, Topa G. 2007. Estimating dynamic local interactions models. J. Econom. 140(1):282–303
Cooley J. 2008. Desegregation and the achievement gap: Do diverse peers help? Unpublished manu-
script, Univ. Wisconsin, Madison
Cooley J. 2009. Alternative mechanisms of peer spillovers in the classroom: implications for identifi-
cation and policy. Unpublished manuscript, Univ. Wisconsin, Madison
Corcoran M, Gordon R, Laren D, Solon F. 1992. The association between men’s economic status and
their family and community origins. J. Hum. Resour. 27(4):575–601
Costa D, Kahn M. 2007. Surviving Andersonville: the benefits of social networks in POW camps.
Am. Econ. Rev. 97(4):1467–87
Crane J. 1991. The epidemic theory of ghettos and neighborhood effects on dropping out and teenage
childbearing. Am. J. Sociol. 96(5):1226–59
Datcher L. 1982. Effects of community and family background on achievement. Rev. Econ. Stat.
64(1):32–41
Davezies L, d’ Haultfoeuille X, Foughre D. 2006. Identification of peer effects using group size
variation. Inst. Study Labor (IZA) Discuss. Pap. 2324
de Paula A. 2009. Inference in a synchronization gamewith social interactions. J. Econom. 148(1):56–71
Develops statistical
mechanics models
that parallel social
interactions models.
Develops identification
conditions for linear
social networksmodels.
Along with Brock &
Durlauf 2001c and
2007, develops theory
and econometrics of
discrete choice models
of social interactions.
Develops a well-
delineated general
equilibrium model of
neighborhood choice
that is estimable.
Pioneering empirical
paper in the effort to
measure social
interactions effects.
www.annualreviews.org � Social Interactions 475
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
Drewianka S. 2003. Estimating social effects in matching markets: externalities in spousal search. Rev.
Econ. Stat. 85:408–23
Durlauf S. 1997. Statistical mechanics approaches to socioeconomic behavior. In The Economy as
an Evolving Complex System II, ed. WB Arthur, SN Durlauf, D Lane, pp. 81–104. Menlo Park,
CA: Addison-Wesley
Durlauf S. 2004. Neighborhood effects. In Handbook of Urban and Regional Economics, Vol. 4,
ed. HJ Vernon, J-F Thisse, pp. 2173–242. Amsterdam: North-Holland
Durlauf S, Seshadri A. 2003. Is assortative matching efficient? Econ. Theory 21(2–3):475–93
Durlauf S, Tanaka H. 2008. Regression versus variance tests for social interactions. Econ. Inq.
46(1):25–28
Epple D, Sieg H. 1999. Estimating equilibrium models of local jurisdictions. J. Polit. Econ.
107(4):645–81
Evans W, Oates W, Schwab R. 1992. Measuring peer group effects: a study of teenage behavior.
J. Polit. Econ. 100(5):966–91
Falk A, Ichino A. 2006. Clean evidence on peer effects. J. Labor Econ. 24:39–57
Fryer R, Torelli P. 2006. An empirical analysis of ‘acting white.’Work. Pap., Dep. Econ., Harvard Univ.
Gaviria A, Raphael S. 2001. School-based peer effects and juvenile behavior. Rev. Econ. Stat.
83(2):257–68
Glaeser E, Sacerdote B, Scheinkman J. 1996. Crime and social interactions.Q. J. Econ. 111(2):507–48
Glaeser E, Sacerdote B, Scheinkman J. 2003. The social multiplier. J. Eur. Econ. Assoc. 1:345–53
Graham B. 2008. Identifying social interactions through conditional variance restrictions.
Econometrica 76(3):643–60
Graham B, Imbens G, Ridder G. 2008. Measuring the average outcome and inequality effects of
segregation in the presence of social spillovers. Unpublished manuscript, Univ. Calif., Berkeley
Guryan J, Kroft K, Notowidigdo M. 2007. Peer effects in the workplace: evidence from random
groupings in professional golf tournaments. NBERWork. Pap. 13422
Hanushek E, Kain J, Markman J, Rivkin S. 2003. Does peer ability affect student achievement?
J. Appl. Econ. 18(5):527–44
Heckman J. 1979. Sample selection bias as a specification error. Econometrica 47:153–61
Heckman J. 2001. Micro data, heterogeneity, and the evaluation of public policy: Nobel Lecture.
J. Polit. Econ. 109(4):674–748
Henry G, Rickman D. 2007. Do peers influence children’s skill development in preschool? Econ. Educ.
Rev. 26(1):100–12
Hirano K, Hahn J. 2009. Design of randomized experiments to measure social interaction effects.
Unpublished manuscript, Univ. Arizona
Horst U, Scheinkman J. 2006. Equilibria in systems of social interactions. J. Econ. Theory 130:44–77
Hoxby C. 2000. Peer effects in the classroom: learning from gender and race variation. NBERWork.
Pap. 7867
Hoxby C, Weingarth G. 2006. Taking race out of the equation: school reassignment and the structure
of peer effects. Work. Pap., Dep. Econ., Stanford Univ.
Imbens G, Newey W. 2009. Identification and estimation of triangular simultaneous equations models
without additivity. Econometrica 77(5):1481–512
Ioannides Y. 2006. Topologies of social interactions. Econ. Theory 28:559–84
Ioannides Y. 2008. Full solution of an endogenous sorting model with contextual and income effects.
Work. Pap., Dep. Econ., Tufts Univ.
Ioannides Y, Loury LD. 2004. Job information networks, neighborhood effects, and inequality.
J. Econ. Lit. 42(4):1056–93
Ioannides Y, Zabel J. 2008. Interactions, neighborhood selection, and housing demand. J. Urban Econ.
63:229–52
Irwin E, Bockstael N. 2002. Interacting agents, spatial externalities and the evolution of residential
land use patterns. J. Econ. Geogr. 2:31–54
Develops the idea of a
social multiplier and
discusses ways to
empirically measure.
Develops equilibrium
properties of social
interactions models
for very general
specifications of
interactions structures.
476 Durlauf � Ioannides
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
Jackson M. 2008. Social and Economic Networks. Princeton, NJ: Princeton Univ. Press
Kling J, Liebman J, Katz L. 2007. Experimental analysis of neighborhood effects. Econometrica.
75(1):83–119
Kling J, Ludwig J, Katz L. 2005. Neighborhood effects on crime for female and male youth: evidence
from a randomized housing voucher experiment. Q. J. Econ. 120(1):87–130
Krauth B. 2005. Peer effects and selection effects on smoking among Canadian youth. Can. J. Econ.
38(3):735–57
Krauth B. 2006. Simulation-based estimation of peer effects. J. Econom. 133(1):243–71
Lavy V, Paserman D, Schlosser A. 2008. Inside the black of box of ability peer effects: evidence from
variation in low achievers in the classroom. NBERWork. Pap. 14415
Lee L-f. 2007. Identification and estimation of econometric models with group interactions, contex-
tual factors and fixed effects. J. Econom. 140(2):333–74
Lee L-f, Liu X, Lin X. 2010. Specification and estimation of social interaction models with network
structure, contextual factors, correlation and fixed effects. Econom. J. In press
Lin X. 2008. Identifying peer effects in student academic achievement by spatial autoregressive models
with group unobservables. Work. Pap., Dep. Econ., Tsinghua Univ.
Ludwig J, Duncan G, Hirschfield P. 2001. Urban poverty and juvenile crime: evidence from a random-
ized housing-mobility experiment. Q. J. Econ. 116:655–80
Manski C. 1988. Identification of binary choice models. J. Am. Stat. Assoc. 83(403):729–38
Manski C. 1993. Identification of endogenous social effects: the reflection problem. Rev. Econ. Stud.
60:531–42
Manski C. 2000. Economic analysis of social interactions. J. Econ. Perspect. 14(3):115–36
Mas A, Moretti E. 2009. Peers at work. Am. Econ. Rev. 99(1):112–45
McKelvey R, Palfrey T. 1995. Quantal response equilibria for normal form games. Games Econ.
Behav. 10:6–38
Nakajima R. 2007. Measuring peer effects on youth smoking behaviour. Rev. Econ. Stud. 74:897–935
Nesheim L. 2002. Equilibrium sorting of heterogeneous consumers across locations: theory and
empirical implications. CeMMAP Work. Pap. CWP08/02, Univ. College London
Oreopoulos P. 2003. The long-run consequences of living in a poor neighborhood. Q. J. Econ.
118(4):1533–75
Sacerdote B. 2001. Peer effects with random assignment: results for Dartmouth roommates.
Q. J. Econ. 116:681–704
Searle S, Casella G, McCullough C. 2006. Variance Components. New York: Wiley
Sims C. 1980. Macroeconomics and reality. Econometrica 48:1–48
Sirakaya S. 2006. Recidivism and social interactions. J. Am. Stat. Assoc. 101(475):863–75
Siow A. 2009. Testing Becker’s theory of positive assortative matching. Work. Pap. 356, Dep. Econ.,
Univ. Toronto
Sobel M. 2006. Spatial concentration and social stratification: does the clustering of disadvantage
‘beget’ bad outcomes? In Poverty Traps, ed. S Bowles, SN Durlauf, K Hoff, pp. 204–29.
Princeton, NJ: Princeton Univ. Press
Soetevent A, Kooreman P. 2007. A discrete choice model with social interactions: with an application
to high school teen behavior. J. Appl. Econ. 22:599–624
Solon G, PageM, Duncan G. 2000. Correlations between neighboring children and in their subsequent
educational attainment. Rev. Econ. Stat. 82(3):383–92
Streufert P. 2000. The effect of underclass isolation on school choice. J. Public Econ. Theory
2(4):461–82
Tamer E. 2003. Incomplete simultaneous discrete response model with multiple equilibria. Rev. Econ.
Stud. 70:147–65
Topa G. 2001. Social interactions, local spillovers, and unemployment. Rev. Econ. Stud. 68(2):261–95
Weinberg B. 2007. Social interactions with endogenous associations. NBERWork. Pap. 13038
Pioneering paper on
the identification of
social interactions,
introduces the
reflection problem to
the literature.
www.annualreviews.org � Social Interactions 477
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
Weinberg B, Reagan P, Yankow J. 2004. Do neighborhoods affect hours worked? Evidence from
longitudinal data. J. Labor Econ. 22(4):891–924
Wilson W. 2009.More than Just Race. New York: Norton
Young P. 2009. Innovation diffusion in heterogeneous populations: contagion, social influence, and
social learning. Am. Econ. Rev. 99(5):1899–924
Young P, Burke M. 2001. Competition and custom in economic contracts: a case study of Illinois
agriculture. Am. Econ. Rev. 91(3):559–73
RELATED RESOURCE
Bisin A, Benhabib J, Jackson M, eds. 2010. Handbook of Social Economics. Amsterdam: North-
Holland. In press. Presents a definitive collection of studies of different aspects of social
interactions.
478 Durlauf � Ioannides
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
Annual Review of
Economics
Contents
Questions in Decision Theory
Itzhak Gilboa. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Structural Estimation and Policy Evaluation in Developing Countries
Petra E. Todd and Kenneth I. Wolpin. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Currency Unions in Prospect and Retrospect
J.M.C. Santos Silva and Silvana Tenreyro . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Hypothesis Testing in Econometrics
Joseph P. Romano, Azeem M. Shaikh, and Michael Wolf . . . . . . . . . . . . . . 75
Recent Advances in the Empirics of Organizational Economics
Nicholas Bloom, Raffaella Sadun, and John Van Reenen . . . . . . . . . . . . . 105
Regional Trade Agreements
Caroline Freund and Emanuel Ornelas . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Partial Identification in Econometrics
Elie Tamer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
Intergenerational Equity
Geir B. Asheim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
The World Trade Organization: Theory and Practice
Kyle Bagwell and Robert W. Staiger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
How (Not) to Do Decision Theory
Eddie Dekel and Barton L. Lipman . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Health, Human Capital, and Development
Hoyt Bleakley . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
Beyond Testing: Empirical Models of Insurance Markets
Liran Einav, Amy Finkelstein, and Jonathan Levin. . . . . . . . . . . . . . . . . . 311
Inside Organizations: Pricing, Politics, and Path Dependence
Robert Gibbons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
Volume 2, 2010
vi
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
Identification of Dynamic Discrete Choice Models
Jaap H. Abbring. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
Microeconomics of Technological Adoption
Andrew D. Foster and Mark R. Rosenzweig . . . . . . . . . . . . . . . . . . . . . . 395
Heterogeneity, Selection, and Wealth Dynamics
Lawrence Blume and David Easley . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
Social Interactions
Steven N. Durlauf and Yannis M. Ioannides. . . . . . . . . . . . . . . . . . . . . . . 451
The Consumption Response to Income Changes
Tullio Jappelli and Luigi Pistaferri . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479
Financial Structure and Economic Welfare: Applied
General Equilibrium Development Economics
Robert Townsend . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
Models of Growth and Firm Heterogeneity
Erzo G.J. Luttmer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 547
Labor Market Models of Worker and Firm Heterogeneity
Rasmus Lentz and Dale T. Mortensen . . . . . . . . . . . . . . . . . . . . . . . . . . . 577
The Changing Nature of Financial Intermediation and the
Financial Crisis of 2007–2009
Tobias Adrian and Hyun Song Shin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603
Competition and Productivity: A Review of Evidence
Thomas J. Holmes and James A. Schmitz, Jr. . . . . . . . . . . . . . . . . . . . . . . 619
Persuasion: Empirical Evidence
Stefano DellaVigna and Matthew Gentzkow . . . . . . . . . . . . . . . . . . . . . . 643
Commitment Devices
Gharad Bryan, Dean Karlan, and Scott Nelson . . . . . . . . . . . . . . . . . . . . 671
Errata
An online log of corrections to Annual Review of Economics
articles may be found at http://econ.annualreviews.org
Contents vii
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.
AnnuAl Reviewsit’s about time. Your time. it’s time well spent.
AnnuAl Reviews | Connect with Our expertsTel: 800.523.8635 (us/can) | Tel: 650.493.4400 | Fax: 650.424.0910 | Email: [email protected]
New From Annual Reviews:
Annual Review of Statistics and Its ApplicationVolume 1 • Online January 2014 • http://statistics.annualreviews.org
Editor: Stephen E. Fienberg, Carnegie Mellon UniversityAssociate Editors: Nancy Reid, University of Toronto
Stephen M. Stigler, University of ChicagoThe Annual Review of Statistics and Its Application aims to inform statisticians and quantitative methodologists, as well as all scientists and users of statistics about major methodological advances and the computational tools that allow for their implementation. It will include developments in the field of statistics, including theoretical statistical underpinnings of new methodology, as well as developments in specific application domains such as biostatistics and bioinformatics, economics, machine learning, psychology, sociology, and aspects of the physical sciences.
Complimentary online access to the first volume will be available until January 2015. table of contents:•What Is Statistics? Stephen E. Fienberg•A Systematic Statistical Approach to Evaluating Evidence
from Observational Studies, David Madigan, Paul E. Stang, Jesse A. Berlin, Martijn Schuemie, J. Marc Overhage, Marc A. Suchard, Bill Dumouchel, Abraham G. Hartzema, Patrick B. Ryan
•The Role of Statistics in the Discovery of a Higgs Boson, David A. van Dyk
•Brain Imaging Analysis, F. DuBois Bowman•Statistics and Climate, Peter Guttorp•Climate Simulators and Climate Projections,
Jonathan Rougier, Michael Goldstein•Probabilistic Forecasting, Tilmann Gneiting,
Matthias Katzfuss•Bayesian Computational Tools, Christian P. Robert•Bayesian Computation Via Markov Chain Monte Carlo,
Radu V. Craiu, Jeffrey S. Rosenthal•Build, Compute, Critique, Repeat: Data Analysis with Latent
Variable Models, David M. Blei•Structured Regularizers for High-Dimensional Problems:
Statistical and Computational Issues, Martin J. Wainwright
•High-Dimensional Statistics with a View Toward Applications in Biology, Peter Bühlmann, Markus Kalisch, Lukas Meier
•Next-Generation Statistical Genetics: Modeling, Penalization, and Optimization in High-Dimensional Data, Kenneth Lange, Jeanette C. Papp, Janet S. Sinsheimer, Eric M. Sobel
•Breaking Bad: Two Decades of Life-Course Data Analysis in Criminology, Developmental Psychology, and Beyond, Elena A. Erosheva, Ross L. Matsueda, Donatello Telesca
•Event History Analysis, Niels Keiding•StatisticalEvaluationofForensicDNAProfileEvidence,
Christopher D. Steele, David J. Balding•Using League Table Rankings in Public Policy Formation:
Statistical Issues, Harvey Goldstein•Statistical Ecology, Ruth King•Estimating the Number of Species in Microbial Diversity
Studies, John Bunge, Amy Willis, Fiona Walsh•Dynamic Treatment Regimes, Bibhas Chakraborty,
Susan A. Murphy•Statistics and Related Topics in Single-Molecule Biophysics,
Hong Qian, S.C. Kou•Statistics and Quantitative Risk Management for Banking
and Insurance, Paul Embrechts, Marius Hofert
Access this and all other Annual Reviews journals via your institution at www.annualreviews.org.
Ann
u. R
ev. E
con.
201
0.2:
451-
478.
Dow
nloa
ded
from
ww
w.a
nnua
lrev
iew
s.or
gby
Uni
vers
ity o
f W
isco
nsin
- M
adis
on o
n 09
/10/
14. F
or p
erso
nal u
se o
nly.