Full text of DOI:10.1162/003355300554935

PARTICIPATION IN HETEROGENEOUS COMMUNITIES*
ALBERTO ALESINA AND ELIANA LA FERRARA
This paper studies what determines group formation and the degree of
participation when the population is heterogeneous, both in terms of income and
race or ethnicity. We are especially interested in whether and how much the degree
of heterogeneity in communities influences the amount of participation in different
types of groups. Using survey data on group membership and data on U. S.
localities, we find that, after controlling for many individual characteristics,
participation in social activities is significantly lower in more unequal and in more
racially or ethnically fragmented localities. We also find that those individuals who
express views against racial mixing are less prone to participate in groups the
more racially heterogeneous their community is. These results are consistent with
our model of group formation.
I. INTRODUCTION
Many observers, including economists, are convinced of the
importance of the complex stock of social norms, trust, and
networks of civic engagement that has been grouped under the
1
term ‘‘social capital.’’ As forcefully argued by Putnam [1993,
1
995a, 1995b], social capital may produce several positive socioeco-
nomic effects which can spur economic success.
But, what determines social capital? Figure I displays the
distribution across U. S. states of an index of social capital that
embodies information on trust, membership in groups, and voting
2
behavior. Overall, this index is highest in the North/Northwest
*
We are grateful to Abhijit Banerjee, Eli Berman, Fran c¸ ois Bourguignon,
Edward Glaeser, Claudia Goldin, Luigi Guiso, Lawrence Katz, Massimiliano
Marcellino, Caroline Minter Hoxby, Daniele Paserman, Robert Putnam, Andrei
Shleifer, Antonio Rangel, Howard Rosenthal, Sidney Verba, Leeat Yaariv, two
anonymous referees, and to seminar participants at Harvard University (Saguaro
Seminar, October 1998), Stanford University, Ente Einaudi, CORE Louvain, 1999
EEA Conference, and 1999 NEUDC Conference for very useful comments. We also
thank Erzo Luttmer for sharing data and Spencer Glendon for help with data
sources. Mario Centeno and Daniel Altman provided excellent research assistance.
This research is supported by a National Science Foundation grant to the National
Bureau of Economic Research. We are grateful to both organizations for their
support. Alesina also gratefully acknowledges financial support from the Wheater-
head Center for International Affairs at Harvard University.
1
. See Coleman [1990] for an extensive discussion of the foundations of ‘‘social
capital’’ in social theory.
. We constructed this index extracting the principal components from three
2
variables obtained from the General Social Survey: the percentage of people in the
state who belong to a group, the percentage who trust others, and the percentage
who voted in the last presidential election. A detailed description of the data is
given in the empirical part of the paper. Principal components analysis has been
used to construct an index of social capital also by Putnam and Yonish [1998]. Our
index is very highly correlated with theirs.
2
000 by the President and Fellows of Harvard College and the Massachusetts Institute of
௠
Technology.
The Quarterly J ournal of Economics, August 2000
8
47

8
48
QUARTERLY J OURNAL OF ECONOMICS
FIGURE I
Social Capital Index
and lowest in the South/Southeast. The former regions are
characterized by a racially homogeneous population and rela-
tively low income inequality, while the latter have the opposite
features. In particular, the top five U. S. states in our social
capital ranking are North Dakota, Utah, Minnesota, Wyoming,
and Montana. All of them are very homogeneous: they are almost
one standard deviation below the national mean for both racial
fragmentation and income inequality.
The relationship between homogeneity and social capital may
not be only a U. S. phenomenon. For instance, according to the
international ranking reported by Knack and Keefer [1997], the
five countries with highest levels of trust are Norway, Finland,
Sweden, Denmark, and Canada; the same countries rank among
the top ones for associational activity and norms of civic coopera-
tion. These countries have an ethnically homogeneous population
and very low levels of income inequality.
This paper shows that there is something systematic about
the relationship between heterogeneity of communities (in terms

PARTICIPATION IN HETEROGENEOUS COMMUNITIES
849
of income and race) and the level of social capital: more homoge-
neous communities have a higher level of social interactions
leading to more social capital. Even though this paper focuses on
U. S. cities, our results have more general implications. Racial
heterogeneity and income inequality vary greatly in different
countries, and, through their effect on social capital, they may
influence economic outcomes and public policies.
One of the common criticisms of the notion of social capital is
that it is very hard to ‘‘measure,’’ hence difficult to use in empirical
analysis. Rather than focusing on a broad index of social capital,
we study a critical component of it that can be measured fairly
precisely, namely the participation in associational activities,
such as religious groups, hobby clubs, youth groups, sport groups,
etc. Our interest in these activities is motivated by Putnam [1993,
1
995a, 1995b] who suggests that these types of social interactions
are particularly conducive to generating the beneficial effects of
social capital. More specifically, participation in social groups may
lead to the transmission of knowledge and may increase aggregate
human capital, and the development of ‘‘trust,’’ which improve the
3
functioning of markets. In addition, social interactions and
networks may influence individual outcomes, from criminal activi-
4
ties, to fertility, to the labor supply. Also, participation in groups
is known to be highly correlated with political participation, and
5
the latter has critical implications for policy choices. If the
wealthy or more educated have a disproportionate propensity to
join groups and engage in political action, then public policies may
be tilted in their favor. This may lead to vicious circles, in which
disadvantaged minorities participate less, have less ‘‘voice,’’ and
become even more disadvantaged, leading to a variety of social
problems.6
3
. On the effects of positive spillovers in the transmission of human capital,
see Romer [1986], Lucas [1988], and B e´ nabou [1996]. On trust see La Porta et al.
1997].
. For theoretical work on the effects of transmission of information in group
[
4
and informational cascades, see, for instance, Banerjee [1992] and Ellison and
Fudenberg [1995]. An early empirical contribution on the importance of networks
is Case and Katz [1991].
5
6
. See Verba and Nie [1987] and Verba, Schlozman, and Brady [1995].
. Even though participation is typically associated with ‘‘positive’’ socioeco-
nomic outcomes, social networks may also transmit ‘‘negative’’ norms. For
example, the so-called ‘‘culture of poverty and welfare’’ may find its roots in social
networks propagating incentives to search for welfare rather than work. See, in
particular, Cutler and Glaeser [1997] and Bertrand, Luttmer, and Mullainathan
[
2000] for recent empirical work on this important question. A different theoretical
perspective on groups and social norms is given by Berman [2000].

8
50
QUARTERLY J OURNAL OF ECONOMICS
In our model individuals prefer to interact with others who
are similar to themselves in terms of income, race, or ethnicity. If
preferences are correlated with these characteristics, then our
assumption is equivalent to saying that individuals prefer to join
groups composed of individuals with preferences similar to their
7
own. Given this setup, one may expect that diffuse preferences
for homogeneity may decrease total participation in a mixed
group if fragmentation increases. However, individuals may choose
to sort into homogeneous groups. Therefore, it is not clear a priori
under which conditions more heterogeneity in the population
would lead to more or less participation. In the theoretical part of
the paper, we investigate this issue. Our model departs from
standard club theory, since our groups do not require contribu-
tions, do not have congestion effects, and are based on free entry
and exit of individuals. We make these assumptions because we
want to focus on how the composition of the group affects
individual choices of participation.
Our empirical results on U. S. localities suggest that income
inequality and racial and ethnic heterogeneity reduce the propen-
sity to participate in a variety of social activities including
8
recreational, religious, civic, and educational groups. Among the
various forms of heterogeneity, racial fragmentation seems to
have the strongest negative effect on participation. Furthermore,
and consistent with our model, these results are more marked for
the groups in which direct contact among members is important,
like churches and youth clubs, while heterogeneity matters less or
not at all in groups with a low degree of interaction, for instance
professional associations. Finally, our model predicts that individu-
als relatively more averse to mixing with different types should be
those more negatively influenced by heterogeneity in the commu-
nity. We successfully test this more stringent implication of the
model, by exploiting individual data on attitudes toward race
7
. Theoretical results by Conley and Wooders [1996] are consistent with this
assumption. They show that when agents can be crowded (positively or negatively)
by the skills of other people in their jurisdictions, taste-homogeneous jurisdictions
are optimal. To the extent that tastes are correlated with income and race, our
assumption follows.
8
. We define by ‘‘race’’ the census classification of black, white, Asian,
American Indian, and other. We define by ethnicity the classification by ancestry,
like Italian, Irish, etc. Throughout this paper we will use the terms ‘‘white’’ and
‘‘black’’ instead of Caucasian and African-American, for the sake of brevity.

PARTICIPATION IN HETEROGENEOUS COMMUNITIES
851
relations. In summary, we find that social capital is lower in more
9
unequal and heterogeneous communities.
Recent research has highlighted a positive bivariate correla-
tion between inequality and social capital measures at the state
level.10 Our multivariate analysis, conducted with individual level
data on participation and community level measures of income
inequality, sheds light on this issue. As for race, much empirical
research has studied the effects on public policy of ethnic and
racial heterogeneity. For instance, Alesina, Baqir, and Easterly
[1999, 2000] show that the supply of ‘‘core’’productive public goods
is lower and measures of patronage are higher in more racially
fragmented localities. Glaeser, Scheinkman, and Shleifer [1995],
Cutler and Glaeser [1997], Poterba [1996], Luttmer [1997], and
Goldin and Katz [1999] study the role of racial conflict as a
determinant of education policies and several other characteris-
tics of U. S. cities. Alesina, Baqir, and Hoxby [1999] link racial and
ethnic fragmentation to the number of jurisdictions in the United
States.
This paper is organized as follows. Section II presents a model
that generates predictions linking heterogeneity of the population
and the level of participation in social activities. Section III
describes our empirical strategy and data. Section IV highlights
some simple correlations at the state level between income
inequality, racial and ethnic fragmentation, and measures of
participation and social capital. Section V presents our economet-
ric results. The last section concludes.
II. THE MODEL
II.1. Setup
Consider a community populated by two types, ‘‘blacks’’ and
‘‘whites,’’their size being B Ͼ 0 and W Ͼ 0, respectively. For ease of
exposition, we shall think about ‘‘race’’ as what identifies the
9
. To the extent that one views social capital as a public good, this result is
consistent with the findings of Alesina, Baqir, and Easterly [1999]. That paper
focuses on public goods provided by means of public policies (and taxes) chosen by a
median voter who ‘‘decides’’ for the entire community. The present paper instead
focuses on a public good (social capital) not generated by policies but by the
interaction of private individuals in private groups. One may think of public
policies that increase the incentive to participate and create social capital, but this
issue is not explicitly addressed in this paper.
1
0. See Putnam’s presentation at the Saguaro Seminar, October 1998,
Harvard University.

8
52
QUARTERLY J OURNAL OF ECONOMICS
types; below we discuss other discriminating characteristics,
especially income. The population is uniformly distributed on a
line, and both types have a uniform distribution on the interval
[
0,1].11
Each individual decides whether or not to participate in a
group. We do not allow for participation in more than one group.
There are no congestion costs and no economies of scale: thus,
group size does not influence individual utility. Members of the
group cannot exclude new members, and entry into (exit from) the
group is free and costless. The reservation utility from nonpartici-
1
2
pation is u for everybody. The utility from participating in the
group depends on the composition of the group and on the distance
between the individual’s and the group’s location. Let P (P ) be
B
W
the proportion of blacks (whites) within the group. The utility
from participation for an individual of type j ϭ W,B located at a
distance l from the group’s headquarters (HQ) is
(
1) U ϭ u(␣,P ) ϩ v(l)
for j ϭ B,W
u␣(·) Ͻ 0,
j
Ϫj
uP (·) Ͻ 0,
vl(·) Ͻ 0
(
A1)
A2)
u␣ P (·) Յ 0
(
u␣(·)0Pϭ0 ϭ 0,
uP (·)0␣ϭ0 ϭ 0,
where PϪj is the proportion of group members whose type is
different from j’s type, and ␣ is a taste indicator that varies across
individuals and captures the intensity of an individual’s aversion
to the opposite race. The functions u (·) and v(·) are continuously
differentiable in their arguments; u (·) and u (·) represent the
␣
P
partial derivatives of u (·) with respect to ␣ and P , respectively,
Ϫj
and v (·) is the derivative of v(·) with respect to l. Underlying
l
assumption u (·) Ͻ 0 is the preference for participating in a social
P
activity with members of one’s own type. Given our assumption
that the cross derivative of u (·) with respect to ␣ and PϪj is
(
weakly) negative, an increase in the proportion of members of the
opposite type decreases an individual’s utility more, the higher is
. We label ␣ the ‘‘degree of intolerance’’ or ‘‘aversion.’’Assumption
␣
1
1. This assumption is obviously restrictive and does not address the pres-
ence of racial segregation in housing. The nature of our arguments remains valid
as long as B and W have the same distribution (even if it is not uniform).
1
2. The reservation utility may of course be a function of individual character-
istics. While this will be important for the empirical analysis that follows, we
abstract from it in the theoretical discussion.

PARTICIPATION IN HETEROGENEOUS COMMUNITIES
853
A2 states that utility, when evaluated at PϪj ϭ 0, is independent of
, and when evaluated at ␣ ϭ 0, is independent of P . In other
␣
Ϫj
words, when the group is totally composed of one’s own type,
intolerance does not play any role, and when an individual has
zero intolerance, the composition of the group does not affect her
utility. In what follows, we will assume that u (0,P ) ϭ 0 purely to
Ϫj
1
3
simplify the notation.
Each individual is therefore characterized by two parameters
(in addition to the type B or W): aversion and distance. We assume
that the aversion parameter ␣ and individual location are indepen-
dently distributed on [0,1]. We denote the density function for ␣
and for the distance l with f (·) and f (·), respectively, and assume
␣
l
that these densities are continuously differentiable on their
support.
In summary, our model has three basic components. The first
is that people prefer to interact with members of their own type.
The second component is that people’s utility from participation
decreases the farther they have to travel to go to group meetings,
etc. In the absence of traveling costs there is no reason why people
should form mixed groups (everyone is better off by being with his
own type, no matter how far away the group is located), nor is
there a reason why more than two groups should form (or, in
general, as many as the different types in the population), given
that there is no congestion. We require the groups to have a
minimum size in order to avoid the degenerate outcome in which
every individual forms his or her own group. Finally, our model
departs from club theory and from the local public goods literature
in that no contributions are paid by the members and utility is
derived directly from participation.
II.2. Equilibrium with One Group
We begin by assuming that the minimum size of the group is
greater than half of the population, so that at most one group can
1
4
form in equilibrium. We relax this assumption below. Consider
an individual of type j ϭ B,W and aversion ␣ who has the choice
between joining a group with composition PϪj and distance l on the
1
3. If we remove this assumption, the only change is that the extreme of
Ϫ1
integration with respect to l in expression (3) below becomes v (u Ϫ u (0,PϪj)),
which, anyway, is independent of PϪj given assumption A2.
1
4. For a simpler version of this model with one group and no travel costs, see
Alesina and La Ferrara [1999].

8
54
QUARTERLY J OURNAL OF ECONOMICS
one hand, and staying out on the other. This individual partici-
pates in the group if and only if
(2)
u (␣,P ) ϩ v(l) Ն u.
Ϫj
In what follows, we concentrate on symmetric equilibria in which
1
5
the HQ is in the middle of the group.
The probability that the participation constraint (2) is satis-
fied can be written as
vϪ1(u)
g(uϪv(l),PϪj
)
(3)
͟ (P ,u) ϵ 2 ͐0 ͐0
f (␣) f (l) d␣ dl,
␣ l
Ϫj
where g(·) is obtained by inverting u (·). Our assumptions imply
that the partial derivative of g(·) with respect to PϪj is negative
and hence that Ѩ⌸(P ,u)/ѨP Ͻ 0. Expression (3) states that, for
Ϫj
Ϫj
given group composition P , the individuals of type j ϭ B,W for
Ϫj
whom the participation constraint holds are those who are located
Ϫ1
no farther than v (u) from the group headquarters and whose
aversion parameter is no greater than g(u Ϫ v(l),P ).
Ϫj
The total mass of individuals of type B and type W willing to
participate in the social activity is then, respectively,
˜
(
4)
5)
B ϭ ⌸(P ,u) · B
W
˜
(
W ϭ ⌸(P ,u) · W.
B
DEFINITION 1. An equilibrium is a group composition (P *,P * )
B
W
such that for both types none of the members wishes to leave
the group and none of the nonmembers wishes to join.
In equilibrium the proportion of individuals of type B in the
group, P , must be equal to the ratio of the mass of the partici-
B
˜
˜
˜
pants B to the total mass of participants B ϩ W. The two
conditions defining the equilibrium are therefore
͟
(P ,u) · B
W
(
6)
7)
P ϭ
B
͟
(P ,u) · B ϩ ͟(P ,u) · W
W
B
(
PW ϭ 1 Ϫ PB,
1
5. The symmetric case is the natural reference point, both from a normative
and from a positive point of view. A utilitarian social planner would always locate
the HQ in the middle of the group, to minimize total travel costs. Alternatively, if
after the group were formed the members decided by majority vote where to locate
the HQ, they would choose the middle of the group (which is the median, given our
assumptions). We can also handle the nonsymmetric case. Details are available.

PARTICIPATION IN HETEROGENEOUS COMMUNITIES
855
which together give us the ‘‘fixed point’’ equilibrium condition
contained in the following.
PROPOSITION 1. There exists at least one equilibrium P * ʦ [0,1]
B
which solves
͟
(1 Ϫ P ,u)
B
(8)
P ϭ
B
.
͟
(1 Ϫ P ,u) ϩ ͟(P ,u)W/B
B
B
Proof. The proof of this proposition like all the other proofs is
in Appendix 1.16
In what follows, for notational convenience we will suppress
the term u from the arguments of ⌸(·). We will restrict our
attention to locally stable equilibria. A formal definition and a
necessary and sufficient condition for local stability are provided
in Appendix 1.
Figure II provides the intuition. The equilibrium value(s) of
PB is (are) given by the intersection of the function in the
right-hand side of (8) with the 45° line. The right-hand side of (8)
represents the fraction of members type B in the group that is
‘‘generated by the reactions’’ of both types to a given composition
(
P ,1 Ϫ P ). The intersection(s) with the identity line give(s) the
B
B
value(s) of PB at which both reactions are consistent with the
actual proportions. Our stability condition requires that the slope
of the above function at the point of intersection with the 45° line
be less than one.17
Figure II depicts various possible configurations of equilibria.
In panel (a) we have a unique interior equilibrium, i.e., the group
that forms is ‘‘mixed,’’ with a proportion P * ʦ (0,1) of blacks and
B
(
1 Ϫ P *) of whites. This equilibrium is stable. Suppose that you
B
add one black member to the group, so the composition becomes
PB Ͼ P *. The shape of the curve in panel (a) suggests that this
B
‘‘more favorable’’ composition for types B does not trigger enough
participation of new B members, nor does it induce enough W
members to exit, so the group goes back to the initial equilibrium.
In panel (d) the opposite occurs; that is, any slight increase
1
6. Underlying Proposition 1 is the condition, ⌸(P *W,u ) · B ϩ ⌸(P *B,u ) · W Ͼ
5(B ϩ W), namely that the utility and distribution functions are parameterized so
that the group that forms in equilibrium meets the minimum size requirement.
.
1
7. Notice that—given our assumptions on u (·), v(·), f␣(·), and fl(·)—the
slope of the ‘‘reaction function’’ defined by the right-hand side of (8) is always
nonnegative.

8
56
QUARTERLY J OURNAL OF ECONOMICS
FIGURE II
Equilibrium Configurations
(
reduction) in the fraction of B members from P * will trigger an
B
inflow (outflow) of B types and outflow (inflow) of W types, so that
the composition of the group moves to complete homogeneity with
P * ϭ 1(P * ϭ 1). While in panel (d) either P * ϭ 0 or P * ϭ 1 can in
B
W
B
B

PARTICIPATION IN HETEROGENEOUS COMMUNITIES
857
principle be stable equilibria18 in panel (b) only P * ϭ 0 is. Finally,
B
1
9
panel (c) illustrates the case of multiple equilibria.
We are interested in two features of the equilibrium. The first
is how the composition of the group relates to the composition of
the total population; the second is who among the heterogeneous
individuals of a given type will choose to participate and who will
stay out.
LEMMA 1. Let (P *,P * ) be a unique stable equilibrium. Then B Ͻ
B
W
W ⇔ P * Ͻ P * .
B
W
COROLLARY 1. If B Þ W and (P *,P * ) is a unique stable equilib-
B
W
rium, then either (P */P * ) Ͻ (B/W) Ͻ 1 or 1 Ͻ (B/W) Ͻ
B
W
(
P */P * ).
B W
COROLLARY 2. Let (P *,P * ) be a unique stable equilibrium. A
B
W
necessary condition for an individual type j ϭ B,W to partici-
j
pate is that her ␣ falls in the interval [0,␣ ], and her distance
j
in the interval [0,l], where ␣ and l solve, respectively,
j
(9)
u (␣ ,P * ) ϩ v(0) ϭ u
Ϫj
(10)
v(l) ϭ u.
B
W
In particular, B Ͻ W ⇔ ␣ Ͻ ␣ .
Lemma 1 and Corollary 1 show that the unbalance between
the types in the population is magnified within the group: not only
is the minority type underrepresented in the social activity, but it
is less than proportionately represented compared with its weight
in the population. Thus, the social group is more homogeneous
than the whole population.
Corollary 2 can help us understand this result. In equilib-
rium, the individuals participating in a mixed group are those
who are ‘‘not too averse’’ to the opposite race, and ‘‘not too far ’’ from
the location of the group. Corollary 2 tells us exactly ‘‘how far ’’ we
can go in these two dimensions. The maximum geographical
distance compatible with the participation constraint is such that
1
8. As a matter of fact, given our assumptions that B Ͻ W and that the
1
minimum size of the group is greater than
⁄2
, only P W ϭ 1 can be a stable
*
equilibrium. As we will see, this will no longer be the case in the multiple-groups
model.
1
9. Of the five equilibria depicted in the figure, 0, P *B2, and 1 are stable, while
P *B1 and P *B3 are unstable. The issue of the selection among multiple equilibria will
not be addressed in this paper.

8
58
QUARTERLY J OURNAL OF ECONOMICS
an individual who receives the lowest possible disutility from
heterogeneity (i.e., an individual with ␣ ϭ 0) is indifferent be-
tween joining the group or not. Given A2, this distance is
independent of the composition of the group and of the individu-
al’s type. The dimension we are most interested in is ‘‘aversion’’ to
the opposite type. The maximum intolerance compatible with the
participation constraint is that of the individual located exactly
j
where the group’s headquarters are, and is given by the ␣ solution
of (9). We can refer to these individuals as the ‘‘most averse’’
members. Corollary 2 tells us that the most averse member of type
B will be less ‘‘tolerant’’than that of type W if and only if blacks are
a minority in the population. In fact, although the two types have
the same ex ante distribution of ␣, we cannot observe the same
degree of ‘‘aversion’’ for B and W in the equilibrium composition of
the group. If we did, the same ␣ should be indifferent between
participating as a majority or as a minority. Instead, the fact that
B is underrepresented in the population induces even some
relatively ‘‘moderate’’ B individuals (low ␣) to stay out, while W
individuals manage to keep some relatively ‘‘participation averse’’
(high ␣) members in the group. This coupled with the sheer
unbalance in the numbers B and W produces the ‘‘magnification’’
effect described above.
II.3. Heterogeneity and Participation
We are now ready to study how a change in the heterogeneity
of the population influences the total mass of participants.
DEFINITION 2. The degree of heterogeneity is the probability that
two randomly drawn individuals from the population belong
to different types.
This is the same definition of our empirical analysis. Obvi-
ously, in our case of two types, a 50-50 split has the maximum
level of heterogeneity. Denote by w the fraction of whites in the
population; i.e.,
(11)
w ϵ W/(W ϩ B).
1
An increase in w represents a decrease in heterogeneity if w Ն ⁄
2
1
and an increase if w Ͻ ⁄
2
.

PARTICIPATION IN HETEROGENEOUS COMMUNITIES
859
DEFINITION 3. The aggregate level of participation S is the share of
the total population who belongs to a group:
˜
˜
B ϩ W
ϭ ͟ (1 Ϫ P *)(1 Ϫ w) ϩ ͟ (P *)w.
B ϩ W
(12)
S ϵ
B
B
Under very mild sufficient conditions on ⌸(·), described in
Appendix 1, the following holds.
PROPOSITION 2. If a unique stable equilibrium exists, an increase
(decrease) in heterogeneity reduces (increases) total participa-
tion; i.e.,
1
2
dS
w Ͻ
w Ն
⇒
Ͻ 0
dw
1
dS
⇒
Ͼ 0.
2
dw
The intuition is simple. We can have two kinds of equilibria.
In the first one the group is homogeneous: by the minimum size
requirement this means that all the individuals of the majority
type participate, and none of the minority type does. In the second
kind of equilibrium the group is mixed, with proportion 0 Ͻ P * Ͻ 1
B
identified by Proposition 1. To examine the impact of heterogene-
1
ity, suppose that whites are the majority (w Ͼ ⁄
2
), and consider a
decrease in w leaving the size of the total population un-
changed—an increase in heterogeneity. In the first kind of equilib-
rium, all that happens is that the total number of whites is lower,
and since all W types were participating, the size of the group
2
0
decreases, and so does aggregate participation. More interesting
is the case of the ‘‘mixed’’ group. We have established that if B Ͻ W
then P * Ͻ P * , and that ѨP * /Ѩw Ͼ 0. Consider what happens to S
B
W
W
defined in (12) when whites are the majority and w decreases.
First of all, since ⌸(P *) Ͼ ⌸(1 Ϫ P *), the fall in w creates an
B
B
absolute loss in participants greater than the gain created by the
increase in (1 Ϫ w). Furthermore, the fractions of the two types
2
0. One could devise a decrease in w such that B types become the majority.
However, as long as the probability that two randomly drawn individuals belong to
different types has increased (our definition of heterogeneity), then the total mass
of B types must be lower than the total mass of W was before the change. This
implies that the new (homogeneous) group will be smaller than before and hence
confirms our conclusion.

8
60
QUARTERLY J OURNAL OF ECONOMICS
who participate change with the new value of heterogeneity. The
sufficient condition mentioned in the text ensures that the in-
crease in the fraction of blacks is not so overwhelmingly larger
than the decrease in the fraction of whites to overcome the
first effect. Appendix 1 reports a simple example to highlight
the critical features of our results in the one-group and in
the multiple-group context, and provides an intuitive graphical
illustration.
II.4. Multiple Groups
We now set a minimum size such that at most N Ն 2 groups
˜
k
˜
k
can exist in equilibrium. If we denote by B and W the mass of B
and W participants to group k, this amounts to requiring that
˜
k
˜
k
(A3)
B ϩ W Ͼ (B ϩ W)/(N ϩ 1).
The main differences with the results obtained so far arise
with respect to the possibility that multiple homogeneous groups
form; i.e., that B and W types sort into perfectly segregated
groups. For clarity of exposition, we will in turn analyze the case
of mixed and that of homogeneous groups, although it should be
clear that the same set of preferences can be consistent with some
groups being mixed and some homogeneous if multiple equilibria
for P * exist, as depicted for example in Figure II, panel (c).
B
In equilibrium a member ’s utility from participating in a
group must exceed not only the reservation utility u, but also the
utility that the same individual could get from joining another
group. As we will see, this implies that if two adjacent groups
exist, the individual located at the ‘‘border ’’ between the two
groups must be indifferent between them. Any group will there-
fore fall into one of two categories: one in which no neighboring
group exists and the ‘‘geographic coverage’’ of the group (i.e., the
distance between the two extreme members of the group) is
maximum and is determined by a participation constraint like (2);
and one in which there is at least one adjacent group and the
border(s) of the group is (are) characterized by the indifference
condition mentioned above. Before we turn to the analysis of these
two configurations, we define our equilibrium concept in the
multiple-group context.2
1
2
1. In the definition of an n-group equilibrium, we employ a slight abuse of
notation in using the intersection operator to express that two groups cannot

PARTICIPATION IN HETEROGENEOUS COMMUNITIES
861
DEFINITION 4. An n-group equilibrium consists of n couples
k*
k
k*
(
P ,L ), k ϭ 1, . . . , n Յ N, with P denoting the fraction of
B
B
k
members type B in group k, and L denoting the distance
between the two most distant members of group k, with
n
i
n
i
k*
k
⌺
L Յ 1 and ʝ L ϭ 0, such that for each (P ,L ) none of
iϭ1
iϭ1 B
the members wishes to leave the group and none of the
nonmembers wishes to join.
In what follows, we start by analyzing the case of multiple
mixed groups. We consider the case in which the utility function
admits a unique stable interior equilibrium, P * ʦ (0,1), and
B
investigate how the conclusions of the one-group model are
2
2
affected by the existence of multiple groups. We start by
considering an equilibrium with n disjoint mixed groups.
LEMMA 2. If an equilibrium with n disjoint mixed groups exists,
all groups must be identical, and each of them is defined by
the equilibrium conditions of the single-group model.
The fact that the groups are disjoint implies that the mem-
bers at the ‘‘frontier ’’ of the group are held down to their
reservation utility u. In this case the existence of other groups
does not affect the conditions characterizing the borders of the
group: the geographic coverage of the group is maximum and is
determined by condition (10) in Corollary 2. Furthermore, the
uniform distribution of B and W types on the line ensures that
every segment of it is characterized by the same proportion B/W in
the population. This, coupled with condition (10), implies that all
groups will be identical n-replicas of the single group analyzed in
2
3
the previous section. As a consequence, the following holds.
‘
‘serve’’ the same segment of the population (i.e., no individual can belong to two
groups at the same time).
2
2. When the utility function admits multiple interior equilibria in P *B, the
analysis is complicated by the fact that the quantitative impacts of an increase in
heterogeneity for two groups with different P *B’s will not be the same. An extensive
treatment of this case would not add much to our analysis.
2
3. Notice that the existence conditions for the multiple-group case are a
straightforward extension of the one-group model. In fact, as shown in the proof of
Proposition 1, the existence of an equilibrium proportion P *B depends on the
functional forms of f␣(·), fl(·), u (·), and v(·), which are the same in the single- and in
the multiple-group case. We just have to modify the condition that the minimum
size is met in equilibrium with the following one: ⌸(P *W,u ) · B ϩ ⌸(P B,u ) · W Ͼ
k
*
k
k
k
(
B ϩ W)/(N ϩ 1), k ϭ 1, . . . , n Յ N, where B (W ) is the mass of B(W) types in the
k
k
k
segment 2l identified by condition v(l ) ϭ u for disjoint groups, and by 2l ϭ 1/n
for adjacent groups.

8
62
QUARTERLY J OURNAL OF ECONOMICS
COROLLARY 3. An increase in heterogeneity decreases participa-
tion in an equilibrium with n mixed disjoint groups under the
same conditions under which it does so in an equilibrium with
one mixed group.
The consideration of multiple groups therefore does not alter
the conclusions of the one-group model in the case in which all
groups are disjoint and each of them behaves as an isolated
community. What if the groups are ‘‘adjacent’’? First of all, the
‘
‘geographic coverage’’ of each group can be below the maximum
identified by condition (10). Second, the members located at the
‘frontier ’’ between two groups must be indifferent between joining
‘
ˆ
k
one or the other. If we denote by l the distance of the farthest
member of group k from its HQ, this implies that for any two
neighboring groups k and k ϩ 1 (with k ϭ 1, . . . , n Ϫ 1), the
following holds for any member of group k with aversion pa-
j
ˆ
k
rameter ␣ , j ϭ B,W, and distance l :
u (␣ ,P ) ϩ v(l ) ϭ u (␣ ,P *) ϩ v( lˆ kϩ 1).
j
k*
ˆ
k
j
kϩ1
(13)
Ϫj Ϫj
Consider the case in which the utility function admits a
ˆ
k
unique stable interior equilibrium, P * ʦ (0,1), and let (P *,l ), k ϭ
B
B
1
, . . . , n be an equilibrium with n adjacent mixed groups. Then it
ˆ
k
ˆ
kϩ1
must be l ϭ l
for all k ϭ 1, . . . , n Ϫ 1. In other words, all
groups are identical. For this case, we still have that an increase
in heterogeneity decreases aggregate participation if it does so in
the one-group model.24 In the example presented in Appendix 1 we
show analytically why this is the case.
We now turn to the case of multiple homogeneous groups,
which yields the results most in contrast with the one-group
model. We can have multiple homogeneous groups when the
utility function gives corner equilibria like those depicted in
Figure II, panel (d). In this case the impact of heterogeneity on
participation depends on the relationship between the minimum
size requirement and the mass of the minority type, as stated in
the following proposition. Without additional conditions on the
2
4. Intuitively, what is different here is that the border members of each
k
group are no longer identified by the condition v(l ) ϭ u (indeed, the utility of the
border members with ␣ ϭ 0 is now strictly greater than u). This implies that the
extremes of integration with respect to l in expression (3) will be different,
although still independent of PϪj. However, the impact of heterogeneity on
participation depends on how changes in w influence ⌸(P *Ϫj,u ), via the impact on
P *Ϫj, and this basically depends on the shape of u (·). Due to the separability of the
utility function, changes in the extreme distance do not affect the sign of dS/dw.

PARTICIPATION IN HETEROGENEOUS COMMUNITIES
863
number of groups, Proposition 3 holds when there are an equal
number of all-B and all-W groups. In Appendix 1, however, we
develop the analysis with respect to a generic number N of all-B
B
groups and N of all-W groups. Denote by ⌬S/⌬(1 Ϫ w) the change
W
in aggregate participation due to a change in the fraction of B
types in the population.
PROPOSITION 3. Let w Ͼ 1⁄
. For any stable equilibrium in disjoint
2
homogeneous groups,
⌬
S
1
Ͻ 0
Ͼ 0
ϭ 0
for 2vϪ1(u)(1 Ϫ w) Ͻ
n ϩ 1
⌬
(1 Ϫ w)
⌬S
1
for 2vϪ1(u)(1 Ϫ w) ϭ
n ϩ 1
⌬
(1 Ϫ w)
⌬S
1
for 2vϪ1(u)(1 Ϫ w) Ͼ
.
⌬
(1 Ϫ w)
n ϩ 1
The intuition underlying Proposition 3 is the following. In an
equilibrium with disjoint homogeneous groups, all individuals
Ϫ1
within distance v (u) from the HQ of a group will join it,
regardless of their aversion parameter ␣. Starting from a situa-
tion where type W is an overwhelming majority and homogeneous
groups type B cannot meet the minimum size requirement, an
increase in heterogeneity decreases aggregate participation be-
cause it reduces the mass of the all-W groups without inducing
participation by B. At some point, however, the increase in (1 Ϫ w)
makes the minimum size feasible for B types, so that the decrease
in the size of existing W groups is more than compensated by the
formation of new B groups, and there is a discrete increase in
aggregate participation. Beyond this ‘‘switch’’ point, when equal
numbers of all-B and all-W groups exist, any increase in heteroge-
neity leaves aggregate participation unchanged because the de-
crease in the size of W groups equals the increase in that of B
groups.
In summary, many combinations are possible depending on
the number of groups and on the adjacent versus disjoint nature of
the groups. However, the main point is that compared with
mixed-group equilibria in which the effect of heterogeneity on
participation is ‘‘smooth’’ and (under mild sufficient conditions)
negative, the case of multiple homogeneous groups yields a

8
64
QUARTERLY J OURNAL OF ECONOMICS
discontinuous effect, whereby increased heterogeneity can actu-
ally increase participation.
II.5. Discussion
We have shown, thus far, that an increase in heterogeneity
can reduce participation, especially in the presence of mixed
groups. Before moving to the empirical evidence, it is worth
mentioning two points. First, the empirical literature on political
participation suggests that, after controlling for socioeconomic
status, blacks have a higher propensity to participate in groups
and to vote than whites.2 In our empirical analysis below, we also
find the same result. The explanation offered by political scien-
tists is that blacks are more conscious of being a minority and
have an extra incentive to engage in political action to preserve
their identity and foster their political and civil rights. In our
5
i
i
model this could be accommodated by assuming that ␣ ϭ ␣ (W/B),
with i ϭ B,W. That is, the propensity to participate is a function of
B
the distribution of types in the population, with Ѩ␣ /Ѩ(W/B) Ͼ 0
W
B
and Ѩ␣ /Ѩ(W/B) Ͻ 0. Therefore, if B Ͻ W, then we must have ␣ Ͼ
␣
W
. This implies that after controlling for all other determinants
of participation, one should obtain the empirical finding that
blacks participate more, since they are a minority.
Second, thus far we have focused on differences across types
not based on income. In the empirical analysis we are also
interested in the effect of an increase in income inequality on
participation. A vast literature in local public finance addresses
2
6
the issue of group formation and income levels. Our model is not
a contribution game. Differences in income matter only to the
extent that they are correlated with preferences and culture. In
this case our formalization could be reinterpreted in terms of
income rather than race; i.e., individuals would prefer to partici-
pate in social activities with people from their own income
bracket. A complication, however, is that income is a ‘‘continuous’’
variable, while race is much less so: in modeling income disper-
sion as a dispersion of ‘‘types,’’ we are therefore simplifying the
analysis.
2
2
5. See, for instance, Verba and Nie [1987].
6. See, for instance, Epple and Romer [1991]. Particularly related to our
analysis is the work by Fernandez and Rogerson [1996] and La Ferrara [2000],
since both examine the effects of changes in income inequality.

PARTICIPATION IN HETEROGENEOUS COMMUNITIES
865
III. EMPIRICAL STRATEGY AND DATA
In the remaining part of the paper, our aim is to focus on the
link between individual aversion to different types and the
decision to join groups, and to estimate the impact of increased
heterogeneity in the community on participation. For our basic
specification, we assume that at any point in time the ‘‘latent
variable’’ measuring the expected utility from participation in a
group for individual i in community c can be modeled as
(14)
Y* ϭ X ␤ ϩ H ␥ ϩ S ␦ ϩ T␭ ϩ ⑀ ,
ic ic c c ic
where X is a vector of individual characteristics; H is a vector of
ic
c
community variables (including heterogeneity), S is a dummy for
c
the state where the individual lives, T is a year dummy, and ⑀ is
ic
an error term normally distributed with mean 0 and variance ␴c.
The vectors ␤, ␥, ␦, and ␭ are parameters. We do not observe the
latent variable Y* but only the choice made by the individual,
ic
which takes value 1 (participate in a group) if Y* is positive, and 0
ic
(not participate) otherwise,
P ϭ 1
if Y* Ͼ 0
ic
ic
(15)
P ϭ 0
ic
if Y* Յ 0.
ic
We estimate the Probit model (14)–(15) using individual level data
and taking Metropolitan Sampling Areas (MSA) and Primary
Metropolitan Sampling Areas (PMSA) as ‘‘community’’ dimension.
We are especially interested in the vector of coefficients ␥,
although many of the components of ␤ will also be important to
gain insights into the determinants of participation.
The main source of data for our regressions is the General
Social Survey (from now on, GSS) for the years 1974–1994. This
survey interviews approximately 1500 individuals every year
from a nationally representative sample, and contains informa-
tion on a variety of sociopolitical indicators, as well as on
2
7
demographic and income characteristics of the respondents. In
particular, the questionnaire includes questions regarding the
2
7. Note that the survey was not conducted in 1979, 1981, and 1992.
Moreover, in 1982 and 1987 black individuals were oversampled; therefore, in our
regressions we will use the weights provided by the cumulative GSS file 1972–
1
994 to correct for this oversampling. The original 1972–1994 cumulative file
contains nominal income data for all years and real income up to 1993. In order to
maximize the number of observations, we have constructed real income figures for
1
994 following the same procedure that had been used for the previous years,

8
66
QUARTERLY J OURNAL OF ECONOMICS
respondents’membership in organizations such as political groups,
religious groups, unions, school associations, service groups,
fraternities, sports and hobby clubs, etc. We use the answers to
these questions to construct our dependent variables. The GSS
also contains information about individual attitudes toward race
relations and racial mixing. This will allow us to construct proxies
for our parameter of ‘‘participation aversion’’ (␣) in order to test
2
8
the implications of our model.
Among the explanatory variables we include individual con-
trols taken from the GSS, as well as community variables
capturing heterogeneity in race, ethnicity, and income in the place
where the individual lives. All variables are described in Appendix
2
. The remainder of this section illustrates our procedure for
constructing community level variables.
It is possible to match approximately two-thirds of the
respondents from the GSS 1972–1994 with the MSA/PMSA where
they live.29 We have used Census data to build community level
variables, adopting the MSA/PMSA as our geographic notion of
‘‘community.’’ Our measure of income inequality is the Gini
coefficient for the MSA/PMSA computed using family income
figures from the 1970, 1980, and 1990 Censuses. The values of
Gini for the remaining years were obtained by linear interpolation
and extrapolation. Our results are not sensitive to the interpola-
tion procedure. Moreover, we computed Gini coefficients at the
state level from the Current Population Survey (CPS) every year
between 1974 and 1994, and the correlation between the CPS and
the Census interpolated Gini’s was .65. The state level Gini’s from
the CPS are those we use in Figures IV, VII, and VIII.
Our racial fragmentation index (Race) is constructed from the
Census 1990 according to the following formula:
2
(16)
Race ϭ 1 Ϫ
s ,
i
͚
ki
k
where i represents a given MSA/PMSA and k the following races:
i) White; ii) Black; iii) American Indian, Eskimo, Aleutian; iv)
which is described in Ligon [1989]. For more detailed information about the GSS,
the reader is referred to Davis and Smith [1994].
2
8. In our model, traveling costs are also important. The GSS data, however,
do not allow us to know the distance between individuals’ residence and the
location of the group’s headquarters.
2
9. For about 40 percent of the individuals who can be matched with their
MSA/PMSA, however, the membership data are missing. This is why the number
of observations in our regressions will be smaller than the full GSS sample.

PARTICIPATION IN HETEROGENEOUS COMMUNITIES
867
Asian, Pacific Islander; v) other. Each term s is the share of race k
ki
in the population of MSA/PMSA i. The index (16) measures the
probability that two randomly drawn individuals in area i belong
to different races. Therefore, higher values of the index represent
3
0
more racial fragmentation.
The ethnic fragmentation index (Ethnic) is computed by a
formula analogous to (16), using ancestry instead of race. In other
words, ski in the formula now represents the fraction of people in
area i whose first ancestry is type k. The original ancestries
reported by the 1990 Census (35 categories) have been aggregated
into 10 different groups on the basis of common language, culture,
and geographic proximity (see Appendix 2 for a precise definition).
We have chosen to aggregate these data in order not to give the
same ‘‘weight’’ in the definition of Ethnic to very similar countries
of origin, say Norway and Sweden, and two very different ones,
say India and Ireland. Our results are not unduly sensitive to
reasonable changes in our aggregation rules.
Note that we use the values of Race and Ethnic in 1990 for the
whole sample. Our reasons for not interpolating are twofold. First,
we believe that racial and ethnic fragmentation within MSAs are
sensibly more stable over time than, say, income inequality.
Second, and most importantly, in order to get variation over time,
we should have resorted to the 1970 and 1980 Censuses, which
contained fewer categories. For example, all Censuses before 1990
distinguished only three races: white, black, and other. Relying on
years earlier than 1990 would thus have meant sacrificing the
precision of our heterogeneity measures to a considerable extent.
We felt that the loss in explanatory power due to this oversimplifi-
cation outweighed the potential gain from time variation of the
above indexes. Hence we chose to adopt the 1990 measures as our
best proxies for racial and ethnic fragmentation.
Before proceeding, we need to justify our choice of MSA/
PMSA as our geographical units. To begin with, we cannot use a
smaller unit such as the PUMA because of a lack of respondent
identifiers at this level in the GSS. However, we feel comfortable
with MSA/PMSA for several reasons. First, one may argue that by
3
0. The Census did not identify ‘‘Hispanic’’ as a separate racial category.
However, Alesina, Baqir, and Easterly [1999], who use the same measure of racial
fragmentation, note that the category ‘‘Hispanic’’ (which they obtain from a
different source) has a correlation of more than 0.9 with the category ‘‘other ’’ in the
Census data. Thus, for all practical purposes, the category ‘‘other ’’ in the Census is
virtually a measure of the Hispanic population.

8
68
QUARTERLY J OURNAL OF ECONOMICS
using a relatively large geographical unit one may bias our
regressions against finding an effect of fragmentation on participa-
tion. Second for several of the groups that we consider, it is
reasonable to assume that direct interaction among group mem-
bers occurs at the MSA level (think of unions, sport groups, and
boy scouts, for instance). Third, and most importantly, we checked
the correlation between our measures of fragmentation at the
MSA/PMSA and at the PUMA level. The results are comforting.
The correlation for the Gini coefficient is 0.65; the correlation for
3
1
Race is 0.75; and the one for Ethnic is 0.75. Finally, we have
checked that a large proportion of the respondents in the GSS
lives in nonsegregated areas, and therefore potentially interacts
with the opposite race. It turns out that 71 percent of white
respondents say that ‘‘there are black families living close’’ to
them, and 48 percent of white respondents say that black families
live ‘‘on this block, a few doors/houses away.’’
IV. DESCRIPTIVE RESULTS
We begin by presenting summary statistics and a few simple
correlations among membership rates and our measures of hetero-
geneity. Summary statistics and definitions of the full set of
variables are in Appendix 2.
The top panel of Table I shows the sample characteristics of
some of our data. Participation rates are on average very high:
overall, 71 percent of the respondents are members of at least one
group, the average number of group memberships being 1.8 per
person. Also, there is considerable variation in participation rates,
both across individuals and across groups: the standard deviation
of our basic membership variable is 0.45. The fraction of partici-
pants in the various groups ranges from 0.02 for farmers’ associa-
tions to 0.34 for religious groups. Sport groups are the second most
popular category, with a participation rate of 0.20, followed by
professional associations (0.17), unions (0.15), and school service
groups (0.14). Literary groups, hobby clubs, fraternities, and
service groups (Rotary, Lions, etc.) have participation rates of
0
.09–0.10. Most notable is the low enrollment in political associa-
tions: only 5 percent of the respondents are members of a political
3
1. These correlations are weighted by the population share of each PUMA in
the MSA/PMSA.

PARTICIPATION IN HETEROGENEOUS COMMUNITIES
869
TABLE I
DESCRIPTIVE STATISTICS
Panel A: Summary statistics
Mean
Std. dev.
Member of any group
Member of church group
Member of fraternity
Member of service group
Member of hobby club
Member of sport club
Member of youth group
Member of literary group
Member of school service group
Member of school fraternity
Member of veterans’ group
Member of political group
Member of nationality group
Member of union
Member of professional association
Member of farmers’ group
Member of other group
Gini
.71
.34
.09
.10
.10
.20
.10
.10
.14
.05
.06
.05
.04
.15
.17
.02
.11
.41
.36
.67
.45
.47
.29
.30
.29
.40
.30
.30
.35
.23
.24
.21
.20
.36
.37
.14
.31
.03
.15
.07
Racial fragmentation (Race)
Ethnic fragmentation (Ethnic)
Panel B: Correlations
Membership
Gini
Race
Gini
Race
Ethnic
Ϫ.06**
Ϫ.05**
Ϫ.04**
.34**
.09
.56**
*
Denotes significance at the 5 percent level; **at the 1 percent level.
group. Nationality groups, which we will consider in more detail
below, are joined by about 4 percent of the respondents.
The last three variables in Panel A of the table are measures
of heterogeneity in income, race, and ethnicity in the MSAs where
the respondents live. The mean of the Gini coefficient is 0.41, with
a standard deviation of 0.03. Our racial fragmentation index has a
mean of 0.36 (standard deviation 0.15), while ethnic fragmenta-
tion is higher at 0.67 (standard deviation 0.07). The correlation
among these three measures of heterogeneity is quite high, as
shown in the bottom panel of Table I: Gini is correlated 0.34 with

8
70
QUARTERLY J OURNAL OF ECONOMICS
FIGURE III
Average Membership Rate, 1973–1994
Race and 0.09 with Ethnic; the correlation between Race and
Ethnic is 0.56.32 On the other hand, the simple correlation
between average membership in the MSA and the Gini index is
not quantitatively significant; this will no longer be true when we
turn to multivariate analysis.
Figures III, IV, V, and VI illustrate the geographic distribu-
tion of our variables of interest by reporting sample averages at
the state level.
Figure III shows the distribution of participation rates from
the GSS data set, that is, the percentage of respondents in each
3
2. We also explored the correlation among our heterogeneity variables and
the measures of racial or income segregation of Cutler, Glaeser, and Vigdor [1999].
It turns out that segregation by income is positively correlated with income
inequality, and that the various measures of racial segregation are positively
correlated with racial and ethnic fragmentation. This may suggest that segrega-
tion is ‘‘valued’’ relatively highly in places where heterogeneity is higher, an
observation consistent with the argument that individuals prefer contact with
people similar to themselves.

PARTICIPATION IN HETEROGENEOUS COMMUNITIES
871
FIGURE IV
Gini Coefficient, 1972–1994
state who are members of at least one group (average from 1974 to
994). As one can see, this percentage is highest in states of the
1
north central and northwest regions, and lowest in the south and
southeast. Figures IV, V, and VI show the distribution of the Gini
coefficient (average 1974–1994 from CPS data), of Racial fragmen-
tation, and of Ethnic fragmentation (both measured in 1990), all
calculated at the state level. These maps show a rather striking
pattern, when compared with Figure III: racial and ethnic fragmen-
tation, as well as income inequality, are highest in the southeast
and lowest in the northeast; i.e., those regions where participation
is, respectively, lowest and highest.
Similar implications can be gathered from the three panels of
Figure VII, which plot state level participation rates against the
Gini index, Race fragmentation, and Ethnic fragmentation. In all
three panels a negative correlation between membership in
groups and heterogeneity clearly emerges.

8
72
QUARTERLY J OURNAL OF ECONOMICS
FIGURE V
Racial Fragmentation, 1990
V. THE ECONOMETRIC EVIDENCE
V.1. Basic Regressions
Table II displays our basic probit regression using the GSS
data set and including only individual controls. The dependent
variable takes the value 1 if the respondent belongs to at least one
group, and 0 otherwise. The regressors include a set of individual
characteristics that, in our model, may influence either the
individual’s reservation utility if not participating, u, or the
3
3
preference for participation captured by the parameter ␣.
3
3. The political science literature has generally looked at these individual
determinants of participation in isolation, i.e., correlating one or two variables at a
time with membership rates. See, for instance, Verba and Nie [1987] and Verba,
Schlozman, and Brady [1995]. A multivariate analysis with demographic controls
similar to those we include is in Glaeser and Glendon [1997], but their dependent
variable is church attendance rather than group participation.

PARTICIPATION IN HETEROGENEOUS COMMUNITIES
873
FIGURE VI
Ethnic Fragmentation, 1990
The estimates in the first column of Table II are marginal
probit coefficients evaluated at the means; in the second column
we report heteroskedasticity-corrected standard errors adjusted
for intra-MSA clustering of the residuals. First of all, the cohort
variable suggests a decline in participation by younger cohorts.
Second, the age distribution variables show a dip in participation
for individuals in their thirties. Child-rearing activities reduce
the time available for participation: in fact, the coefficient on the
variable that captures whether the respondent has children below
the age of five is negative and significant. Note that the dummy for
age group 30–39 and that for children below age 5 are highly
correlated. More generally, both variables capture a period of
individual lifetime that is particularly ‘‘busy’’because of marriage,
having children, setting up new households, etc. There is some
weak evidence that older people participate more, probably
because they have more time if they are retired, although health

FIGURE VII
Heterogeneity and Participation in Groups

PARTICIPATION IN HETEROGENEOUS COMMUNITIES
875
TABLE II
INDIVIDUAL DETERMINANTS OF PARTICIPATION
Marg. Probit coeff.a
Std. errorb
Cohort
Ϫ.002*
Ϫ.035
Ϫ.029*
.004
(.001)
(.028)
(.018)
(.016)
(.027)
(.011)
(.011)
(.011)
(.013)
(.012)
(.014)
(.012)
(.014)
(.006)
(.011)
(.016)
Age Ͻ30
Age 30–39
Age 50–59
Age Ն60
Married
Female
.035
Ϫ.001
Ϫ.046**
.045**
Black
Educ Ͻ12 yrs
Educ Ͼ16 yrs
Children Յ5 yrs
Children 6–12
Children 13–17
ln (real income)
Full-time
Ϫ.122**
.144**
Ϫ.035**
.071**
Ϫ.005
.074**
.025**
.067**
Yes
Part-time
STATES
YEARS
Yes
No. obs.
Pseudo R
10534
.09
2
Observed P
Predicted P
.72
.74
*
Denotes significance at the 10 percent level; **at the 5 percent level.
a. Marginal probit coefficients are calculated at the means.
b. Standard errors are corrected for heteroskedasticity and clustering of the residuals at the MSA level.
considerations may work the other way. This result, together with
the cohort effect, accounts for the notion of ‘‘older civic generation’’
emphasized by Putnam [1995a, 1995b].
Years of schooling are positively associated with participa-
tion: high school dropouts participate significantly less, while
college graduates significantly more. The coefficients on the
education variables remain highly significant and stable through-
out all specifications. Among the possible explanations for this
strong association, Verba and Nie [1987] suggest that more
education is generally combined with a higher evaluation of one’s
own ability to influence sociopolitical outcomes, and with a higher
level of social interaction. In Table II we also see that women
participate significantly less than men. Our interpretation is that
they often carry the weight of a job plus a preponderant share of

8
76
QUARTERLY J OURNAL OF ECONOMICS
household chores and child-rearing activities: this heavy load
leaves women with less time for leisure and participation. We
checked, in fact, that women do not participate less in voting, an
act of participation which does not require a significant amount of
time. The probability of being a member of a group is increasing in
family income of the respondent, suggesting that participation is a
‘‘normal good.’’ We investigate in more detail the nonlinear effects
of income on participation in Section V.
Consider now the time spent at work. The omitted category
captures people who are not working, including homemakers,
retirees, students, and unemployed. After controlling for the level
of income, the effect of time spent at work could be twofold. On the
one hand, a constraint on time may decrease participation; on the
other hand, socialization in the workplace may increase social
interaction, incentives, and ability to participate. Our results on
this point are consistent with basic economic principles. The
coefficient on part-time workers is larger (and significantly so)
than the one on full-time workers. This suggests that, even
though socialization in the workplace helps (in fact, full-time
workers participate more than those out of the labor force), the
time constraint is binding for people who work full time. It is less
binding for part-time workers who, on the other hand, still get the
3
4
benefits of social interactions in the workplace. When we control
for all these variables, marital status does not seem to affect
participation (contrary to the common notion in sociological and
political analyses based on partial correlations, which indicate
that married people participate significantly more).
For our purposes, a particularly interesting variable is ‘‘black.’’
As we can see from Table II, ceteris paribus, members of this
racial group participate significantly more, a result which emerges
clearly when we control for the other individual determinants of
participation but which is obscured if we only look at partial
correlations. Note that this result is not driven by the higher
church attendance of blacks in the south: in fact, it survives if
churches are left out of the definition of groups and if the south is
omitted from the regression. As discussed above, our model could
be extended to incorporate a feature of group consciousness, in
which the minority type participates more to preserve identity
3
4. This interpretation is confirmed by additional sensitivity analysis. Inter-
estingly, the unemployed participate less even after controlling for income.