University of Southern California Dissertations and Theses

Three essays on cooperation, social interactions, and religion

(USC Thesis Other)

Three essays on cooperation, social interactions, and religion

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content THREE ESSAYS ON COOPERATION,
SOCIAL INTERACTIONS, AND RELIGION
by
Arya B. Gaduh
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulllment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)
August 2013
Copyright 2013 Arya B. Gaduh
Acknowledgments
Through the journey that was graduate school, I have accrued debts to many. Fore-
most of these is to my adviser and mentor, John Strauss, who began as a teacher and
ended up a friend. John was there from the journey's very beginning to its very end.
In between, he prodded, pushed, and never pulled his punches when discussing my
work in our internal meetings; yet at the same time, was my strongest advocate to
the outside world, perpetually connecting me to his wide array of academic networks.
When push came to shove during the job market process, he spent time to prepare
each of his students { he had ve { and helped guide us through each step of the way.
I cannot imagine that I could be where I am today as an applied economist were it
not for his support.
I am also indebted to Juan Carrillo and Jerey Nugent. Juan piqued my interest
in experimental economics, took a gamble by investing in my unconventional idea,
and helped develop it into a full-blown experiment. In the process, he has shaped
my approach toward economic experiments in important ways, and has become a
mentor, a colleague, and a friend. Meanwhile, Je was most generous in his friendship,
wise advice, and willingness to provide detailed feedback on the various drafts of my
research.
I am thankful to Geert Ridder who served in my qualifying and Gary Painter
who have served in both the qualifying and the dissertation committees. They have
provided valuable inputs to my research. Meanwhile, Menno Pradhan also lent his
support before this journey began and during the job market process, for which I am
truly grateful.
ii
The journey would have felt longer were it not for a great many friends. Saurabh
Singhal and James Ng are (almost) always ready to drop what they are doing for
random discussions over a cup of coee. Together with Osman Abbaso glu, we were
brought together by that bane of all rst-year students' existence. Meanwhile, chats
with Tirta Susilo, Philips Vermonte, and Teguh Yudo Wicaksono are always stimulat-
ing { or, at the very least, entertaining. I have also been very fortunate to be working
with Samuel Bazzi, Alexander Rothenberg, and Maisy Wong. In big and small ways,
fragments of these conversations have sneaked in this nal work and made it all the
better.
In addition, I want to acknowledge Young Miller, Morgan Ponder, and Christopher
Frias who have helped make the various administrative processes throughout my
years at USC as painless as anyone could expect. I am also grateful for the dierent
fellowships from USC's College of Letters and Science, Department of Economics, and
Graduate School during the 2010-2013 period, as well as partial funding provided by
the American-Indonesian Cultural and Educational Foundation (AICEF) during the
2010-2012 period.
Last, but not least, I thank my family for their love and support. After my father
passed away when I was seven, my mother gave us an abundance of love and single-
handedly provided me with the best possible education any Indonesian could expect,
and then some. She also taught me the importance of perseverance and excellence.
Her advice has carried me through graduate school { and, indeed, through most of
my life. Meanwhile, my siblings have always been great companies, and each helps
me appreciate the diverse ways religious identities can shape attitudes. Finally, I
thank my wife, Ellen Tjahja, whose love and support have always been among the
few constants in my life throughout this journey.
iii
Table of Contents
Acknowledgments ii
List of Tables xi
List of Figures xii
Abstract xiii
1 Introduction 1
1.1 Experimental evidence on network formation . . . . . . . . . . . . . . . 2
1.2 Religion: Uniter or Divider? Evidence from Indonesia . . . . . . . . . . 4
1.2.1 Individual religious identities and social cooperation . . . . . . . 6
1.2.2 Social interactions, religion, and cooperation . . . . . . . . . . . 7
Chapter bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 The Dynamic Evolution of Social Networks: Experimental
Evidence
1
14
2.1 Network environment and basic denitions . . . . . . . . . . . . . . . . 17
2.2 Experimental setting and procedures . . . . . . . . . . . . . . . . . . . 20
2.2.1 The basic conguration . . . . . . . . . . . . . . . . . . . . . . . 20
2.2.2 Treatments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2.3 Implementation details . . . . . . . . . . . . . . . . . . . . . . . 25
1
This chapter is co-authored with Juan Carrillo.
iv
2.3 Network outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.1 Network convergence: eciency vs. stability . . . . . . . . . . . 27
2.3.2 Network stability: myopic vs. farsighted . . . . . . . . . . . . . . 31
2.3.3 Payos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4 Individual decisions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4.1 Descriptive statistics . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4.2 Regression analysis . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.4.3 Model predictions . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.4.4 Extensions and alternative representations . . . . . . . . . . . . 47
2.4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.5 Heterogeneity across subjects . . . . . . . . . . . . . . . . . . . . . . . 49
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Chapter bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Appendices
2.A Additional analyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.A.1 Geodesic distance to ecient and PWS networks . . . . . . . . 60
2.A.2 Myopic rationality by treatment and decision problem . . . . . . 62
2.A.3 Pooled FE LPM on probability of myopic rational choices . . . . 62
2.A.4 Logit model for out-of-sample predictions . . . . . . . . . . . . 62
2.A.5 Logit estimation . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.A.6 Eect of experience . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.B Experimental instructions . . . . . . . . . . . . . . . . . . . . . . . . . 70
v
3 Religious Identities and Cooperative Attitudes in Indonesia:
Evidence from IFLS 84
3.1 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.2 Data and measurements . . . . . . . . . . . . . . . . . . . . . . . . . . 89
3.2.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
3.2.2 Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.3.1 Individual and household characteristics . . . . . . . . . . . . . . 103
3.3.2 Religiosity, religious education, and attitudes . . . . . . . . . . . 110
3.3.3 Does the religion matter? . . . . . . . . . . . . . . . . . . . . . . 117
3.3.4 Gender dierences in the religiosity correlates . . . . . . . . . . . 121
3.3.5 Religiosity and political preference . . . . . . . . . . . . . . . . . 123
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
Chapter bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Appendix
3.A Additional analyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
3.A.1 Robustness check: Household xed-eects model . . . . . . . . . 132
3.A.2 Inter-religion dierences in religiosity eects for the very religious 132
3.A.3 Pooling test of inter-gender dierences . . . . . . . . . . . . . . . 135
4 Religious Diversity, Segregation and Cooperative Attitudes in
Indonesia 139
4.1 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
4.2 Data and measurements . . . . . . . . . . . . . . . . . . . . . . . . . . 142
4.2.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
4.2.2 Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
vi
4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
4.3.1 Does segregation matter? . . . . . . . . . . . . . . . . . . . . . . 148
4.3.2 Diversity and segregation . . . . . . . . . . . . . . . . . . . . . . 149
4.3.3 Heterogenous responses to community heterogeneity . . . . . . . 151
4.3.4 Political preference in heterogeneous communities . . . . . . . . 154
4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Chapter bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Appendices
4.A Additional analyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
4.A.1 Community characteristics . . . . . . . . . . . . . . . . . . . . . 163
4.A.2 Inter-religion dierences in cooperative attitudes . . . . . . . . . 163
4.A.3 The role of the majority status . . . . . . . . . . . . . . . . . . . 166
4.A.4 Religiosity and community heterogeneity: Muslim v. non-Muslim 170
4.A.5 Community heterogeneity and the very religious . . . . . . . . . 175
4.B Data appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
4.B.1 Data sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
4.B.2 Merging the datasets . . . . . . . . . . . . . . . . . . . . . . . . 178
4.C A note on segregation measures . . . . . . . . . . . . . . . . . . . . . . 181
4.C.1 Criteria for evaluation . . . . . . . . . . . . . . . . . . . . . . . . 181
4.C.2 Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
4.C.3 Usage in the literature . . . . . . . . . . . . . . . . . . . . . . . 187
5 Conclusion 188
5.1 Social networks, cooperation, economic incentives . . . . . . . . . . . . 188
5.2 Religious identities and social cooperation . . . . . . . . . . . . . . . . 190
Chapter bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
vii
Comprehensive bibliography 194
viii
List of Tables
2.1 Ecient and PWS networks . . . . . . . . . . . . . . . . . . . . . . . 26
2.2 Summary of network outcomes . . . . . . . . . . . . . . . . . . . . . . 28
2.3 Network convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4 Summary of network outcomes conditional on convergence . . . . . . 30
2.5 Summary of network values . . . . . . . . . . . . . . . . . . . . . . . 33
2.6 Myopic rationality of individual decisions . . . . . . . . . . . . . . . . 34
2.7 Movements to and from PWS networks . . . . . . . . . . . . . . . . . 37
2.8 The regression coecients and the types of decision problems . . . . . 38
2.9 FE LPM on myopic rationality: eect of type of decision . . . . . . . 41
2.10 FE LPM on myopic rationality: eect of marginal payo . . . . . . . 43
2.11 FE LPM on the likelihood of myopic rational action . . . . . . . . . . 45
2.12 Clustering based on myopic rational behavior . . . . . . . . . . . . . 51
2.13 Average distances from outcomes . . . . . . . . . . . . . . . . . . . . 61
2.14 Average distance from outcomes conditional on convergence . . . . . 61
2.15 Pooled FE LPM on the likelihood of myopic rational action . . . . . . 64
2.16 Logit regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.17 FE Logit on myopic rationality: eect of type of decision . . . . . . . 66
2.18 FE Logit on the likelihood of myopic rational action . . . . . . . . . . 67
2.19 Pooled FE LPM on the likelihood of myopic rational action . . . . . . 69
3.1 Summary Statistics: Willingess to Help, Trust and Tolerance . . . . . 95
3.2 Distribution of Religiosity . . . . . . . . . . . . . . . . . . . . . . . . 96
ix
3.3 Share of Practicing Individuals for a Given Religiosity . . . . . . . . . 98
3.4 Share Participating in Religious Activities in the Village
y
. . . . . . . 99
3.5 Summary Statistics: Individual and Household Regressors . . . . . . 100
3.6 Community Cohesion & Trust Beliefs . . . . . . . . . . . . . . . . . 104
3.7 Discriminative Trust & Tolerance . . . . . . . . . . . . . . . . . . . . 105
3.8 Religiosity and religious education . . . . . . . . . . . . . . . . . . . . 112
3.9 Selection on observables to assess potential bias from unobservables . 118
3.10 Inter-Religion Dierences in the Associations between Religiosity and
Attitudes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
3.11 Religiosity by Gender . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
3.12 Religiosity and District Head Criteria . . . . . . . . . . . . . . . . . . 125
3.13 Community Cohesion & Trust Beliefs . . . . . . . . . . . . . . . . . 133
3.14 Discriminative Trust & Tolerance . . . . . . . . . . . . . . . . . . . . 134
3.15 Inter-Religion Dierences in the Associations between Religiosity and
Attitudes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
3.16 Pooling Test: Inter-gender Dierences in Community Cohesion &
Trust Beliefs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
3.17 Pooling Test: Inter-gender Dierences in Discriminative Trust &
Tolerance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
4.1 Summary Statistics: Community Regressors . . . . . . . . . . . . . . 146
4.2 Diversity, Segregation and Attitudes . . . . . . . . . . . . . . . . . . . 150
4.3 Diversity, Segregation, and Community Cohesion & Trust Beliefs . . 152
4.4 Diversity, Segregation, and Discriminative Trust & Tolerance . . . . 153
4.5 Diversity, Segregation and District Head Criteria . . . . . . . . . . . . 155
4.6 Community Compositions, Religiosity, and District Head Criteria . . 156
x
4.7 Community characteristics & Community Cohesion & Trust Beliefs . 164
4.8 Community characteristics & Discriminative Trust & Tolerance . . . 165
4.9 Inter-religion Dierences in Community Cohesion . . . . . . . . . . . 167
4.10 Inter-religion Dierences in Trust Beliefs . . . . . . . . . . . . . . . . 168
4.11 Inter-religion Dierences in Tolerance . . . . . . . . . . . . . . . . . 169
4.12 Community Compositions, Religiosity and Community Cohesion by
Muslim/Non-Muslim . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
4.13 Community Compositions, Religiosity and Trust Beliefs by
Muslim/Non-Muslim . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
4.14 Community Compositions, Religiosity and Tolerance by
Muslim/Non-Muslim . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
4.15 Diversity, Segregation, and Community Cohesion & Trust Beliefs . . 175
4.16 Diversity, Segregation, and Discriminative Trust & Tolerance . . . . 176
4.17 Variables by data sources . . . . . . . . . . . . . . . . . . . . . . . . . 179
xi
List of Figures
2.1 User interface for the linking game. . . . . . . . . . . . . . . . . . . . . 23
2.2 Out-of-sample logit prediction by treatment . . . . . . . . . . . . . . . 47
2.3 Empirical CDF of myopic rationality by treatment . . . . . . . . . . . 50
2.4 Improving paths in Treatment 1 . . . . . . . . . . . . . . . . . . . . . 53
2.5 Improving paths in Treatment 2 . . . . . . . . . . . . . . . . . . . . . 54
2.6 Improving paths in Treatment 3 . . . . . . . . . . . . . . . . . . . . . 55
2.7 Improving paths in Treatment 4 . . . . . . . . . . . . . . . . . . . . . 56
2.8 Myopic rationality by treatment and decision problem . . . . . . . . . 63
3.1 Trust game . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.1 Distributions of the diversity and segregation measures . . . . . . . . . 145
xii
Abstract
This dissertation provides empirical evidence on how economic incentives and religious
identities can shape social cooperation. It consists of three essays. The rst essay
(Chapter 2) uses a laboratory experiment to examine how economic motives shape
the strategic formation of social networks in a six-player linking game where link
creation requires mutual consent. The experiment provides empirical evidence on
the extent to which the strategic network formation theory proposed by Jackson and
Watts (2002) is able to predict the nal network from this linking process and the
validity of its underlying behavioral assumptions.
Meanwhile, the second and third essays (in Chapters 3 and 4) use observational
data to evaluate the role of religious identities in social cooperation. Chapter 3
utilizes the Indonesia Family Life Survey (IFLS) data to investigate how individual-
level variations in religious denominations and religiosity are linked to particularized
and generalized trust, and inter-group discrimination and tolerance in contemporary
Indonesia. I nd that religiosity is positively associated with particularized trust
and in-group preference, and negatively with religious tolerance. These ndings are
consistent with the notion that religion may facilitate \parochial altruism", which is
altruism toward members of one's own group combined with hostility toward members
of the out-groups. In Indonesia, this link is strongest for Muslims.
Chapter 4 extends the analysis to the community level and examines the in
uence
of social interactions among people with dierent religious identities on cooperative
attitudes. Using a dataset that combines the IFLS data with the 2000 Indonesian
National Census microdata, the 2000 Poverty Map and the 2005 Village Potential
xiii
statistics (or Podes), it examines how the religious diversity and segregation in the
communities where people reside are linked to their cooperative attitudes. I nd that
people tend to be more cooperative and trusting in more religiously homogeneous
communities, but exhibit a stronger in-group trust and are less tolerant of members
of the religious out-groups. On the other hand, segregation matters for some outcomes
{ and when it does, its eects tend to be opposite those of diversity. Overall, the
evidence suggests that social interactions may have had a role in amplifying and
ameliorating the parochial eects of religion.
xiv
Chapter 1
Introduction
[How] could this oddly cooperative animal,Homosapiens, ever have come to be?. . . [We]
advance two propositions. First, people cooperate not only for self-interested reasons
but also because they are genuinely concerned about the well-being of others, try to
uphold social norms, and value behaving ethically for its own sake. . . Second, we came
to have these \moral sentiments" because our ancestors lived in environments, both
natural and socially constructed, in which groups of individuals who were predisposed
to cooperate and uphold ethical norms tended to survive and expand relative to other
groups, thereby allowing these prosocial motivations to proliferate.
{ Samuel Bowles and Herbert Gintis, A Cooperative Species (2011), Chapter 1, p.1
This dissertation is an empirical investigation into how economic incentives and so-
cial identities can shape social cooperation. It comprises three essays. The rst essay
utilizes an economic experiment to study how economic incentives can shape social
cooperation through its in
uence on the formation of social networks. Meanwhile, the
second and third essays use observational data to explore the link between religious
identities and social cooperation in contemporary Indonesia. Each of these essays
examine how religious identities operate at dierent levels { to wit, at individual and
community levels. More specically, the second essay looks at how individual reli-
gious identities may in
uence cooperative attitudes. Religion predisposes individuals
to cooperate and uphold ethical norms (at least, its own community), and this essay
looks at its likely role in shaping both intra- and inter-group cooperative attitudes.
Meanwhile, the third essay explores how intra- and inter-group interactions may in-
1

uence individual attitudes toward social cooperation. In particular, it examines how
variations in the level of religious diversity and segregation in the various communities
are associated with cooperative attitudes. The following sections elaborate each of
these essays.
1.1 Experimental evidence on network formation
The rst essay (Chapter 2) focuses on understanding how economic incentives can
shape social networks { and hence, social cooperation.
1
Social networks can often
provide an imperfect substitute for market institutions, especially in less advanced
societies. Social networks facilitate information exchange when formal channels are
unavailable (Munshi, 2007). They also work as a means to smooth consumption in
the absence of formal insurance markets (Udry, 1994; Fafchamps and Lund, 2003).
Even markets can be seen not as anonymous institutions of exchange but as networks
that facilitate exchange between buyers and sellers (Kranton and Minehart, 2001).
Sociologists have, for a while now, recognized the importance of networks and social
structures, and more recently, economists too have joined the fray (Granovetter, 2005;
Jackson, 2005).
More recently, theoretical work in economics has begun to explore how networks
are formed and the role of socio-economic incentives in shaping them.
2
This literature
provided social scientists with a set of important models of network formation. In
particular, Jackson and Wolinsky (1996) analyze cases where links between two agents
in a network are formed cooperatively. More specically, their model analyzes an
underlying game where links between two agents are formed by mutual agreement,
1
This chapter is co-authored with Juan Carrillo from the Department of Economics, USC.
2
In particular, the two seminal theoretical papers are Bala and Goyal (2000) and Jackson and
Wolinsky (1996).
2
but can be broken unilaterally. Their theory provides a set of predictions regarding
the nal shape of the network given how its participants value dierent network
connections under a particular behavioral assumption.
Given the model's prominence, it is important to examine the extent to which its
predictions are empirically supported.
3
However, there are at least three reasons why
it would be dicult to test this theoretical model with observational data. First, the
data necessary for such an analysis are hard to nd. The researcher would need to
observe the complete set of links in the network over time to examine the process of
network formation. To my knowledge, a full-network longitudinal dataset does not
yet exist. Second, a test of these models requires the researcher to observe individual
valuations of their network connections. Again, this information is very dicult to
obtain, especially since individuals often have overlapping objectives in forming a
network. Finally, the econometric methodology necessary to identify and estimate
the dierent parameters under dierent types of network formation games is still in
its infancy.
4
To get around these diculties, we turn to economic laboratory experiments.
In an experiment, the values that individuals would receive in an economic game
across multiple turns can be controlled. One can therefore calculate the shape of the
network that is predicted by theory and compare the experimental outcomes with the
theoretical predictions. In this experiment, we test a theoretical model of cooperative
network formation proposed by Jackson and Watts (2002) where the network evolves
one link at a time. In doing so, this study joins a small number of recent experimental
studies that have investigated this particular question (Pantz, 2006; Kirchsteiger et
al., 2011).
3
According to Google Scholar, Jackson and Wolinsky (1996) has been cited by more than 1,500
papers as of March 2013.
4
See Sheng (2012) and the literature therein.
3
Under the assumption that agents' linking decisions are myopically rational { i.e.,
agents take into account only the net benet of the immediate link and not on the
option value of forming or severing links in the future { then the theoretical model
predicts that this process will converge to a pairwise-stable (PWS) network when
one exists and will not converge if none exists.
5
The ndings of our experiment oer
qualied support for the predictions of their theoretical models. We nd evidence that
the network formation process tends to converge to the PWS network when it exists
and not to converge to any network when a PWS network does not exist. However,
convergence is by no means guaranteed, especially when the PWS network involves
asymmetric payos among subjects. Moreover, we also nd that with multiple PWS,
subjects often converge to the higher-payo PWS network that, in a model with
purely myopic-rational agents, should not have been reachable.
1.2 Religion: Uniter or Divider? Evidence from Indonesia
Meanwhile, Chapters 3 and 4 focus on the role of religious identities on social coop-
eration. For believers, religious beliefs shape other attitudes that can determine the
individual behaviors that ultimately aect welfare and economic outcomes (Akerlof
and Kranton, 2000; Deaton, 2009; Lehrer, 2009). The propensity for social coopera-
tion is an instance of a behavior linked to religiously-shaped attitudes. Two aspects
of religion may in
uence cooperation through its eects on attitudes. First, all world
religions urge their believers to extend benevolence to others, including strangers
(Neusner and Chilton, 2005). At the same time, however, many religious traditions
emphasize the importance of religious communities (Iannaccone, 1992; Berman, 2000).
5
A pairwise-stable (PWS) network is a network that satises the following: (i) all existing links
are weakly preferred by both agents in the link and are strictly preferred by at least one of them;
and (ii) all non-existing links are such that at least one of the agents on the non-existing link strictly
prefers its absence (Jackson and Wolinsky, 1996).
4
This emphasis endows a believer with a social identity while at the same time creates
a categorical distinction between believers and non-believers. In religiously diverse
societies, therefore, the net eects of religion on social cooperation in religiously di-
verse societies are ambiguous: The exhortation to charity may improve the general
propensity to cooperate, but the emphasis on the religious community may focus
that cooperation toward members of one's own religious community at the expense
of outsiders.
I examine how religious identities can in
uence attitudes toward cooperation and
how these eects may be amplied by social interactions. The analysis focuses on
the following outcomes: willingness to help others, trust, and religious tolerance. The
social capital literature provides macro- and micro-level evidence of the link between
generalized trust and institutional quality (e.g., Putnam et al., 1993; Knack and
Keefer, 1997; La Porta et al., 1997), while historical evidence suggests that intolerance
may impose a long-term cost on growth (e.g., Landes, 1998; Chaney, 2008). At the
same time, observational studies using cross-country and rich-country datasets as well
as a number of experimental studies suggest that religion do in
uence trust as well
as other cooperative attitudes.
6
However, we nd very few national-level studies in
developing countries that examine the determinants of trust and tolerance and their
links to religious identities.
Combining the latest Indonesia Family Life Survey (IFLS) { which contains new
modules on religion, trust, and tolerance { with a number of national-level datasets,
these chapters hope to ll this gap with a pair of studies on contemporary Indone-
sia. Indonesia presents an excellent opportunity for a country-level analysis of this
question, given its religious diversity: It is a Muslim-dominant country, albeit with
a considerable share of believers from other major world religions { Protestantism,
6
See the literature review in Chapter 3.
5
Catholicism, Hinduism and Buddhism { represented in its population. With this
dataset, I am able to examine how variations in religious denominations, religiosity,
and the religious diversity and segregation of such groups within communities are
linked to local level attitudes toward intra- and inter-group cooperation. It also pro-
vides measures of cooperative attitudes that are specic and multi-dimensional. For
trust, for instance, the dataset allows for distinctions between particularized and gen-
eralized trust.
7
It also allows for distinctions between trust of neighbors, strangers,
as well as trust of coreligionists and coethnics. Chapter 3 focuses on the individual
aspects of religious identities, while Chapter 4 focuses on the community aspects and
the potential role of social interactions.
1.2.1 Individual religious identities and social cooperation
Chapter 3 explores how variations in an individual's religious denominations and re-
ligiosity are associated with cooperative attitudes. As a measure of religiosity, I use
a simple self-reported measure of religiosity that is highly correlated with individu-
als' observance of religious rituals. My ndings provide evidence of the link between
religious identities and particularized cooperative attitudes in Indonesia. Religiosity
is positively associated with trust { but these associations are only robust for partic-
ularized, but not generalized trust. Moreover, religiosity is positively associated with
discriminative trust of coethnics and coreligionists and is negatively associated with
tolerance. These results are consistent with existing empirical evidence from separate
studies in social psychology, sociology, and economics on the link between religiosity
and altruism (Batson et al., 1993; Anderson et al., 2010), intra- and inter-group trust
(Guiso et al., 2003; Anderson et al., 2010; Tan and Vogel, 2008; Sosis and Rue,
7
At the extremes, \particularized trust" refers to trust of people from their own group while
\generalized trust" refers to trust of others beyond the borders of one's own group and includes that
of strangers (Uslaner and Conley, 2003).
6
2004), and tolerance (Allport and Kramer, 1946; Hall et al., 2010; Guiso et al., 2003).
I also nd that for all measures of in-group preferences and religious tolerance, the
magnitudes of the religiosity coecients are larger than those of gender, an additional
level of education, or a standard deviation change in log per-capita expenditure. For
these measures, they are also strongest among Muslims, the dominant majority in
Indonesia. These results provide evidence of the link between religion and \parochial
altruism", which is altruism toward members of one's own group with hostility to-
ward members of the out-groups (Bernhard et al., 2006; Choi and Bowles, 2007). In
Indonesia, this link is strongest for Muslims.
1.2.2 Social interactions, religion, and cooperation
Armed with these ndings, Chapter 4 extends the analysis to look at how community-
level social interactions among people of dierent religious identities may in
uence
attitudes toward social cooperation. I examine two important and related hypotheses
on how social interactions might amplify the in
uence of religiosity on social coop-
eration. First, numerous studies show that community heterogeneity lowers trust,
social capital, and the ability to implement collective action (Costa and Kahn, 2003;
Alesina and La Ferrara, 2002; Vigdor, 2004; Miguel and Gugerty, 2005). One explana-
tion for this result is network eects: Information travels faster within homogeneous
networks; and similarly, more homogeneous communities are better able to enforce
intra-group norms (Fafchamps, 2004; Habyarimana et al., 2007). This result, how-
ever, somewhat contrasts with the second hypothesis on inter-group interactions.
Allport (1979 [1954]) argues that under optimal conditions contacts with dissimilar
people will break down stereotypes and reduce prejudice { which may foster better
inter-group cooperation.
7
Chapter 4 examines how religious diversity and segregation aect intra- and inter-
group cooperative attitudes. While many studies have examined how diversity is as-
sociated with cooperation, very few have looked at the role of segregation.
8
However,
on its own, diversity may not adequately capture the notion of within-community
networks. Indeed, my ndings suggest that both diversity and segregation play im-
portant, albeit opposite, roles in aecting cooperative attitudes.
Similar to the ndings of Alesina and La Ferrara (2002) for the case of ethnic
diversity in the United States, I nd negative associations between the generalized
trust of strangers and religious diversity. I also nd that, people tend to be more
trusting of each other (and of strangers) in more religiously segregated communities.
However, religious diversity is positively correlated with tolerance, while segregation
is negatively correlated with tolerance. These ndings are suggestive of the role of
network eects in sustaining discriminative attitudes. At the same time, they are also
evidence supportive of the optimal inter-group contact hypothesis of Allport (1979
[1954]) which posits that, under the right circumstances, frequent interactions with
those who are dissimilar may reduce prejudice.
8
The exceptions are Alesina and Zhuravskaya (2011); Uslaner (2010); Rothwell (2010).
8
Chapter bibliography
Akerlof, George A. and Rachel E. Kranton, \Economics and Identity," Quar-
terly Journal of Economics, 2000, 115 (3), 715{753.
Alesina, Alberto and Ekaterina Zhuravskaya, \Segregation and the Quality of
Government in a Cross Section of Countries," American Economic Review, August
2011, 101 (5), 1872{1911.
and Eliana La Ferrara, \Who trusts others?," Journal of Public Economics,
August 2002, 85 (2), 207{234.
Allport, Gordon W., The nature of prejudice, Reading Mass.: Addison-Wesley
Pub. Co., 1979 [1954].
and Bernard M. Kramer, \Some roots of prejudice," The Journal of general
psychology, 1946, 22, 9.
Anderson, Lisa, Jennifer Mellor, and Jerey Milyo, \Did the Devil Make
Them Do It? The Eects of Religion in Public Goods and Trust Games," Kyklos,
2010, 63 (2), 163{175.
Bala, Venkatesh and Sanjeev Goyal, \A Noncooperative Model of Network For-
mation," Econometrica, September 2000, 68 (5), 1181{1229.
Batson, Charles, Patricia Schoenrade, and W. Larry Ventis, Religion and
the Individual : A Social-Psychological Perspective, New York: Oxford Univ. Press,
1993.
Berman, Eli, \Sect, Subsidy, and Sacrice: An Economist's View of Ultra-Orthodox
Jews," Quarterly Journal of Economics, August 2000, 115, 905{953.
9
Bernhard, Helen, Urs Fischbacher, and Ernst Fehr, \Parochial altruism in
humans," Nature, August 2006, 442 (7105), 912{915.
Bowles, Samuel and Herbert Gintis, A cooperative species: human reciprocity
and its evolution, Princeton: Princeton University Press, 2011.
Chaney, Eric, \Tolerance, Religious Competition and the Rise and Fall of Muslim
Science," 2008.
Choi, Jung-Kyoo and Samuel Bowles, \The Coevolution of Parochial Altruism
and War," Science, October 2007, 318 (5850), 636 {640.
Costa, Dora L. and Matthew E. Kahn, \Civic Engagement and Community
Heterogeneity: An Economist's Perspective," Perspectives on Politics, 2003, 1 (01),
103{111.
Deaton, A. S, \Aging, religion, and health," NBER Working Paper 15271, National
Bureau of Economic Research 2009.
Fafchamps, M. and S. Lund, \Risk-sharing networks in rural Philippines," Journal
of Development Economics, 2003, 71 (2), 261{288.
Fafchamps, Marcel, Market institutions in Sub-Saharan Africa : theory and evi-
dence, Cambridge Mass.: MIT Press, 2004.
Granovetter, Mark, \The Impact of Social Structure on Economic Outcomes,"
Journal of Economic Perspectives, January 2005, 19 (1), 33{50.
Guiso, Luigi, Paola Sapienza, and Luigi Zingales, \People's opium? Religion
and economic attitudes," Journal of Monetary Economics, January 2003, 50 (1),
225{282.
10
Habyarimana, James, Macartan Humphreys, Daniel N. Posner, and
Jeremy M. Weinstein, \Why Does Ethnic Diversity Undermine Public Goods
Provision?," American Political Science Review, November 2007, 101 (04).
Hall, Deborah L., David C. Matz, and Wendy Wood, \Why Don't We Practice
What We Preach? A Meta-Analytic Review of Religious Racism," Personality and
Social Psychology Review, February 2010, 14 (1), 126 {139.
Iannaccone, Laurence R., \Sacrice and Stigma: Reducing Free-riding in Cults,
Communes, and Other Collectives," The Journal of Political Economy, April 1992,
100 (2), 271{291.
Jackson, M and A Wolinsky, \A Strategic Model of Social and Economic Net-
works," Journal of Economic Theory, October 1996, 71 (1), 44{74.
Jackson, M. O, \A Survey of Network Formation Models: Stability and Eciency,"
in Gabrielle Demange and Myrna Wooders, eds., Group formation in economics :
networks, clubs and coalitions, Cambridge UK, New York: Cambridge University
Press, 2005, pp. 11{57.
Jackson, Matthew O. and Alison Watts, \The Evolution of Social and Economic
Networks," Journal of Economic Theory, October 2002, 106 (2), 265{295.
Kirchsteiger, Georg, Marco Mantovani, Ana Mauleon, and Vincent Van-
netelbosch, \Myopic or farsighted? An experiment on network formation," Dis-
cussion Paper 8263, Centre for Economic Policy Research, London February 2011.
Knack, Stephen and Philip Keefer, \Does Social Capital Have An Economic
Payo? A Cross-Country Investigation," Quarterly Journal of Economics, 1997,
112 (4), 1251{1288.
11
Kranton, Rachel E and Deborah F Minehart, \A Theory of Buyer-Seller Net-
works," American Economic Review, June 2001, 91 (3), 485{508.
La Porta, Rafael, Florencio Lopez de Silanes, Andrei Shleifer, and
Robert W. Vishny, \Trust in Large Organizations," The American Economic
Review, May 1997, 87 (2), 333{338.
Landes, David, The Wealth and Poverty of Nations : Why Some Are So Rich and
Some So Poor, New York Ny: W.W. Norton, 1998.
Lehrer, Evelyn L, Religion, economics, and demography : the eects of religion on
education, work, and the family, London: Routledge, 2009.
Miguel, E and M Gugerty, \Ethnic diversity, social sanctions, and public goods
in Kenya," Journal of Public Economics, December 2005, 89 (11-12), 2325{2368.
Munshi, Kaivan, \Information Networks in Dynamic Agrarian Economies," in
T. Paul Schultz and John A. Strauss, eds., Handbook of development economics,
Vol. 4 of Handbook of economics, Elsevier, 2007, pp. 3085{3113.
Neusner, Jacob and Bruce Chilton, Altruism in world religions, Washington,
D.C.: Georgetown University Press, 2005.
Pantz, Katinka, The strategic formation of social networks : experimental evidence,
Aachen: Shaker, 2006.
Putnam, Robert, Robert Leonardi, and Raaella Y. Nanetti, Making Democ-
racy Work : Civic Traditions in Modern Italy, Princeton N.J.: Princeton University
Press, 1993.
Rothwell, Jonathan T., \Trust in Diverse, Integrated, Cities: A Revisionist Per-
spective," SSRN Electronic Journal, 2010.
12
Sheng, Shuyang, \Identication and Estimation of Network Formation Games,"
Job market paper, University of Southern California December 2012.
Sosis, Richard and Bradley J. Rue, \Ideology, religion, and the evolution of
cooperation: Field experiments in Israeli Kibbutzim," in \Research in Economic
Anthropology," Vol. 23, Bingley: Emerald (MCB UP ), 2004, pp. 89{117.
Tan, Jonathan H.W. and Claudia Vogel, \Religion and trust: An experimental
study," Journal of Economic Psychology, December 2008, 29 (6), 832{848.
Udry, Christopher, \Risk and Insurance in a Rural Credit Market: An Empirical
Investigation in Northern Nigeria," The Review of Economic Studies, July 1994, 61
(3), 495{526.
Uslaner, Eric M., \Segregation, mistrust and minorities," Ethnicities, December
2010, 10 (4), 415{434.
and Richard S. Conley, \Civic engagement and particularized trust," American
Politics Research, 2003, 31 (4), 331{360.
Vigdor, Jacob L., \Community Composition and Collective Action: Analyzing
Initial Mail Response to the 2000 Census," Review of Economics and Statistics,
2004, 86 (1), 303{312.
13
Chapter 2
The Dynamic Evolution of Social Networks:
Experimental Evidence
1
Abstract
We use a laboratory experiment to explore dynamic network formation in a
six-player game where link creation requires mutual consent. The analysis
of network outcomes suggests that the process converges to the pairwise-
stable (PWS) network when a PWS network exists and not to converge to
any network when a PWS network does not exist. When multiple Pareto-
rankable PWS networks exist, subjects coordinate on the high-payo one.
The analysis at the single choice level indicates that the percentage of my-
opically rational behavior is high. Deviations from myopic rationality are
more prevalent in early turns, when marginal payo losses are small and
when deviations lead to redundant links that may be removed unilaterally
later on.
Within the economic literature, one question of interest is how networks are formed.
Two seminal theoretical papers advanced explorations of this question for dierent
link formation rules. Bala and Goyal (2000) study ecient and Nash-stable networks
in non-cooperative games where links are made and broken unilaterally. Jackson and
Wolinsky (1996) address a similar problem with an underlying game that requires links
between two agents to be mutually agreed, although they can be broken unilaterally.
In addition, there is a small but growing experimental literature that builds on
these theories.
2
However, the bulk of this literature focuses on the former type of rule,
1
This chapter is co-authored with Juan Carrillo.
2
There is a somewhat larger experimental literature on equilibrium selection in network games
14
namely when links can be made and broken unilaterally (see e.g. the seminal paper
by Callander and Plott (2005) as well as Berninghaus et al. (2006), Berninghaus et
al. (2007), Falk and Kosfeld (2003) and Goeree et al. (2009)).
This chapter reports a network formation experiment that tests a theory based
on the latter type rule, in which link creation requires mutual consent. Following
Jackson and Watts (2002), we look at a sequential game where the network evolves
one link at a time.
3
Two experimental studies { Pantz (2006) and Kirchsteiger et
al. (2011) { examine a similar question. Our experiment diers methodologically
from these two studies in two important respects. First, we consider networks of six
rather than four subjects. This dramatically enlarges the individual choice set and
the number of possible outcomes.
4
Second, in those experiments a pairwise-stable
(PWS) network always existed and was the empty network, the full network, or both.
Instead, we consider four treatments with unique, multiple, or no PWS network. In
the two treatments with a unique PWS network, it is neither the empty nor the full
network.
5
The results of the experiment can be summarized as follows. The analysis of
outcomes suggests that the network formation process tends to converge to the PWS
network when it exists and not to converge to any network when a PWS network
does not exist. However, convergence is by no means guaranteed, especially when
the PWS network involves asymmetric payos among subjects. When multiple PWS
(see e.g. Fatas et al., 2010; Charness et al., 2012).
3
Alternatively, Goyal and Joshi (2006) examine a linking game with mutual consent where at
every turn, each agent simultaneously proposes to all other agents. See, e.g., Di Cagno and Sciubba
(2008) and Burger and Buskens (2009) for experimental studies of this type of game.
4
For instance and as discussed below, with 4 players there are 64 possible networks and 6
minimally-connected architectures whereas with 6 players there are 32,768 possible networks and 20
minimally-connected architectures.
5
Other dierences include a benet function from network membership that depends exclusively
on the size of the network, a long sequence of choices with an average of 102 individual decisions
per match, and a reshuing of partners after each match.
15
networks exist, subjects often manage to reach the high-payo one.
To understand better the choices of our subjects, we then study each single de-
cision. Overall, the majority of these decisions are myopically rational. However,
descriptive and regression analyses indicate four instances where deviations from my-
opic rationality are more likely. First, at the beginning of the match. This makes
intuitive sense, since subjects play at least 12 rounds and they know that decisions
made early in the game are likely to be reversible later on. Second, deviations are
also more likely when they are necessary to escape the low-payo PWS network. This
result reinforces one of the results from our analysis of the network outcome, to wit
that subjects look for the high-payo network. Third, deviations are also more likely
when the marginal payo losses are small. This would be consistent with a theory of
`imperfect choice' where deviations from myopic rationality are inversely proportional
to the cost of doing so. Fourth, deviations are also more likely to occur when the
myopic rational decision resulted in the subject having a smaller number of links {
perhaps, because removing a link does not require the consent of other subjects and
therefore can be taken unilaterally later on if the opportunity presents.
A cluster analysis performed at the subject level reveals heterogeneous behavior
across individuals. More than half of our subjects exhibit a sustained increase in my-
opic rationality as the end of the match approaches, and low myopic rationality when
required to reach the high-payo PWS network. About a quarter of the subjects have
a rather erratic behavior, in which they deviate substantially from myopic rationality
without dierentiating between treatments or turns. The rest exhibit an interesting
pattern, with initially low myopic rationality (consistent with costless experimenta-
tion) and a dramatic increase in myopic rationality (around 20%) when choices are
most likely to be irreversible.
Finally, we brie
y compare our results with the literature. First, despite our game
16
being more complex than Pantz (2006) and Kirchsteiger et al. (2011), we still nd
support for forward-looking behavior in the treatment with multiple PWS networks,
just like they do. Second, the treatment with no PWS network allows us to establish
that convergence is greatly facilitated by the existence of a PWS network, as predicted
by Jackson and Watts (2002). Third, we show that convergence crucially depends on
the architecture of the (unique) PWS network. In particular, it is more dicult to
reach PWS in the treatment where dierent members of the network have dierent
payos (e.g., the extremes vs. the center in components with 3 subjects) than in the
treatment where all subjects have identical payos (e.g., components with 2 subjects).
Fourth, our setting permits a detailed study of the determinants of myopic rational
choices. We show that deviations from myopic rationality are rare but, at the same
time, more prevalent when one would expect them to occur: in early turns, when
marginal losses are small, and for choices that are easier to reverse later on.
The chapter is organized as follows. In Section 2.1, we present the conceptual
framework and network models that are pertinent to our experiment. Section 2.2
elaborates the experimental design and introduces our treatments. Then, in the fol-
lowing three sections, we present our analysis at the nal network level (Section 2.3),
single decision level (Section 2.4) and subject level (Section 2.5). Section 2.6 con-
cludes.
2.1 Network environment and basic denitions
A network is a collection of links connecting \nodes" which, in our case, represent
agents. A link between two agents can form only if both decide that it is worth form-
ing. Each link is costly for both agents and this cost is non-transferable. Meanwhile,
the benet depends on and is a strictly increasing function of the size of the network
17
component that an agent belongs to. We distinguish between \network" and \com-
ponent". A network describes the link congurations that include the full graph (i.e.,
all six agents) while a component is a sub-graph in which there exists a path linking
any two agents. We assume that all agents in a component receive the same benet.
Payos are computed as the dierence between benets and costs.
Given that link formation needs mutual consent while link destruction does not,
what network is likely to emerge? A candidate is the pairwise-stable network. A
network is pairwise-stable (PWS) if: (i) all existing links are weakly preferred by
both agents in the link and are strictly preferred by at least one of them; and (ii)
all non-existing links are such that at least one of the agents on the non-existing
link strictly prefers its absence (Jackson and Wolinsky, 1996). One other plausible
outcome arises when agents maximize the total value of the network, that is, the sum
of payos received by all agents. This is referred to as the ecient network (Jackson
and Wolinsky, 1996).
A network consists of minimally-connected components, also called a minimally-
connected network, if the removal of any existing link increases the number of compo-
nents (Bala and Goyal, 2000). In our setting, stable networks and ecient networks
always consist of minimally-connected components. To see why, notice that when
benets are strictly a function of the component size as in our game, in any non-
minimally-connected network removing a link that does not split a component will
increase the net payo of two individuals without aecting the net payos of any
other subject.
Pairwise stability, however, does not provide a prediction of which network out-
come will be reached. Jackson and Watts (2002) model a dynamic process of link
formation and destruction by assuming that pairs of agents meet at random and de-
cide whether to propose a link to each other when none exists, or sever an existing
18
one. We call a myopic rational choice one in which agents make their decisions based
solely on the marginal payo resulting from the potential link they are considering
(and not on the option value of forming or severing links in the future). Under my-
opic rationality, a link is created if both agents are weakly better-o with the link
and at least one is strictly better-o. A link is severed if at least one agent is strictly
better-o without the link. If all agents follow a myopic rational behavior, then the
network evolves following an improving path.
Starting from any network, Jackson and Watts (2002, Lemma 1) show that im-
proving paths lead to either a PWS network or a cycle. A set of networks forms a
cycle if there is an improving path from any network to any other network in the set.
A set of networks is in a closed cycle if no network in the set is on the improving
path of a network that is not in the set. Since improving paths assume myopic ratio-
nal choices, they predict for example that players will be stuck in an empty network
when developing a network is initially costly but eventually benecial. Also, there
might be multiple PWS networks, some more attractive than others. In that case,
depending on the starting network conguration, the improving path will lead to one
PWS network or another, independently of their payo properties.
To deal with predictions under multiplicity, we consider the possibility of far-
sighted behavior. We call PS-PWS a Pareto-Superior pairwise stable network in
which the payos of all agents are weakly higher (strictly for at least one of them)
than in any other PWS network. If agents follow myopic rational choices, they will
be stuck in any PWS network they reach, but if agents realize the long run benets
of myopically suboptimal choices, they may move away from one PWS network into
a PS-PWS network.
6
6
This denition is dierent from the notion of pairwise farsightedly stable network developed by
Herings et al. (2009) which explicitly considers the notion of farsighted improving paths.
19
Finally but crucial for our experiment, notice that dierent agents in a given
component may obtain dierent payos (they all have the same benets but some
may have more direct links than others). This poses another multiplicity problem:
even within a PWS network, agents have incentives to make or break links in or-
der to change their nal position in the network (typically trying to keep the same
component size but have other agents bear most of the link costs).
2.2 Experimental setting and procedures
2.2.1 The basic conguration
Our experiment examines how well ecient networks and PWS networks predict
the outcome of a dynamic linking game. We are interested in environments with a
large number of network congurations where links are costly and mutual pairwise
consent is needed to form a link but not to break it. To this end, we implement
a stochastic dynamic linking game that slightly modies the procedure proposed by
Jackson and Watts (2002). We consider networks with n = 6 players. This means
n(n 1)=2 = 15 possible bilateral undirected links between dierent players, and
therefore 2
n(n1)=2
= 32; 768 possible networks. Notice that many networks dier
from each other only on the identity of subjects in the dierent positions (e.g., which
players are at the center of the component and which ones are at the extremes). We
say that two networks have the same network architecture if they are identical up to
a permutation of the identity of subjects. With 6 players there are 20 minimally-
connected network architectures.
7
We specically look at a setting where all agents in a component receive the same
7
Although six players may not seem like a large network, it is the biggest size we could manage.
For comparison, Pantz (2006) and Kirchsteiger et al. (2011) have four players, which means 6 bilateral
undirected links, 64 possible networks and just 6 minimally-connected network architectures.
20
benets. One example of such a setting is the formation of risk-sharing insurance
networks, where individuals commit to share monetary holdings equally with linked
partners (Bramoull e and Kranton, 2007). With repeated interactions between linked
partners, they formally show that such an arrangement is tantamount to equal sharing
within network components. Another example would be the case of club goods (such
as those provided by religious or social groups, see e.g., Berman (2000)) without
centralized coordination. In this case, all members benet from the participation from
an additional member, but to ensure participation may require individuals to maintain
costly links with other members. We therefore deviate from much of the network
experimental literature that implements the connections model where the benets of
indirect links decay with distance (Jackson and Wolinsky, 1996). One advantage of
our approach is that payos from decisions are straightforward to calculate. This is
made even simpler by the fact that we maintain the unit cost of a link to be constant
both within and across treatments. By doing so, we dramatically reduce the likelihood
that some decisions are made out of payo miscalculations.
Each match consists of multiple turns and starts with an empty network. At
each turn, the computer randomly assigns the six players into three pairs. Once
paired, players choose their actions with respect to their partner in the pair. A new
turn begins after all players have taken their actions. If all players are satised with
the network outcome, they can collectively end the game. If we were to mimic the
dynamic linking procedure of Jackson and Watts (2002), we should require matches
to end only when all players have agreed on the outcome. However, this would
make sessions unmanageably long. Furthermore, some agents could sabotage the
experiment and choose actions randomly knowing that they always have the option
to change it in the future. It would also favor those who stubbornly refuse to give
up until a specic position in a specic component is reached. We therefore decided
21
to implement a match-ending rule that provides enough opportunities for players to
converge but, at the same time, allows decisions to be meaningful and the experiment
to be time manageable: Subjects play for 12 turns unless all players are satised with
the network; afterwards, each turn is the last one with probability p = 0:2, providing
an additional 1=p = 5 turns on average. With this probabilistic match-end rule, we
hope to mitigate the last-round eects. At the same time, it allows for an interesting
comparison of behavior before and after the 12
th
turn. Finally, notice that each turn
is composed of six decisions, one for each player, providing a fairly large number of
individual decisions per match (17 6 = 102 on average, unless subjects decide to
stop before).
Figure 2.1 shows the user interface. At each turn, players make decisions by
clicking on one of the action buttons. If a player is not linked to his partner, he
chooses whether to \Propose" a link or \Pass Turn". If he is linked, he chooses
whether to \Remove" a link or \Pass Turn". Once a pair of partners have taken their
actions, the result is immediately displayed on the screen. Hence, when each player
makes his decision, he observes the latest state of the network. Showing the latest
network conguration within a turn allows us to cleanly determine whether each
individual decision re
ects a myopic rational behavior or otherwise. On the other
hand, it may encourage a war of attrition, where players wait to see what others do
in a turn before choosing their own action.
8
If a player is not only satised with the relationship with his partner but also with
the overall network, instead of choosing \Pass Turn", he can choose \Network OK".
9
As mentioned above, the match immediately ends if all players within a turn choose
8
There is no evidence in our data of such war of attrition type of behavior.
9
Once a player chooses \Network OK", he does not need to choose further actions until the
network changes. To avoid mistakes, all of his action buttons become inactive. These buttons are
immediately reactivated following a change in the network.
22
Figure 2.1: User interface for the linking game.
\Network OK".
The user interface displays all the pertinent information: the subject's role, the
role of the person he is currently matched with, whether the current turn is a potential
terminal turn and, naturally, the current network conguration. It also displays the
benet of the subject as a function of the size of the component he is in, the cost as
a function of his number of direct links, and his net payo given the current network
conguration. This succinct but comprehensive visual display allows the subject to
compute rather easily not only the net value of adding or removing an existing link
(i.e., the improving path) but also his payo in any other network conguration. Fi-
nally, notice that the node representing the subject is always located at the center
and labeled \You". The nodes representing the other players in a match are labeled
by their roles and surround the subject's node in clockwise order at an equal dis-
tance from it. By always putting the subject's node at the center, even though the
underlying connections between subjects in a match are identical, each subject sees
a dierent graphical representation. We therefore avoid leading participants towards
23
focal networks such as the star or wheel network.
2.2.2 Treatments
Since we are primarily interested in comparing observed and predicted outcomes
under dierent scenarios, the benets are not based on any particular functional
form. With this freedom (we only impose a benet strictly increasing in component
size and a constant unit link cost), we can set the payos such that we have treatments
with unique PWS network architectures that are dierent from the empty or full
network, with no PWS network, and with multiple PWS networks. We also facilitate
computations by maintaining the same unit link cost across treatments and only vary
the benets.
Our experiment includes four treatments. They are graphically depicted as Treat-
ments 1, 2, 3 and 4 and are inserted at the end of the chapter for reference. These
gures illustrate how we determine the dierent PWS networks. First, we draw a
\supernetwork", comprising the 20 possible minimally-connected six-node network
architectures (labeledfAg tofTg). Then we add all the arcs connecting pairs of
networks that dier from one another by a single link. We put all network archi-
tectures with identical number of links in the same row and order them from top
row (network with no links) to bottom row (networks with 5 links). Each network is
therefore connected by an arc to one or more networks in the row above and the row
below it.
10
The direction of the arc represents the improving path: forming a new
link (arrow pointing to the row below) or removing an existing link (arrow pointing
to the row above).
The four treatments we consider dier on existence and number of PWS net-
10
Networks that are not minimally connected are necessarily o the improving paths. They are
omitted unless a match ends in one of them.
24
works. Treatment 1 has no PWS network. There is a unique closed cycle comprised
of networksfB;C;D;F;G;H;Ng within which the dynamic process is expected to
stay (shaded area). Treatment 2 has four PWS networks,fA;I;J;Kg, and network
fKg (in light shade) is Pareto-Superior to the other three. If players exhibit my-
opic rational behavior and follow the improving path, they will stay on the empty
networkfAg. If for some reason (a perturbation, a mistake, an experimentation,
etc.) they reach any of the other PWS networks,fI;J;Kg, they will also stay there.
However, with some form of forward-looking behavior, players might be able to reach
the Pareto-Superior PWS networkfKg. Treatments 3 and 4 have unique PWS net-
works (shaded),fHg andfLg, with three components of two subjects each and two
components of three subjects each respectively.
Meanwhile, in all treatments, the ecient networks comprise all of the minimally-
connected full networks, namely,fO;P;Q;R;S;Tg. We choose to have the same
ecient networks in all treatments to facilitate comparisons. Also, we choose several
ecient networks and one of them focal (fTg, the line that comprises all agents) in
order to give a fair chance to the ecient outcome. The information is summarized
in Table 2.1.
2.2.3 Implementation details
The experiment was conducted in the California Social Science Experimental Labo-
ratory at UCLA in August 2010. All experimental subjects were UCLA students. We
conducted 8 experimental sessions with 12 subjects in each session. Subjects played
two sets of four treatments for a total of 8 matches in each session. We shued the
order of the treatments such that: (i) the orders of the treatments in the rst half
and the second half of each session are dierent; (ii) no two sessions have identical
25
Table 2.1: Efficient and PWS networks
Treatment
Benet in a
component of size
Cost
per link
Ecient
network(s)
PWS
network(s)
1 2 3 4 5 6
1 0 20 30 39 42 43 15 [6]
fO;P;Q;R;S;Tg
None
2 0 10 17 22 38 44 15 [6]
fO;P;Q;R;S;Tg
[1-1-1-1-1-1]
fAg
[5-1]
fI;J;K

g
3 0 29 36 41 43 44 15 [6]
fO;P;Q;R;S;Tg
[2-2-2]
fHg
4 0 19 36 42 44 45 15 [6]
fO;P;Q;R;S;Tg
[3-3]
fLg
The number in brackets refers to the size of each component.

is the Pareto-Superior PWS network (PS-PWS).
treatment sequences; and (iii) each treatment is implemented in exactly two sessions
for each order in the sequence. This shuing was done to neutralize the possible
eects from the ordering of the treatments within a session. In total, we obtained 128
match observations (32 matches for each treatment) from 96 subjects (12 subjects in
8 sessions).
With 12 subjects, there are always 2 groups in each session. After each match
we reshued subjects into new groups and assigned a new role (1 to 6) to each
subject to introduce anonymity in game play. Each session lasted for between 90
and 120 minutes. No subject took part in more than one session. Participants
interacted exclusively through computer terminals without knowing the identities of
the subjects they played with. Before the paid matches, instructions were read aloud
and two practice matches were played to familiarize participants with the computer
interface and procedure. After that, participants had to complete a quiz to ensure
they understood the rules of the experiment.
At the end of each match, subjects obtained a payo based on the size of the
component they were in (benet) and the number of direct links (cost). Participants
were endowed with experimental tokens and they could earn or lose tokens. At the end
26
of the session, the payos in tokens accumulated from all experimental games were
converted into cash, at the exchange rate of 4 tokens = $1. Participants received
a show-up fee of $5, plus the amount they accumulated during the paid matches.
Payments were made in cash and in private. Matches lasted for between 13 and 36
turns, with an average of 16.8 turns. There was a signicant spread in winnings:
including the show-up fee, participants earned between $11 and $43 with an average
of $29. A copy of the instructions is included in Appendix 2.B.
2.3 Network outcomes
We rst analyze the general properties of the network formation game, most notably
convergence and stability of the nal conguration.
2.3.1 Network convergence: eciency vs. stability
Result 1 Subjects rarely end up in components that are not minimally connected.
We begin by noting that players show understanding of the basic tenets of the exper-
imental game. Table 2.2 summarizes the network outcomes of the four treatments.
We rst focus on whether players end up in networks with components that are not
minimally connected. Remember that removing a link that does not reduce the com-
ponent size strictly increases the payos of both agents involved without aecting
any other. The second column in Table 2.2 suggests that players understand this
principle. Only 5 matches out of 128 (or 4%) end up in a network that is not mini-
mally connected. The result is all the more remarkable that non-minimally connected
networks are sometimes necessary intermediate steps towards farsighted goals.
27
Table 2.2: Summary of network outcomes
Treatment
Not Min.
Conn.
Ecient PWS PS-PWS Closed cycle
1 2 3
- -
21
(6.3%) (9.4%) (65.6%)
2 1 5 3

14
y
-
(3.1%) (15.6%) (9.4%) (43.8%)
3 0 0 15
- -
(0.0%) (0.0%) (46.9%)
4 2 2 10
- -
(6.3%) (6.3%) (31.3%)
N = 32 for each treatment;

networksfA;I;Jg; y networkfKg.
Result 2 The process tends to converge to the PWS network when a PWS exists and
not to converge to any network when no PWS network exists. It rarely leads to an
ecient network.
Before discussing the result, we need an operational denition of convergence. Callan-
der and Plott (2005) argue that a network has converged if it maintains the same
conguration in the last T turns before the end of the game. Alternatively, we can
use the \Network OK" action. In this case, a network is said to have converged if
all six players chose \Network OK" or if ve players did so and the sixth was either
unable or unwilling to unilaterally change the network based on her previous actions.
The rst denition may \over-detect" convergence since it includes cases where the
lack of change is merely the result of how pairs were randomly assigned in the last T
turns. The second denition may \under-detect" convergence. Indeed, a player may
delay choosing \Network OK" hoping to benet from a possible mistake or deviation
by another player.
28
Table 2.3 presents the rate of convergence under the rst denition withT = 3 and
under the second denition.
11
Under the rst denition, the convergence rate is lowest
in Treatment 1 (around 20%), where there is no PWS network. For the treatments
with unique or multiple PWS networks, convergence is above 50%. Convergence
under the second denition is always low (16% or less) indicating that this notion is
probably too strict.
Table 2.3: Network convergence
Treatment
No change in
last 3 turns
y
Network OK
1 7 1
(21.9%) (3.1%)
2 20 3
(62.5%) (9.4%)
3 17 5
(53.1%) (15.6%)
4 18 3
(56.3%) (9.4%)
N = 32 for each treatment;
y
Includes \Network OK".
We now go back to Table 2.2 and look at how well pairwise stability predicts the
nal network. The graphical illustration of Treatments 1 to 4 elaborates on this infor-
mation: it displays the number of nal outcomes in each network architecturefAg to
fTg both conditional on no change in the last 3 turns (labeled C) and unconditional
on convergence (labeled U). In the absence of a PWS network (Treatment 1), 66%
of the matches end in a network within the closed cycle.
12
More interestingly and
11
T = 3 is arbitrary. It corresponds to 18 individual decisions which seems reasonably large. With
larger T , convergence decreases but the qualitative conclusions remain largely the same (T = 5 is
not presented for brevity but it is available from the authors).
12
Matches stay within the closed cycle most of the time: in 56% of all the moves after Turn 6
subjects remain in a network within the closed cycle. The result should not be overemphasized as
29
ignoring convergence for now, 53%, 47% and 31% of the nal networks in Treatments
2, 3 and 4 are PWS. Finally, the dynamic formation process rarely leads to any of
the ecient networks.
13
We next examine the following question: When stable networks exist, how well
do they predict the outcome conditional on convergence? Hereafter, we employ the
operational denition of convergence as the lack of change in the last T = 3 turns.
Table 2.4 suggests a mixed picture. For Treatments 2 and 3, more than half of the
convergent networks are PWS networks. For Treatment 4, only 22% are (Treatment 1
is not included in this analysis as it predicts no convergence). The dierence between
Treatments 3 and 4 is intriguing and deserves further study. A possible reason is the
asymmetry of players' payos within components. Indeed, with [2-2-2] all subjects
earn equal amounts and have little room for improvement. By contrast, with [3-3]
the players in the center of the component may deviate, hoping to reach the same
network but with someone else bearing the two link costs.
Table 2.4: Summary of network outcomes conditional on convergence
Treatment Ecient PWS PS-PWS N
2 4 1

11 20
(20.0%) (5.0%) (55.0%)
3 0 9 - 17
(0.0%) (52.9%)
4 1 4 - 18
(5.6%) (22.2%)

Includes networksfA;I;Jg.
the closed cycle is rather large (7 out of 20 network architectures).
13
The treatment that ends with the ecient network the most is Treatment 2 (16%). However,
in this treatment, all ecient networks that became the nal outcome are just one link away from
the PS-PWS.
30
In Appendix 2.A.1 we provide a complementary analysis based on the distance
between the observed and predicted (ecient, closed cycle, PWS and PS-PWS) out-
comes, both conditional and unconditional on convergence. The conclusions are sim-
ilar. The distance is smaller between the observed outcome and the networks in
the closed cycle (Treatment 1), the PS-PWS network (Treatment 2) and the PWS
network (Treatment 3) than between the observed outcome and any of the ecient
networks in Treatments 1, 2 and 3. Again, the result is less clearcut for Treatment 4.
2.3.2 Network stability: myopic vs. farsighted
Result 3 Agents are able to \coordinate" away from the myopic improving path to
reach the higher return, PS-PWS network.
As elaborated in the previous section, the strict use of myopic improving paths may
lead to networks with low returns for all subjects, especially in the presence of scale
economies. Treatment 2 implements such scale economies: agents with one link obtain
positive payos for components of size 3 and above and agents with two links obtain
positive payos for components of size 5 and above. Thus, starting from the empty
networkfAg, the improving path will stay there. Meanwhile,fKg, the PS-PWS
network has a [5-1] architecture.
Tables 2.2 and 2.4 show that the dynamic link formation process did lead by and
large to the high return network. Indeed, 44% (unconditional on convergence) and
55% (conditional on convergence) of matches in Treatment 2 matches ended in the
PS-PWS network. By contrast, only 9% (unconditional) and 5% (conditional) ended
in one of the other three PWS networks.
The results are in line with Pantz (2006) who found that many of her games did
reach the forward-looking network architecture. Our setup, however, is more complex
31
for players. First, we have 20 minimally-connected architectures. Starting from the
empty PWS networkfAg, players must go through three non-improving paths before
they reach networksfF;Gg, and only from there they may converge to the PS-PWS
networkfKg.
14
And second, the sequence of link decisions in Pantz (2006) is xed
and known ex ante by all players, so subjects can compute the subgame-perfect
equilibrium of the game. Instead, our subjects face random pairing at each round.
2.3.3 Payos
Table 2.5 presents the average sum of payos obtained by subjects in each treatment,
also called \network value". These values are compared with the average value of
the networks in the closed cycle (for Treatment 1), with the values of the PWS and
PS-PWS networks (for Treatments 2, 3 and 4), and with the values of the ecient
networks (for all treatments). We nd substantial losses compared to the payos
that could be collectively obtained: the empirical payos are between 46% and 71%
of the payos generated by any of the ecient networks. In fact, the observed payos
are smaller but close to the payos in the PS-PWS network for Treatment 2 and to
the payos in the unique PWS network for Treatments 3 and 4. For Treatment 1,
the payo is 50% higher than the average payo in the closed cycle. The results are
similar when we consider only the empirical payos of the convergent networks.
2.3.4 Summary
Subjects in our game understand the strategic nature of network formation and avoid
non-minimally connected components. Network eciency does not appear to be
14
As can be seen from the graphical illustration of Treatment 2, all the arcs representing the
improving paths in rows 2, 3 and 4 are directed upwards (towardsfAg and away fromfKg), hence
the need for at least three non-myopic rational choices.
32
Table 2.5: Summary of network values
Treatment
Obtained
All
Obtained
Convergent
PWS PS-PWS Closed cycle Ecient
1 63.13 - - - 43:1
y
108
(26.58)
2 52.91 62.70 0
z
70 - 114
(44.68) (41.53)
3 81.47 79.35 84 - - 114
(12.33) (12.80)
4 85.38 84.78 96 - - 120
(17.88) (16.51)
Standard deviations in parenthesis
y
Unweighted average payos of all networks in the cycle.
z
We only consider the empty networkfAg.
an important motivation in any treatment. The existence of one or several PWS
networks is necessary for convergence: when no PWS network exists (Treatment 1)
the process hardly converges. At the same time, it is not sucient: convergence is
weaker in Treatment 4 than in Treatments 2 or 3. Finally, subjects are reasonably
forward-looking when this is needed to reach the high-paying PS-PWS network.
2.4 Individual decisions
Having analyzed aggregate outcomes, we now study each individual decision. At each
turn, each subject in a pair must choose to either \act" or \pass". If subjects in the
pair are initially unlinked, acting implies proposing a link and passing implies remain-
ing unlinked. If, on the contrary, subjects are initially linked, acting implies removing
a link and passing implies remaining linked. We are interested in the extent to which
decisions are myopic rational in each of these four cases and for each treatment.
33
2.4.1 Descriptive statistics
Table 2.6 summarizes the proportion of myopic rational actions across turns. We
organize the data into four groups of turns. We use the last certain turn that players
get unless everyone agrees on the network outcome (Turn 12) as a natural point to
partition and further split each of these partitions into two.
Table 2.6: Myopic rationality of individual decisions
Turns
y
Mean dierence t-test
z
[1-6] [7-12] [13-18] 19
[7-12] v.
[1-6]
[13-18] v.
[7-12]
A. All 0.741 0.819 0.894 0.910 0.000 0.000
(0.438) (0.385) (0.308) (0.286)
B. By decision problem
i. Stay unlinked (Pass) 0.374 0.392 0.571 0.647 0.292 0.000
(0.484) (0.489) (0.496) (0.485)
ii. Stay linked (Pass) 0.882 0.878 0.877 0.846 0.613 0.505
(0.323) (0.328) (0.328) (0.362)
iii. Remove link (Act) 0.663 0.837 0.916 0.940 0.000 0.000
(0.473) (0.369) (0.277) (0.238)
iv. Propose link (Act) 0.956 0.967 0.979 0.967 0.144 0.109
(0.206) (0.179) (0.144) (0.180)
C. By treatment
Treatment 1 0.829 0.826 0.885 0.944 0.566 0.000
(0.377) (0.379) (0.320) (0.229)
Treatment 2 0.563 0.856 0.905 0.862 0.000 0.001
(0.496) (0.351) (0.294) (0.345)
Treatment 3 0.764 0.788 0.895 0.919 0.082 0.000
(0.425) (0.409) (0.306) (0.274)
Treatment 4 0.811 0.805 0.890 0.969 0.644 0.000
(0.392) (0.397) (0.313) (0.175)
y
Standard deviations in parenthesis.
z
P-values of one-sided t-test of whether mean in latter turns is greater than in earlier turns.
This split captures behaviors at dierent stages. First, players attempt to get fa-
34
miliar with the current match and try dierent strategies which, with high probability,
can be reversed later if desired (Turns [1-6]). Then, players adjust their behavior as
the last certain turn approaches (Turns [7-12]). After that, players enter the random
stopping phase where, presumably, they behave under the assumption that matches
can be terminated at any time (Turns [13-18]). Finally, Turns 19 and above are set in
a another category because the sample size is dramatically reduced as turns advance
and the sample becomes non-representative of the population.
15
Although formal tests are presented in the regression analysis of Section 2.4.2,
Table 2.6 is instructive. Panel A suggests that decisions are more myopic rational as
players get closer to the end of the match, with jumps statistically signicant at the
0.1% level between [1-6] and [7-12] and also between [7-12] and [13-18].
Panel B investigates myopic rationality by type of decision. We examine choices
under four mutually exclusive conditions, namely when the rational action is: (i) to
pass and remain unlinked; (ii) to pass and remain linked; (iii) to remove an existing
link; and (iv) to propose a new link. By comparing conditions (i) with (ii), and (iii)
with (iv), we nd evidence that individuals deviate more from improving paths in
decisions that reduce the number of links than in decisions that increase the number
of links. By comparing conditions (i) with (iii), and (ii) with (iv), individuals appear
to deviate more from improving paths by failing to act when they should than by
acting when they should not.
16
However, our regressions below suggest that this last
result does not hold once we control for individual xed eects and marginals payo
from myopic rational choices.
Panel C displays myopic rationality across treatments. In all four treatments,
15
The split is arbitrary. Similar results are obtained if the rst and third partition are changed
marginally. The key issue is to introduce a separation around turn 12.
16
A set of t-tests (not reported here) conrms that for each turn group, the mean dierences
in myopic rationality both between conditions (i) and (iii) and between conditions (ii) and (iv) are
negative and statistically signicant at the 0.1% level.
35
players are signicantly less myopic rational before Turn 12 than after it at the 0.1%
level. Interestingly, the dierence in myopic rationality between [1-6] and [7-12] de-
scribed in Panel A is entirely driven by Treatment 2. This supports the hypothesis of
forward-looking behavior since in Treatment 2 and only in that treatment, subjects
need to play against myopic rationality at least three times in order to escape the
basin of attraction of the zero payo PWS networkfAg. A graphical illustration of
the results in panels B and C is presented in Appendix 2.A.2.
Finally, Table 2.7 presents the number of instances in which individuals choose to
\enter", \stay" and \leave" a PWS network, broken down by treatment and turn. The
last column reports the total number of turns in that set. We observe that individuals
are prone to leave the PWS network in Turns [1-12], but this tendency is dramatically
reduced when each turn can be nal, especially for Treatment 3. Surprisingly, even
though Treatment 2 has a predictable outcome (44% of matches end up in the PS-
PWS network, see Table 2.2), individuals move away from the PS-PWS networkfKg
after Turn 12 more frequently than they do in the other treatments, possibly due to
the existence of other PWS networks as well. Also, subjects leave the zero payo
PWS networkfAg in turn 1 and never come back to it.
2.4.2 Regression analysis
To study more thoroughly individual decisions, we regress the probability that an
individual chooses the myopic rational action on the attributes of the problem. That
is, for each treatment we estimate a linear probability model (LPM) for the following
specication:
P(Y
ij
nt
= 1j X
ij
nt
;c
n
) =
0
+ X
ij
nt
+c
n
(2.1)
36
Table 2.7: Movements to and from PWS networks
Turns Enter Stay Leave Total turns
Turns [1-12]
Treatment 2 -fAg
y
0 { { 384
Treatment 2 -fKg 28 51 19 384
Treatment 3 30 69 24 384
Treatment 4 7 17 4 384
Turns [ 13]
Treatment 2 -fAg 0 { { 177
Treatment 2 -fKg 14 67 9 177
Treatment 3 9 44 0 143
Treatment 4 8 19 1 128
y
Excludes the initial choice
whereY
ij
nt
indicates whether the action is myopic rational, X
ij
nt
captures the attributes
that move individualn from networki toj in the supernetwork at turnt. Meanwhile,
c
n
captures the unobservable characteristics of the individualn which may aect how
she makes decisions. We do not assume that the unobservable individual character-
istics are independent from the attributes of the decisions, and hence, implement an
individual xed eects specication. We also cluster standard errors by session.
We choose the xed-eects linear probability model (LPM) instead of a logit model
because it is easier to interpret the marginal eects for the former, especially with
regards to the interaction terms (Ai and Norton, 2003).
17
At the end of the section,
we brie
y discuss some extensions and alternative representations.
We can use the regression framework to investigate the four types of decisions
described in Panel B of Table Table 2.6. Consider the following specication:
E(Y
ij
nt
j X) =
0
+
1
morelink
ij
+
2
act
ij
+
3
(morelink
ij
act
ij
) +"
(2.2)
17
We do not consider the xed-eects probit model given its known bias (Greene, 2004).
37
where morelink
ij
and act
ij
are dummy variables and " is the residual. The variable
morelink
ij
takes on a value of 1 if between networks i and j the network with more
links gives the individual a higher payo; the variable act
ij
takes on a value of 1 if
the myopic rational choice is to act.
Under the LPM, the interpretation of these -coecients is straightforward. The
coecient
0
captures the probability that an individual does not propose a link in
accordance to the myopic rational strategy (M. rat.). Similarly,
0
+
1
captures the
probability that an individual does not remove a link in accordance to the myopic
rational strategy. Table 2.8 provides interpretations for the dierent combinations of
coecients.
Table 2.8: The regression coefficients and the types of decision problems
Interpretation
more
act
ij
Function
link
ij
i. P(M. rat. = 1j M. rat. = Stay unlinked) 0 0
0
ii. P(M. rat. = 1j M. rat. = Stay linked) 1 0
0
+
1
iii. P(M. rat. = 1j M. rat. = Remove) 0 1
0
+
2
iv. P(M. rat. = 1j M. rat. = Propose) 1 1
0
+
1
+
2
+
3
This specication allows us to explore how the nature of the problem aects
game play. We also include three sets of additional variables (and the individual
xed eects) to explore individual strategies. The extended models are, therefore,
variations based on the following specication:
E(Y
ij
nt
j X) =
0
+
1
morelink
ij
+
2
act
ij
+
3
(morelink
ij
act
ij
)
+
mpay
ij
+
q
chdist(q)
ij
+
P
4
t=1

t
turn sp(t) +c
n
+"
(2.3)
38
First, we investigate whether the size of marginal payos aects the deviations
from the improving path. The variable mpay
ij
contains the marginal payo from
making a myopic rational choice between networks i and j. With a myopic strategy,
the sign of the decision (a loss vs. a gain) should matter but not the magnitude of
the loss or the gain. However, if we assume imperfect choices (like in the Quan-
tal Response Equilibrium model of McKelvey and Palfrey (1995) for example) it is
reasonable to think that deviations are less likely to occur when marginal losses are
large.
Second, we explore the possibility that individuals follow the shortest distance
towards the ecient line network (E.Line) or the Pareto-Superior PWS network
(PS). The variablechdist(q)
ij
denotes the change in the geodesic distance to network
q2fE.Line, PSg if an individual takes the myopic rational decision in choosing
between networks i and j. Each of these variables can take a value of 1, 0, or -1 and
they are included one at a time in the regression. A negative coecient on chdist(q)
indicates that all else the same, players are more likely to choose a myopic rational
action if it moves them closer to network q.
Finally, we control for possible turn eects using a linear spline on the turn vari-
ables, turn sp, with knots at turns 6, 12, and 18.
18
The knot choices mimic the turn
grouping we did for the descriptive analysis.
Result 4 Individual decisions are more myopic rational after Turn 13 than before
Turn 12.
We rst perform a test for pooling for all of our specications to investigate whether
there is a structural change after Turn 12. In all but one specication, we can reject
18
Hence, the variable turn sp(1) is the spline for Turns [1-6], turn sp(2) is for Turns [7-12],
turn sp(3) is for Turns [13-18] and turn sp(4) is for Turns 19 and greater.
39
the null hypothesis that the coecients before and after Turn 12 are equal at 1%
signicance; in all cases, we can reject the hypothesis at 5% signicance (see Table 2.15
in Appendix 2.A.3 for details).
We therefore analyze the two sets of turns separately. We begin by examining the
extent to which improving paths drive individual behaviors. If improving paths were
the sole driver of network evolution, the constant terms in all specications would be
one and the coecients on all other variables would be zero. Table 2.9 presents the
regression results of our basic model with individual xed eects. The constant terms
are high but signicantly lower than one, and the coecients of the other variables
are signicantly dierent from zero, suggesting deviations from the improving paths.
To better study deviations, we included estimates of the linear combinations of
the coecients for the constant term, morelink, act, and morelinkact. These
linear combinations are derived from Table 2.8 to allow immediate comparisons of the
probabilities that individuals make myopic rational choices for the dierent decision
problems.
Pairwise comparisons of estimates conrm that, all else the same, individuals are
more myopic rational in Turns [ 13] than in Turns [1-12]. Of the 16 combinations of
treatments and decision problems, the point estimates are always larger in later turns
with only one exception: Treatment 2 when the myopic rational choice is to propose.
Result 5 In early turns, individuals deviate from improving paths by maintaining
excessive links (over-proposing and not removing redundant links). In later turns,
individuals deviate mainly by not removing redundant links.
Panel B of Table 2.6 suggests that subjects keep, if anything, too many links. We
hypothesized that the asymmetry of the linking game may explain this behavior.
40
Table 2.9: FE LPM on myopic rationality: effect of type of decision
Turns [1-12] Turns [ 13]
Tr. 1 Tr. 2 Tr. 3 Tr. 4 Tr. 1 Tr. 2 Tr. 3 Tr. 4
(1) (2) (3) (4) (5) (6) (7) (8)
morelink [
1
] 0.168

0.256

0.233

0.144

0.040 0.015 0.097

0.072

(6.51) (25.39) (22.96) (6.51) (1.28) (1.25) (2.96) (2.43)
act [
2
] -0.265

-0.267

-0.277

-0.512

-0.374

-0.074 -0.168

-0.418

(-4.28) (-2.64) (-5.96) (-18.75) (-3.56) (-1.35) (-3.65) (-5.45)
morelink act [
3
] 0.173

0.182 0.263

0.444

0.287 -0.068 0.201

0.335

(3.81) (1.86) (4.79) (17.47) (2.21) (-0.95) (4.02) (5.38)
Constant [
0
] 0.809

0.664

0.731

0.816

0.942

0.916

0.878

0.921

(46.84) (69.11) (127.02) (65.97) (46.10) (57.06) (89.13) (51.57)
Individual xed eects Yes Yes Yes Yes Yes Yes Yes Yes
Linear combinations:
i.
0
0.809

0.664

0.731

0.816

0.942

0.916

0.878

0.921

(46.84) (69.11) (127.0) (65.97) (46.10) (57.06) (89.13) (51.57)
ii.
0
+
1
0.977

0.920

0.964

0.960

0.981

0.931

0.976

0.993

(48.56) (64.59) (114.4) (77.36) (70.81) (56.75) (32.78) (36.17)
iii.
0
+
2
0.544

0.397

0.454

0.303

0.568

0.842

0.711

0.503

(11.57) (4.240) (10.26) (16.29) (6.020) (18.73) (16.16) (6.990)
iv.
0
+
1
+
2
+
3
0.885

0.835

0.950

0.892

0.894

0.789

1.009

0.909

(57.83) (35.82) (49.59) (60.15) (30.18) (14.40) (25.63) (19.83)
N 2304 2304 2304 2304 972 1062 858 768
Adj. R
2
0.142 0.125 0.256 0.267 0.264 0.171 0.242 0.228
t statistics in parentheses. Standard errors are clustered at the session level.

p< 0:05,

p< 0:01,

p< 0:001
41
Since link formation requires mutual agreement while removal does not, one possible
strategy would be to form and maintain some redundant links early on. As the game
approaches the end, individuals begin to unilaterally remove some of them.
We nd some evidence supporting this hypothesis in our regressions. As shown in
Table 2.9, in Turns [1-12] the coecient for myopic rationality in all four treatments
is highest when the myopic rational action is to stay linked (ii), followed by propose
a link (iv), stay unlinked (i), and remove a link (iii). For Turns [ 13], the least
myopic rational decision is still by far to remove a link (iii), except for Treatment 2.
The order of the other coecients are somewhat perturbed although it is dicult to
make strong conclusions since most coecients are very high (90% and above).
Result 6 The size of marginal payos aects the likelihood of a deviation from my-
opic rationality in early turns for Treatments 1, 3 and 4 and in all turns for Treatment
2.
With a myopic rational strategy, the size of the marginal payos should be irrelevant:
individuals would choose actions that give them non-zero gain, irrespective of their
size. To investigate whether this is the case, we implement the extended specication
of (2.3). We then interact the payo variable with the interactions betweenmorelink
and act to capture dierential eects of marginal payos across dierent decision
problems.
The results of the regressions are presented in Table 2.10. We linearly combine
the coecients for the payo variables to explore the heterogeneity of the payo-size
eects across decision problems. We use a strategy similar to the way we linearly
combined in Table 2.8 the coecients of the morelink, act and morelinkact to
examine the myopic rationality of the dierent decision problems. Hence, for example,

0
measures how the size of the marginal payo aects the probability that individuals
42
Table 2.10: FE LPM on myopic rationality: effect of marginal payoff
Turns [1-12] Turns [ 13]
Tr. 1 Tr. 2 Tr. 3 Tr. 4 Tr. 1 Tr. 2 Tr. 3 Tr. 4
(1) (2) (3) (4) (5) (6) (7) (8)
morelink 0.311

0.428

0.303

0.455

0.120 0.053 0.124 0.127

(5.66) (4.32) (5.15) (3.20) (1.38) (1.14) (1.39) (2.40)
act -0.309

0.249

-0.381

-0.613

-0.063 0.227 -0.390

0.084
(-2.98) (3.39) (-6.18) (-4.02) (-0.37) (2.19) (-2.71) (0.27)
morelink act 0.213 -0.251

0.454

0.579

-0.043 -0.531

0.524

-0.241
(1.90) (-3.83) (9.54) (3.83) (-0.24) (-3.01) (2.78) (-0.69)
mpay [
0
] 0.011

0.034

0.015

0.026

0.009 0.011

0.010 0.006
(2.89) (5.24) (3.35) (2.74) (2.24) (4.65) (2.09) (2.19)
mpay morelink [
1
] -0.012

-0.024

-0.014

-0.024

-0.008 -0.002 -0.009 -0.005
(-3.20) (-3.37) (-3.30) (-2.45) (-1.59) (-0.64) (-1.73) (-1.31)
mpay act [
2
] 0.008 -0.048

0.011 0.017 -0.033 -0.027 0.027 -0.043
(0.95) (-4.86) (1.75) (1.57) (-1.33) (-2.28) (1.90) (-1.50)
mpay morelink act [
3
] -0.008 0.045

-0.013 -0.018 0.035 0.038

-0.032 0.048
(-0.82) (4.07) (-2.23) (-1.47) (1.46) (2.64) (-2.11) (1.55)
Constant 0.659

0.012 0.534

0.408

0.767

0.736

0.793

0.798

(10.74) (0.29) (9.33) (2.94) (11.60) (18.08) (16.89) (16.57)
turn sp(1) & turn sp(2) Yes Yes Yes Yes No No No No
turn sp(3) & turn sp(4) No No No No Yes Yes Yes Yes
chdist(ELine) Yes No Yes Yes Yes No Yes Yes
chdist(PS) No Yes No No No Yes No No
Individual xed eects Yes Yes Yes Yes Yes Yes Yes Yes
Linear combinations:
i.
0
0.011

0.034

0.015

0.026

0.009 0.011

0.010 0.006
(2.89) (5.24) (3.35) (2.74) (2.24) (4.65) (2.09) (2.19)
ii.
0
+
1
-0.001 0.010

0.001 0.003

0.001 0.009

0.001 0.001
(-0.94) (5.34) (1.18) (2.87) (0.52) (4.32) (0.43) (0.36)
iii.
0
+
2
0.019

-0.014 0.026

0.044

-0.024 -0.016 0.037

-0.037
(3.16) (-1.39) (4.17) (5.86) (-0.91) (-1.23) (3.03) (-1.31)
iv.
0
+
1
+
2
+
3
-0.000 0.007

-0.002 0.003

0.003 0.020

-0.004 0.005
(-0.03) (5.29) (-0.82) (2.79) (1.21) (5.36) (-0.64) (1.14)
N 2304 2304 2304 2304 972 1062 858 768
Adj. R
2
0.154 0.375 0.300 0.294 0.291 0.282 0.270 0.246
t statistics in parentheses. Standard errors are clustered at the session level.

p< 0:05,

p< 0:01,

p< 0:001
43
myopic-rationally stay unlinked;
0
+
1
measures how the payo size in
uences the
probability that individuals myopic-rationally stay linked, and so on.
For Turns [1-12] in Treatments 1, 3 and 4, myopic rational choices are positively
correlated with the size of marginal payos if the myopic rational choice is to reduce
the number of links (cases (i) and (iii)). Most of these coecients lose their signi-
cance in Turns [ 13]. Meanwhile, for Treatment 2, all of the payo coecients are
signicant in both [1-12] and [ 13] except when the myopic rational choice is to
remove a link.
These results provide additional insights on how individuals deviate from improv-
ing paths. For Treatments 1, 3 and 4, the evidence suggests that individuals take
the opportunity loss from removing a link more seriously than that from staying
unlinked. For Treatment 2, subjects take all losses into consideration. As explored
next, they presumably see those (potential though not necessarily realized) losses as
an intermediary step towards the PS-PWS network.
Result 7 Individual strategies are strongly suggestive of forward-looking behavior.
Finally, we examine the crucial question of forward-looking behavior in Treatment 2.
The analysis at the network outcome level suggests that individuals tend to reach the
PS-PWS network (Result 3). We examine whether this conclusion is also supported
at the individual-level. Table 2.11 presents the extended regression based on (2.3).
Two pieces of evidence corroborate the notion that individuals are forward looking.
First, remember that individuals in Treatment 2 would need to violate myopic
rationality in the rst few turns to escape the zero-payo PWS networkfAg. However,
once these initial \barriers" have been overcome, individuals can reach the PS-PWS
networkfKg by following the improving paths. We therefore expect an increase in
myopic rationality within the rst six turns. As shown in Table 2.11, the spline for
44
Table 2.11: FE LPM on the likelihood of myopic rational action
Turns [1-12] Turns [ 13]
Tr. 1 Tr. 2 Tr. 3 Tr. 4 Tr. 1 Tr. 2 Tr. 3 Tr. 4
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
morelink [
1
] 0.156

0.121

0.092

0.129

0.120

0.018 0.027 0.016 0.016 0.061
(6.22) (7.94) (5.41) (6.56) (4.82) (0.80) (1.90) (1.16) (0.25) (2.12)
act [
2
] -0.242

-0.254 -0.206 -0.258

-0.469

-0.349

0.007 0.018 -0.149

-0.410

(-4.32) (-2.18) (-2.00) (-6.77) (-15.60) (-2.99) (0.12) (0.37) (-3.54) (-5.12)
morelink act [
3
] 0.161

0.232 0.186 0.293

0.438

0.265 -0.160 -0.174 0.188

0.319

(4.39) (1.93) (1.68) (4.96) (15.49) (1.92) (-1.96) (-2.19) (3.01) (5.32)
turn sp(1)
y
-0.006 0.079

0.079

0.012 0.003
(-1.01) (10.96) (10.86) (2.03) (0.30)
turn sp(2)
y
0.012

-0.000 0.000 0.009 0.005
(2.73) (-0.10) (0.08) (2.00) (1.85)
turn sp(3)
y
0.017

0.002 0.003 0.006 0.016
(2.81) (0.64) (0.73) (0.91) (1.82)
turn sp(4)
y
0.003 -0.007 -0.008 -0.007 0.001
(1.23) (-2.29) (-2.35) (-1.80) (0.09)
chdist(E. Line) -0.025

-0.086

-0.056

-0.037

-0.018 -0.036

-0.034 0.006
(-3.42) (-5.37) (-5.23) (-2.71) (-1.18) (-2.60) (-1.93) (0.60)
chdist(PS) -0.071

-0.038

(-5.81) (-2.98)
Constant [
0
] 0.773

0.157

0.197

0.621

0.707

0.843

0.693

0.713

0.857

0.837

(19.37) (7.01) (7.17) (15.74) (14.06) (28.07) (24.16) (32.38) (17.34) (18.81)
mpay Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Individual xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Linear combinations:

0
0.773

0.157

0.197

0.621

0.707

0.843

0.693

0.713

0.857

0.837

(19.37) (7.009) (7.167) (15.74) (14.06) (28.07) (24.16) (32.38) (17.34) (18.81)

0
+
1
0.929

0.278

0.289

0.750

0.827

0.861

0.720

0.730

0.872

0.898

(23.24) (9.824) (9.472) (21.81) (22.47) (53.26) (29.77) (32.58) (8.398) (15.55)

0
+
2
0.531

-0.097

-0.009

0.363

0.238

0.495

0.700

0.732

0.707

0.426

(16.28) (-0.831) (-0.0975) (5.535) (5.785) (5.184) (14.97) (17.31) (11.76) (5.469)

0
+
1
+
2
+
3
0.848

0.257

0.268

0.785

0.796

0.778

0.566

0.574

0.911

0.807

(45.19) (6.557) (6.374) (30.34) (37.74) (24.07) (7.394) (7.862) (7.650) (11.60)
N 2304 2304 2304 2304 2304 972 1062 1062 858 768
Adj. R
2
0.186 0.381 0.388 0.319 0.311 0.353 0.341 0.342 0.344 0.336
t statistics in parentheses
y
Spline coecients are for the slope of the intervals. Standard errors are clustered at the session level.

p< 0:05,

p< 0:01,

p< 0:001
45
the rst six turns, turn sp1, is positive and signicant for Treatment 2 and only for
that treatment. All other turn spline variables for Treatment 2 are not signicant.
Second, the coecient for the distance change to the PS-PWS network, chdist(PS),
is negative and signicant in both early and late turns (columns (2) and (7) in the
table). It suggests that individuals may be implementing strategies that move them
closer to the PS-PWS network. Notice, however, that the impact of this variable is
similar to that of the distance to the ecient-line network variable, chdist(E.Line)
(columns (3) and (8) in the table). Because the two networks dier by a mere one
link, we cannot identify which network is the individuals' ultimate goal. However the
evidence regarding the convergent network outcomes for Treatment 2 (in Table 2.4)
suggests that the PS-PWS architecture is the ultimate goal.
2.4.3 Model predictions
We next examine the capacity of our empirical model to predict actions. For this
exercise, we include the regressors of the basic model and the marginal payo variable
interacted with morelink, in line with our previous ndings. All the variables are
interacted with turn> 12 to account for coecient dierences before and after Turn
12. We also include only the spline for the rst six turns, and interact it with a
Treatment 2 indicator variable to account for players' forward-looking behavior in
that treatment. For the prediction model, we do not include the individual xed
eects.
The results of the Logit regressions used for out-of-sample predictions are relegated
to Appendix 2.A.4. Figure 2.2 graphically depicts the plot of the actual (dashed line)
and out-of-sample predicted (solid line) proportion of myopic rational choices for
each treatment in the rst 18 turns. The model predicts aggregate behavior well. As
46
expected, for Treatment 2 it overestimates myopic rationality at early stages, when
deviations are necessary to reach the PS-PWS network, but it predicts actual behavior
accurately after Turn 6. The model slightly under-predicts myopic rationality in early
turns of Treatment 3 and slightly over-predicts myopic rationality in late turns of
Treatment 4.
1 .8 .6 .4 .2 0
Myopic rationality
0 6 12 18
Turn
Actual Predicted
Treatment 1
1 .8 .6 .4 .2 0
Myopic rationality
0 6 12 18
Turn
Actual Predicted
Treatment 2
1 .8 .6 .4 .2 0
Myopic rationality
0 6 12 18
Turn
Actual Predicted
Treatment 3
1 .8 .6 .4 .2 0
Myopic rationality
0 6 12 18
Turn
Actual Predicted
Treatment 4
Figure 2.2: Out-of-sample logit prediction by treatment
2.4.4 Extensions and alternative representations
We have focused on the LPM in the regression. As a robustness check, we also esti-
mated xed-eects logit models. The results for the analogue of Tables 2.9 and 2.11
47
in these alternative estimations are presented in Appendix 2.A.5. For the most part,
the estimates are qualitatively similar, except for the coecients of the interactions
between morelink and act.
We also entertain the possibility that experience matters. Remember that each
treatment is played twice in a session. We examine dierences in behavior between
the rst and second time a treatment is encountered by running a test of pooling.
The results are presented in Appendix 2.A.6. In general, we nd little evidence of
dierences in behavior. The most notable dierence is Treatment 2 where, in their
second encounter, subjects are less myopic rational when it prescribes to remove or
propose a link.
2.4.5 Summary
The individual-choice analysis suggests that subjects play less myopically rational in
turns with sure future than in random ending turns and also less myopically rational
when deviations are necessary to escape the low-payo PWS network. Deviations tend
to take the form of taking up excessive links, possibly because they can be removed
unilaterally, although proving this hypothesis would require further work. Finally,
deviations are also more prevalent the smaller the marginal payo losses, as expected
in a behavioral theory where mistakes depend inversely on loss magnitudes. Overall
and with some interesting exceptions, the analysis provides support for convergence to
PWS networks through myopic rational choices, except when multiple PWS networks
exist in which case subjects exhibit the forward-looking attitude necessary to reach
high payos.
48
2.5 Heterogeneity across subjects
So far we have studied choices at the network outcome and single decision levels. One
question that remains unanswered is the degree of heterogeneity between subjects. A
simple way to address this question is to determine how often each subject plays the
myopic rational strategy.
Figure 2.3 plots the cumulative distribution function (CDF) of myopic rational
behavior of each subject by treatment. A steep CDF re
ects homogeneity across
subjects whereas a right shift captures a more myopic rational behavior on aggregate.
In Treatments 1, 3 and 4, behavior is myopic rational and homogeneous: 80% of
subjects play the myopic rational strategy 75% of the time or more. In Treatment 2,
behavior is slightly less myopically rational and more heterogeneous. A Kolmogorov-
Smirnov test conrms this observation: the CDF of Treatment 2 is dierent from the
CDF of Treatments 1, 3 and 4 at the 1% level, whereas no statistical dierence is
observed between the CDFs in Treatments 1, 3 and 4 at the 10% level.
Heterogeneity can also be studied by searching for clusters of people (as in Camerer
and Ho (1999) for example). This allows us to quantify the degree of similarity of
subjects' behavior within and between clusters. Given the documented dierence in
behavior between early and late turns and also between Treatment 2 and Treatments
1, 3 and 4, we use these four variables to cluster the subjects. There are many
clustering methods but they usually require the number of clusters and the clustering
criterion to be set ex-ante rather than endogenously optimized. Mixture models, on
the other hand, treat each cluster as a component probability distribution. Thus,
the choice of the model and the number of clusters is made using Bayesian statistical
methods (Fraley and Raftery, 2002). We implement model-based clustering analysis
with the Mclust package in R (Fraley and Raftery, 2009). A maximum of nine clusters
49
0
.2
.4
.6
.8
1
Cumulative Probability
.2 .4 .6 .8 1
Myopic rationality
Treat. 1 Treat.2 Treat. 3 Treat. 4
Figure 2.3: Empirical CDF of myopic rationality by treatment
are considered for up to ten dierent models and the combination that yields the
maximum Bayesian Information Criterion (BIC) is chosen.
For our multidimensional data, the model that maximizes the BIC provides ve
clusters. Table 2.12 shows the frequencies of subjects in each cluster, listed from low
to high according to the general frequency of their myopic rational behavior. It also
displays their average earnings.
Result 8 There is substantial heterogeneity across individuals. Some subjects consis-
tently follow improving paths except when deviations are needed to reach the PS-PWS
network whereas others exhibit a rather erratic behavior.
Clusters are clearly dierentiated in: (i) total level of myopic rationality, (ii) dier-
ence in behavior between Treatment 2 and the other treatments, and (iii) dierence
in behavior between early and late turns. Subjects in cluster 1 are lowest on all three
dimensions, followed by subjects in cluster 2. Their erratic behavior and undieren-
tiated choice across treatments and turns indicates that one-quarter of our subjects
50
Table 2.12: Clustering based on myopic rational behavior
Cluster
Treatments 1-3-4 Treatment 2
Earnings N
[1-12] [ 13] [1-12] [ 13]
1 0.678 0.735 0.617 0.645 86.5 10
(0.132) (0.178) (0.134) (0.226)
2 0.786 0.850 0.708 0.814 95.8 13
(0.060) (0.060) (0.099) (0.082)
3 0.779 0.955 0.645 0.844 83.7 19
(0.125) (0.026) (0.147) (0.099)
4 0.843 0.909 0.746 1.000 96.6 46
(0.076) (0.076) (0.113) (0.000)
5 0.870 0.975 0.771 0.881 113.8 8
(0.038) (0.034) (0.059) (0.054)
Total 0.808 0.897 0.710 0.897 96
(0.104) (0.104) (0.127) (0.145)
Standard deviations in parenthesis
have diculties in understanding some of the strategic aspects of the game. Subjects
in cluster 3 start with low myopic rationality but have the sharpest increase after turn
12 (around 20%). The near perfectly myopic rational behavior in later stages of Treat-
ments 1-3-4 suggests that, contrary to subjects in clusters 1 and 2, these individuals
understand the incentives in our game and use early turns to look for an advantageous
position. Clusters 4 and 5, which comprise more than half of the population, are a
portrait of rational individual maximization in a strategic game: very high myopic
rationality in later turns of all treatments, a somewhat lower myopic rationality in
early turns of Treatments 1-3-4 (possibly due to costless experimentation), and a sig-
nicantly lower myopic rationality in early turns of Treatment 2 (necessary to escape
the low-payo PWS network). The main dierence between the two clusters lies in
the late turn treatments where myopic rationality is highest: Treatment 2 for cluster
4 and Treatments 1-3-4 for cluster 5.
51
Earnings are correlated with behavior, with clusters 1 and 5 near the bottom
and top of the distribution. However, the mapping between payos and myopic
rationality is not monotonic. This can be expected for several reasons. First, because
PWS networks do not necessarily generate the highest payos. Second, because a
subject's payo depends on the behavior of the ve other players in the network and
their nal position in the component. Rational subjects may end up bearing a larger
number of links.
2.6 Conclusion
The chapter studied the dynamic formation of social networks. We found that sub-
jects rarely consider the total value of the network as a key criterion when making
their decisions. Instead, choices are roughly consistent with individual maximiza-
tion of payos. Many subjects exhibit forward-looking behavior if (and only if) this
strategy leads to a PS-PWS conguration. Interestingly, myopic rationality is less
prevalent at the margin when actions are reversible (early turns), marginal payo
losses are smaller, and when deviations lead subjects to have excessive links that
can be removed unilaterally later on. There is, however, signicant heterogeneity in
behavior across subjects.
52
M L K I
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
N
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Non minimally-connected network
J
A
B
T
S R Q P
O
H G F E
D C
Treatment 1
Benefit 0,20,30,39,42,43
Cost per link 15
  
  
  
  
U: 1 (3.1%)
C: -
U: 2 (6.3%)
C: -
U: 3 (9.4%)
C: -
U: 5 (15.6%)
C: -
U: 4 (12.5%)
C: 1 (14.3%)
U: 2 (6.3%)
C: 1 (14.3%)
U: 3 (9.4%)
C: -
U: 5 (15.6%)
C: 1 (14.3%)
U: 2 (6.3%)
C: 1(14.3%)
U: 2 (6.3%)
C: 2 (28.6%)
U: 1 (3.1%)
C: -
U: 1 (3.1%)
C: -
U: 1 (3.1%)
C: 1 (14.3%)
Notes. Letters refer to all minimally-connected network architectures. Numbers next to each network refer to the
frequency (and percentage) that the process ends in that network: U = Unconditional on convergence, C =
Conditional on no change in the last 3 turns. Networks that are part of the closed cycle are inside the shaded region.
Figure 2.4: Improving paths in Treatment 1
53
M L I
  
  
  
  
N
  
  
K
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Non minimally-connected network
J
A
B
T
S R Q P
O
H G F E
D C
  
  
  
  
  
  
  
  
Treatment 2
Benefit 0,10,17,22,38,44
Cost per link 15
  
  
  
  
  
  
  
  
  
  
U: 2 (6.3%)
C: -
U: 2 (6.3%)
C: -
U: 3 (9.4%)
C: -
U: 2 (6.3%)
C: 2 (10%)
U: 1 (3.1%)
C: 1 (5%)
U: 1 (3.1%)
C: 1 (5%)
U: 1 (3.1%)
C: -
U: 1 (3.1%)
C: -
U: 4 (12.5%)
C: 4 (20%)
U: 14 (43.8%)
C: 11 (55%)
U: 3 (9.4%)
C: 1 (5%)
Notes. Letters refer to all minimally-connected network architectures. Numbers next to each network refer to the
frequency (and percentage) that the process ends in that network: U = Unconditional on convergence, C =
Conditional on no change in the last 3 turns. PWS networks are shaded (light shade for the PS-PWS network).
Figure 2.5: Improving paths in Treatment 2
54
M L K I
  
  
  
  
N
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Non minimally-connected network
J
A
B
T
S R Q P
O
H G F E
D C
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Treatment 3
Benefit 0,29,36,41,43,44
Cost per link 15
U: 4 (12.5%)
C: 3 (17.7%)
U: 7 (21.9%)
C: 3 (17.7%)
U: 2 (6.3%)
C: 1 (5.9%)
U:15 (46.9%)
C: 9 (52.9%)
U: 1 (3.1%)
C: 1 (5.9%)
U: 2 (6.3%)
C: -
U: 1 (3.1%)
C: -
Notes. Letters refer to all minimally-connected network architectures. Numbers next to each network refer to the
frequency (and percentage) that the process ends in that network: U = Unconditional on convergence, C =
Conditional on no change in the last 3 turns. PWS network is shaded.
Figure 2.6: Improving paths in Treatment 3
55
M L K I
  
  
  
  
N
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Non minimally-connected network
J
A
B
T
S R Q P
O
H G F E
D C
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Treatment 4
Benefit 0,19,36,42,44,45
Cost per link 15
  
  
  
  
U: 1 (3.1%)
C: -
U: 3 (9.4%)
C: 3 (16.7%)
U: 5 (15.6%)
C: 2 (11.1%)
U: 2 (6.3%)
C: 1 (5.6%)
U: 10 (31.2%)
C: 4 (22.2%)
U: 7 (21.9%)
C: 6 (33.3%)
U: 1 (3.1%)
C: 1 (5.6%)
U: 1 (3.1%)
C: -
U: 1 (3.1%)
C: 1 (5.6%)
U: 1 (3.1%)
C: -
Notes. Letters refer to all minimally-connected network architectures. Numbers next to each network refer to the
frequency (and percentage) that the process ends in that network: U = Unconditional on convergence, C =
Conditional on no change in the last 3 turns. PWS network is shaded.
Figure 2.7: Improving paths in Treatment 4
56
Chapter bibliography
Ai, Chunrong and Edward C. Norton, \Interaction terms in logit and probit
models," Economics Letters, July 2003, 80, 123{129.
Bala, Venkatesh and Sanjeev Goyal, \A Noncooperative Model of Network For-
mation," Econometrica, September 2000, 68 (5), 1181{1229.
Berman, Eli, \Sect, Subsidy, and Sacrice: An Economist's View of Ultra-Orthodox
Jews," Quarterly Journal of Economics, August 2000, 115, 905{953.
Berninghaus, Siegfried K., Karl-Martin Ehrhart, and Marion Ott, \A net-
work experiment in continuous time: The in
uence of link costs," Experimental
Economics, September 2006, 9 (3), 237{251.
, , , and Bodo Vogt, \Evolution of networks|an experimental analysis,"
Journal of Evolutionary Economics, January 2007, 17 (3), 317{347.
Bramoull e, Y. and R. Kranton, \Risk-sharing networks," Journal of Economic
Behavior & Organization, 2007, 64 (3-4), 275{294.
Burger, Martijn J. and Vincent Buskens, \Social context and network forma-
tion: An experimental study," Social Networks, January 2009, 31 (1), 63{75.
Callander, S. and C. R. Plott, \Principles of network development and evolution:
An experimental study," Journal of Public Economics, 2005, 89 (8), 1469{1495.
Camerer, Colin and Teck Hua Ho, \Experience-weighted Attraction Learning in
Normal Form Games," Econometrica, 1999, 67 (4), 827{874.
57
Charness, Gary, Francesco Feri, Miguel A. Mel endez-Jim enez, and
Matthias Sutter, \Equilibrium Selection in Experimental Games on Networks,"
Working Paper February 2012.
Di Cagno, Daniela and Emanuela Sciubba, \The Determinants of Individual
Behaviour in Network Formation: Some Experimental Evidence," in Mohammed
Abdellaoui, John D. Hey, W. Leinfellner, G. Eberlein, and S. H. Tijs, eds., Advances
in Decision Making Under Risk and Uncertainty, Springer Berlin Heidelberg, 2008,
pp. 219{241.
Falk, Armin and Michael Kosfeld, \It's all about connections: Evidence on
network formation," Working paper, Institute for Empirical Research in Economics,
University of Zurich March 2003.
Fatas, Enrique, Miguel Mel endez-Jim enez, and Hector Solaz, \An experi-
mental analysis of team production in networks," Experimental Economics, 2010,
13 (4), 399{411.
Fraley, Chris and Adrian E. Raftery, \Model-Based Clustering, Discriminant
Analysis, and Density Estimation," Journal of the American Statistical Association,
June 2002, 97 (458), 611{631.
and , \MCLUST Version 3 for R: Normal Mixture Modeling and Model-Based
Clustering," Technical Report 504, Department of Statistics, University of Wash-
ington December 2009.
Goeree, Jacob K., Arno Riedl, and Alja z Ule, \In search of stars: Network for-
mation among heterogeneous agents," Games and Economic Behavior, November
2009, 67 (2), 445{466.
58
Goyal, Sanjeev and Sumit Joshi, \Unequal connections," International Journal
of Game Theory, 2006, 34 (3), 319{349.
Greene, William, \The behaviour of the maximum likelihood estimator of limited
dependent variable models in the presence of xed eects," Econometrics Journal,
2004, 7 (1), 98{119.
Herings, P. Jean-Jacques, Ana Mauleon, and Vincent Vannetelbosch, \Far-
sightedly stable networks," Games and Economic Behavior, November 2009, 67 (2),
526{541.
Jackson, M and A Wolinsky, \A Strategic Model of Social and Economic Net-
works," Journal of Economic Theory, October 1996, 71 (1), 44{74.
Jackson, Matthew O. and Alison Watts, \The Evolution of Social and Economic
Networks," Journal of Economic Theory, October 2002, 106 (2), 265{295.
Kirchsteiger, Georg, Marco Mantovani, Ana Mauleon, and Vincent Van-
netelbosch, \Myopic or farsighted? An experiment on network formation," Dis-
cussion Paper 8263, Centre for Economic Policy Research, London February 2011.
McKelvey, Richard and Thomas Palfrey, \Quantal Response Equilibria for
Normal Form Games," Games and Economic Behavior, July 1995, 10 (1), 6{38.
Pantz, Katinka, The strategic formation of social networks : experimental evidence,
Aachen: Shaker, 2006.
59
Appendix 2.A Additional analyses
2.A.1 Geodesic distance to ecient and PWS networks
In this section, we provide a complementary study of the dierence between observed
and predicted outcomes to the one presented in section 2.3.1. More precisely, we
calculate the shortest (or \geodesic") distance between the resulting networks and
the closest network in the closed cycle (for Treatment 1) or the PWS networks (for
Treatments 2, 3 and 4) as well as the distance between the resulting networks and
the ecient networks. For the latter, we separately calculate the distance to the
closest of all the ecient networksfO;P;Q;R;S;Tg and to the line networkfTg.
19
Table 2.13 shows that for Treatments 1 and 3, the distance to the closed cycle and the
PWS network respectively is substantially shorter than the distance to the ecient
networks. For Treatment 2, the distance is shorter to the PS-PWS network but
longer to the PWS empty networkfAg. For Treatment 4, however, the distance from
the PWS network is equal to the distance from the ecient line network and longer
than the distance from the closest ecient network, suggesting a larger dispersion in
behavior.
Table 2.14 presents the average distance between outcomes and predicted net-
works conditional on convergence. For Treatments 2 and 3, the results provide further
support for convergence to the PS-PWS network and the unique PWS network re-
spectively. For Treatment 4, the distance from the convergent network to the ecient
network is lower than to the PWS network. This result (as well as that in Table 2.13)
19
If agents were to aim at the ecient network, the line network is the most likely outcome since
it distributes payos most equally. For examplefOg, which is never played in our experiment, is
ecient but requires one player to form 5 links and therefore bear signicant payo losses (30 to 32
tokens depending on the treatment).
60
Table 2.13: Average distances from outcomes
Treatment PWS PS-PWS
Closed
cycle
Shortest
ecient
Ecient
line
1 - - 0:41
y
1.66 2.03
2 3:66
z
0.91 - 1.41 1.66
3 1.00 - - 1.94 2.06
4 1.41 - - 1.28 1.41
y
We calculate the distance to the closest network in the cycle.
z
We only consider the empty networkfAg.
comes from the fact that most of the network outcomes in this treatment, both con-
ditional and unconditional on convergence, are split almost equally between the PWS
networkfLg and networkfNg. Since the distance between these two networks is two
and both of them are at a distance of one to the ecient networks, the distance from
the stable network and the ecient networks are similar. It is dicult to infer from
the outcomes alone where the formation processes is leading toward. In our analysis
of individual decisions (Result 5), we provide a plausible explanation for our ndings
here.
Table 2.14: Average distance from outcomes conditional on convergence
Treatment PWS PS-PWS
Shortest
ecient
Ecient
line
2 3:75

0.75 1.25 1.35
3 0.88 - 2.06 2.18
4 1.61 - 1.28 1.39

We only consider the empty networkfAg.
61
2.A.2 Myopic rationality by treatment and decision problem
Figure 2.8 further illustrates the results in Panels B and C of Table 2.6. We plot for
each treatment the proportion of myopic rational behavior across turns when passing
is rational (B1) and when acting is rational (B2). In all treatments, players maintain
more links in all turns than would be observed if they played myopically rational all
the time. The gap is bigger and the variation larger for decisions where acting is
myopic rational (gures on the right), although the dierence narrows as the match
nears its end.
2.A.3 Pooled FE LPM on probability of myopic rational choices
Table 2.15 presents the xed eects pooled linear probability model to investigate if
there is a dierence in myopic rational choices before and after Turn 12.
2.A.4 Logit model for out-of-sample predictions
Table 2.16 presents the logit model used for the out-of-sample predictions.
2.A.5 Logit estimation
In Tables 2.17 and 2.18 we present the analogue of Tables 2.9 and 2.11 using a xed-
eects logic estimation. Note that the sign and signicance of the morelinkact
interaction-term coecients often dier between the LPM and logit models. This,
however, does not necessarily indicate that the two models contradict each other
because, unlike in linear models (such as LPM), the coecients on the interaction
terms in non-linear models (such as logit) do not easily translate into their marginal
eects (for a detailed discussion, see Ai and Norton (2003)).
62
1 .8 .6 .4 .2 0
Myopic rationality
6 12 18
Turn
(i) Stay unlinked (ii) Stay linked
B1. Myopic rational to pass
1 .8 .6 .4 .2 0
Myopic rationality
6 12 18
Turn
(iii) Remove (iv) Propose
B2. Myopic rational to act
Treatment 1
1 .8 .6 .4 .2 0
Myopic rationality
6 12 18
Turn
(i) Stay unlinked (ii) Stay linked
B1. Myopic rational to pass
1 .8 .6 .4 .2 0
Myopic rationality
6 12 18
Turn
(iii) Remove (iv) Propose
B2. Myopic rational to act
Treatment 2
1 .8 .6 .4 .2 0
Myopic rationality
6 12 18
Turn
(i) Stay unlinked (ii) Stay linked
B1. Myopic rational to pass
1 .8 .6 .4 .2 0
Myopic rationality
6 12 18
Turn
(iii) Remove (iv) Propose
B2. Myopic rational to act
Treatment 3
1 .8 .6 .4 .2 0
Myopic rationality
6 12 18
Turn
(i) Stay unlinked (ii) Stay linked
B1. Myopic rational to pass
1 .8 .6 .4 .2 0
Myopic rationality
6 12 18
Turn
(iii) Remove (iv) Propose
B2. Myopic rational to act
Treatment 4
Figure 2.8: Myopic rationality by treatment and decision problem
63
Table 2.15: Pooled FE LPM on the likelihood of myopic rational action
Treatment 1 Treatment 2 Treatment 3 Treatment 4
(1) (2) (3) (4) (5) (6) (7) (8) (9)
1(turn>12) 0.130

0.063 0.263

0.486

0.472

0.169

0.204

0.105

0.018
(5.54) (1.31) (31.72) (9.16) (9.58) (4.07) (2.67) (6.32) (0.32)
morelink 0.163

0.153

0.253

0.142

0.109

0.236

0.138

0.145

0.123

(7.13) (6.62) (27.72) (7.94) (5.36) (22.35) (5.92) (6.81) (5.72)
. . . 1(turn>12) -0.121

-0.125

-0.231

-0.108

-0.089

-0.162

-0.119

-0.088

-0.069

(-4.16) (-5.49) (-15.03) (-4.45) (-3.75) (-4.23) (-2.60) (-4.73) (-3.66)
act -0.274

-0.255

-0.273

-0.234 -0.182 -0.294

-0.277

-0.524

-0.483

(-4.38) (-4.45) (-2.77) (-2.12) (-1.91) (-6.56) (-7.61) (-18.38) (-15.46)
. . . 1(turn>12) -0.087 -0.084 0.194 0.233

0.198 0.077 0.079 0.103 0.076
(-1.06) (-0.85) (2.32) (2.56) (2.34) (1.26) (1.54) (1.63) (1.18)
morelink act 0.185

0.186

0.188 0.202 0.152 0.265

0.291

0.458

0.454

(3.83) (4.65) (1.96) (1.77) (1.45) (4.90) (5.37) (17.77) (16.66)
. . . 1(turn>12) 0.079 0.056 -0.291

-0.416

-0.384

-0.046 -0.078 -0.130

-0.142

(0.86) (0.50) (-3.07) (-4.03) (-3.57) (-0.70) (-1.26) (-2.52) (-3.12)
mpay 0.003

0.015

0.013

0.005

0.005

(2.67) (9.26) (10.84) (2.88) (4.94)
. . . 1(turn>12) -0.000 -0.002 -0.001 -0.002 -0.003
(-0.06) (-0.52) (-0.23) (-0.91) (-1.40)
turn 0.005 0.030

0.031

0.010

0.004
(1.23) (9.00) (10.97) (3.20) (1.93)
. . . 1(turn>12) 0.002 -0.034

-0.035

-0.007 0.007
(0.51) (-7.53) (-8.83) (-1.47) (2.05)
chdist(E. Line) -0.024

-0.096

-0.055

-0.034

(-3.68) (-5.96) (-5.21) (-2.39)
. . . 1(turn>12) 0.010 0.046

0.030

0.040

(0.61) (2.92) (2.65) (3.02)
chdist(PS) -0.087

(-6.94)
. . . 1(turn>12) 0.039

(2.89)
Constant 0.811

0.736

0.663

0.286

0.327

0.732

0.635

0.818

0.704

(49.21) (21.15) (70.98) (11.99) (12.77) (59.12) (20.63) (57.31) (21.23)
Individual xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes
P(pooling) 0.004 0.002 0.000 0.000 0.000 0.023 0.004 0.003 0.001
N 3276 3276 3366 3366 3366 3162 3162 3072 3072
Adj. R
2
0.169 0.176 0.154 0.331 0.336 0.245 0.270 0.253 0.264
t statistics in parentheses. Standard errors are clustered at the session level.

p< 0:05,

p< 0:01,

p< 0:001
64
Table 2.16: Logit regression
Treat. 1 Treat. 2 Treat. 3 Treat. 4
(1) (2) (3) (4)
1(turn>12) 1.831

1.669

2.193

1.487

(4.35) (4.24) (3.67) (3.46)
morelink 2.817

3.761

3.246

3.090

(7.95) (9.62) (9.13) (8.47)
. . . 1(turn>12) -1.961

-1.323

-2.179

-1.737

(-2.67) (-1.97) (-3.16) (-3.39)
act -1.861

-1.763

-1.499

-1.465

(-12.04) (-21.76) (-8.52) (-7.87)
. . . 1(turn>12) -0.139 -0.515 -0.705

-0.433
(-0.57) (-1.80) (-2.49) (-1.34)
morelink act 1.244

-0.0600 0.803

0.595

(4.97) (-0.18) (3.80) (2.45)
. . . 1(turn>12) -1.353

-0.363 -0.578 -0.725

(-2.99) (-1.13) (-1.43) (-2.19)
turn sp(1)
y
0.233

0.0902

0.206

0.243

(8.62) (2.11) (6.90) (7.43)
. . . 1(Treatment 2) -0.0520

-0.0479

-0.0749

(-2.95) (-2.36) (-7.72)
mpay 0.117

0.0825

0.143

0.109

(6.93) (5.33) (5.52) (6.36)
. . . 1(turn>12) -0.0686

-0.0511 -0.0756 -0.0291
(-2.14) (-1.56) (-1.46) (-0.76)
mpay morelink -0.0863

-0.0876

-0.125

-0.0944

(-3.29) (-3.70) (-3.93) (-3.01)
. . . 1(turn>12) 0.112

0.0823

0.110

0.0701
(2.10) (2.64) (2.27) (1.54)
Constant -1.206

-0.0529 -1.459

-1.114

(-6.44) (-0.22) (-7.40) (-6.06)
Observations 9600 9510 9714 9804
Pseudo R
2
0.206 0.173 0.190 0.183
t statistics in parentheses. Standard errors clustered at the session level.
y
Spline coecients are for the slope of the intervals.

p< 0:05,

p< 0:01,

p< 0:001
65
Table 2.17: FE Logit on myopic rationality: effect of type of decision
Turns [1-12] Turns [ 13]
Tr. 1 Tr. 2 Tr. 3 Tr. 4 Tr. 1 Tr. 2 Tr. 3 Tr. 4
(1) (2) (3) (4) (5) (6) (7) (8)
morelink 2.878

1.674

5.044

1.921

2.126

0.329 -
z
1.808

(5.61) (8.54) (5.00) (6.15) (2.04) (0.93) (2.39)
act -1.385

-1.199

-1.365

-2.608

-2.896

-0.964

-1.619

-2.854

(-8.46) (-7.27) (-7.50) (-14.13) (-8.05) (-2.02) (-3.96) (-7.68)
morelink act -0.910 0.547 -1.677 1.321

0.062 -0.651 3.481

0.651
(-1.67) (1.89) (-1.61) (3.52) (0.06) (-1.06) (4.93) (0.76)
Constant 1.573

0.719

1.204

1.682

3.322

2.958

3.080

2.804

(14.88) (9.66) (9.08) (13.85) (11.17) (12.10) (9.59) (11.08)
Individual xed eects Yes Yes Yes Yes Yes Yes Yes Yes
Ln
2
u
Constant -0.966

-1.636

0.114 -0.456 0.426 0.279 0.972

-0.303
(-3.36) (-4.96) (0.50) (-1.74) (1.16) (0.79) (2.64) (-0.53)
N 2304 2304 2304 2304 972 1062 858 768
t statistics in parentheses.
y
Spline coecients are for the slope of the intervals.
z
Dummy variable dropped due to perfect collinearity.

p< 0:05,

p< 0:01,

p< 0:001
66
Table 2.18: FE Logit on the likelihood of myopic rational action
Turns [1-12] Turns [ 13]
Tr. 1 Tr. 2 Tr. 3 Tr. 4 Tr. 1 Tr. 2 Tr. 3 Tr. 4
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
morelink 2.869

1.177

0.998

4.420

1.939

1.883 0.840

0.605 -
z
1.899

(5.56) (5.28) (4.42) (4.35) (5.92) (1.78) (2.02) (1.42) (2.47)
act -1.220

-1.326

-1.114

-1.305

-2.249

-2.587

-0.022 0.195 -1.214

-2.699

(-7.00) (-7.00) (-5.77) (-6.70) (-11.41) (-6.88) (-0.04) (0.36) (-2.70) (-6.81)
morelink act -1.020 1.035

0.802

-1.421 1.174

-0.257 -1.974

-2.137

2.372

0.271
(-1.84) (3.16) (2.42) (-1.35) (2.91) (-0.23) (-2.74) (-2.94) (3.13) (0.31)
mpay 0.029

0.105

0.094

0.046

0.074

0.051

0.171

0.161

0.064 0.049

(2.31) (8.50) (7.65) (3.17) (4.83) (2.35) (7.36) (7.01) (1.96) (1.98)
turn spline: before 6
y
-0.051 0.361

0.354

0.097 -0.012
(-0.95) (7.74) (7.59) (1.83) (-0.20)
turn spline: b/w 6 & 12
y
0.107

0.045 0.050 0.103

0.063
(2.94) (1.20) (1.34) (2.71) (1.63)
chdist(E. Line) -0.162

-0.471

-0.398

-0.265

-0.280 -0.540

-0.449

0.154
(-2.38) (-6.80) (-4.75) (-3.47) (-1.89) (-3.82) (-2.27) (0.95)
chdist(PS) -0.308

-0.468

(-4.59) (-3.09)
turn spline: b/w 12 & 18
y
0.227

0.038 0.043 0.059 0.308

(2.49) (0.45) (0.52) (0.62) (3.07)
turn spline: after 18
y
0.040 -0.068 -0.072 -0.079 0.220
(0.72) (-1.61) (-1.70) (-0.83) (0.63)
Constant 1.238

-2.038

-1.783

0.323 0.540 2.049

0.798 1.013

2.856

1.355

(4.55) (-10.59) (-8.89) (1.14) (1.74) (4.74) (1.81) (2.35) (5.34) (2.84)
Individual xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Ln
2
u
Constant -0.802

-0.791

-0.658

0.447

-0.269 0.473 0.923

0.887

1.483

-0.399
(-2.94) (-3.00) (-2.59) (2.02) (-1.06) (1.31) (2.93) (2.78) (3.99) (-0.65)
N 2304 2304 2304 2304 2304 972 1062 1062 858 768
t statistics in parentheses
y
Spline coecients are for the slope of the intervals.
z
Dummy variable dropped due to perfect collinearity.

p< 0:05,

p< 0:01,

p< 0:001
67
2.A.6 Eect of experience
Table 2.19 presents the results regarding the eect of experience on behavior. For the
basic specications, we only nd evidence of a dierence in behavior for Treatment
2. With additional variables, the hypothesis of no dierence in behavior is rejected
for Treatment 4. Looking at the individual coecients, we nd that in Treatment 2,
coecients that are signicantly dierent in the second half of the sessions are those
for act and morelinkact. For Treatment 4, none of the individual coecients are
signicantly dierent in the second half of the sessions at 5% signicance, and only
one coecient, namely turn, is signicantly dierent at 10% signicance.
68
Table 2.19: Pooled FE LPM on the likelihood of myopic rational action
Treatment 1 Treatment 2 Treatment 3 Treatment 4
(1) (2) (3) (4) (5) (6) (7) (8) (9)
1(second half) -0.051 -0.073 -0.025 -0.024 -0.022 -0.020 0.009 -0.031 -0.094
(-1.19) (-1.54) (-0.95) (-0.37) (-0.37) (-0.88) (0.21) (-0.76) (-1.24)
morelink 0.098

0.090

0.175

0.121

0.087

0.184

0.098

0.098

0.075

(4.65) (3.19) (10.07) (6.39) (3.29) (7.83) (2.88) (3.25) (3.09)
. . . 1(second half) 0.063 0.050 0.023 0.010 0.012 0.013 0.013 0.046 0.044
(1.53) (1.18) (0.77) (0.33) (0.38) (0.50) (0.35) (1.19) (1.35)
act -0.300

-0.289

-0.136 -0.021 0.024 -0.255

-0.250

-0.491

-0.457

(-3.43) (-3.25) (-1.98) (-0.31) (0.45) (-6.17) (-6.03) (-6.52) (-5.89)
. . . 1(second half) 0.003 0.028 -0.222

-0.279

-0.261

-0.069 -0.021 -0.032 -0.034
(0.03) (0.24) (-3.71) (-5.73) (-4.13) (-0.96) (-0.24) (-0.28) (-0.29)
morelink act 0.197

0.200

0.031 -0.111 -0.159

0.224

0.261

0.404

0.389

(2.80) (2.89) (0.46) (-1.48) (-2.80) (6.05) (8.67) (4.91) (5.06)
. . . 1(second half) 0.020 0.005 0.196

0.313

0.299

0.085 0.044 0.065 0.104
(0.21) (0.04) (3.17) (5.45) (4.66) (1.25) (0.60) (0.51) (0.84)
mpay 0.000 0.020

0.018

0.005

0.004

(0.28) (9.95) (9.21) (2.49) (2.43)
. . . 1(second half) 0.003

-0.006 -0.006 -0.004 -0.001
(2.41) (-1.59) (-2.22) (-1.08) (-0.41)
turn 0.007

0.010

0.010

0.012

0.003
(8.07) (3.56) (3.53) (4.02) (1.39)
. . . 1(second half) -0.002 0.005 0.005 0.000 0.007
(-1.25) (0.87) (0.97) (0.12) (2.32)
chdist(E. Line) -0.015 -0.104

-0.033 -0.027
(-1.00) (-4.57) (-1.65) (-1.77)
. . . 1(second half) -0.020 -0.001 -0.032 -0.002
(-0.69) (-0.08) (-0.82) (-0.12)
chdist(PS) -0.100

(-6.83)
. . . 1(second half) -0.005
(-0.34)
Constant 0.875

0.799

0.754

0.391

0.440

0.791

0.649

0.864

0.779

(39.30) (28.32) (57.19) (17.64) (20.91) (56.57) (16.69) (33.78) (15.19)
Individual xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes
P(pooling) 0.300 0.127 0.042 0.000 0.001 0.148 0.260 0.120 0.021
N 3276 3276 3366 3366 3366 3162 3162 3072 3072
Adj. R
2
0.160 0.173 0.103 0.301 0.304 0.219 0.259 0.243 0.259
t statistics in parentheses. Standard errors are clustered at the session level.

p< 0:05,

p< 0:01,

p< 0:001
69
Appendix 2.B Experimental instructions
Welcome. This is an experiment on individual decision making in groups, and you
will be paid for your participation in cash at the end of the experiment. The entire
experiment will take place through computer terminals, and all interactions between
participants will take place through the computers. You will remain anonymous to
me and to all the other participants during the entire experiment; the only person
who will know your identity is the Lab Manager who is responsible for paying you
in the end. Moreover, it is important that you do not talk or in any way try to
communicate with other participants during the experiment.
We will start with a brief instruction period. During the instruction period, you
will be given a complete description of the experiment and will be shown how to use
the computers. You must take a quiz after the instruction period, so it is important
that you listen carefully. If you have any questions during the instruction period, raise
your hand and your question will be answered so everyone can hear. If any diculties
arise after the experiment has begun, raise your hand, and an experimenter will come
and assist you. Please note that you are not being deceived and you will not be
deceived: everything I tell you is true.
Your earnings during the experiment are denominated in tokens. Depending on
your decisions, you can earn more tokens or lose some tokens. At the end of the
experiment, we will count the number of tokens you have earned in all of the matches
and you will receive $1.00 for every 4 tokens. You will be paid this amount plus
the show-up fee of $5. Dierent participants may earn dierent amounts. Everyone
will be paid in private and you are under no obligation to tell others how much you
earned.
The experiment will consist of 8 matches. In each match, you will be put in a
70
group with 5 other participants in the experiment. Since there are 12 participants in
today's session, there will be 2 groups in each match. You are not told the identity
of the participants in your group. Your payo in each match depends only on your
decisions, the decisions of the other 5 participants in your group and on chance. What
happens in the other group has no eect on your payo and vice versa. Your decisions
are not revealed to participants in the other group.
We will now explain how each match proceeds. At the beginning of the match,
the computer randomly assigns each of you to a group consisting of 6 participants.
Next, the computer randomly assigns with equal probability a role to each of the
participants as \Subject 1", \Subject 2" and so on up to \Subject 6". Then, the
match begins.
Each match consists of several turns. At the beginning of each turn, the computer
randomly pairs all subjects within each group with one another. We shall call the
subject that you are paired with at each turn as your \Current Partner". Once
everyone receives a Current Partner, a turn begins.
71
At the beginning of each turn, you will see a screen similar to that shown here.
The top panel provides the information and interface that you will use to interact
with other subjects within your group. Meanwhile, the bottom panel lists your payo
history throughout the experiment. Payo information in each match, including the
practice matches, is recorded here.
This is the top panel. On the top-left is your role in this match. In this example,
you are Subject 1. The computer also informs you of your Current Partner at each
turn. In this turn, your Current Partner is Subject 2.
In the middle of the left panel, you will see a network representation of the connec-
tions between all subjects in your group. Other subjects in your group are represented
by nodes with their role ID numbers. Meanwhile, you are always represented by the
center node labeled \YOU". In each turn, the node for your Current Partner is col-
ored YELLOW unlike the rest of the subjects. From the color, you can see here that
your Current Partner is Subject 2.
The lines connecting the nodes represent the links between subjects in your group.
72
Everyone in your group sees the same sets of links. In this example, you have direct
links to Subjects 5 and 6. Through Subject 6, your are also indirectly connected with
Subject 4. Subjects who are either directly or indirectly connected belong in the same
\Set". In this example, there are two sets. The rst consists of You, Subjects 4, 5,
and 6. The second set consists of Subjects 2 and 3.
At each turn, the joint actions of you and your current partner aect how the two
of you are linked. You take actions by clicking one of the action buttons below the
network representation. Through your actions, you can either propose a link, remove
a link, or maintain how you are connected with your partner.
In this rst example, since you are not linked to Subject 2, only three actions are
available: \Propose", \Pass Turn", and \Network OK". The \Remove" button is not
active. Clicking \Propose" lets the computer know that you would like to propose
a link with your Current Partner. If your partner does the same, the computer will
create a link between you and your partner. Otherwise, no link will be created. In
other words, a link is created if and only if BOTH partners propose a link to each
other.
If you don’t want to link with your Current Partner, you can either click \Pass
Turn" or \Network OK". In either case, a link will not be created. However, notice
the dierence between the two actions. When you pass a turn, you tell the computer
that you want to keep the way you are linked with your current partner in this turn.
However, you may still want to change how you are linked with some of the other
subjects. So, your buttons will remain active in the next turn
Meanwhile, if you choose \Network OK", you tell the computer that as long as
the network doesn’t change, you are happy with the way you are linked with everyone
in your group. Therefore, if you click \Network OK", you won’t need to take further
actions until the network changes. Your buttons will therefore be inactive. However,
73
these buttons will immediately become active once the decisions of other pairs either
break or make a link. If all active subjects choose \Network OK" in the same turn,
then the match ends.
The turn ends once everyone in your group has taken an action. The computer
then begins a new turn, and you will be randomly assigned a new Current Partner.
Please note that since pairs are selected randomly, you may be paired with the same
partner in consecutive turns.
This gure illustrates a new turn in which you are paired with Subject 6. Now,
since you are already directly linked with this subject, the three actions available
to you are: \Remove", \Pass Turn" and \Network OK". The \Propose" button is
deactivated in this turn.
Your link with Subject 6 will remain intact only if BOTH you and Subject 6 don’t
want to remove it. If at least one subject in the pair wants to remove it, your direct
link with your Current Partner will be broken at the end of the the turn. Obviously,
the link will also be broken if both subjects in a pair choose to remove it.
74
In each match, the computer will continue to generate new turns for at least
12 turns unless all subjects choose \Network OK". However, if a match does not
end after 12 turns, the match enters the random-end stage. In the random-end
stage, at each turn, the computer randomly decides whether it will end the match
or generate a new turn. Each time, there is a 20% probability that it will decide
to end the match. On average, this implies about 5 additional turns in each match.
The number of remaining turns before this random-end stage is displayed above the
network representation.
The network representation updates links that are made and broken in real time.
You can see changes to the network immediately after each pair makes their decisions
within each turn. Similarly, you can also keep track of changes within each turn
through the \Status" indicator on the lower right panel. This status indicator resets
at each new turn.
We will next discuss about the payo. Your payo depends on the size of your set
and the number of direct links at the end of the match. Your set size, which is the
number of subjects who are either directly or indirectly connected to you, determines
your revenue. Meanwhile, your cost is determined by the number of direct links you
have.
The right panel provides you with all of the information necessary to calculate
your payo. The table on the left gives you the revenue schedule for dierent set
sizes. Above it, you can see the list of subjects in your set. In this example, your set
consists of You and Subjects 4, 5, and 6. Therefore, as part of a set of size 4, your
revenue is 35.
Next to the revenue table is the cost schedule for dierent numbers of direct links.
Each direct link incurs a constant cost. In this particular example, the cost for each
link is 10 and, therefore, the total cost is 10 times the number of subjects with whom
75
you are directly linked. Above that table, you can see that you are directly linked to
Subjects 5 and 6. Since you have two direct links, the current total cost is 20 tokens.
Your current revenue and cost at any stage of the game are highlighted in YEL-
LOW. They are updated in real time as the actions of subjects make and break links
within each turn. The rightmost box entitled \Current Payo" calculates your payo
at each stage of the game. The current payo is simply the revenue minus cost, which
in this case is 15. This payo information is also updated in real time. Note that the
revenue and cost tables may change from match to match.
This gure illustrates what you will see at the end of a match. Below the status
indicator, you will see your payo for this match. At the end of the match, please click
\Continue to the Next Match". In each new match, you will be randomly assigned
to a new group. A new match will begin only after all groups have completed their
matches. This continues for 8 matches, after which the experiment ends.
At the end of the nal match in the experiment, you will see the following screen.
76
This nal screen tells you the total payo that you will receive for this experiment.
When you see this screen, don’t click OK until you have written down your total payo
on the payo sheet provided. After you have written down your total payo, click
OK to conclude the session. (*)
The following slides summarize the rules of the experiment:
77
We will now begin the Practice session and go through two practice matches to
familiarize you with the computer interface and the procedures. During these practice
matches, please do not hit any keys until you are asked to. Remember, you are not
78
paid for these matches. At the end of the practice matches you will have to answer
some review questions.
Throughout the session, pay attention to the network representation display and
status indicators. Also, notice the movements of the yellow highlights on your Revenue
and Cost tables, as well as updates to your Current Payo.
[START GAME]
You have just received a new turn. First, pay attention to your role. If you are
Subject 1, 2, or 3, please click \Propose". For Subject 1, 2, or 3, notice a link has
just been created between you and your partner if your partner is also Subject 1, 2
or 3.
Now, if you are Subject 4, 5, or 6, please click the \Pass Turn" button. Notice
here that a link is created if and only if BOTH partners propose a link. If only one
partner proposes a link, no link is created.
You have moved to a new turn. We will now see how the \Network OK" action
works. If you are either Subject 5 or 6, please click \Network OK". For the rest of
the group, please click \Pass Turn".
You have moved to a new turn. For Subjects 5 or 6, since the network has not
changed after you clicked \Network OK", all of your buttons are now inactive. Notice
that they will become active following a change in the network.
For others, please check your Current Partner. If your partner is not Subject 5
or 6, click the \Remove" button if it’s active or \Propose" otherwise. For Subjects 5
and 6, notice how a change in the network activates your buttons.
If you are not Subject 5 or 6 and your buttons are still active, please click \Pass
Turn". If you are Subject 5 or 6 and your buttons are active, please click \Pass
Turn". Notice here that if your buttons are inactive due to a \Network OK" action
79
in a previous turn, a change in the network will immediately activate your buttons.
In the following, we will do the same exercise for Subjects 1 to 4.
You have moved to a new turn. If you are Subject 3 or 4, please click \Network
OK". For the rest of the group, please click \Pass Turn".
You have moved to a new turn. Subjects 3 and 4, notice that your buttons are
inactive. If the network changes in this turn, your buttons will become activated.
For all others, check your Current Partner. If your partner is not Subject 3 or 4,
click \Remove" if it’s active or click \Propose" otherwise. If you are not Subject 3
or 4 and your buttons are still active, click \Pass Turn". Now, if you are Subject 3
or 4, please click \Pass Turn".
You have moved to a new turn. If you are either Subject number 1 or 2, please
click \Network OK". For the rest, please click \Pass Turn".
You have moved to a new turn. For Subject 1 or 2, your buttons are now inactive.
For all others, if your Current Partner is not Subject 1 or 2, click the \Remove"
button if it’s active, or click \Propose" otherwise. For everyone else who has not
taken an action, please click \Pass Turn".
You have moved to a new turn. Notice from the message above the network display
that this is the last turn before the random-end stage. During the paid match, you
will have 12 turns before entering this stage. If the match has not ended after 12
turns, the computer will randomly decide the end of the match.
We will now deliberately end the match. If your buttons are active, please click the
\Network OK" button. This ends the rst practice match. The bottom part of your
screen contains a table summarizing the results for all matches you have participated
in. This is called the history screen. It will be lled out as the experiment proceeds.
80
Now click \Continue to the Next Match". We will now begin with the second practice
match.
[NEXT MATCH]
You are in a new match. Note here that the revenue and cost tables have changed as
they may during the real matches. We’ll now examine the behavior of the \Remove"
action.
If you are either Subject 2, 4, or 6, please click \Remove". For Subjects 1, 3, and
5, please click \Pass Turn". Hence, notice that a link is broken if at least one of the
partners chooses to remove it.
You have moved to a new turn. Next, we’ll see what will happen if the network
changes within the turn in which you click \Network OK". If you are Subject number
1, 3, or 5, please click the \Network OK" button. For all others, please click your
\Remove" button.
You have moved to a new turn. For Subjects 1, 3, or 5 notice that if in the previous
turn the network changed after you clicked \Network OK", your action buttons are
active in this turn. If the network did not change after you clicked \Network OK",
your buttons remain inactive. Now, if you are either Subject 2, 4, or 6, click \Network
OK". For all others, if you haven’t taken an action in this turn, please click the
\Remove" button if it’s active, or \Propose" otherwise.
You have moved to a new turn. Similarly for Subjects 2, 4, and 6, notice that
if in the previous turn the network changed after you clicked \Network OK", your
buttons are now active. If the network did not change after you clicked \Network
OK", your buttons are still inactive. If the network changes in the same turn and
after you choose \Network OK", your buttons stay active in the following turn.
81
We will now end the match. If your buttons are active, please click \Network
OK". This ends the second practice match.
*** END OF PRACTICE SESSION ***
The practice matches are over. Please click \Continue to the next match" and
complete the quiz. It has 8 questions in two pages. You will move to the next page
once everyone in your group has completed the questions in that page correctly. On
your table, you will nd the screenshots that you will need to answer these questions.
Raise your hand if you have any questions.
[WAIT for everyone to nish the quiz]
Are there any questions before we begin with the paid session? We will now begin
with the 8 paid matches. Please pull out your dividers. If there are any problems
or questions from this point on, raise your hand and an experimenter will come and
assist you.
[START MATCH 1]
[After MATCH 8, read:]
This was the last match of the experiment. Now, please write down your ID on
the payment sheet. Your ID is located on top of your physical monitor and it began
with CASSEL. At this point, if you haven’t clicked \Continue to the next match",
please do so. Your total payo is displayed on your screen. Please record this payo
in the earned column of your sheet and sign it. Once you have written it down, please
click OK.
82
Your Total Payo will be this amount rounded up to the nearest dollar plus the
show-up fee of $5. We will pay each of you in private in the next room. Remember
you are under no obligation to reveal your earnings to the other subjects.
If you are done, please line up behind the yellow line until the lab manager calls
you to be paid. Do not converse with the other subjects or use your cell phone.
Thank you for your cooperation.
83
Chapter 3
Religious Identities and Cooperative Attitudes in
Indonesia: Evidence from IFLS
Abstract
I investigate how variations in individuals' religion and religiosity are linked
to trust and religious tolerance in contemporary Indonesia using the Indone-
sia Family Life Survey (IFLS). I nd that religiosity is positively associated
with particularized trust and in-group preference, and negatively with re-
ligious tolerance. The strengths of the associations between measures of
in-group preference (including political preference) and individual religiosity
are much stronger than those from gender, education, or per-capita expen-
diture; they are also strongest among Muslims, the dominant majority in
Indonesia. These associations are robust to various identication strategies.
Using selection on observables to benchmark the potential bias from selec-
tion on unobservables, I nd that the selection on unobservables needs to be
multiple times that on observables to explain away these results.
This chapter studies how individual religious identities are associated with the will-
ingness to help others, trust and religious tolerance in the ethnically- and religiously-
diverse country of Indonesia. Ample evidence from other contexts suggests positive
associations between generalized trust and economic development, often through the
link between trust and institutional quality.
1
Historical evidence also supports the
1
Positive associations between generalized trust and institutional quality have been shown in case
studies (Putnam et al., 1993) as well as quantitative analyses using cross-country data (Knack and
Keefer, 1997; La Porta et al., 1997) and household-level data (Narayan and Pritchett, 1999; Maluccio
et al., 2000; Carter and Castillo, 2011). Furthermore, Carter and Castillo (2011) provided empirical
evidence of the important role of altruistic sharing norms in improving household well-being in South
African communities.
84
idea that religious intolerance can impede development by limiting innovations.
2
Nonetheless, very few national-level studies examine the determinants of trust and
tolerance in developing countries, in part due to the lack of data. This chapter is an
attempt to address this gap. I use the new religion, trust, and religious tolerance
modules of the latest Indonesian Family Life Survey (IFLS) to examine individual
correlates of trust and religious tolerance in contemporary Indonesia. These modules
also allow for a study of how religious identities are linked to attitudes toward intra-
and inter-group cooperation. There are at least two contributions of this chapter.
First, it oers a rare country-level examination of the links between variations in
religiosity and religion, and cooperative attitudes for the dierent world religions
in a developing country.
3
Guiso et al. (2003) examine a similar question on inter-
religion dierences in attitudes, albeit using cross-country regression with its well-
known econometric issues, such as omitted variables and the crudeness of its measures.
The richness of the dataset used in this paper helps address these problems to a great
extent. Additionally, this chapter also oers a careful examination of the individual
correlates of intra- and inter-group trust and tolerance in Indonesia.
Indonesia's religious diversity provides a perfect context for this investigation. In-
donesia is a non-secular state in which the major world religions, i.e., Islam, Catholi-
cism, Protestantism, Hinduism, Buddhism and Confucianism, are represented.
4
Even
though Muslims are the dominant majority, accounting for 87.2% of the population
2
Landes (1998), for instance, argued that religious intolerance was responsible for scientic
regress in many (Catholic) European countries and Chaney (2008) has made a similar claim regarding
Islam in medieval Muslim societies.
3
A previous study on Indonesia asking a similar question is by Mujani (2004), who examined
correlates of trust and tolerance among Indonesian Muslims. However, in addition to its limited
focus on Muslims, his data have a much smaller sample size, fewer individual- and household-level
variables, and none of the community-level variables.
4
I use the term \non-secular state" because, even though the state does not adhere to any
particular religion, the rst principle in its ideology is \[Belief] in the one and only God" { with a
fairly loose interpretation of the term \the one and only God".
85
in 2010, there are considerable shares of believers of other religions, such as Protes-
tants (7.0%), Catholics (2.9%), Hindus (1.7%), Buddhists (0.7%), and Confucianists
(0.05%).
The chapter is organized as follows. The next section provides a review of the
literature linking religion and dierent types of cooperative attitudes. Section 3.2
discusses the data and measurements used for the analysis. Section 3.3 describes
the empirical strategy and the results on individuals' religion/religiosity variables.
Section 3.4 concludes.
3.1 Literature review
Does religion facilitate attitudes conducive to social cooperation? First, consider al-
truism. One of the common denominators across all religions is the emphasis on
benevolence (Neusner and Chilton, 2005). Intuitively, it would be reasonable to
expect more religious people to be more altruistic. Existing evidence, however, is
mixed. Sociological surveys based on self reports often provide evidence that people
who attend religious services and pray more are more likely to contribute to char-
ity. Social psychology studies, however, question some of these ndings. Batson et
al. (1993) compared between studies that used self-reports measures and those using
behavioral ones to examine the link between helpfulness (or altruism) and religious
involvement. They found that the positive associations often found using the former
measures disappeared when behavioral measures were used. Using economic exper-
iments, Anderson et al. (2010) did not nd religious involvement to be a signicant
predictor of contributions in public goods games. Further evidence suggests that the
positive ndings based on self-reports may have been driven by stronger reputational
concerns, instead of actual willingness to help, among the religious (Batson et al.,
86
1993; Norenzayan and Shari, 2008).
Meanwhile, religious teachings also put a lot of emphasis on trustworthiness. In
Islam, it is captured in the notion of \amanah" { which is to render trust to whom
it was due (an-Nisa, 4:58) { and in Christianity, in the notion of \stewardship",
illustrated among others in the parable of the talents (Matthew 25:14-30; Luke 19:12-
28).
5
However, they do not seem to advocate unconditional (generalized) trust.
6
This
distinction between trustworthiness and trusting behaviors may explain why, as we
shall see below, the overall evidence on the link between religion and trust has been
mixed.
Analyses of observational data provides the evidence for the link between reli-
giosity and trust. Using the generalized trust question from the World Value Survey
(WVS) data for 66 countries, Guiso et al. (2003) found that religious people trust
others more than the non-religious (although not compared to atheists). Among the
religious, trust toward others is positively correlated with current religious participa-
tion, but not by whether a person is brought up religiously. Using a similar question
on generalized trust, Mujani (2004) found that participation in the various Islamic
rituals was positively correlated with interpersonal trust in Indonesia.
However, the evidence from economic experiments is more mixed. Using the stan-
dard experimental trust game, Anderson et al. (2010) did not nd a link between the
intensity of religious participation and trust toward anonymous partners. However,
information about the partner's religious norms appears to in
uence trust. When the
5
The parable of the talents tells a story of how a master who, coming from a journey, dierentially
rewarded servants who made productive use of the possessions that he entrusted them and punished
the one who did not.
6
In Islam, the Qur'an (al-Hujurat, 49:12) advises Muslims not to have unfounded suspicions
toward each other. A similar advice can be found in the Judeo-Christian tradition { \you shall not
hate your brother in your heart" (Leviticus 19:17) and \Do not act vengefully or bear a grudge
against a member of your nation" (Leviticus 19:19). In both cases, however, such trust is extended
primarily to members of the in-group, and not to strangers (Levy and Razin, 2012).
87
same game is implemented among (mainly Judeo-Christian) German subjects, Tan
and Vogel (2008) nd that information about the otherwise anonymous partner's re-
ligiosity aects behavior. The religious are trusted more, particularly by the religious
others. Moreover, the religious trustees are also more trustworthy. The importance
of information on partner's religion (or ideology) is echoed in studies using a dierent
experimental game between kibbutzim and non-kibbutzim members in both religious
and secular kibbutzims (Sosis and Rue, 2004; Rue and Sosis, 2006). Sosis and
Rue (2004) nd that members of religious kibbutzims in Israel are more willing to
cooperate when anonymously paired with a member of the kibbutzim than with a
city resident.
The one relationship in which both observational and experimental evidence align
is that between religiosity and inter-group tolerance. Results based on observational
as well as behavioral evidence since Allport and Kramer (1946) rst found the positive
association between religious aliation and racial prejudice are strongly in favor of
nding a positive link between religiosity and intolerance (Batson et al., 1993; Hall et
al., 2010; Guiso et al., 2003). More recently, experimental evidence using priming of
religious concepts provide further evidence that when one's religious identity is made
salient, there is greater intolerance toward members of the out-group { both in terms
of religion and ethnicity (McCauley, 2009; Johnson et al., 2010; Parra, 2011).
7
Are there inter-religion dierences in cooperative behavior? The role of (multiple)
interpretations and institutions on dierent religions makes it very dicult ex ante to
predict these attitude dierences. Instead, we turn to the empirical literature to look
for empirical regularities. Benjamin et al. (2010) used priming to examine the impact
of the salience of religious identities among Catholics, Protestants, Jews, and non-
7
Moreover, McCauley (2009) also found that the eects of salient religious identities on inter-
group discrimination are stronger than those of tribal ones.
88
believers. After receiving religious priming, subjects were asked to play experimental
games to measure their contributions to the public goods and dictator games. Among
Catholics, religious priming decreased public good contributions and expectations of
other's contributions, while among Protestants, it increased contributions. However,
religious identity did not aect generosity in the dictator game.
With respect to trust, the cross-country analysis of observational data by Guiso
et al. (2003) found that participation in religious services increases trust only among
Christians. Among the Christian denominations, Putnam et al. (1993) has argued
that because of its hierarchical structure, Catholicism tend to breed less interpersonal
trust than Protestantism. Observational analyses using cross-country data found sup-
port for this conjecture, although this dierence was smaller among younger Chris-
tians (La Porta et al., 1997; Guiso et al., 2003).
8
However, such a dierence is not
found in the analysis using United States data (Alesina and La Ferrara, 2002).
Meanwhile, the link between religion and intolerance are present across all religious
denominations, with a notable except of Buddhists, who are on average more tolerant
than non-religious people. The least intolerant toward immigrants and other races
were Hindus and Muslims, followed by Jews, Catholics and Protestants (Guiso et al.,
2003).
3.2 Data and measurements
3.2.1 Data
The analysis below uses the Fourth Wave of the Indonesia Family Life Survey (IFLS).
IFLS is a longitudinal, socio-economic household survey and its fourth wave, IFLS4,
8
Guiso et al. (2003) show that Catholics born after the Second Vatican Council are more trusting
and tolerant than their older cohorts, even though their moral values did not signicantly dier from
older Catholics.
89
contains a set of questions about religion, trust, and religious tolerance. The survey
also collects a rich set of information on all members of each sample household, as well
as the communities that they live in, and the facilities that are available to them. For
this cross-sectional analysis, I use IFLS4's adult sample, which comprises household
members who are 15 years or older. A total of 29,054 adults in 12,692 households
were interviewed for the religion, trust, and tolerance modules in IFLS4. Moreover,
as elaborated below, I use the panel structure of IFLS to construct the educational
history of these sampled individuals.
3.2.2 Measures
Measures of cooperative attitudes
IFLS4 contains multiple questions that measure dierent aspects of attitudes toward
cooperation in the community. This proves very useful for the analysis. For trust,
the multiple questions allow for the distinction between generalized and particular-
ized trust, and between trusting behaviors and beliefs. I will argue below why the
distinctions matter. In addition, it also has separate measures of religious tolerance.
In all of these attitude questions, answers to the questions are on a four-point scale.
Regarding trust, one criticism of many existing survey-based studies of trust is
inadequate specicity, given their heavy reliance on a single question on the gener-
alized trust (Nannestad, 2008).
9
IFLS4 addresses this criticism to a great extent.
There are seven questions in IFLS4 on individual measures of trust attitudes. These
questions allow a distinction of the trust concept in two dimensions: Between beliefs
and behaviors; and between generalized and particularized trust.
9
The question that is often the based of such studies is the one used in the American General
Social Surveys, to wit, \Generally speaking, would you say that most people can be trusted or that
you can't be too careful in dealing with people?" or its variations.
90
First, the distinction between beliefs and behaviors. As a behavior, trust is the
willingness to place one's resources in the hands of another party without any legal
commitment from the latter (Fehr, 2009). A rational principal will trust an agent if
the expected payo from that action exceeds the alternative action, which is not to
trust. This expected payo depends on both the payo that the principal will receive
from that transaction and his beliefs regarding the trustworthiness of the agent.
To illustrate, consider a simple version of extensive-form trust game (following
Berg et al. (1995)) between a principal and an agent depicted in Figure 3.1. A
principal chooses whether to entrust his resources,p, to an agent in his community or
otherwise receives nothing. If the principal chooses to trust, the agent must choose
between behaving honestly or dishonestly. If the agent behaves honestly, the principal
will receive a return of P > 0 and the agent, a; otherwise, he will lose all of his
investment and the agent will receive A that depends on her type. There are two
types of agents: the high type (H) and the low type (L). We assume that high-type's
payo from taking dishonest actions (A
H
) are lower than that of the low types (A
L
).
10
To simplify, assume that this payo dierence is signicant enough that the high types
are always honest, and the low types are always dishonest. In the community, the
share of the high types is. The principal does not know the agent's type and instead
only knows the share of the dierent types of agents.
The principal's decision to trust depends on the stake (p), potential payos (P )
and his beliefs regarding the value of , to wit, the trustworthiness of the agent
population.
11
Most surveys on trust typically ask respondents to rate the statement
10
This may be due dierences in individual norms or the quality of the dierent institutions
between groups.
11
An natural extension to this model is to incorporate the quality of local institutions to punish
breaches of trust. We can incorporate this notion in the game depicted in Figure 3.1 by adding a
branch following the agent's decision to behave dishonestly. In this branch, with some probability
(that depends on institutional quality), \Nature" would nd out and punish the dishonest behavior.
This probability of capture will enter into the principal's optimization problem and in
uence her
91
P
A
H
A
L
Trust
(T)
Not trust
(NT)
(0; 0)
Dishonest
(D)
Honest
(H)
(0; 0)
T
NT
D
H
(p;A
H
(:))
(P;a)
(p;A
L
(:))
(P;a)
P;p> 0
A
L
>a>A
H
Nature

1
Figure 3.1: Trust game
\In general, one can trust people", which can plausibly be interpreted as a measure of
other people's trustworthiness { a proxy for .
12
However, in some cases, we may be
more interested in the trusting behaviors rather than beliefs and Glaeser et al. (2000)
suggests that the usual trust question is correlated with the latter, not the former.
13
In addition to the distinction between behaviors and beliefs, there is also the
distinction between particularized and generalized trust. The former refers to a more
narrow type of trust, namely that toward similar others (in terms of gender, race,
ethnicity, and so on) while the latter refers to a broad type of trust that is, for
instance, embodied in an armation to the statement that \most people can be
decision to trust.
12
Fehr (2009), however, questioned this interpretation. He found that individuals' preference
parameters are associated with their responses to this trust question in the German Socio-Economic
Panel (SOEP) data, suggesting that individuals may introspect on their own behaviors when an-
swering this question.
13
The ndings of Glaeser et al. (2000), however, may not be general across societies. See Nannes-
tad (2008) for a review.
92
trusted". Results from the literature suggests that it is the latter type of trust, and
not the former, that is positively associated with social, economic, and governance
outcomes (Putnam et al., 1993; Glaeser et al., 2000; Uslaner and Conley, 2003; Guiso
et al., 2011).
With multiple questions on trust in IFLS4, I can disentangle some of these aspects.
First, on behaviors that correspond to the particularized trust of a known neighbor,
respondents were asked to rate on a four-point Likert-type scale { from \strongly
disagree" to \strongly agree" { the following statements:
(i) \I would be willing to leave my children with my neighbors for a few hours if I
cannot bring my children with along";
(ii) \I would be willing to ask my neighbors to look after my house if I leave for a
few days".
Second, respondents were also asked questions about their beliefs regarding the
trustworthiness of dierent types of an anonymous other. They were asked to imagine
a scenario where they lost a wallet or a purse containing Rp. 200,000 (approximately
US$20, almost half of the average monthly per-capita expenditure of the IFLS4 re-
spondents) along with an identity card. They were then asked to assess how likely
they would get the wallet back with the money intact if it were found by: (i) someone
who lives close by; (ii) a stranger; and (iii) a policeman. Respondents can respond on
a 4-scale measure from \very unlikely" to \very likely". Responses to (i) and (ii) can
be interpreted as particularized and generalized trust beliefs respectively. Meanwhile,
responses to (iii) can be interpreted as trust beliefs of the authorities.
Finally, IFLS4 allows for further distinction of particularized trust with regards
to religion and ethnicity. Respondents were asked to rate on a four-point Likert-
type scale the following statements about trust of people of the same ethnicity and
93
religion: \Taking into account the diversity of ethnicities (religions) in the village, I
trust people with the same ethnicity (religion) as mine more". The interpretation of
these questions along the belief-behavior dimension is somewhat ambiguous. At any
rate, I take them as measures of in-group (or discriminative) trust.
Meanwhile, to measure religious tolerance, I use a set of questions regarding re-
spondent attitudes toward non-coreligionists. In particular, IFLS4 asked whether
respondents object to having non-coreligionists live in their village, neighborhood,
or house. It also asked whether respondents would object if a relative was going to
marry a non-coreligionist and if people of a dierent religion were to build a house of
worship. In all these questions, respondents can respond on a 4-scale measure, from
\no objection at all" to \not acceptable".
Finally, as a proxy of altruism, I use the responses to the following statement: \I
am willing to help people in this village if they need it". In addition, respondents
were asked to assess how safe their villages were; and how safe it was to walk around
at night. Overall, this set of questions, in combination with particularized trust
questions above, indicate the extent to which respondents nd their communities to
be cohesive.
Table 3.1 presents the summary statistics for these outcome variables. In gen-
eral, respondents report a high level of willingness to help and trust their neighbors.
They are also more willing to entrust their properties than their children to their
neighbors. With regards to their beliefs of the trustworthiness of others, respon-
dents believe neighbors and, to a lesser extent, the police are trustworthy. However,
their generalized trust belief { often seen as the type of trust that matters most in
facilitating economic outcomes { is much lower.
Meanwhile, based on the averages of the tolerance measures, we can rank the
94
Table 3.1: Summary Statistics: Willingess to Help, Trust and Tolerance
Num.
of obs.
Mean
Std.
dev.
Median IQR Min Max
Community cohesion
Willingness to help 29037 3.15 0.38 3 0 1 4
Village is [. . . ]
generally safe 29034 3.07 0.37 3 0 1 4
safe at night 29032 2.99 0.38 3 0 1 4
Trust neighbor to watch [. . . ]
children 21842 2.68 0.57 3 1 1 4
house 29035 2.87 0.46 3 0 1 4
Trust beliefs
Trust [. . . ] to return lost wallet
neighbors 28425 3.03 0.94 3 1 1 4
strangers 27498 1.52 0.78 1 1 1 4
police 26916 2.81 0.99 3 2 1 4
Trust [. . . ] more
coreligionist 29036 2.80 0.58 3 1 1 4
coethnic 29036 2.65 0.58 3 1 1 4
Tolerance
Tolerate non-coreligionist to live in [. . . ]
village. 29037 2.80 0.54 3 0 1 4
neighborhood. 29037 2.75 0.58 3 0 1 4
house. 29035 2.43 0.73 3 1 1 4
Tolerate non-coreligionist to [. . . ]
marry a relative. 29035 1.77 0.81 2 1 1 4
build house of worship. 29035 2.26 0.79 2 1 1 4
95
issues captured by these measures from the most to the least contentious. Interfaith
marriage is the most contentious, followed by the issue of allowing non-coreligionists
to build a place of worship. Relative to these two issues, respondents are much more
tolerant about allowing non-coreligionists live in the same village or neighborhood,
but not so much in the same house.
Religion, religion-based education, and the religiosity measure
Our analysis focuses on examining how religion and religious intensity correlate with
social and civic capital in Indonesia. In IFLS4, each respondent was asked about his
or her religion and can choose between Islam, Catholicism, Protestantism, Hindu,
Buddhism, and Confucianism. Our analysis focuses on the rst ve of these religions
since there are only two Confucians in the sample.
14
Each respondent was also asked
to evaluate his or her own religiosity out of a 4-scale measure { \not religious",
\somewhat religious", \religious" and \very religious". These two variables and their
interactions will be our main regressors of interest. Table 3.2 presents the distribution
of religiosity overall and for each religion.
Table 3.2: Distribution of Religiosity
Degree of religiosity
Num.
of obs. Not
religious
Somewhat
religious
Religious
Very
religious
All religions 0.03 0.19 0.73 0.06 28973
Islam 0.03 0.19 0.73 0.05 25890
Catholic 0.03 0.15 0.72 0.09 447
Protestant 0.02 0.15 0.76 0.07 1157
Hindu 0.01 0.05 0.77 0.17 1392
Buddhist 0 0.21 0.70 0.09 87
14
To focus on the ve main world religions, 17 observations were dropped either because they
refused to answer (10 observations), listed \other" as their religion (5 observations) or are Confucians
(2 observations).
96
The religiosity question in IFLS is a self-assessment question; it is therefore useful
to examine how answers to these questions relate to observed behavior. For adherents
of each religion, IFLS4 asked a pair of questions on an individual's religious practices.
Muslims were asked how many times they prayed every day and whether they observed
the halal food requirement. Christians were asked how often did they pray or read the
bible and whether they actively participated in activities such as religious fellowships.
Meanwhile Buddhists and Hindus were asked whether they meditated in the temple
and whether they observed certain religion-related diets. I use these data to validate
respondents' self-assessments of their religiosity.
Table 3.3 presents the share of individuals that follow a particular religious practice
for a given level of religiosity and for each religion. The pattern suggests strong
correlations between self-assessment of one's religiosity and his or her adherence to
religious practices across dierent religions. For Muslims, the more religious a person,
the more likely that he or she follows (and go beyond) the mandatory number of
prayers of ve times a day. However, there does not seem to be much variation with
respect to keeping the halal diet across dierent religious intensities, except among the
non-religious muslims. Similarly among Christians, the more religious tend to pray
more frequently during the day. In addition, they are also more likely to participate
actively in religious activities such as prayer fellowships. Meanwhile, more religious
Hindus are more likely to frequent temples daily, and are more likely to maintain
follow the no beef/red meat dietary restrictions. Similarly, more religious Buddhists
are more likely to pray in the temple daily and be a vegetarian.
To further validate this measure, I also consider a question from IFLS's community
participation module { which is a module that is separate from the religion module.
In the community participation module, respondents were asked whether they knew
of a particular activity in the village, and if they do so, whether they participated.
97
Table 3.3: Share of Practicing Individuals for a Given Religiosity
Not
religious
Somewhat
religious
Religious
Very
religious
Refused to
answer
Muslim
How many times do you pray each day?
[
2
(9; 25856) = 8:9e + 03;p = 0:00]

Do not practice 0.66 0.25 0.04 0.01 0.19
Between 0 and 5 0.25 0.43 0.11 0.09 0.07
5 times 0.08 0.29 0.73 0.65 0.47
More than 5 0.01 0.02 0.11 0.25 0.07
Refused to answer 0.00 0.00 0.00 0.00 0.21
Do you eat halal food?
[
2
(3; 25856) = 140:4;p = 0:00]

Yes 0.91 0.96 0.98 0.98 0.95
Num. of obs. 712 5034 18793 1352 58
Christian
How often do you pray/read the bible?
[
2
(12; 1601) = 319:9;p = 0:00]

Do not practice 0.27 0.02 0.01 0.01 0.00
Sometimes 0.41 0.31 0.12 0.05 0.00
Morning and evening 0.10 0.20 0.08 0.05 0.00
Once a day 0.15 0.17 0.28 0.23 0.50
Before each activities 0.07 0.29 0.51 0.67 0.50
Refused to answer 0.00 0.00 0.00 0.00 0.00
Do you actively participate in religious activities?
[
2
(3; 1601) = 151:8;p = 0:00]

Yes 0.27 0.62 0.85 0.91 0.5
Num. of obs. 41 244 1205 120 2
Hindu
Do you practice meditation in the temple?
[
2
(9; 1392) = 118:1;p = 0:00]

Do not practice 0.38 0.04 0.01 0.00 0.00
On holy days 0.25 0.41 0.28 0.19 0.33
During the full moon 0.38 0.17 0.25 0.20 0.67
Every day 0.00 0.38 0.46 0.61 0.00
Do you have religious-related dietary restrictions?
[
2
(9; 1392) = 27:1;p = 0:00]

No dietary restrictions 0.75 0.80 0.70 0.61 0.67
Some dietary restriction 0.13 0.01 0.02 0.00 0.00
No beef/red meat 0.13 0.17 0.27 0.36 0.33
Vegetarian/vegan diet 0.00 0.01 0.01 0.02 0.00
Num. of obs. 8 71 1068 242 3
Buddhist
Do you practice meditation in the temple?
[
2
(4; 86) = 11:49;p = 0:02]

Do not practice - 0.56 0.16 0.25 0.00
On 1st & 15th of each Chinese month - 0.22 0.39 0.25 1.00
Every day - 0.22 0.43 0.50 0.00
Are you a vegetarian?
[
2
(2; 86) = 3:93;p = 0:14]

Yes 0.00 0.13 0.25 1.00
Num. of obs. 18 61 8 1

2
calculations exclude respondents who refuse to answer the religiosity question.
98
Table 3.4: Share Participating in Religious Activities in the Village
y
Degree of religiosity Statistics
Not
religious
Somewhat
religious
Religious
Very
religious
Num.
of obs.
P-val

2
All religions 0.29 0.41 0.61 0.70 25917 0.000
Muslims 0.28 0.40 0.59 0.66 23291 0.000
Catholics 0.27 0.57 0.80 0.76 383 0.000
Protestants 0.45 0.62 0.76 0.88 1016 0.000
Hindus 0.67 0.66 0.79 0.82 1173 0.053
Buddhists - 0.00 0.26 0.57 54 0.035
y
Responses to whether respondents participate in any religious activity held in the village in the past 12
months.
Included in the list of activities inquired is a religious activity. Table 3.4 presents a
summary on responses for dierent levels of religiosity. Participation tends to increase
in religiosity, and the
2
tests reject statistical independence between religiosity and
participation.
Meanwhile, many religion-based educational institutions often function as a source
of oblique socialization of religious values and beliefs. The values transmitted through
these institutions in the past may aect cooperative attitudes at present. To capture
this, I employ data on each individuals' educational history. IFLS contains informa-
tion on the types of institution managing the schools attended by the respondents,
including whether it is a religion-based { to wit, Catholic, Protestant, or Buddhist, but
not Hindu { institution. With this information, I construct an indicator of whether
the respondent receives an education from an institution of her religion (or a \core-
ligion education") or a religion-based institution that is not of her religion (or a
\non-coreligion education").
15
15
To obtain this information, I made use the panel nature of the dataset to trace the education
history from the rst wave of IFLS (IFLS1). This introduced a minor problem, since IFLS1 con
ated
Buddhists and Protestant schools into a single category. In these cases, I assume that the respondent
is attending a Protestant-managed school. The potential misclassication from this last assumption
is miniscule, since even if all of these schools assumed to be Protestant-managed are Buddhist-
99
Having received an education from religion-based institution is correlated with an
individual's religiosity. Therefore, including these variables will likely absorb some of
the correlations between religiosity and outcomes. I therefore do not include these
variables in the base specication. Nonetheless, I believe that the question of the role
of religion-based education is in itself an interesting and important one. I therefore
implement a separate set of regressions to examine the question.
Table 3.5: Summary Statistics: Individual and Household Regressors
Num.
of obs.
Mean
Std.
dev.
Median IQR Min Max
Individual-level variables
Basic specication
Religiosity 28973 2.82 0.56 3 0 1 4
Male 29037 0.48 0.50 0 1 0 1
Age 29034 36.87 15.62 34 22 13 100
Years of education 29023 7.40 4.02 9 3 0 18
Risk aversion 29029 2.72 1.46 3 2 0 4
Patience 29020 1.48 0.93 1 1 0 4
Married 29037 0.70 0.46 1 1 0 1
Extended specication
Received corlgn edu. 29037 0.33 0.47 0 1 0 1
Received non-corlgn edu. 29037 0.04 0.20 0 0 0 1
Household-level variables
Monthly per-capita expenditure 29014 484193.48 706572.53 326482.3 348159.5 0 5.8e+07
Log (1 + PCE) 29014 12.74 0.79 12.7 1.00 0 18
Other regressors
I implement the same set of control variables across outcomes, which is summarized in
Table 3.5. For the base specication, I include the standard individual characteristics
such as sex, age, married status, and years of education. To address potential non-
linear eects of age, I include dummy variables indicating whether the individual's
managed, at most I would have misclassied 59 individuals (49 Protestants and 10 Buddhists).
100
age is greater or equal to 25, 45, and 65 years old. Similarly for education, I also
introduce a set of dummy variables to indicate whether an individual has received
some junior high, senior high, or college-level education.
The decision to cooperate can be a risky act and risk preference may aect co-
operative behaviors. Indeed, Schechter (2007) shows that failing to include the risk
preference parameters in trust regressions in experimental games may signicantly
alter the coecients of important regressors. IFLS elicits risk aversion by asking re-
spondents to choose payos with dierent risk levels, which I used to create an ordinal
ordering of risk aversion.
16
This variable can take a value of between 0 and 4 where a
larger number indicates greater risk aversion.
17
Risk aversion is elicited using without
real payos and there are some concerns about potential biases from this approach.
However, the experience from the Mexican Family Life Survey suggests that such
biases may not be so severe (Hamoudi, 2006).
Moreover, an individual's discount factor may aect local cooperative attitudes
through its eects on social capital investment (Glaeser et al., 2002). IFLS elicits a
measure of the individual discount factor by asking respondents to choose dierent
payos that give returns at dierent times from today. Similar to the measure of risk
aversion, the discount factor is elicited without real payos.
At the household level, I include the level spline of the log per-capita expenditure
(PCE), with a knot point at the median.
16
IFLS4 elicited risk preference using two sets of questions on risk aversion. The hypothetical
sure payo in rst set is Rp.800.000, almost twice the average monthly per-capita expenditure of
the IFLS4 respondents. Meanwhile, the sure payo in second set is ve times that in the rst set.
The amounts of relative risk in the two sets are also dierent. I use the rst set of questions as a
measure of the risk preference parameter.
17
I code as \4" individuals whose \strong dislike" for risk cannot be explained by the standard
utility theory: They prefer a sure payo over a 50-50 gamble even though the smaller payo in the
gamble equals the sure payo.
101
3.3 Results
Individual religiosity and attitudes are likely to be endogeneous. I try to address
this problem in the following ways. First, I include a rich set of control variables at
various levels of aggregation. Second, to further ameliorate the omitted variable bias,
I estimated xed eects models. With IFLS data, I can include xed-eects up to the
household level; however, there are potential trade-os between bias reduction and
information loss from the \over-inclusion" of controls. Finally, following the strategy
similar to that of Altonji et al. (2005), I conduct an exercise to assess the likelihood
that the entirety of these ndings come from omitted variable biases.
To estimate the association between religiosity and attitudes, I employ the follow-
ing specication:
Y
ijk
= +
X
r=3;4

r
1
:1(rlgs
i
r) + X
i
:
i
+ X
j
:
j
+ X
k
:
k
+
r
+
d
+"
ijk
(3.1)
whereY is the outcome variable, 1(rlgs
i
r) is a dummy variable that is equal to one
if individuali reports a religiosity greater or equal tor forr = 3; 4 and zero otherwise,
X is the vector of control variables, " is the residual, and j, k index households,
and communities respectively. For the analysis, I combine the rst two religiosity
categories { i.e., \not religious" and \somewhat religious" { into one (default) category
of \less religious". The coecients of interest,
3
1
and
4
1
are the marginal eects of
being religious and very religious respectively. To account for heterogeneous response
of religiosity in dierent religions, I included the religion xed eects (
d
) in the
model. In addition, for reasons explained below, I also included the community xed
eects (
d
) as the preferred specication.
102
Why community xed eects? In this specication, including household xed-
eects would provide the most reduction in the omitted variable bias possible in this
dataset. However, 8,387 out of 12,680 households (and 5,444 out of 9,737 households
with more than one members) in the sample are homogeneous in their religiosity. The
inclusion of the household xed-eects will remove the eects of religiosity that have
been \institutionalized" in the household. Since individuals living in homogeneous-
religiosity households tend to be more religious, results would tend to discount the
eects coming from them.
18
As such, I will report results that are estimated using
the community xed-eects specication. For robustness, I include in the appendix
results estimated using the household xed-eects model. In both cases, standard
errors that are robust-clustered at the level of the xed-eects.
3.3.1 Individual and household characteristics
The results are presented in Tables 3.6 and 3.7. Before addressing the role of religiosity
in cooperative behaviors, however, I will rst examine the links between the dierent
plausibly exogenous regressors and the dierent outcomes in this subsection.
Gender
Men exhibit greater willingness to help and trust than women. The coecients for the
indicator variable male in Table 3.6 are positive and signicant on the willingness to
help neighbors and to trust neighbors to watch their children and their house. These
trusting behaviors may be borne out of the fact that, compared to women, men are
more likely to assess the trustworthiness of their close neighbors more favorably. Men
are also more likely to perceive the village to be safe than women.
18
Among individuals living in a multiple-member, homogeneous-religiosity household, 90.2% con-
sider themselves either religious or very religious, compared to 64.7% among those living in a
heterogeneous-religiosity household.
103
Table 3.6: Community Cohesion & Trust Beliefs
Willing
to help
Village is safe [. . . ] Trust nbr. to watch Trust [. . . ] to return lost wallet
generally at night kid(s) house neighbors strangers police
(1) (2) (3) (4) (5) (6) (7) (8)
Religiosity:
Religious/ very religious 0.013

0.026

0.008 0.011 -0.000 0.109

0.027

0.139

(1.98) (4.81) (1.26) (0.93) (-0.00) (6.52) (2.08) (7.05)
Very religious 0.162

0.171

0.090

0.063

0.062

0.097

-0.055

0.172

(11.64) (10.73) (5.81) (2.94) (4.08) (3.28) (-2.30) (6.22)
Male 0.043

0.024

0.109

0.055

0.051

0.028

0.032

-0.119

(8.69) (4.87) (15.75) (7.58) (9.41) (2.61) (3.06) (-8.74)
Age:
25 years old -0.004 0.021

0.053

0.049

0.030

0.017 0.009 -0.087

(-0.58) (3.51) (7.16) (3.89) (3.86) (1.11) (0.74) (-4.74)
45 years old 0.002 0.012

0.027

0.069

0.025

0.022 0.044

0.042

(0.36) (1.94) (4.57) (6.49) (3.17) (1.38) (3.20) (2.69)
65 years old -0.041

-0.025

-0.012 -0.023 -0.031

-0.003 0.059

-0.024
(-4.04) (-2.91) (-1.45) (-1.39) (-2.50) (-0.13) (2.64) (-0.88)
Education:
Some junior high school 0.032

-0.009 -0.020

-0.049

-0.023

0.053

0.027

0.047

(4.45) (-1.25) (-2.66) (-4.44) (-2.93) (3.11) (1.96) (2.72)
Some senior high school 0.019

-0.006 -0.005 -0.045

-0.017

0.045

0.045

-0.000
(2.56) (-0.89) (-0.80) (-3.45) (-2.01) (2.62) (3.06) (-0.01)
Some college 0.017

0.004 0.018

-0.051

-0.003 0.078

0.142

-0.055

(1.69) (0.45) (1.96) (-3.12) (-0.32) (3.65) (6.84) (-2.34)
Risk aversion 0.003 0.000 -0.001 -0.010

-0.004

-0.005 -0.005 0.001
(1.44) (0.13) (-0.42) (-2.77) (-1.77) (-1.02) (-1.24) (0.16)
Patience 0.014

0.000 -0.007

0.006 0.001 0.009 0.012

0.042

(4.64) (0.06) (-2.97) (1.17) (0.15) (1.33) (2.05) (5.35)
Married 0.007 0.005 0.005 -0.047

-0.019

0.008 -0.013 -0.015
(1.26) (0.97) (0.89) (-4.10) (-2.78) (0.58) (-1.07) (-0.97)
Log. (1 + PCE) Spline:
Below median
y
0.016

0.005 0.005 0.003 0.003 0.020 0.027

-0.004
(2.11) (0.72) (0.66) (0.26) (0.33) (1.10) (1.81) (-0.18)
Above median
y
0.000 -0.011

-0.005 -0.007 -0.021

-0.037

0.004 -0.012
(0.07) (-1.88) (-0.72) (-0.65) (-2.77) (-2.69) (0.28) (-0.78)
Constant 2.859

2.954

2.837

2.649

2.834

2.606

1.094

2.778

(30.95) (32.90) (29.17) (17.82) (27.20) (11.73) (6.07) (10.58)
Community xed eects Yes Yes Yes Yes Yes Yes Yes Yes
Religion xed eects Yes Yes Yes Yes Yes Yes Yes
N 28932 28929 28927 21759 28930 28333 27411 26832
Adj. R
2
0.084 0.088 0.088 0.097 0.050 0.100 0.059 0.082
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
y
Spline coecients are for the slope of the interval. The omitted education category is "Some primary or no school".
The omitted age category is "15-24 years old". Standard errors are robust and clustered at the community level.
104
Table 3.7: Discriminative Trust & Tolerance
Trust [. . . ] more Tolerate non-corlgn living in [. . . ] Tolerate non-corlgn to [. . . ]
corelgn coethnics village neighbor house marry rltv. bld h. wrshp
(1) (2) (3) (4) (5) (6) (7)
Religiosity:
Religious/ very religious 0.087

0.068

-0.059

-0.081

-0.120

-0.132

-0.101

(8.48) (6.33) (-6.10) (-8.50) (-10.89) (-9.79) (-8.56)
Very religious 0.160

0.070

-0.007 -0.011 -0.089

-0.138

-0.103

(8.72) (3.71) (-0.43) (-0.64) (-4.32) (-5.79) (-4.60)
Male -0.016

-0.041

0.013

0.010 0.032

0.059

-0.001
(-2.44) (-6.27) (2.09) (1.40) (3.95) (6.87) (-0.07)
Age:
25 years old 0.008 -0.020

-0.007 -0.016

-0.040

-0.065

-0.027

(0.82) (-2.13) (-0.95) (-1.99) (-3.46) (-4.78) (-2.14)
45 years old 0.029

0.009 -0.023

-0.034

-0.061

-0.049

-0.006
(3.01) (0.96) (-3.08) (-3.81) (-5.24) (-3.79) (-0.49)
65 years old -0.000 0.024

-0.029

-0.027

-0.041

0.038

-0.008
(-0.01) (1.91) (-2.19) (-1.97) (-2.37) (2.10) (-0.40)
Education:
Some junior high school -0.054

-0.068

0.070

0.081

0.050

-0.027

0.014
(-5.09) (-6.90) (6.87) (8.39) (4.06) (-1.82) (1.06)
Some senior high school -0.074

-0.098

0.026

0.021

0.005 -0.047

0.025

(-7.58) (-9.65) (3.01) (2.29) (0.43) (-3.57) (1.95)
Some college -0.073

-0.118

0.044

0.035

0.022 -0.071

0.037

(-5.15) (-8.23) (4.15) (3.30) (1.33) (-4.06) (2.19)
Risk aversion 0.012

0.013

0.001 0.001 -0.005 0.008

-0.002
(4.03) (4.53) (0.18) (0.18) (-1.61) (2.01) (-0.46)
Patience 0.006 -0.012

0.008

0.006 -0.012

-0.013

-0.008
(1.36) (-2.96) (2.17) (1.57) (-2.78) (-2.49) (-1.59)
Married -0.009 0.011 -0.014

-0.005 -0.044

-0.056

-0.027

(-1.16) (1.38) (-1.95) (-0.67) (-4.55) (-4.65) (-2.58)
Log. (1 + PCE) Spline:
Below median
y
-0.002 -0.028

0.029

0.025

0.018 -0.027

0.018
(-0.19) (-3.01) (2.67) (2.25) (1.36) (-1.82) (1.36)
Above median
y
-0.014 -0.017

0.011 0.013

0.003 -0.010 0.024

(-1.53) (-1.92) (1.45) (1.67) (0.27) (-0.84) (2.13)
Constant 2.795

3.058

2.417

2.431

2.333

2.260

2.105

(22.90) (26.06) (17.79) (17.15) (14.06) (12.20) (12.31)
Community xed eects Yes Yes Yes Yes Yes Yes Yes
Religion xed eects Yes Yes Yes Yes Yes Yes Yes
N 28931 28931 28932 28932 28931 28931 28931
Adj. R
2
0.151 0.174 0.222 0.252 0.263 0.242 0.265
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
y
Spline coecients are for the slope of the interval. The omitted education category is "Some primary or no school".
The omitted age category is "15-24 years old". Standard errors are robust and clustered at the community level.
105
Men are also generally more tolerant toward non-coreligionists than women (Ta-
ble 3.7). Again, there may be a belief component to tolerant behavior, as men tend
to rate the trustworthiness of strangers higher than women. Consistent with this,
men also exhibit less discriminative trust either with regards to ethnicity and reli-
gion. There is, however, an exception to the gender dierence in tolerance. Men are
not more tolerant { although neither are they less tolerant { than women on allowing
non-coreligionists build their house of worship in the village. Moreover, men also tend
to be less trusting of the police.
19
These eects are robust to the both the community
and household xed eects specications.
The ndings on helpfulness and interpersonal trust broadly align with what is
known about gender dierences in social preference. On helping behavior, the meta-
analytic studies of the psychology literature by Eagly and Crowley (1986) found that
men helped more than women. Meanwhile on trust, using U.S. data, Alesina and
La Ferrara (2002) nd that women exhibit less generalized trust. Similarly, in their
survey paper of gender dierences in the experimental literature, Croson and Gneezy
(2009) nd that in trust games, women tend to trust less or the same than men, and
that their decisions to trust are more sensitive to the experimental context and social
distance.
Age
With only a cross-section dataset, we cannot disentangle between age and cohort
eects. Hence, the analysis below will confound both eects. Below, we refer to
the age groups as \young adulthood" (15-24 years old), \early adulthood" (25-44
years old), \middle adulthood" (45-64 years old), and \late adulthood" (65 years old
19
Guiso et al. (2003) also found negative, albeit insignicant, coecient of being a male and trust
toward the police.
106
and older). The results suggest that in most cases, there are non-linear relationships
between age and cooperative attitudes.
The willingness to help neighbors hardly varies by age except of the slight decline
in late adulthood. The willingness to trust neighbors to watch one's children or house
when away increases with age up until middle adulthood, perhaps partly due to the
positive associations between age and trust beliefs of strangers and close neighbors
(although the latter association is not statistically signicant) among early adults.
On average, there is a marginal increase in religious discriminative trust between
those in early and middle adulthood, and a similar increase for ethnic discriminative
trust between middle and late adulthood. Consistent with this result, the tolerance
of having non-coreligionists in the village also decreases between early and middle
adulthood. For other residential tolerance measures, tolerance is negatively associated
with age across all age groups. Interestingly, in terms of inter-faith marriage, those
in their middle adulthood are the most intolerant { perhaps, because it is at around
this age group that parents marry their children.
20
Meanwhile, in terms of allowing
other believers to build their house of worship, it appears that the young adults in
the sample are the most intolerant compared those in the other age groups.
Education
Additional education is positively associated willingness to help, but negatively with
trusting behaviors. Interestingly, however, education is positively associated with
inter-personal trusting beliefs: More educated people are more likely to assess neigh-
bors and strangers to be trustworthy. Perhaps, education alters an individual's un-
derstanding of potential sources of risks associated with trusting behaviors other than
20
For instance, the average age of fathers to the once-married adults in the IFLS4 sample that
were married between 1997 and 2007 was 56.9 (with a median at 55) years old and that of mothers
to be 50.1 (with a median at 49) years old.
107
the trustworthiness of her \agent", such as the quality of local institutions to punish
breaches of trust.
21
Indeed, education up until high school is negatively associated
with a lower assessment of village safety. Meanwhile, the assessment of the trustwor-
thiness of the police increases with junior secondary education, but decreases with
college education.
Education is negatively associated with religious and ethnic discriminative trust
and this negative associations at dierent levels of education are stronger for ethnic
discrimination. It is also positively correlated across all residential tolerance measures.
Education, particularly at the college level, is also positively associated with tolerance
of other believers' house of worship. However, more education { even beyond high
school { is associated with less tolerance of interfaith marriages.
Risk and time preferences
Trusting behaviors of neighbors are negatively associated by risk preference, but not
trusting beliefs. More risk averse individuals are less likely to entrust their children
and house to their neighbor's watch, although in the case of the house, this associa-
tion is not robust to the household xed-eects specication. However, their beliefs
of the trustworthiness of their neighbors and strangers are not associated with risk
preference. Similarly, their trust beliefs of the police are not aected by risk prefer-
ence.
More risk-averse individuals also tend to trust people who are similar to them
more. Risk aversion is positively correlated with discriminative trust with respect
to both religion and ethnicity. It is negatively correlated with tolerance in allowing
non-coreligionists stay in their house, and positively with inter-faith marriage, but
21
See footnote 11 for a sketch of model that incorporates institutional quality in the trust game
framework.
108
these correlations are weak and not robust to the household xed eects.
Meanwhile, Glaeser et al.'s (2002) static model suggest that (local) social capital
should increase with the individual discount factor. We therefore expect the discount
factor to be positively correlated with community trust and altruism. We nd some
support for this prediction: A higher discount factor is positively associated with a
willingness to help and a higher trust belief of strangers. It is also positively associated
with the tolerance of having non-coreligionists live in the village. At the same time,
however, a higher discount factor is negatively associated with the tolerance of having
non-coreligionists live at home and of interfaith marriage.
Household expenditure
A higher per-capita expenditure of the is associated with the willingness to help neigh-
bors among households whose per-capita expenditure (PCE) is below the median (or
the \poorer households") but those with above median not those with above-median
PCE (or \richer households") in the community xed-eects model. Among richer
households, a higher PCE is correlated with less willingness to entrust one's property
to a neighbor and a lower belief of the trustworthiness on neighbors. Meanwhile,
among poorer households, a higher PCE is (weakly) associated with greater trust of
strangers.
Moreover, a higher PCE is associated with less in-group preference. In all house-
holds, a higher PCE is associated with less discriminative trust with respect to eth-
nicity. It is also associated with less religious discrimination, but only among the
richer households. Among richer households, a higher PCE is associated with more
tolerance regarding allowing non-coreligionists to live in the village or neighborhood
as well as tolerance of non-coreligionists' house of worship. Among poorer households,
PCE is also positively associated with tolerance at the village and neighborhood level;
109
however, it is negatively correlated with tolerance on inter-faith marriage.
3.3.2 Religiosity, religious education, and attitudes
Religiosity is positively correlated with cooperative attitudes involving members of
the community and the in-groups, but not the out-groups. Furthermore, for most
outcomes, the correlations are monotonic in religiosity. More religious people exhibits
more willingness to help neighbors. They are also more willing to trust neighbors with
their children or property. This behavior may have partly arisen out of their more
favorable beliefs regarding the trustworthiness of others: Religious people are more
trusting of neighbors, strangers, and the police compared to less religious people.
22
However, in the case of the trust belief of stranger, the relationship is not monotonic
in religiosity: The very religious are less trusting of strangers compared to the less
religious.
At the same time, religiosity is also positively correlated with religion-based and
ethnic-based discriminative trust. It is negatively correlated with all measures of
tolerance.
The religiosity coecients are robust to both community and household xed
eects, although they are smaller in the latter specication.
23
Since homogeneously
religious households tend to be more religious (see footnote 18), the lower magnitudes
of these coecients in the household xed-eects specication come partly from re-
moving the eects from these more religious individuals. In a separate analysis that
is not reported here, I nd non-linearity in the relationships between religiosity and
22
\Less religious" includes individuals reporting themselves to be \not religious" and \somewhat
religious".
23
I also consider the possibility that the risk and time preference measures are endogenous by
estimating a model without these preference measures (not reported). Comparisons of the point
estimates suggest that the inclusion of these measures in the model hardly has any eect on the
religiosity coecients.
110
some of the outcomes. Nonetheless, qualitatively the results in that analysis are
identical to the linear case presented here.
Columns (3) - (6) of Table 3.8 present estimates from the extended specication
that includes indicators for individuals' religious educational background. Including
these indicators reduce the magnitude of the religiosity coecients, albeit only very
slightly. Overall, the results suggest that religious educational background mainly
plays a role in aecting inter-group cooperative attitudes. Coreligion educational
background has a weak negative impact on helpfulness while non-coreligion has no
impact. The only signicant eect of religious educational background on community
trust behaviors and beliefs comes from that of non-coreligion education on trusting
neighbors to watch one's children.
However, coreligion education is associated with more trust of coreligionists, and
less religious tolerance across all measures. In contrast, having been educated by a
non-coreligion religious institution reduces religious discriminative trust, and increases
tolerance across all of the measures. This latter result may be interpreted as support
for Allport's (1979 [1954]) contact hypothesis.
The evidence so far points to statistically signicant correlations between reli-
giosity and attitudes. How meaningful are these correlations in real terms?
24
To
answer this question, I examine compare coecients on the religiosity dummies with
the coecients on three other regressors: gender, PCE, and education. I will mostly
focus on the coecients of the (religious/very religious) dummy { i.e., the marginal
response of the religious relative to the less religious. Hence, unless noted dierently,
any reference to the religiosity variable in the remainder of this section refers to this
variable. Moreover, keep in mind that I am making comparisons of correlative, not
causal, relationships. The analysis is based on results presented in Tables 3.6 and 3.7.
24
I would like to thank Larry Iannaccone for suggesting this line of inquiry.
111
Table 3.8: Religiosity and religious education
Basic Extended
Religious/
Very
Very
Religious
Religious/
Very
Very
Religious
Corelgn.
education
Non-corelgn.
education
(1) (2) (3) (4) (5) (6)
Willing to help 0.013

0.162

0.013

0.162

-0.010

-0.016
(1.98) (11.64) (2.01) (11.64) (-1.73) (-1.09)
Vilage is safe [...]
generally 0.026

0.171

0.026

0.170

-0.003 -0.008
(4.81) (10.73) (4.82) (10.72) (-0.61) (-0.64)
at night 0.008 0.090

0.008 0.090

-0.002 0.023

(1.26) (5.81) (1.30) (5.82) (-0.32) (1.86)
Trust neighbor to watch [...]
kid(s) 0.011 0.063

0.010 0.063

0.021

-0.028
(0.93) (2.94) (0.85) (2.92) (2.26) (-1.43)
house -0.000 0.062

-0.000 0.062

0.007 -0.005
(-0.00) (4.08) (-0.04) (4.06) (0.92) (-0.34)
Trust [...] to return wallet
neighbors 0.109

0.097

0.109

0.097

0.004 0.024
(6.52) (3.28) (6.54) (3.29) (0.31) (0.69)
strangers 0.027

-0.055

0.027

-0.055

0.002 0.010
(2.08) (-2.30) (2.08) (-2.30) (0.19) (0.28)
police 0.139

0.172

0.139

0.171

-0.015 -0.040
(7.05) (6.22) (7.04) (6.20) (-1.10) (-1.20)
Trust [...] more
coreligionists 0.087

0.160

0.085

0.159

0.025

-0.039

(8.48) (8.72) (8.37) (8.67) (3.07) (-1.99)
coethnics 0.068

0.070

0.067

0.070

0.019

-0.006
(6.33) (3.71) (6.28) (3.68) (2.30) (-0.35)
Tolerate non-coreligionist to live in [...]
village -0.059

-0.007 -0.058

-0.006 -0.038

0.020
(-6.10) (-0.43) (-5.98) (-0.37) (-5.06) (1.52)
neighborhood -0.081

-0.011 -0.079

-0.010 -0.045

0.030

(-8.50) (-0.64) (-8.37) (-0.57) (-5.89) (2.22)
house -0.120

-0.089

-0.118

-0.087

-0.059

0.055

(-10.89) (-4.32) (-10.72) (-4.20) (-6.06) (2.78)
Tolerate non-coreligionist to [...]
marry relative -0.132

-0.138

-0.130

-0.136

-0.038

0.061

(-9.79) (-5.79) (-9.68) (-5.72) (-3.29) (2.32)
build house of worship -0.101

-0.103

-0.098

-0.101

-0.063

0.087

(-8.56) (-4.60) (-8.33) (-4.51) (-5.56) (3.82)
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01.
Each row presents results from separate regressions with community and religion xed-eects. Standard errors are robust and clustered
at the community level. Included variables not shown: sex, dummy variables for age and education categories, risk and time preference,
married status, linear spline for log PCE, and a constant. Education institution dummies are relative to non-religious public and private
education.
112
Given a lot of interest among social scientists to see gender dierences in attitudes,
the gender variable serves as a useful benchmark. The religiosity coecient tend to
be negligible for the willingness to help or trust neighbors. However, the coecient
on the very religious dummy variable are almost four times the eect of gender for
the willingness to help neighbors, and they are comparable to the eect of gender for
the willingness to trust. The ratio of the religiosity and the gender coecients in the
trust belief of neighbor regression is slightly less than four. Meanwhile, for the trust
beliefs of strangers, the magnitude of the two coecients are comparable.
The ratio of the magnitude of the religiosity and the gender coecients is more
than ve times for the discriminative trust of coreligionists, and almost ve-thirds for
the discriminative trust of coethnics. These ratios are also generally large { between
2.2 and 4.5 { for tolerance measures whose gender eects are signicantly dierent
from zero.
Next, I compare the magnitudes of the religiosity coecient with the eects of a
standard deviation change in log PCE (hereafter, the \log PCE eect").
25
Among
poorer households, the two coecients are comparable for the willingness to help
neighbors while the religiosity eect is slightly larger than the log PCE eect for the
trust belief of strangers. The eect of religiosity on trust belief of neighbors is almost
four times the log PCE eect among richer households.
In general, the eects of religiosity on inter-group cooperation are much larger
than the log PCE eects. Among poorer households, the magnitude of the religiosity
coecient is three times the log PCE eect on trust of coethnics and between 2.6
and 6.2 times the log PCE eect on various measures of tolerance. Among richer
households, where the log PCE eect is mostly absent or imprecisely estimated for
25
As shown in Table 3.5, the standard deviation of log PCE is 0.79. As such, the \log PCE eect"
equals to 0.79 the log PCE coecient.
113
most outcomes, these ratios tend to be larger except for tolerance regarding allowing
non-coreligionists build their house of worship.
Relative to the education coecients, the religiosity coecient for the willingness
to help neighbors is smaller than that from an additional level of education (i.e.,
from primary to junior/senior secondary to college). It is, however, larger when we
consider the coecient on the very religious variable. The religiosity coecients for
trusting behaviors also tend to be larger than any of the education level coecients.
The religiosity coecient for trust belief of neighbors is also larger than any of the
education coecients.
Meanwhile, we nd that in most cases, the magnitudes of the eects of religiosity
on in-group preferences are stronger than those from an additional level of education.
In all cases, except for inter-faith marriage, the religiosity and education coecients
have dierent signs. To simplify exposition for these coecients, I will compare here
two hypothetical persons from the sample, namely, a very religious person with some
college education vs. a less religious person with primary school or less.
For discriminative trust of coreligionists, the eect of religiosity is much stronger
than education such that a highly-educated highly-religious individual, on average,
will be more discriminative than a less religious individual with a primary school
education. Similarly, the former will be less tolerant in allowing non-coreligionists live
in his or her house, and much more intolerant of allowing non-coreligionists to build
their house of worship. On the other hand, the total eects of education is stronger
for ethnic discriminative trust and tolerance of having non-believers in one's village
and neighborhood. For intolerance regarding inter-faith marriage, the magnitude of
the religiosity coecient is larger than the marginal eects of education at all levels,
but unlike those for other measures of tolerance, they are of the same sign.
114
Selection on observables as a benchmark for omitted variable bias
Even with the household xed eects, potential latent variables problems may nonethe-
less remain.
26
Altonji et al. (2005) suggest a way to informally benchmark potential
omitted variable bias using selection on the observable characteristics for a bivariate
normal model and Bellows and Miguel (2008) develop a similar test for linear models
without the assumption of joint normality. These authors derive measures to quantify
how important the omitted variable bias needs to be in order to explain away the en-
tire eects. This paper follows the approach of Bellows and Miguel, whose derivation
is reproduced below.
27
The objective of the derivation is to quantify how much stronger the relationship
between the unobservable and religiosity relative to the relationship between the
observable and religiosity in order for all of the eects to come from the omitted
variable bias. To this end, consider the specication of interest:
Y =R +q +" (3.2)
where q is the index of the full control variables, including both observables and
unobservables. If we estimate using OLS without q, we have the following omitted
variable bias:
plimb
NC
= +
Cov(R;q)
Var(R)
(3.3)
where NC indicates the \No control" estimate.
26
For instance, household xed eects may have absorbed some of the dierences that are inherent
to a family (such as genetic dierences), but may not have completely eliminated intra-household
unobservables such as personality dierences.
27
See Bellows and Miguel (2008, Appendix A). Nunn and Wantchekon (2011) also utilize this
approach in their examination of the eects of living in regions that were heavily raided for slaves
in the past on current levels of trust in Africa.
115
Now, suppose that there are a set of control variables X andq is linearly correlated
with these variables:
q = X
0

+ ~ q (3.4)
Plugging this into (3.3), we obtain:
Y =R + X
0

+~ q +" (3.5)
In this case, our estimate of yields:
plimb
C
= +
Cov(R; ~ q)
Var(R)
(3.6)
where C denotes \Control". Given the linear relation between q and X
0

, we have
the following:
b
NC
b
C
=

Cov(R;q)
Var(R)

Cov(R; ~ q)
Var(R)

=

Cov(R; X
0

)
Var(R)
+
Cov(R; ~ q)
Var(R)

Cov(R; ~ q)
Var(R)

=
Cov(R; X
0

)
Var(R)
(3.7)
We can now nd an estimate of the measure of omitted variable bias necessary to
explain away the entire religiosity eects. Suppose there is no religiosity eect and
we set =0. Dividing (3.5) with (3.7), we have:
b
C
b
NC
b
C
=
Cov(R; ~ q)
Cov(R; X
0

)
(3.8)
The term on the left-hand side can be estimated. Meanwhile, the right-hand side
116
term gives the ratio between the religiosity-unobservable and religiosity-observable
covariances, which captures how much stronger the covariance between religiosity
and the unobservable variable relative to its covariance with the observable variables
needs to be to explain away the entire eect of religiosity.
Table 3.9 presents the calculations of this ratio. I consider the basic and extended
specications (i.e., without and with the religious education background). In both
specications, I included community and religions xed eects. In the basic specica-
tion, for outcomes in which the religious=very religious coecients are statistically
signicant, the magnitudes of these ratios lie between 1.2 and 83.4. Meanwhile, for
outcomes where the very religious coecients are statistically signicant, the mag-
nitudes of these ratios lie between 4.6 and 88.7. The ratios are quite similar for the
extended specication. In sum, in most cases, the selection on the unobservables
needs to be multiples of that on observables in order for the results to come entirely
from the omitted variable bias.
3.3.3 Does the religion matter?
Next, we look at inter-religion dierences. Before we begin the analysis, however, two
caveats are in order. First, as is the case in many multiethnic, multireligion coun-
tries, ethnicity and religion are not easily separable in Indonesia. In this particular
sample, two adherents of two of the religions are ethnically very homogeneous: 88%
of Hindus are Balinese and 81% of Buddhists in the sample are of Chinese descent.
In addition, 86% of Hindus live in the province of Bali. Hence, the analysis cannot
rule out confounding ethnicity eects. The second caveat relates to the small sample
of Buddhists. In this sample, there are only 88 Buddhist respondents in the sample
{ and therefore, the estimates of its coecients have low power.
117
Table 3.9: Selection on observables to assess potential bias from
unobservables
Basic Extended
Religious/
Very
Very
Religious
Religious/
Very
Very
Religious
(1) (2) (3) (4)
Willing to help -1.17 -50.21 -1.17 -50.67
Vilage is safe [...]
generally 24.60 88.65 25.20 85.69
at night -3.83 13.66 -3.48 14.54
Trust neighbor to watch [...]
kid(s) 1.36 4.60 1.11 4.37
house -0.03 11.83 0.16 11.36
Trust [...] to return wallet
neighbors -42.31 -15.84 -40.37 -15.20
strangers 7.01 63.45 7.07 56.87
police 83.43 -69.75 89.29 -78.43
Trust [...] more
coreligionists 12.34 22.76 10.42 19.71
coethnics 6.60 6.63 6.14 6.33
Tolerate non-coreligionist to live in [...]
village 5.27 1.03 4.58 0.77
neighborhood 6.31 1.40 5.46 1.06
house 5.62 28.20 4.95 17.30
Tolerate non-coreligionist to [...]
marry relative 9.75 -8.05 8.47 -8.78
build house of worship -69.53 -24.47 70.76 -56.77
Full-control specication includes:
Religious education controls No No Yes Yes
Individual-level controls Yes Yes Yes Yes
Community xed eects Yes Yes Yes Yes
Religion xed eects Yes Yes Yes Yes
Each cell calculates the following measure:
^

C
=(
^

NC

^

C
) where
^

C
is the estimated religiosity
coecient in the full-control specication and
^

NC
is the coecient in the no-control specication.
Estimates are made using OLS and xed eects are implemented in estimating both control and
no-control specications. The individual-level controls are sex, dummy variables for age and education
categories, risk and time preference, married status. The household-level controls are the linear
splines for log PCE.
118
Table 3.10: Inter-Religion Differences in the Associations between
Religiosity and Attitudes
Religious/
Very
religious
Very
religious
(Religious/ Very religious) [. . . ] Statistics
Catholic Protestant Hindu Buddhist
Num.
of obs.
P-val of
joint test:
Religions
(1) (2) (3) (4) (5) (6) (7) (8)
Willing to help 0.011

0.162

0.043 0.028 -0.012 0.070 28932 0.381
(1.66) (11.61) (1.04) (1.18) (-0.28) (1.40)
Vilage is safe [...]
generally 0.027

0.171

0.020 -0.011 -0.061 -0.060 28929 0.572
(4.95) (10.75) (0.43) (-0.45) (-1.51) (-0.57)
at night 0.008 0.090

-0.017 -0.011 0.015 -0.079 28927 0.939
(1.33) (5.81) (-0.31) (-0.37) (0.35) (-0.64)
Trust neighbor to watch [...]
kid(s) 0.012 0.063

-0.064 0.009 -0.062 0.092 21759 0.789
(1.00) (2.94) (-0.85) (0.17) (-0.84) (0.51)
house 0.002 0.062

0.015 -0.028 -0.030 -0.112 28930 0.799
(0.20) (4.08) (0.31) (-0.73) (-0.63) (-0.87)
Trust [...] to return wallet
neighbors 0.111

0.098

-0.123 0.095 -0.185 -0.231 28333 0.300
(6.57) (3.31) (-1.16) (0.98) (-1.44) (-0.91)
strangers 0.027

-0.055

-0.032 0.030 -0.002 -0.150 27411 0.911
(2.03) (-2.29) (-0.30) (0.45) (-0.03) (-0.83)
police 0.140

0.172

0.297

-0.094 -0.133 -0.040 26832 0.048
(6.81) (6.24) (2.77) (-0.94) (-1.16) (-0.14)
Trust [...] more
coreligionists 0.087

0.160

0.071 -0.036 -0.002 -0.062 28931 0.631
(8.08) (8.71) (1.07) (-0.95) (-0.03) (-0.49)
coethnics 0.069

0.070

0.030 -0.062 0.046 0.042 28931 0.400
(6.22) (3.69) (0.43) (-1.51) (1.10) (0.29)
Tolerate non-coreligionist
to live in [...]
village -0.066

-0.008 0.167

0.046

0.133

0.154 28932 0
(-6.42) (-0.49) (3.12) (1.71) (3.86) (1.19)
neighborhood -0.089

-0.012 0.190

0.078

0.154

0.167 28932 0
(-8.93) (-0.71) (3.66) (2.94) (4.70) (1.27)
house -0.132

-0.090

0.193

0.151

0.219

0 28931 0
(-11.20) (-4.43) (3.46) (4.28) (4.27) (0.00)
Tolerate non-coreligionist
to [...]
marry relative -0.141

-0.139

-0.051 0.141

0.212

0.283

28931 0.001
(-9.93) (-5.88) (-0.65) (2.26) (2.77) (2.50)
build house of worship -0.112

-0.105

0.073 0.128

0.262

0.212 28931 0
(-8.89) (-4.65) (1.18) (2.73) (4.01) (1.53)
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01.
Each row presents results from separate regressions for model with community and religion xed eects. Standard errors are robust and
clustered at the community level. Included variables not shown: sex, dummy variables for age and education categories, risk and time
preference, married status, linear spline for log PCE, and a constant.
119
Below, I examine inter-religion dierences in how attitudes are correlated with re-
ligiosity. I estimate the following specication to examine the inter-religion dierences
that allows for dierent intercepts and religiosity coecients for dierent religions, to
wit:
Y
r
ijk
=
1
+
5
X
d=2

d
:1(x
ri
=d) +
X
r=3;4

r
1
:1(rlgsr) +
5
X
d=2

d
:1(x
ri
=d):1(rlgs
i
3)
+ X
i
:
i
+ X
j
:
j
+ X
k
:
k
+
r
+
d
+"
r
ijk
(3.9)
For tractability, I will focus on the interactions between religion and thereligious=very
religious dummy variable. Based on the same rationale used to justify the basic re-
ligiosity specication above, I included the religion and community xed-eects and
estimated the regressions with robust standard errors that are clustered at the com-
munity level.
The coecients on the interaction terms describe the inter-religion dierences.
Assuming that people rarely switch religion in Indonesia, religion can be treated as
an exogenous attribute that is inter-generationally transmitted. However, individuals
choose the level of religiosity for their given religion. Religiosity is, therefore, likely
to be endogenous to attitudes. Take an example of tolerance. If an unobserved
preference parameter or personality trait, say sociability, aects both tolerance and
the choice of religiosity for a given religion, a sociable person will be more tolerant
and, at the same time, choose a low level of religiosity if he is \given" an intolerant
religion.
28
Therefore, the inter-religion dierences in the religiosity coecients can
28
Here, we also assume that the characteristics of a religion responds very slowly, if at all, to
individuals' religiosity choice.
120
be interpreted as the relative extent to which one religion is more likely to encourage
(or discourage) the attitude in question.
The results are shown in Table 3.10.
29
For the associations between religiosity and
measures of community cohesiveness, there do not appear to be many inter-religion
variations in the association between religiosity and attitudes. Moreover, we also do
not nd inter-religion variations with respect to trust beliefs and discriminative trust:
Across all of these measures, except with regards to trust belief of the police, the joint
F-tests suggest we cannot reject the hypotheses that each of the religion interaction
terms is equal to zero.
However for tolerance measures, in almost all cases the negative links between
religiosity and intolerance are strongest among Muslims. In fact, they appear to be
mostly absent in all other religions. Exceptions to this are tolerance of having non-
coreligionists in the house among Buddhists (where the link between religiosity and
intolerance is as strong as among the Muslims) and tolerance of interfaith marriage
among Catholics (where the link between religiosity and intolerance is potentially
stronger than Muslims, even though the interaction term is not statistically signi-
cant).
3.3.4 Gender dierences in the religiosity correlates
Finally, I decompose the analysis to look at inter-gender dierences in the behavioral
responses to religion. The analysis utilizes the basic specication with the commu-
nity xed-eects described in Equation 3.1. Tests of pooling by gender show for all
outcomes rejected the pooling hypotheses { in most cases, at the 1% critical value.
30
29
Table 3.15 provide the results for cases with religionvery religious interactions.
30
The critical value for rejection is at 1% except for trust beliefs of neighbors and tolerance of
non-coreligionists' house of worship at 5% and trust of coreligionists at 10%. See Tables 3.16 and 3.17
in the appendix
121
Table 3.11: Religiosity by Gender
Male Female
Religious/
Very
Very
Religious
Religious/
Very
Very
Religious
(1) (2) (3) (4)
Willing to help 0.017

0.163

0.004 0.158

(1.81) (7.86) (0.51) (9.08)
Vilage is safe [...]
generally 0.028

0.181

0.024

0.157

(3.67) (8.46) (2.89) (7.73)
at night 0.006 0.134

0.018

0.056

(0.82) (6.54) (1.89) (2.64)
Trust neighbor to watch [...]
kid(s) 0.015 0.066

0.003 0.056

(0.98) (2.01) (0.18) (2.19)
house 0.005 0.069

-0.010 0.050

(0.59) (3.27) (-0.87) (2.42)
Trust [...] to return wallet
neighbors 0.123

0.096

0.082

0.098

(5.78) (2.38) (3.51) (2.56)
strangers 0.040

-0.045 0.008 -0.065

(2.19) (-1.23) (0.44) (-1.87)
police 0.136

0.199

0.126

0.143

(5.15) (5.21) (4.94) (3.77)
Trust [...] more
coreligionists 0.102

0.139

0.058

0.167

(7.80) (5.37) (4.33) (6.80)
coethnics 0.092

0.056

0.034

0.080

(7.12) (2.02) (2.35) (3.53)
Tolerate non-coreligionist to live in [...]
village -0.073

0.024 -0.042

-0.032
(-6.41) (0.89) (-3.22) (-1.44)
neighborhood -0.099

0.020 -0.059

-0.035
(-8.39) (0.72) (-4.63) (-1.57)
house -0.123

-0.071

-0.105

-0.091

(-8.61) (-2.23) (-6.68) (-3.30)
Tolerate non-coreligionist to [...]
marry relative -0.132

-0.099

-0.108

-0.142

(-7.27) (-2.98) (-5.61) (-4.51)
build house of worship -0.108

-0.050

-0.077

-0.138

(-6.92) (-1.65) (-4.66) (-4.42)
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01.
Each cell presents the religiosity coecient from a separate regression for the community xed-eects model.
Standard errors are robust and clustered at the community level. Included variables not shown: sex, dummy
variables for age and education categories, risk and time preference, married status, linear spline for log PCE,
and a constant. Education institution dummies are relative to non-religious public and private education.
122
Table 3.11 presents the religiosity coecients when we estimated the data sep-
arately by gender. Overall, the links between religiosity and trusting attitudes are
stronger for men than women. Religiosity is associated with trusting behaviors among
men, but not women. The extent to which religiosity are associated with trusting
beliefs are also stronger for men.
I nd similar results regarding inter-group cooperative attitudes. The positive
association between religiosity and discriminative trust is stronger for men; similarly,
the extent to which men become more intolerant as they become more religious is
larger { although to dierent degrees for dierent measures of tolerance { compared
to that of women. If we were to take a causal interpretation, then one can interpret
this as suggesting that religious commitment alters cooperative attitudes more among
men than women.
3.3.5 Religiosity and political preference
A natural extension to this analysis is to ask whether the eects of religion go beyond
preferences regarding interpersonal interactions, into the political sphere. Using data
on the characteristics that respondents nd important in a political leader, I examine
whether a stronger commitment to the religious identity is also positively associated
with an in-group bias in politics. Upon nding a bias, I also examine the \cost" of
such a bias by examining which other important characteristic(s) (if any) in a political
leader gets \crowded out" by this in-group bias. Answers to these questions can help
understand the implications of the use (and misuse) of religion in political discourse.
IFLS4 contains a set of questions that elicit each respondent's opinion regarding
factors that are important when electing a district head or a mayor. The interviewer
begins by asking a series of \Yes/No" questions to determine whether a particular fac-
123
tor is important to a respondent. Nine factors were considered, to wit, the candidate's
appearance, popularity, program quality, similarity in political aliation, similarity
in religious aliation, similarity in ethnicity, governing experience, gender, and the
amount of money (or \campaign gifts") that the candidate gives out during his/her
campaign stops. Once the respondent had answered all nine questions, he/she was
asked to rank the most, second- and third-most important factors for his/her choice
of a district head.
For each of the factors, I generated three binary outcome variables. The rst is
whether that factor is an important factor for the respondent. It is equal to one
for a particular factor if the respondent answers \Yes" to it during the rst series
of the \Yes/No" questions. The second variable captures the notion of whether the
respondents consider the factor to be among the top-three most important factors in
a political candidate. It is equal to one for a factor if the respondent included it in
his/her list of the three most important factors in a district head candidate. Finally,
I generated a variable that is equal to one for a factor if it is listed as the respondent's
most important factor in a district head candidate.
Table 3.12 presents the results. The rst column suggests that religiosity is posi-
tively correlated with assigning importance on almost all factors except for program
quality and the campaign-stop gifts. On the second column, as expected, we nd that
religiosity is associated with a greater likelihood of putting the candidate's religion
among the top three most important factors. However, religiosity is negative asso-
ciated with the likelihood of including the candidate's popularity, program quality,
or \gift money" as the three most important factors in a candidate. Finally, in the
third column, religiosity is positively associated with the likelihood of considering the
candidate's religion to be the most important criterion, and negatively with the like-
lihood of considering the candidate's appearance, experience, program quality and
124
Table 3.12: Religiosity and District Head Criteria
Characteristic is [. . . ] in district head
Important Three-most Important Most Important
Religious/
Very
Very
Religious
Religious/
Very
Very
Religious
Religious/
Very
Very
Religious
(1) (2) (3) (4) (5) (6)
Religion 0.048

0.034

0.082

0.032

0.068

0.040

(6.50) (3.53) (9.74) (2.76) (9.49) (3.44)
Ethnicity 0.033

0.057

0.000 0.002 -0.001 -0.001
(3.93) (4.61) (0.08) (0.21) (-0.37) (-0.34)
Appearance 0.027

0.039

-0.006 0.032

-0.006

0.006
(3.16) (3.29) (-0.83) (2.63) (-1.70) (0.93)
Popularity 0.024

0.046

-0.015

-0.011 -0.005 -0.002
(2.93) (3.62) (-2.12) (-1.06) (-1.52) (-0.35)
Program quality 0.001 0.013

-0.018

-0.046

-0.026

-0.047

(0.30) (1.91) (-2.91) (-3.71) (3.29) (-4.23)
Political A. 0.027

0.013

0.004 -0.046

0.002 0.002
(3.00) (1.91) (0.80) (-3.71) (0.91) (0.42)
Experience 0.013

0.014

-0.001 -0.002 -0.018

0.005
(2.43) (2.02) (-0.14) (-0.22) (-2.15) (0.42)
Gender 0.029

0.051

0.002 0.013

-0.000 0.002
(3.28) (3.63) (0.43) (2.10) (-0.05) (0.85)
Gift -0.018

0.043

-0.020

0.010 -0.007

0.005
(-2.43) (3.49) (-4.89) (1.48) (-3.06) (0.91)
Community FE Yes Yes Yes Yes Yes Yes
Religion FE Yes Yes Yes Yes Yes Yes
Each cell shows the estimate for the religiosity coecient from a separate community xed eects regression.
Standard errors are robust and clustered at the community level. Included variables not shown: sex, dummy
variables for age and education categories, risk and time preference, married status, linear spline for log PCE,
and a constant.
125
campaign gifts to be the most important criteria. For the most religious, the religion
of the candidate gains increasing importance at the expense of program quality.
3.4 Conclusion
Using data on contemporary Indonesia, this paper provides evidence on the positive
association between religion and cooperative attitudes. More religious people tend to
trust neighbors and members of their communities more, but at the same time, exhibit
more in-group preference in terms of their trust and political preferences. Religiosity
is also negatively correlated with tolerance and these correlations are strongest for
among Muslims.
These ndings suggest that individual religious identity may have an important
role in aecting dierent aspects of cooperative attitudes. It is likely that these ef-
fects may manifest themselves at a more aggregate level { such as at the village or
subdistrict level { especially in the religiously diverse societies in Indonesia. The
next chapter, which follows up on the ndings here, empirically examines how vari-
ations in the religious compositions of communities may be linked to dierent levels
of helpfulness, trust, and religious tolerance in Indonesia.
126
Chapter bibliography
Alesina, Alberto and Eliana La Ferrara, \Who trusts others?," Journal of Public
Economics, August 2002, 85 (2), 207{234.
Allport, Gordon W., The nature of prejudice, Reading Mass.: Addison-Wesley
Pub. Co., 1979 [1954].
and Bernard M. Kramer, \Some roots of prejudice," The Journal of general
psychology, 1946, 22, 9.
Altonji, J. G, T. E Elder, and C. R Taber, \Selection on observed and un-
observed variables: Assessing the eectiveness of Catholic schools," Journal of
Political Economy, 2005, 113 (1), 151{184.
Anderson, Lisa, Jennifer Mellor, and Jerey Milyo, \Did the Devil Make
Them Do It? The Eects of Religion in Public Goods and Trust Games," Kyklos,
2010, 63 (2), 163{175.
Batson, Charles, Patricia Schoenrade, and W. Larry Ventis, Religion and
the Individual : A Social-Psychological Perspective, New York: Oxford Univ. Press,
1993.
Bellows, John and Edward Miguel, \War and Local Collective Action in Sierra
Leone," Working Paper December 2008.
Benjamin, Daniel J., James J. Choi, and Georey W. Fisher, \Religious
Identity and Economic Behavior," NBER Working Paper 15925 April 2010.
Berg, Joyce, John Dickhaut, and Kevin McCabe, \Trust, Reciprocity, and
Social History," Games and Economic Behavior, July 1995, 10 (1), 122{142.
127
Carter, Michael R. and Marco Castillo, \Trustworthiness and Social Capital
in South Africa: Analysis of Actual Living Standards Data and Artifactual Field
Experiments," Economic Development and Cultural Change, July 2011, 59 (4),
695{722.
Chaney, Eric, \Tolerance, Religious Competition and the Rise and Fall of Muslim
Science," 2008.
Croson, Rachel and Uri Gneezy, \Gender Dierences in Preferences," Journal
of Economic Literature, May 2009, 47, 448{474.
Eagly, Alice H. and Maureen Crowley, \Gender and helping behavior: A meta-
analytic review of the social psychological literature.," Psychological Bulletin, 1986,
100 (3), 283{308.
Fehr, Ernst, \On The Economics and Biology of Trust," Journal of the European
Economic Association, April 2009, 7, 235{266.
Glaeser, Edward L., David I. Laibson, and Bruce Sacerdote, \An Economic
Approach to Social Capital," The Economic Journal, 2002, 112 (483), F437{F458.
, , Jose A. Scheinkman, and Christine L. Soutter, \Measuring Trust,"
Quarterly Journal of Economics, August 2000, 115 (3), 811{846.
Guiso, Luigi, Paola Sapienza, and Luigi Zingales, \People's opium? Religion
and economic attitudes," Journal of Monetary Economics, January 2003, 50 (1),
225{282.
, , and , \Civic Capital as the Missing Link," in Jess Benhabib, Alberto
Bisin, and Matthew Jackson, eds., Handbook of social economics., Vol. Volume 1,
North-Holland, 2011, pp. 417{480.
128
Hall, Deborah L., David C. Matz, and Wendy Wood, \Why Don't We Practice
What We Preach? A Meta-Analytic Review of Religious Racism," Personality and
Social Psychology Review, February 2010, 14 (1), 126 {139.
Hamoudi, A., \Risk Preferences in Households and Families," Technical Report,
UCLA working paper 2006.
Johnson, Megan K., Wade C. Rowatt, and Jordan LaBou, \Priming Chris-
tian Religious Concepts Increases Racial Prejudice," Social Psychological and Per-
sonality Science, April 2010, 1 (2), 119 {126.
Knack, Stephen and Philip Keefer, \Does Social Capital Have An Economic
Payo? A Cross-Country Investigation," Quarterly Journal of Economics, 1997,
112 (4), 1251{1288.
La Porta, Rafael, Florencio Lopez de Silanes, Andrei Shleifer, and
Robert W. Vishny, \Trust in Large Organizations," The American Economic
Review, May 1997, 87 (2), 333{338.
Landes, David, The Wealth and Poverty of Nations : Why Some Are So Rich and
Some So Poor, New York Ny: W.W. Norton, 1998.
Levy, Gilat and Ronny Razin, \Religious Beliefs, Religious Participation, and
Cooperation," American Economic Journal: Microeconomics, August 2012, 4 (3),
121{151.
Maluccio, John, Lawrence Haddad, and Julian May, \Social capital and
household welfare in South Africa, 1993-98," Journal of Development Studies, Au-
gust 2000, 36 (6), 54{81.
129
McCauley, John F., \Distinguishing Religious and Ethnic Politics: Evidence from
a Field Experiment in West Africa," 2009.
Mujani, Saiful, \Religious Democrats: Democratic Culture and Muslim Political
Participation in Post-Suharto Indonesia." PhD dissertation, Ohio State University
2004.
Nannestad, P., \What have we learned about generalized trust, if anything?," Annu.
Rev. Polit. Sci., 2008, 11, 413{436.
Narayan, Deepa and Lant Pritchett, \Cents and Sociability: Household In-
come and Social Capital in Rural Tanzania," Economic Development and Cultural
Change, July 1999, 47 (4), 871{897.
Neusner, Jacob and Bruce Chilton, Altruism in world religions, Washington,
D.C.: Georgetown University Press, 2005.
Norenzayan, Ara and Azim F. Shari, \The Origin and Evolution of Religious
Prosociality," Science, October 2008, 322 (5898), 58 {62.
Nunn, Nathan and Leonard Wantchekon, \The Slave Trade and the Origins of
Mistrust in Africa," American Economic Review, December 2011, 101 (7), 3221{
3252.
Parra, Juan Carlos, \The eects of religion on social cooperation: Results from a
eld experiment in Ghana," 2011.
Putnam, Robert, Robert Leonardi, and Raaella Y. Nanetti, Making Democ-
racy Work : Civic Traditions in Modern Italy, Princeton N.J.: Princeton University
Press, 1993.
130
Rue, Bradley J. and Richard Sosis, \Cooperation and the in-group-out-group
bias: A eld test on Israeli kibbutz members and city residents," Journal of Eco-
nomic Behavior & Organization, June 2006, 60 (2), 147{163.
Schechter, L., \Traditional trust measurement and the risk confound: An experi-
ment in rural Paraguay," Journal of Economic Behavior & Organization, 2007, 62
(2), 272{292.
Sosis, Richard and Bradley J. Rue, \Ideology, religion, and the evolution
of cooperation: Field experiments in Israeli Kibbutzim," Research in Economic
Anthropology, 2004, 23, 89{117.
Tan, Jonathan H.W. and Claudia Vogel, \Religion and trust: An experimental
study," Journal of Economic Psychology, December 2008, 29 (6), 832{848.
Uslaner, Eric M. and Richard S. Conley, \Civic engagement and particularized
trust," American Politics Research, 2003, 31 (4), 331{360.
131
Appendix 3.A Additional analyses
3.A.1 Robustness check: Household xed-eects model
Tables 3.13 and 3.14 present results of the regression with a specication similar to
Equation 3.1, but with household, instead of community, xed eects. The use of
household xed-eects address potential time-invariant unobservable variables at the
household level. However, as elaborated before, since in 5,444 out of 9,737 households
with more than one members, religiosity is homogeneous within the household, the
inclusion of household xed eects might remove too much of the variation.
Nonetheless, we nd that the relations between religiosity and cooperative atti-
tudes are robust to the household xed eects. When we compare the religiosity
coecients in both specications, the pattern of the ndings from the community
xed-eects model generally holds. With household xed eects, however, the reli-
gious dummy on trust of strangers is no longer positive and statistically signicant;
instead, the coecient is negative, albeit small and statistically insignicant.
3.A.2 Inter-religion dierences in religiosity eects for the
very religious
To complement our analysis of inter-religion dierences of the associations between
religiosity and attitudes, we estimated a regression similar to that specied in Equa-
tion 3.9, but included interactions with the \very religious" instead of the \religious"
dummy. These interactions provided an estimate of inter-religion dierences in the
marginal eects of being very religious on cooperative attitudes. The results are
presented in Table 3.15.
132
Table 3.13: Community Cohesion & Trust Beliefs
Willing
to help
Village is safe [. . . ] Trust nbr. to watch Trust [. . . ] to return lost wallet
generally at night kid(s) house neighbors strangers police
(1) (2) (3) (4) (5) (6) (7) (8)
Religiosity:
Religious/ very religious 0.016 0.020

-0.003 0.019 -0.000 0.076

-0.008 0.108

(1.58) (2.00) (-0.24) (0.87) (-0.01) (3.04) (-0.38) (3.85)
Very religious 0.132

0.153

0.081

0.057 0.063

0.108

-0.076

0.161

(6.39) (7.56) (3.80) (1.56) (2.62) (2.46) (-1.97) (3.36)
Male 0.043

0.021

0.100

0.049

0.051

0.029

0.041

-0.112

(6.80) (3.54) (16.21) (4.32) (6.68) (1.91) (3.05) (-6.63)
Age:
25 years old -0.009 0.014 0.052

0.029 0.033

0.031 -0.001 -0.094

(-0.79) (1.20) (4.40) (1.15) (2.31) (1.11) (-0.02) (-2.97)
45 years old -0.001 0.013 0.022

0.062

0.018 -0.007 0.036 0.002
(-0.05) (1.28) (2.13) (3.06) (1.40) (-0.24) (1.52) (0.08)
65 years old -0.036

-0.027

-0.024 -0.008 -0.015 0.006 0.046 0.022
(-2.11) (-1.74) (-1.46) (-0.28) (-0.77) (0.14) (1.06) (0.42)
Education:
Some junior high school 0.041

-0.005 -0.024

-0.021 -0.011 0.065

0.027 0.041
(3.51) (-0.41) (-2.13) (-0.94) (-0.81) (2.40) (1.11) (1.33)
Some senior high school 0.024

-0.003 -0.003 -0.021 -0.004 0.018 0.031 0.016
(2.01) (-0.24) (-0.24) (-0.90) (-0.26) (0.66) (1.22) (0.52)
Some college 0.026 0.011 0.026 -0.038 0.012 0.031 0.082

-0.115

(1.55) (0.74) (1.61) (-1.22) (0.61) (0.86) (2.27) (-2.83)
Risk aversion 0.003 0.002 0.003 -0.010

-0.002 -0.004 -0.011

-0.003
(1.16) (0.81) (0.88) (-1.90) (-0.52) (-0.59) (-1.71) (-0.45)
Patience 0.009

-0.004 -0.006 0.002 0.000 -0.006 0.004 0.010
(1.82) (-0.84) (-1.27) (0.25) (0.03) (-0.53) (0.47) (0.84)
Married 0.014 0.009 0.014 -0.036 -0.010 -0.002 -0.002 0.006
(1.31) (0.83) (1.25) (-1.47) (-0.76) (-0.06) (-0.08) (0.20)
Constant 3.054

3.020

2.892

2.705

2.834

2.898

1.525

2.786

(141.95) (143.49) (138.33) (57.55) (88.24) (63.00) (33.18) (50.72)
Household xed eects Yes Yes Yes Yes Yes Yes Yes Yes
Religion xed eects Yes Yes Yes Yes Yes Yes Yes Yes
N 28932 28929 28927 21759 28930 28333 27411 26832
Adj. R
2
0.173 0.201 0.177 0.243 0.171 0.255 0.162 0.221
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
y
Spline coecients are for the slope of the interval. The omitted education category is "Some primary or no school". The
omitted age category is "15-24 years old". Standard errors are robust and clustered at the household level.
133
Table 3.14: Discriminative Trust & Tolerance
Trust [. . . ] more Tolerate non-corlgn living in [. . . ] Tolerate non-corlgn to [. . . ]
corelgn coethnics village neighbor house marry rltv. bld h. wrshp
(1) (2) (3) (4) (5) (6) (7)
Religiosity:
Religious/ very religious 0.050

0.040

-0.042

-0.056

-0.075

-0.085

-0.059

(3.17) (2.59) (-3.20) (-4.13) (-4.40) (-4.19) (-3.19)
Very religious 0.113

0.035 0.018 -0.001 -0.050 -0.096

-0.063

(4.03) (1.24) (0.69) (-0.02) (-1.58) (-2.96) (-1.84)
Male -0.034

-0.055

0.019

0.015

0.033

0.048

0.007
(-3.75) (-6.16) (2.36) (1.74) (3.23) (4.14) (0.61)
Age:
25 years old 0.010 -0.017 -0.016 -0.034

-0.036

-0.056

-0.030
(0.57) (-0.98) (-1.07) (-2.23) (-1.90) (-2.54) (-1.43)
45 years old 0.050

0.031

-0.023

-0.040

-0.053

-0.012 -0.030
(3.18) (2.00) (-1.65) (-2.67) (-2.88) (-0.58) (-1.52)
65 years old -0.006 0.016 -0.030 -0.023 -0.048 -0.023 0.008
(-0.23) (0.68) (-1.28) (-0.92) (-1.61) (-0.67) (0.24)
Education:
Some junior high school -0.025 -0.038

0.062

0.070

0.046

-0.006 0.001
(-1.52) (-2.30) (4.20) (4.46) (2.44) (-0.26) (0.06)
Some senior high school -0.078

-0.085

0.023 0.012 0.015 -0.043

0.017
(-4.57) (-4.93) (1.61) (0.81) (0.78) (-2.02) (0.85)
Some college -0.072

-0.093

0.027 0.029 0.058

-0.021 0.019
(-3.01) (-3.99) (1.48) (1.44) (2.37) (-0.77) (0.66)
Risk aversion 0.008

0.012

-0.002 -0.004 -0.005 0.005 -0.006
(1.99) (2.94) (-0.64) (-0.97) (-1.03) (1.02) (-1.11)
Patience 0.003 -0.012

0.003 0.001 -0.003 -0.004 -0.008
(0.53) (-1.79) (0.53) (0.23) (-0.37) (-0.54) (-1.08)
Married -0.016 0.009 -0.019 -0.016 -0.069

-0.113

-0.037

(-1.00) (0.56) (-1.26) (-1.01) (-3.66) (-5.30) (-1.83)
Constant 2.778

2.679

2.801

2.800

2.540

1.937

2.386

(91.46) (89.15) (109.90) (106.06) (80.56) (50.50) (65.80)
Household xed eects Yes Yes Yes Yes Yes Yes Yes
Religion xed eects Yes Yes Yes Yes Yes Yes Yes
N 28931 28931 28932 28932 28931 28931 28931
Adj. R
2
0.260 0.269 0.334 0.351 0.389 0.376 0.374
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
y
Spline coecients are for the slope of the interval. The omitted education category is "Some primary or no school". The
omitted age category is "15-24 years old". Standard errors are robust and clustered at the household level.
134
As in the case of the religious dummy, we do not nd systematic inter-religion
dierences in these associations for community cohesiveness outcomes. Meanwhile,
being very religious is less likely to increase trust of coreligionists among Protestants
and Hindus, compared to Muslims and Catholics. It is also less likely to increase trust
of coethnics among Protestants. Also, as in the case with the interactions with the
religion dummy, we nd that negative associations between being very religious and
religious tolerance among non-Muslims tend to be systematically weaker than those
among Muslims. For tolerance of inter-faith marriages and allowing non-coreligionists
build their house of worship, although in many cases the point estimates of these inter-
religion dierences are similar to those in the \religious dummy" specication, they
are often not statistically signicant.
3.A.3 Pooling test of inter-gender dierences
Tables 3.16 and 3.17 present results of the pooling test for the inter-gender dierences
in the correlates. Tests of pooling by gender show for all outcomes, the pooling
hypotheses are rejected. The critical value for rejection for most outcomes is at
1%, except for trust beliefs of neighbors and tolerance of non-coreligionists' house of
worship (at 5%) and trust of coreligionists at (10%).
135
Table 3.15: Inter-Religion Differences in the Associations between
Religiosity and Attitudes
Religious/
Very
religious
Very
religious
Very religious [. . . ] Statistics
Catholic Protestant Hindu Buddhist
Num.
of obs.
P-val of
joint test:
Religions
(1) (2) (3) (4) (5) (6) (7) (8)
Willing to help 0.013

0.163

-0.013 -0.053 0.015 0.068 28932 0.705
(1.99) (9.88) (-0.18) (-1.22) (0.49) (0.40)
Vilage is safe [...]
generally 0.026

0.174

-0.080 -0.126

0.042 -0.068 28929 0.106
(4.83) (9.57) (-0.80) (-2.30) (0.90) (-0.19)
at night 0.008 0.082

0.021 -0.033 0.073

-0.138 28927 0.419
(1.32) (4.75) (0.22) (-0.49) (1.74) (-0.59)
Trust neighbor to watch [...]
kid(s) 0.012 0.044

-0.137 0.083 0.148

0.599

21759 0.037
(1.01) (1.97) (-1.19) (0.84) (1.97) (2.21)
house 0 0.050

0.037 -0.004 0.076 0.401 28930 0.296
(0.05) (3.00) (0.41) (-0.05) (1.50) (1.61)
Trust [...] to return wallet
neighbors 0.108

0.127

-0.193 -0.233 -0.107 -0.224 28333 0.211
(6.49) (3.76) (-1.43) (-1.54) (-1.46) (-0.52)
strangers 0.027

-0.044

-0.195

0.101 -0.084 -0.298

27411 0.025
(2.05) (-1.70) (-1.82) (0.77) (-1.15) (-2.62)
police 0.138

0.187

0.009 -0.074 -0.107 0.248 26832 0.704
(7.01) (6.14) (0.05) (-0.60) (-1.26) (0.49)
Trust [...] more
coreligionists 0.086

0.180

0.031 -0.129

-0.101

-0.365 28931 0.044
(8.44) (9.09) (0.26) (-1.85) (-1.96) (-1.55)
coethnics 0.068

0.074

-0.059 -0.186

0.053 0.072 28931 0.069
(6.36) (3.44) (-0.49) (-2.57) (0.92) (0.39)
Tolerate non-coreligionist
to live in [...]
village -0.058

-0.035

0.206

0.081 0.149

0.315 28932 0
(-6.01) (-1.90) (3.21) (1.60) (3.99) (1.62)
neighborhood -0.080

-0.035

0.185

0.139

0.092

0.339

28932 0.001
(-8.44) (-1.73) (2.70) (2.46) (2.62) (1.77)
house -0.119

-0.130

0.235

0.282

0.180

0.070 28931 0
(-10.80) (-5.97) (2.88) (4.61) (3.23) (0.63)
Tolerate non-coreligionist
to [...]
marry relative -0.131

-0.174

-0.041 0.138 0.239

0.339 28931 0
(-9.68) (-6.97) (-0.30) (1.24) (4.81) (1.18)
build house of worship -0.101

-0.111

0.067 0.121 -0.010 0.415 28931 0.221
(-8.55) (-4.56) (1.10) (1.57) (-0.12) (1.42)
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01.
Each row presents results from separate regressions for model with community and religion xed eects. Standard errors are robust and
clustered at the community level. Included variables not shown: sex, dummy variables for age and education categories, risk and time
preference, married status, linear spline for log PCE, and a constant.
136
Table 3.16: Pooling Test: Inter-gender Differences in Community Cohesion
& Trust Beliefs
Willing
to help
Village is safe [. . . ] Trust nbr. to watch Trust [. . . ] to return lost wallet
generally at night kid(s) house neighbors strangers police
(1) (2) (3) (4) (5) (6) (7) (8)
Male 0.025 0.190 0.218 -0.243 -0.309

-0.042 -0.414 -0.783

(0.18) (1.09) (1.29) (-1.10) (-2.03) (-0.12) (-1.37) (-1.84)
Religiosity:
Religious/ very religious 0.003 0.026

0.025

0.008 -0.007 0.086

0.019 0.142

(0.40) (3.21) (2.66) (0.46) (-0.54) (3.82) (1.01) (5.56)
. . . Male 0.015 -0.002 -0.030

0.006 0.009 0.034 0.008 -0.017
(1.32) (-0.15) (-2.60) (0.31) (0.68) (1.26) (0.33) (-0.55)
Very religious 0.150

0.156

0.044

0.049

0.053

0.085

-0.077

0.139

(8.48) (7.65) (2.07) (1.98) (2.63) (2.22) (-2.22) (3.56)
. . . Male 0.029 0.036 0.105

0.024 0.021 0.027 0.050 0.072
(1.11) (1.33) (3.80) (0.62) (0.79) (0.55) (0.98) (1.35)
Age:
25 years old -0.013 0.018

0.069

0.060

0.038

0.014 -0.026 -0.082

(-1.61) (2.36) (6.90) (3.72) (3.51) (0.70) (-1.63) (-3.52)
. . . Male 0.025

0.005 -0.046

-0.023 -0.020 0.004 0.074

0.021
(1.74) (0.38) (-3.48) (-0.82) (-1.36) (0.14) (2.63) (0.55)
45 years old -0.001 0.006 0.045

0.055

0.016 0.026 0.043

0.011
(-0.14) (0.74) (5.10) (3.82) (1.43) (1.23) (2.36) (0.49)
. . . Male 0.006 0.011 -0.028

0.025 0.016 -0.012 -0.006 0.052
(0.48) (0.95) (-2.31) (1.29) (1.04) (-0.41) (-0.23) (1.58)
65 years old -0.042

-0.009 0.012 -0.017 -0.032

-0.020 0.045 -0.116

(-3.34) (-0.83) (1.01) (-0.78) (-1.88) (-0.56) (1.45) (-2.96)
. . . Male -0.002 -0.034

-0.036

0.001 0.002 0.033 0.027 0.160

(-0.13) (-2.13) (-2.16) (0.04) (0.07) (0.67) (0.65) (2.93)
Education:
Some junior high school 0.025

-0.021

-0.034

-0.042

-0.028

0.040

0.024 0.069

(2.90) (-2.39) (-3.22) (-2.78) (-2.60) (1.80) (1.27) (2.75)
. . . Male 0.012 0.027

0.037

-0.011 0.011 0.025 0.010 -0.044
(0.93) (2.24) (2.88) (-0.51) (0.73) (0.78) (0.39) (-1.24)
Some senior high school 0.034

-0.005 -0.008 -0.079

-0.020

0.047

0.026 -0.002
(3.55) (-0.54) (-0.73) (-4.62) (-1.75) (2.07) (1.33) (-0.07)
. . . Male -0.031

-0.005 0.006 0.072

0.002 -0.009 0.029 -0.007
(-2.31) (-0.37) (0.43) (3.02) (0.11) (-0.31) (1.07) (-0.19)
Some college 0.007 0.002 0.033

-0.021 0.012 0.067

0.146

-0.051
(0.53) (0.19) (2.38) (-0.88) (0.75) (2.26) (4.94) (-1.62)
. . . Male 0.024 0.006 -0.023 -0.059

-0.029 0.016 0.004 -0.015
(1.38) (0.40) (-1.35) (-1.94) (-1.43) (0.42) (0.10) (-0.36)
Risk aversion 0.002 0.001 -0.000 -0.011

-0.004 0.005 -0.001 -0.001
(0.76) (0.35) (-0.17) (-2.42) (-1.42) (0.69) (-0.25) (-0.15)
. . . Male 0.002 -0.001 0.000 0.002 -0.000 -0.016

-0.005 0.002
(0.52) (-0.32) (0.03) (0.45) (-0.05) (-2.12) (-0.70) (0.27)
Patience 0.009

0.004 -0.013

0.005 -0.003 0.001 0.011 0.050

(2.54) (1.02) (-3.51) (0.71) (-0.69) (0.07) (1.60) (5.24)
. . . Male 0.009

-0.008 0.013

0.001 0.008 0.017 0.001 -0.019
(1.89) (-1.60) (2.65) (0.13) (1.30) (1.49) (0.13) (-1.46)
Married 0.001 0.005 0.043

-0.045

-0.028

0.006 -0.037

-0.066

(0.18) (0.74) (4.91) (-3.12) (-2.83) (0.33) (-2.27) (-3.29)
. . . Male 0.005 -0.000 -0.053

-0.006 0.021 0.001 0.033 0.057
(0.42) (-0.04) (-4.54) (-0.26) (1.56) (0.03) (1.27) (1.64)
Log. (1 + PCE) Spline:
Below median
y
0.018

0.011 0.007 -0.008 -0.010 0.020 0.011 -0.029
(2.06) (1.36) (0.72) (-0.49) (-0.94) (0.87) (0.56) (-1.05)
. . . Male -0.002 -0.014 -0.005 0.023 0.026

0.003 0.028 0.051
(-0.21) (-0.98) (-0.37) (1.31) (2.13) (0.12) (1.15) (1.48)
Above median
y
0.005 -0.008 -0.019

-0.008 -0.024

-0.045

0.011 -0.025
(0.75) (-1.02) (-2.22) (-0.52) (-2.37) (-2.55) (0.65) (-1.22)
. . . Male -0.011 -0.004 0.034

0.000 0.011 0.011 -0.010 0.033
(-1.09) (-0.41) (3.10) (0.00) (0.93) (0.45) (-0.48) (1.29)
Constant 2.862

2.880

2.778

2.792

3.003

2.621

1.337

3.114

(26.80) (28.18) (22.77) (14.10) (23.99) (9.24) (5.68) (9.21)
Community xed eects Yes Yes Yes Yes Yes Yes Yes Yes
Religion xed eects Yes Yes Yes Yes Yes Yes Yes Yes
P-val of joint test of:
Pooling hypothesis 0.000 0.000 0.000 0.000 0.000 0.024 0.005 0.000
N 28932 28929 28927 21759 28930 28333 27411 26832
Adj. R
2
0.085 0.092 0.098 0.102 0.058 0.105 0.062 0.086
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
y
Spline coecients are for the slope of the interval. The omitted education category is "Some primary or no school".
The omitted age category is "15-24 years old". Standard errors are robust and clustered at the community level.
137
Table 3.17: Pooling Test: Inter-gender Differences in Discriminative Trust
& Tolerance
Trust [. . . ] more Tolerate non-corlgn living in [. . . ] Tolerate non-corlgn to [. . . ]
corelgn coethnics village neighbor house marry rltv. bld h. wrshp
(1) (2) (3) (4) (5) (6) (7)
Male 0.009 0.012 -0.118 -0.175 -0.711

-0.404 -0.280
(0.05) (0.06) (-0.56) (-0.82) (-2.96) (-1.58) (-1.03)
Religiosity:
Religious/ very religious 0.061

0.036

-0.044

-0.058

-0.102

-0.114

-0.081

(4.63) (2.45) (-3.25) (-4.59) (-6.86) (-6.23) (-4.87)
. . . Male 0.042

0.054

-0.029

-0.043

-0.036

-0.029 -0.038

(2.55) (3.34) (-2.03) (-2.99) (-2.02) (-1.27) (-1.91)
Very religious 0.165

0.083

-0.031 -0.038

-0.101

-0.158

-0.146

(6.57) (3.49) (-1.36) (-1.73) (-3.71) (-5.29) (-4.85)
. . . Male -0.015 -0.029 0.051 0.053 0.021 0.041 0.095

(-0.45) (-0.87) (1.40) (1.53) (0.51) (1.05) (2.32)
Age:
25 years old 0.029

0.000 0.003 -0.021

-0.060

-0.084

-0.038

(2.17) (0.02) (0.28) (-1.84) (-4.01) (-5.07) (-2.37)
. . . Male -0.045

-0.047

-0.016 0.024 0.049

0.048

0.020
(-2.17) (-2.42) (-0.91) (1.37) (2.25) (2.02) (0.85)
45 years old 0.032

0.007 -0.042

-0.054

-0.069

-0.046

-0.022
(2.68) (0.63) (-3.70) (-4.28) (-4.48) (-2.76) (-1.40)
. . . Male 0.001 0.012 0.035

0.043

0.017 0.004 0.025
(0.04) (0.75) (2.29) (2.72) (0.86) (0.17) (1.14)
65 years old -0.001 0.031

-0.048

-0.028 -0.062

0.020 -0.040
(-0.04) (1.92) (-2.70) (-1.57) (-2.57) (0.74) (-1.61)
. . . Male 0.003 -0.018 0.035 -0.002 0.038 0.037 0.066

(0.12) (-0.73) (1.49) (-0.08) (1.13) (0.97) (1.97)
Education:
Some junior high school -0.064

-0.085

0.094

0.101

0.047

-0.025 0.009
(-4.62) (-6.40) (6.90) (7.75) (2.75) (-1.34) (0.47)
. . . Male 0.021 0.039

-0.049

-0.042

0.006 -0.004 0.007
(1.11) (2.03) (-2.97) (-2.30) (0.25) (-0.14) (0.28)
Some senior high school -0.053

-0.072

0.008 0.007 0.008 -0.060

0.002
(-3.82) (-5.39) (0.63) (0.51) (0.44) (-3.13) (0.11)
. . . Male -0.040

-0.048

0.036

0.027 -0.009 0.025 0.041
(-1.93) (-2.57) (2.06) (1.44) (-0.38) (0.96) (1.63)
Some college -0.072

-0.094

0.048

0.042

0.021 -0.060

0.047

(-3.87) (-4.89) (3.71) (3.24) (1.02) (-2.57) (2.15)
. . . Male 0.006 -0.041

-0.012 -0.020 -0.007 -0.027 -0.021
(0.22) (-1.71) (-0.68) (-1.03) (-0.25) (-0.87) (-0.73)
Risk aversion 0.012

0.011

-0.004 -0.005 -0.010

0.004 -0.003
(3.36) (3.21) (-1.02) (-1.30) (-2.16) (0.87) (-0.53)
. . . Male -0.002 0.001 0.008

0.011

0.009 0.006 0.002
(-0.40) (0.26) (1.82) (2.32) (1.53) (0.98) (0.33)
Patience 0.009 -0.009

0.006 0.002 -0.014

-0.009 -0.009
(1.59) (-2.03) (1.14) (0.34) (-2.21) (-1.33) (-1.59)
. . . Male -0.004 -0.005 0.005 0.009 0.004 -0.009 0.005
(-0.59) (-0.79) (0.76) (1.41) (0.39) (-0.91) (0.54)
Married 0.001 0.025

-0.033

-0.012 -0.051

-0.030

-0.045

(0.13) (2.33) (-3.50) (-1.35) (-4.23) (-2.07) (-3.12)
. . . Male -0.016 -0.018 0.023 -0.007 -0.004 -0.069

0.027
(-0.89) (-1.02) (1.60) (-0.48) (-0.21) (-3.17) (1.24)
Log. (1 + PCE) Spline:
Below median
y
-0.002 -0.026

0.023

0.019 -0.009 -0.046

0.009
(-0.15) (-2.31) (1.71) (1.37) (-0.56) (-2.74) (0.48)
. . . Male 0.000 -0.004 0.010 0.013 0.056

0.038

0.018
(0.01) (-0.29) (0.59) (0.74) (2.87) (1.83) (0.82)
Above median
y
-0.008 -0.035

0.018

0.010 -0.001 -0.030

0.020
(-0.70) (-2.94) (1.67) (0.94) (-0.06) (-1.97) (1.38)
. . . Male -0.006 0.035

-0.010 0.012 0.012 0.036

0.010
(-0.42) (2.34) (-0.77) (0.90) (0.70) (1.87) (0.58)
Constant 2.762

3.022

2.504

2.543

2.731

2.564

2.294

(17.82) (21.78) (14.76) (14.94) (13.60) (12.47) (10.44)
Community xed eects Yes Yes Yes Yes Yes Yes Yes
Religion xed eects Yes Yes Yes Yes Yes Yes Yes
P-val of joint test of:
Pooling hypothesis 0.084 0.000 0.000 0.000 0.000 0.000 0.049
N 28931 28931 28932 28932 28931 28931 28931
Adj. R
2
0.156 0.179 0.221 0.251 0.262 0.246 0.265
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
y
Spline coecients are for the slope of the interval. The omitted education category is "Some primary or no school".
The omitted age category is "15-24 years old". Standard errors are robust and clustered at the community level.
138
Chapter 4
Religious Diversity, Segregation and Cooperative
Attitudes in Indonesia
Abstract
This chapter examines how variations in community-level religious diversity
and segregation are associated with dierent kinds of cooperative attitudes
in contemporary Indonesia. Consistent with previous empirical studies in
economics and political science in the United States and other countries, I
nd in Indonesia that individuals are more cooperative and trusting of their
community members in more religiously homogeneous communities. At the
same time { and in support of the optimal contact hypothesis of Allport
(1979 [1954]) { individuals in more homogeneous communities exhibit more
in-group trust and are less tolerant of members of the religious out-groups.
I also nd that the inclusion of segregation measures can substantially aect
the size of the diversity coecients. Conditional on diversity, the segrega-
tion coecients are signicant and their signs are opposite those of religious
diversity for some of the outcomes.
The ndings of the previous chapter suggest that individuals' religious identities tend
to strengthen particularized or within-group cooperative attitudes. A natural exten-
sion to this examination is to examine the role of these identities at the aggregate
level { namely, whether the religious heterogeneity of communities are correlated with
individuals' overall willingness to cooperate in a way similar to ethnic or income het-
erogeneity. In this chapter, I empirically examine whether this implication is borne
out by the data for Indonesia.
This chapter, henceforth, is related to the empirical literature on the link be-
139
tween ethno-religious fragmentation and social capital. The analysis here, however,
introduces two variations to the standard fragmentation literature. First, its focus is
on religious, instead of ethnic, heterogeneity. The reason for this focus may not be
immediately obvious, given that in addition to being religiously diverse, Indonesia is
also country that is ethnically diverse, with hundreds of ethnic groups spread across
its archipelago. However, religious identity appears to play a more important role
throughout its recent history.
1
As argued by Bertrand (2004, p.110) in his discus-
sion of sources of ethno-religious con
icts in Indonesia: \In the Indonesian context,
one's religious identity is often more important than one's ethnic identity as Javanese,
Ambonese, Madurese, or other."
Meanwhile, the second variation is the inclusion of religious segregation in addi-
tion to the fragmentation (or diversity) measures. Recent literature, reviewed below,
suggests that the degree of segregation may play a more important role than diversity
in in
uencing social capital. In addition, the inclusion of segregation measures allow
for an examination of the likely role of social networks in aecting attitudes.
The next section begins with the literature review. It is followed by Section 4.2
on data and measurements. Section 4.3 discusses the results. Section 4.4 concludes
the chapter.
4.1 Literature review
Existing studies on the associations between community heterogeneity and social
capital documented negative associations between community heterogeneity and the
1
In the words of Bertrand (2004, p.72): \During the New Order period, religious identity emerged
as the most important form of ethnic identication." Here, Bertrand adopts an inclusive denition
of ethnic identity that includes religious identity. This observation is corroborated in our sample:
Table 3.1 suggests that people tend to exhibit greater trust of coreligionists compared to that of
coethnics, and the dierence between the two is statistically signicant.
140
various measures of civic engagements (Alesina and La Ferrara, 2000; Costa and
Kahn, 2003), trust (Glaeser et al., 2000; Alesina and La Ferrara, 2002), and the
willingness to provide public goods (Vigdor, 2004; Miguel and Gugerty, 2005) or
support redistribution policies (Luttmer, 2001). Most of the literature focuses on
diversity, typically measured using the fragmentation index. More recent literature,
however, begins to provide evidence that segregation may play a more important
role than diversity in in
uencing the quality of governance (Alesina and Zhuravskaya,
2011) and social capital (Uslaner, 2010; Rothwell, 2010).
Inter-group discrimination may account for the link between heterogeneity and
lower cooperation (e.g., Alesina and La Ferrara, 2002). Individuals may discrimi-
nate out of either preference or prejudice (or false expectations). Social interactions
can aect discrimination by, among others, facilitating statistical discrimination or
through network eects (Arrow, 1998; Fafchamps, 2004). In the former, if people
do not interact in groups, then those interactions would allow individuals to assess
each other's qualities (or \types") based on their observable characteristics, in which
religion may be one. In this case, statistical intergroup discrimination occurs only if
individuals in dierent groups have dierent hidden characteristics.
On the other hand, if individuals tend to interact more within groups or networks,
these interactions may result in discrimination, even when individuals do not have a
preference for discrimination and there is no dierential hidden characteristics across
groups. Why? For one, within-network (or in-group) interactions facilitate better
transmission of information (Granovetter, 2005; Fafchamps, 2004). As a result, indi-
viduals can screen the \good" from the \bad" types among the in-groups better than
among the out-groups.
2
Moreover, denser networks allow for better enforcements
2
Fafchamps (2004) elaborates a game-theoretic model of trust-based exchanges in which infor-
mation propagated through ethnic-based (or religion-based) social networks can act to sustain an
equilibrium with discrimination among individuals with no preference for discrimination even in the
141
of cooperative norms among the in-groups. Using eld experiments among subjects
from a slum in Kampala, Uganda, Habyarimana et al. (2007) nd that better within-
ethnicity enforcement of cooperative norms may be one of the key explanations for
why ethnic diversity lowers public good provision.
In the presence of network eects, diversity may reduce overall level of coopera-
tion in the community. On the other hand, diversity can also foster better intergroup
cooperation by softening prejudice. The optimal contact hypothesis of Allport (1979
[1954]) suggests that under optimal conditions, contacts with people who are dierent
will break down stereotypes and reduce prejudice. Henceforth, diversity can poten-
tially reduce discriminative trust and intolerance. A large meta-analytic study of
intergroup contacts by Pettigrew and Tropp (2006) provides support for this optimal
contact hypothesis. Meanwhile, making a clever use of the lottery nature of Hajj
visa allocation in Pakistan to identify the eect of pilgrimage on attitudes, Clinging-
smith et al. (2009) found that the pilgrimage increases religious tolerance ve to eight
months after participants returned home. They argue that this increased tolerance is
a result of their interactions with other hajjis from around the world.
4.2 Data and measurements
4.2.1 Data
For this analysis, I combine four national-level datasets. As in the previous chapter,
the outcome, and individual- and household-level variables are obtained from IFLS4.
I merge IFLS4 with three other national-level datasets to obtain the various com-
munity variables. First, from the 2000 Indonesian Population Census microdata, I
constructed various community religious composition measures from the individual
absence of dierential hidden characteristics across groups.
142
religion identiers. Second, from the 2000 Indonesian Poverty Map, I obtained mea-
sures of community-level expenditure inequality. Finally, I add various topographical,
demographical and other village characteristics from the 2005 Village Potential statis-
tics (or Podes 2005).
3
4.2.2 Measurements
Diversity and segregation
For diversity, instead of using the usual fractionalization index, I follow Reardon et
al. (2000) in using the diversity index rst proposed in Theil and Finizza (1971) that
makes use of the entropy of the discrete probability distribution of groups in the
unit of analysis.
4
That is, in community i, the entropy of the discrete probability
distribution of religion in a village is calculated as follows:
H
i
=
R
X
r
s
ir
:ln

1
s
ir

(4.1)
wheres
ir
indicates the share of population with religionr in communityi. The index
can take a value of between zero (perfectly homogeneous) and the natural log of the
number of distinct religious groups in the community.
Meanwhile, we measure segregation using the Mutual Information Index that is
also based on the entropy measure of diversity. Essentially, the Mutual Information
Index measures the dierence in the entropy of the community's religious distribution
with the weighted average of the entropy of the sub-communities. In their comparisons
of the properties of dierent segregation measures, Reardon and Firebaugh (2002)
and Frankel and Volij (2011) conclude that the Mutual Information Index is the most
3
See Section 4.B for the details on these datasets and their merging process.
4
Despite not using the usual fractionalization index, the correlation between the entropy index
and the fractionalization index in my data is 99.1%.
143
well-behaved.
5
Hence, for community i and its subcommunities, indexed by n, the
segregation index is calculated as:
M
i
=H
i

X
n2N

n
H
n
: (4.2)
where
n
is the population weight for subcommunity n. A larger value indicates a
more segregated community. Like the diversity index, the segregation index can take
a value of between zero and the natural log of the number of distinct groups in the
community.
I use the 2000 population census microdata to construct these indices both at
the subdistrict and village level. The subdistrict segregation index compares the
subdistrict entropy with the population-weighted average of the entropy of its villages.
Hence, a more segregated subdistrict is one where people of dierent religions are more
clustered in the dierent villages. Meanwhile, the village segregation index compares
the village entropy with the population-weighted average of the entropy of the census
blocks within the village. A more segregated village, therefore, is one where people of
dierent religions are clustered in the dierent census blocks. As shown in Figure 4.1,
there are signicant variations in village- and subdistrict-level diversity as well as
segregation among the communities included in the dataset.
Other regressors
At the community level, in addition to the community diversity and segregation
measures, I include the urban status of the community and dummy variables of its
5
Frankel and Volij (2011) found that the Mutual Information Index did not satisfy the com-
position invariance property. Composition invariance property states that the segregation of a
community should not change when the number of students from a particular religion in the sub-
communities is multiplied by the same number across the community. However, in this analysis,
segregation is used to analyze the eect of exposure on attitudes. Coleman et al. (1982) argue that
this property is unnecessary in this case.
144
Figure 4.1: Distributions of the diversity and segregation measures
topographical characteristic (i.e., on a coast,
atland, hill, or valley), log population
density in the village, whether the village has recently experienced a natural disaster,
and distance from subdistrict and district capitals. Olken (2009) found that television
and radio reduce social capital. Hence, I also included a dummy variable for whether
the village can receive a broadcast from the national public television and a regional
television. In the base specication, I also include subdistrict PCE Gini obtained
from the 2000 Poverty Map.
4.3 Results
The theory and evidence reviewed above suggest two channels through which commu-
nity diversity and segregation may aect cooperation: Network eects and (optimal)
145
Table 4.1: Summary Statistics: Community Regressors
Num.
of obs.
Mean
Std.
dev.
Median IQR Min Max
Community-level variables
Urban 27844 0.54 0.50 1 1 0 1
Topography:
Plain 27844 0.80 0.40 1 0 0 1
Coast 27844 0.10 0.30 0 0 0 1
Valley 27844 0.01 0.10 0 0 0 1
Hill 27844 0.09 0.28 0 0 0 1
Population density (pop/ha) 27844 51.75 106.49 17.9 44.6 0.0089 2304
Receives broadcast of:
Public TV station 27844 0.92 0.27 1 0 0 1
Private TV station 27844 0.85 0.36 1 0 0 1
Natural disaster in last 5 years 27844 0.53 0.50 1 1 0 1
Distance to:
Subistrict capital (km) 27844 4.61 6.02 3 3.80 0.10 99
District capital (km) 27844 19.85 22.00 12 23.5 0.10 243
Village-level:
Diversity 27844 0.22 0.28 0.094 0.35 0 1.3
Segregation 27844 0.04 0.07 0.015 0.055 0 .61
Subdistrict-level:
Diversity 27844 0.26 0.27 0.18 0.36 0 1.3
Segregation 27844 0.04 0.06 0.014 0.031 0 .52
PCE Gini 27844 0.24 0.03 0.24 0.050 0.14 .36
146
inter-group contact. In the case of the former, under the assumption of mostly trust-
based exchanges, diversity may weaken overall cooperation by weakening intra-group
information transmission and norm enforcement. Moreover, denser networks may
strengthen cooperation by strengthening these intra-group mechanisms. Therefore,
we expect to nd greater community trust in more homogeneous communities, and
we expect this trust to be stronger among people living in segregated communities.
However, on the
ip side, diversity means a higher likelihood of contacts with
those who are dierent, and frequent contacts may break down stereotypes, increase
trust, and reduce prejudice. Diversity, therefore, can lessen prejudice and this po-
tentially facilitate greater inter-group (and overall) cooperation. Under this premise,
residential segregation will likely strengthen inter-group prejudice.
The net eect of heterogeneity on community trust is theoretically unclear and is
therefore an empirical matter. On the other hand, the theoretical prediction based
on Allport's contact hypothesis is clearer: Diversity is expected to reduce inter-group
discriminative trust and increase tolerance while segregation is expected to have the
reverse eect.
Before presenting the results, it is important to note the potential endogeneity
of these community heterogeneity variables. For instance, its is plausible that the
observed religious diversity in the community was in fact the outcome of its more
tolerant residents. On the one hand, these measures of community heterogeneity were
derived from the national census dataset, which was collected eight years prior to the
outcomes of interest. This reduces concerns of contemporaneous reverse causality.
However, this may not solve the issue that potentially arises if both attitudes and
community compositions are persistent over time.
147
4.3.1 Does segregation matter?
To estimate the level eects of community compositions, I estimated the following
specication:
Y
r
ijkl
= +
X
r=3;4

r
1
:1(rlgs
i
r) +
d
:div
k
+
s
:seg
k
+ X
i
:
i
+ X
j
:
j
+ X
k
:
k
+
d
+
l
+"
ijkl
(4.3)
where
l
is the district xed eects.
Table 4.2 present the coecients for the village and subdistrict heterogeneity vari-
ables from this regression.
6
On the left halves of these tables, we have the coecients
of the community diversity variables when the segregation variables are excluded.
Meanwhile, on the right halves are the coecients for both the community and the
segregation variables.
In many cases, failing to include segregation in the regression often changes the
inference on the link between diversity and attitudes.
7
For trust beliefs and inter-
religious tolerance, the exclusion of segregation variables may have introduced biases.
For instance, in case of trust beliefs, the diversity coecient in the trust-of-neighbors
regression became negative when segregation is introduced for regressions with village-
level heterogeneity variables. Similarly, with the inclusion of segregation, diversity
coecients for the regressions on the trust beliefs of strangers and the police became
signicant and its magnitudes increased. In the case of tolerance, the inclusions of the
village segregation variable lead to increases in the magnitude of the village diversity
coecient in all cases except for the tolerance of having non-coreligionists live in
6
For the coecients of other community variables, see Tables 4.7 and 4.8 in the appendix.
7
The correlation between the village diversity and segregation variables is 0.76, while the corre-
lation between the subdistrict diversity and segregation variables is 0.58.
148
the same village, and in all tolerance outcome for the regressions with subdistrict
heterogeneity variables. In none of these cases did the sign change. In the following
section, I will focus on the specication that includes the segregation variable.
4.3.2 Diversity and segregation
Religious diversity appears to be associated with lower trust of strangers and the
police. Meanwhile, the associations between diversity and community cohesiveness
appear to be imprecisely estimated: The signs on the willingness to entrust neighbors
with their house or children, as well as trust beliefs of neighbors are negative, but
they are rarely statistically signicant. Interestingly, however, people living in more
diverse subdistricts tend to nd it safer to walk in their village at night.
Subdistrict diversity is associated with less discriminative trust with regards to
religion. Furthermore, people living in more diverse villages and subdistricts are also
more tolerant across all ve measures of tolerance.
Meanwhile, those living in more religiously segregated subdistricts exhibit less
willingness to help neighbors. In both of the regressions utilizing the village and
subdistrict heterogeneity variables, the segregation variables are positively associated
with belief that one's neighbors are trustworthy. They are also associated with the
belief that the police are trustworthy. In the case of the trust belief of strangers, the
coecient is only signicant for the subdistrict segregation variable.
Segregation does not appear to be correlated with discriminative trust. However,
people in segregated villages tend to be less tolerant of allowing people of dierent
faiths in their home. People in more segregated subdistricts are also less willing to
let non-coreligionists build their house of worship in their village.
Overall, the evidence, therefore, suggests some support for Allport's optimal con-
149
Table 4.2: Diversity, Segregation and Attitudes
Village Heterogeneity Subdistrict Heterogeneity
Model 1 Model 2 Num.
of
obs.
Model 3 Model 4 Num.
of
obs.
Diver-
sity
Diver-
sity
Segre-
gation
Diver-
sity
Diver-
sity
Segre-
gation
(1) (2) (3) (4) (5) (6) (7) (8)
Willing to help 0.001 -0.031 0.144 27751 -0.015 0.031 -0.194

27824
(0.04) (-1.10) (1.53) (-0.64) (0.85) (-1.79)
Vilage is safe [...]
generally -0.004 0.001 -0.024 27748 0.013 -0.008 0.090 27821
(-0.20) (0.03) (-0.33) (0.39) (-0.24) (0.63)
at night 0.055

0.048 0.033 27746 0.092

0.077

0.065 27819
(2.75) (1.54) (0.42) (2.14) (2.40) (0.37)
Trust neighbor to watch [...]
kid(s) -0.078

-0.044 -0.153 20912 -0.047 -0.115

0.280 20972
(-1.90) (-0.74) (-0.69) (-0.99) (-1.75) (1.21)
house -0.021 -0.040 0.084 27749 0 -0.008 0.036 27822
(-0.98) (-1.33) (0.83) (0.01) (-0.20) (0.28)
Trust [...] to return wallet
neighbors 0.006 -0.104 0.505

27178 0.010 -0.132 0.605

27251
(0.11) (-1.56) (2.11) (0.14) (-1.27) (2.07)
strangers -0.044 -0.077

0.153 26278 -0.074

-0.161

0.384

26349
(-1.31) (-1.69) (0.99) (-1.69) (-2.47) (2.18)
police -0.032 -0.153

0.552

25690 0.038 -0.063 0.439

25761
(-0.65) (-2.11) (2.07) (0.56) (-0.65) (1.77)
Trust [...] more
coreligionists -0.069

-0.055 -0.065 27750 -0.090 -0.133

0.182 27823
(-1.87) (-0.85) (-0.37) (-1.51) (-1.83) (1.12)
coethnics -0.115

-0.079 -0.169 27750 -0.052 -0.066 0.060 27823
(-3.16) (-1.18) (-0.89) (-0.95) (-0.91) (0.37)
Tolerate non-corlgn to live in [...]
village 0.111

0.105

0.028 27751 0.113

0.121

-0.034 27824
(3.56) (2.62) (0.19) (2.40) (2.51) (-0.21)
neighborhood 0.129

0.155

-0.118 27751 0.134

0.158

-0.102 27824
(3.49) (2.90) (-0.61) (2.31) (2.85) (-0.51)
house 0.178

0.265

-0.403

27750 0.174

0.212

-0.161 27823
(3.74) (3.93) (-2.00) (2.57) (2.86) (-0.69)
Tolerate non-corlgn to [...]
marry relative 0.192

0.216

-0.111 27750 0.200

0.249

-0.207 27823
(4.04) (3.09) (-0.62) (2.57) (2.25) (-0.82)
build house of worship 0.340

0.439

-0.456 27750 0.360

0.502

-0.595

27823
(5.44) (4.57) (-1.55) (4.01) (4.39) (-2.30)
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01.
Each row presents results from four models with district xed eects. Standard errors are robust and clustered at the
subdistrict level. Included variables not shown for all four models: dummy variables for religiosity, age and education
categories, sex, risk and time preference, married status, linear spline for log PCE, urban/rural status, log population density,
dummy variables for topography, whether village experienced natural disaster in the last ve years, receptions of public and
local television signals, distance from subdistricts and districts, and a constant.
150
tact hypothesis that intergroup contact can reduce prejudice. However, it does not
appear that improved intergroup relations are able to compensate the negative over-
all eects of diversity on particularized and generalized trust, which in part might
be attributable to network eects. At any rate, these results need to be interpreted
carefully given the potential endogeneity of residential choices.
4.3.3 Heterogenous responses to community heterogeneity
The analysis so far provides evidence of positive correlations between religiosity and
in-group preferences. As such, individuals' behavioral and attitudinal responses to
the religious compositions of their communities may vary with their level of religiosity.
In other words, given the association between religiosity and in-group preferences, the
religious may be less willing to cooperate in more religiously diverse communities and,
conversely, more willing to cooperate in more segregated communities. To examine
this possibility, we regress the following specication:
Y
r
ijk
= +
X
r=3;4

r
1
:1(rlgs
i
r) +
3
d
:1(rlgs
i
3)div
k
+
3
s
:1(rlgs
i
3)seg
k
+ X
i
:
i
+ X
j
:
j
+ X
k
:
k
+
d
+
k
+"
ijk
(4.4)
and examine the coecients of the interactions between thereligious=very religious
dummy and community diversity and segregation in the model with community and
religion xed eects.
8
The results in Table 4.3 support this hypothesis. The strength of the association
between religiosity and the willingness to help neighbors and to entrust neighbors
8
See Tables 4.15 and 4.16 for interactions with the very religious dummy.
151
Table 4.3: Diversity, Segregation, and Community Cohesion & Trust Beliefs
Willing
to help
Village is safe [. . . ] Trust nbr. to watch Trust [. . . ] to return lost wallet
generally at night kid(s) house neighbors strangers police
A. Village Heterogeneity (1) (2) (3) (4) (5) (6) (7) (8)
Religious/ very religious 0.013

0.025

0.006 0.012 -0.001 0.108

0.026

0.141

(1.87) (4.54) (1.07) (0.97) (-0.07) (6.30) (1.90) (6.94)
. . . vilage diversity
y
-0.054

0.015 0.007 -0.067 -0.060

-0.169

-0.094 -0.055
(-2.24) (0.51) (0.22) (-1.22) (-1.93) (-2.42) (-1.49) (-0.77)
. . . village segregation
y
0.308

-0.018 0.108 -0.144 0.161 0.388 0.209 0.067
(3.50) (-0.15) (0.88) (-0.64) (1.29) (1.32) (0.94) (0.23)
Very religious 0.164

0.178

0.098

0.060

0.060

0.095

-0.051

0.170

(11.45) (11.17) (6.26) (2.70) (3.78) (3.12) (-2.05) (6.01)
P-val of joint test of:
Village interactions 0.002 0.810 0.238 0.041 0.153 0.045 0.314 0.660
N 28287 28284 28282 21304 28285 27695 26777 26196
Adj. R
2
0.085 0.093 0.090 0.102 0.060 0.106 0.063 0.084
B. Subdistrict Heterogeneity (1) (2) (3) (4) (5) (6) (7) (8)
Religious/ very religious 0.012

0.024

0.007 0.014 -0.000 0.107

0.028

0.141

(1.80) (4.43) (1.16) (1.17) (-0.03) (6.30) (2.11) (6.94)
. . . subdistrict diversity
y
0.000 0.010 -0.008 -0.122

-0.061

-0.161

-0.107

-0.067
(0.01) (0.34) (-0.25) (-2.44) (-2.17) (-2.48) (-1.84) (-1.09)
. . . subdistrict segregation
y
0.135 0.093 0.065 0.413

0.251

0.495 0.138 -0.046
(1.24) (0.92) (0.49) (2.70) (2.09) (1.57) (0.64) (-0.16)
Very religious 0.163

0.178

0.095

0.059

0.062

0.094

-0.052

0.167

(11.53) (11.22) (6.09) (2.67) (3.96) (3.10) (-2.12) (5.96)
P-val of joint test of:
Subdistrict interactions 0.285 0.323 0.884 0.018 0.064 0.046 0.156 0.376
N 28561 28558 28556 21511 28559 27967 27048 26466
Adj. R
2
0.084 0.092 0.090 0.102 0.059 0.105 0.063 0.084
Community xed eects Yes Yes Yes Yes Yes Yes Yes Yes
Religion xed eects Yes Yes Yes Yes Yes Yes Yes Yes
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
y
De-meaned diversity/segregation variables. Standard errors are robust and clustered at the community level. Included variables
not shown: sex, dummy variables for age and education categories, risk and time preference, married status, linear spline for log
PCE, and a constant.
152
Table 4.4: Diversity, Segregation, and Discriminative Trust & Tolerance
Trust [. . . ] more Tolerate non-corlgn living in [. . . ] Tolerate non-corlgn to [. . . ]
corelgn coethnics village neighbor house marry rltv. bld h. wrshp
A. Village Heterogeneity (1) (2) (3) (4) (5) (6) (7)
Religious/ very religious 0.084

0.068

-0.061

-0.081

-0.124

-0.132

-0.105

(7.90) (6.18) (-6.10) (-8.24) (-10.81) (-9.34) (-8.36)
. . . vilage diversity
y
-0.033 -0.066 0.125

0.159

0.180

0.049 0.196

(-0.49) (-1.07) (3.21) (4.31) (3.72) (1.01) (2.90)
. . . village segregation
y
0.094 -0.142 -0.131 -0.202 -0.298

0.085 -0.344
(0.27) (-0.42) (-0.64) (-1.20) (-2.12) (0.51) (-1.55)
Very religious 0.156

0.066

-0.008 -0.014 -0.092

-0.142

-0.102

(8.30) (3.39) (-0.48) (-0.77) (-4.35) (-5.84) (-4.39)
P-val of joint test of:
Village interactions 0.843 0.019 0.000 0.000 0.001 0.182 0.010
N 28286 28286 28287 28287 28286 28286 28286
Adj. R
2
0.155 0.180 0.221 0.249 0.260 0.247 0.256
B. Subdistrict Heterogeneity (1) (2) (3) (4) (5) (6) (7)
Religious/ very religious 0.085

0.068

-0.062

-0.084

-0.125

-0.133

-0.105

(8.05) (6.14) (-6.15) (-8.56) (-10.95) (-9.46) (-8.43)
. . . subdistrict diversity
y
-0.022 -0.077

0.148

0.185

0.148

0.037 0.173

(-0.51) (-2.01) (4.69) (5.59) (3.30) (0.75) (2.77)
. . . subdistrict segregation
y
0.064 0.011 -0.091 -0.207 -0.197 0.255

-0.315
(0.27) (0.05) (-0.57) (-1.27) (-1.07) (1.67) (-1.52)
Very religious 0.157

0.067

-0.009 -0.015 -0.092

-0.137

-0.101

(8.43) (3.51) (-0.55) (-0.86) (-4.39) (-5.61) (-4.43)
P-val of joint test of:
Subdistrict interactions 0.879 0.052 0.000 0.000 0.004 0.026 0.022
N 28560 28560 28561 28561 28560 28560 28560
Adj. R
2
0.155 0.179 0.222 0.251 0.261 0.246 0.261
Community xed eects Yes Yes Yes Yes Yes Yes Yes
Religion xed eects Yes Yes Yes Yes Yes Yes Yes
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
y
De-meaned diversity/segregation variables. Standard errors are robust and clustered at the community level. Included variables
not shown: sex, dummy variables for age and education categories, risk and time preference, married status, linear spline for log
PCE, and a constant.
153
with their house is weaker in more religiously diverse villages, and stronger in more
religiously segregated villages. The associations between religiosity and trust beliefs
toward neighbors is also weaker in more diverse villages. We nd similar results at
the subdistrict level for diversity. Meanwhile, the associations between religiosity and
the willingness to entrust neighbors are stronger in segregated subdistricts.
At the same time, diversity is linked with weaker associations between religiosity
and in-group preferences and intolerance. The magnitude of the positive correlations
between religiosity and trusts of coethnics are weaker in more diverse subdistricts.
Similarly, the magnitude of the negative correlations between religiosity and tolerance
are also weaker in more diverse villages and subdistricts for all tolerance measures,
except regarding inter-faith marriage.
9
In contrast, the magnitude of the negative
association between religiosity and tolerance in allowing non-coreligionists live in the
same house is larger in more segregated villages. Curiously, link between religiosity
and intolerance of interfaith marriage is weaker in more segregated subdistricts.
4.3.4 Political preference in heterogeneous communities
Table 4.5 presents the correlations between the in-group political preferences and
community heterogeneity. Similar to our results on discriminative trust and tolerance,
we nd that in-group-biased political preferences are negatively associated with village
an subdistrict diversity. Meanwhile, subdistrict segregation is positively associated
with assigning religion as the most important criteria in a district head. Like before,
the coecients on thereligiositycommunitydiversity variables (Table 4.6) suggest
that greater diversity (both at the village and subdistrict) is correlated with weaker
associations between religiosity and in-group biases: the link between religiosity and
9
Table 4.16 suggests that diversity also weakens the link between religiosity and intolerance of
inter-faith marriage among the very religious.
154
placing religion as the most important criteria in the political leader is weaker in more
diverse villages and subdistricts.
Consistent with previous results on inter-group cooperative attitudes, Table 4.5
suggests that religion is a more important criteria in more religiously segregated
subdistricts. Moreover, the analysis of the interactions in Table 4.6 indicates that the
associations between religiosity and the likelihood of assigning religion as the most
important criteria of a district head candidate are stronger in more segregated villages
and subdistricts.
Table 4.5: Diversity, Segregation and District Head Criteria
Village Heterogeneity Subdistrict Heterogeneity
Diver-
sity
Segre-
gation
Num.
of obs.
Diver-
sity
Segre-
gation
Num.
of obs.
(1) (2) (3) (4) (5) (6)
[ . . . ] is important
Religion -0.087

0.147 27717 -0.145

0.144 27790
(-1.76) (1.13) (-3.38) (1.50)
Ethnicity -0.094

0.070 27717 -0.109

0.023 27790
(-1.71) (0.47) (-1.96) (0.19)
[ . . . ] is three-most important
Religion -0.100

0.130 27717 -0.188

0.382

27790
(-2.64) (0.96) (-3.60) (2.79)
Ethnicity -0.041

0.066 27717 -0.058

0.134

27790
(-2.46) (1.02) (-2.59) (1.86)
[ . . . ] is the most important
Religion -0.070

0.141 27717 -0.134

0.463

27790
(-2.77) (0.98) (-4.21) (3.40)
Ethnicity -0.003 0.009 27717 -0.004 -0.004 27790
(-0.65) (0.40) (-0.63) (-0.18)
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01.
Each row presents results from two models: The left half is a model with the village heterogeneity variables, and
the right half is a model with the subdistrict heterogeneity variables. Both models are estimated with district
xed eects. Standard errors are robust and clustered at the subdistrict level. Included variables not shown:
dummy variables for religiosity, age and education categories, sex, risk and time preference, married status, linear
spline for log PCE, urban/rural status, log population density, dummy variables for topography, whether village
experienced natural disaster in the last ve years, receptions of public and local television signals, distance from
subdistricts and districts, and a constant.
155
Table 4.6: Community Compositions, Religiosity, and District Head Criteria
Important Three-most Important Most Important
Religion Ethnicity Religion Ethnicity Religion Ethnicity
A. Village Heterogeneity (1) (2) (3) (4) (5) (6)
Religious/ very religious 0.057

0.038

0.094

0.001 0.075

-0.000
(7.71) (4.24) (10.76) (0.22) (10.37) (-0.15)
. . . vilage diversity
y
0.014 -0.041 -0.049 -0.047

-0.116

-0.000
(0.37) (-1.11) (-1.27) (-2.25) (-4.41) (-0.02)
. . . village segregation
y
-0.118 -0.112 0.051 0.059 0.237

-0.014
(-0.96) (-0.76) (0.33) (0.59) (2.40) (-0.90)
Very religious 0.041

0.059

0.039

-0.003 0.043

-0.001
(4.31) (4.63) (3.29) (-0.27) (3.71) (-0.40)
P-val of joint test of:
Village interactions 0.513 0.040 0.242 0.015 0.000 0.525
N 28250 28250 28250 28250 28250 28250
Adj. R
2
0.189 0.167 0.167 0.081 0.116 0.014
B. Subdistrict Heterogeneity (1) (2) (3) (4) (5) (6)
Religious/ very religious 0.056

0.037

0.092

0.001 0.074

-0.000
(7.63) (4.14) (10.49) (0.12) (10.37) (-0.07)
. . . subdistrict diversity
y
0.015 -0.068

-0.043 -0.051

-0.129

-0.001
(0.49) (-1.85) (-1.43) (-2.88) (-5.69) (-0.23)
. . . subdistrict segregation
y
-0.134 0.139 0.129 0.119 0.486

-0.043

(-0.95) (0.99) (1.10) (1.33) (4.72) (-1.80)
Very religious 0.041

0.057

0.039

-0.001 0.043

-0.001
(4.33) (4.55) (3.32) (-0.09) (3.73) (-0.42)
P-val of joint test of:
Subdistrict interactions 0.636 0.178 0.336 0.015 0.000 0.079
N 28523 28523 28523 28523 28523 28523
Adj. R
2
0.188 0.166 0.167 0.081 0.117 0.014
Community xed eects Yes Yes Yes Yes Yes Yes
Religion xed eects Yes Yes Yes Yes Yes Yes
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
y
De-meaned diversity/segregation variables. Standard errors are robust and clustered at the community
level. Included variables not shown: sex, dummy variables for age and education categories, risk and time
preference, married status, linear spline for log PCE, and a constant.
156
4.4 Conclusion
This chapter presents evidence on how the religious compositions of communities are
associated with dierent types of cooperative attitudes. Consistent with previous
literature, we nd that after controlling for segregation, religious diversity tends to
be negatively associated with generalized trust beliefs. We do not, however, nd
similarly signicant results for trust of neighbors. Moreover, in support of the inter-
group contact hypothesis, diversity tends to be positively associated with religious
tolerance. On the other hand, religious segregation tends to have opposite associations
with various measures of inter-group cooperative attitudes.
An examination of the heterogeneous responses of people of dierent religiosity
to to the community's religious compositions conrms our results from the previ-
ous chapter. The positive associations between religiosity and the community cohe-
siveness variables tend to be weaker in more diverse communities, and stronger in
more segregated communities. At the same time, the associations between religiosity
and intolerance also tend to be weaker in more diverse communities, and somewhat
stronger in more segregated communities. Both results are suggestive of the role of
social interactions in in
uencing cooperative attitudes.
157
Chapter bibliography
Alesina, Alberto and Ekaterina Zhuravskaya, \Segregation and the Quality of
Government in a Cross Section of Countries," American Economic Review, August
2011, 101 (5), 1872{1911.
and Eliana La Ferrara, \Participation in Heterogeneous Communities," Quar-
terly Journal of Economics, 2000, 115 (3), 847{904.
and , \Who trusts others?," Journal of Public Economics, August 2002, 85 (2),
207{234.
Allport, Gordon W., The nature of prejudice, Reading Mass.: Addison-Wesley
Pub. Co., 1979 [1954].
Arrow, Kenneth J., \What Has Economics to Say about Racial Discrimination?,"
The Journal of Economic Perspectives, April 1998, 12 (2), 91{100.
Barron, Patrick, Kai Kaiser, and Menno Pradhan, \Understanding Variations
in Local Con
ict: Evidence and Implications from Indonesia," World Development,
March 2009, 37 (3), 698{713.
Bertrand, Jacques, Nationalism and ethnic con
ict in Indonesia, Cambridge, UK;
New York: Cambridge University Press, 2004.
Clingingsmith, David, Asim Ijaz Khwaja, and Michael Kremer, \Estimating
the Impact of The Hajj: Religion and Tolerance in Islam's Global Gathering," The
Quarterly Journal of Economics, 2009, 124 (3), 1133 {1170.
158
Coleman, James, Thomas Hoer, and Sally Kilgore, \Achievement and Seg-
regation in Secondary Schools: A Further Look at Public and Private School Dif-
ferences," Sociology of Education, April 1982, 55 (2/3), 162{182.
Costa, Dora L. and Matthew E. Kahn, \Civic Engagement and Community
Heterogeneity: An Economist's Perspective," Perspectives on Politics, 2003, 1 (01),
103{111.
Cutler, David M., Edward L. Glaeser, and Jacob L. Vigdor, \The Rise and
Decline of the American Ghetto," Journal of Political Economy, June 1999, 107
(3), 455.
, , and , \Is the Melting Pot Still Hot? Explaining the Resurgence of Im-
migrant Segregation," Review of Economics and Statistics, August 2008, 90 (3),
478{497.
, , and , \When are ghettos bad? Lessons from immigrant segregation in the
United States," Journal of Urban Economics, May 2008, 63 (3), 759{774.
Echenique, F. and R. G. Fryer, \A Measure of Segregation Based on Social
Interactions," The Quarterly Journal of Economics, May 2007, 122 (2), 441{485.
Fafchamps, Marcel, Market institutions in Sub-Saharan Africa : theory and evi-
dence, Cambridge Mass.: MIT Press, 2004.
Frankel, David M. and Oscar Volij, \Measuring school segregation," Journal of
Economic Theory, January 2011, 146 (1), 1{38.
Glaeser, Edward L., David I. Laibson, Jose A. Scheinkman, and Chris-
tine L. Soutter, \Measuring Trust," Quarterly Journal of Economics, August
2000, 115 (3), 811{846.
159
Granovetter, Mark, \The Impact of Social Structure on Economic Outcomes,"
Journal of Economic Perspectives, January 2005, 19 (1), 33{50.
Guiso, Luigi, Paola Sapienza, and Luigi Zingales, \People's opium? Religion
and economic attitudes," Journal of Monetary Economics, January 2003, 50 (1),
225{282.
Habyarimana, James, Macartan Humphreys, Daniel N. Posner, and
Jeremy M. Weinstein, \Why Does Ethnic Diversity Undermine Public Goods
Provision?," American Political Science Review, November 2007, 101 (04).
Iceland, John and Melissa. Scopilliti, \Immigrant Residential Segregation in
U.S. Metropolitan Areas, 1990{2000," Demography, 2008, 45 (1), 79{94.
James, F. J., \A New Generalized "Exposure-Based" Segregation Index: Demon-
stration in Denver and Houston," Sociological Methods & Research, February 1986,
14 (3), 301{316.
Luttmer, Erzo F. P., \Group Loyalty and the Taste for Redistribution," Journal
of Political Economy, 2001, 109 (3), 500{528.
Massey, D. S. and N. A. Denton, \The Dimensions of Residential Segregation,"
Social Forces, December 1988, 67 (2), 281{315.
Miguel, E and M Gugerty, \Ethnic diversity, social sanctions, and public goods
in Kenya," Journal of Public Economics, December 2005, 89 (11-12), 2325{2368.
Olken, Benjamin A, \Do Television and Radio Destroy Social Capital? Evi-
dence from Indonesian Villages," American Economic Journal: Applied Economics,
September 2009, 1 (4), 1{33.
160
Pettigrew, Thomas F. and Linda R. Tropp, \A Meta-Analytic Test of Inter-
group Contact Theory.," Journal of Personality and Social Psychology, 2006, 90
(5), 751{783.
Reardon, Sean F. and Glenn Firebaugh, \Measures of Multigroup Segregation,"
Sociological Methodology, 2002, 32 (1), 33{67.
, John T. Yun, and Tamela Mcnulty Eitle, \The Changing Structure of School
Segregation: Measurement and Evidence of Multiracial Metropolitan-Area School
Segregation, 1989-1995," Demography, 2000, 37 (3), 351{364.
Rothwell, Jonathan T., \Trust in Diverse, Integrated, Cities: A Revisionist Per-
spective," SSRN Electronic Journal, 2010.
Strauss, John, Firman Witoelar, Bondan Sikoki, and Anna Wattie, \The
Fourth Wave of the Indonesia Family Life Survey: Overview and Field Report,
Volume 1," RAND Working Paper WR-675/1-NIA/NICHD, RAND April 2009.
Suryadinata, Leo, Evi Arin Nurvidya, and Aris Ananta, Indonesia's popula-
tion : ethnicity and religion in a changing political landscape, Singapore: Institute
of Southeast Asian Studies, 2003.
Suryahadi, Asep, Chris Elbers, and Menno Pradhan, Developing a Poverty
map for Indonesia: an initiatory work in three provinces, SMERU Research Insti-
tute and BPS, 2003.
Theil, Henri and Anthony J. Finizza, \A note on the measurement of racial
integration of schools by means of informational concepts," The Journal of Math-
ematical Sociology, July 1971, 1 (2), 187{193.
161
Uslaner, Eric M., \Segregation, mistrust and minorities," Ethnicities, December
2010, 10 (4), 415{434.
Vigdor, Jacob L., \Community Composition and Collective Action: Analyzing
Initial Mail Response to the 2000 Census," Review of Economics and Statistics,
2004, 86 (1), 303{312.
162
Appendix 4.A Additional analyses
4.A.1 Community characteristics
Tables 4.7 and 4.8 present the community regressors for the results presented in
Table 4.2.
4.A.2 Inter-religion dierences in cooperative attitudes
Here, I examine the average dierences across religions by estimating:
Y
r
ijk
=+
5
X
d=2

d
:1(x
ri
=d)+
X
r=3;4

r
1
:1(rlgs
i
r)+X
i
:
i
+X
j
:
j
+X
k
:
k
+"
r
ijk
(4.5)
where as in above 1 is the indicator function, x
ri
denotes the index of individual
i`s religion and d indexes the dierent religions. In the estimations, \Islam" is the
omitted religion category.
10
For this analysis, I opt for the province xed-eects
specication since in 138 out of 262 sample districts (or 52.7% of the sample districts),
all respondents within these districts adhere to the same religion. This is equal to
36.0% of the sample respondents. As a robustness check, I include in the appendix
estimates using the district xed-eects specication.
Tables 4.9 to 4.11 present the results of the province xed eects estimations.
In the regressions analyzing inter-religion dierences, Islam is the omitted religion
category. Overall, there appears to be very little inter-religion dierences in terms of
community and non-discriminative cooperative attitudes. However, there are signi-
cant inter-religion dierences in terms of discriminative trust and tolerance and these
10
We follow this convention of setting \Islam" as the omitted category for all estimations that
involve religion categories in this paper.
163
Table 4.7: Community characteristics & Community Cohesion & Trust Beliefs
Willing
to help
Village is safe [. . . ] Trust nbr. to watch Trust [. . . ] to return lost wallet
generally at night kid(s) house neighbors strangers police
(1) (2) (3) (4) (5) (6) (7) (8)
Urban -0.016 -0.013 -0.007 -0.017 -0.029

-0.039 0.045

-0.017
(-1.30) (-0.92) (-0.51) (-0.98) (-2.44) (-1.32) (2.19) (-0.71)
Coast 0.018 0.019 0.022 0.008 0.003 -0.084

-0.096

-0.045
(0.93) (0.84) (0.99) (0.28) (0.16) (-2.19) (-3.70) (-1.20)
Valley 0.080

-0.056

0.007 0.060 0.030 0.170

-0.002 0.023
(2.60) (-1.90) (0.37) (0.96) (1.00) (3.27) (-0.06) (0.25)
Hill -0.008 0.003 0.021 -0.006 -0.014 0.012 0.022 -0.015
(-0.67) (0.11) (0.80) (-0.26) (-0.85) (0.34) (0.73) (-0.55)
Log population density 0.003 0.003 -0.001 -0.015

0.001 0.004 -0.013 -0.005
(0.78) (0.61) (-0.11) (-1.89) (0.32) (0.33) (-1.59) (-0.50)
Receive public TV broadcast 0.011 0.015 -0.001 -0.014 -0.025 0.007 -0.003 0.011
(0.73) (0.67) (-0.08) (-0.56) (-1.45) (0.17) (-0.09) (0.24)
Receive private TV broadcast 0.020 0.006 0.037

0.043 0.014 -0.107 -0.006 -0.048
(1.16) (0.29) (1.70) (1.55) (0.77) (-1.43) (-0.14) (-1.12)
Natural disaster in last 5 years -0.001 -0.014 -0.011 0.009 -0.011 -0.008 -0.014 0.008
(-0.09) (-1.64) (-1.49) (0.75) (-1.29) (-0.44) (-1.05) (0.42)
Distance to subdistrict capital -0.000 0.001 0.001 0.002

-0.000 0.003

0.004

-0.002
(-0.07) (1.00) (0.92) (1.86) (-0.72) (2.30) (4.70) (-1.33)
Distance to district capital -0.000 -0.000 -0.000 -0.000 -0.000 0.001 -0.000 -0.000
(-0.48) (-1.01) (-0.88) (-0.73) (-0.58) (0.86) (-0.36) (-0.35)
Subdistrict PCE Gini -0.495

-0.332

-0.287 -0.098 -0.115 -0.783 0.435 -0.469
(-2.68) (-1.71) (-1.55) (-0.28) (-0.54) (-1.56) (1.01) (-0.90)
District xed eects Yes Yes Yes Yes Yes Yes Yes Yes
N 27751 27748 27746 20912 27749 27178 26278 25690
Adj. R
2
0.077 0.073 0.072 0.092 0.047 0.095 0.059 0.078
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
Standard errors are robust and clustered at the subdistrict level. Included variables not shown: dummy variables for religiosity, age and
education categories, sex, risk and time preference, linear spline for log PCE, married status, village religious diversity and segregation.
164
Table 4.8: Community characteristics & Discriminative Trust & Tolerance
Trust [. . . ] more Tolerate non-corlgn living in [. . . ] Tolerate non-corlgn to [. . . ]
corelgn coethnics village neighbor house marry rltv. bld h. wrshp
(1) (2) (3) (4) (5) (6) (7)
Urban 0.003 -0.008 -0.001 0.009 0.006 -0.012 0.042
(0.16) (-0.48) (-0.08) (0.49) (0.24) (-0.49) (1.57)
Coast -0.008 -0.018 0.019 0.039 0.111

0.085

0.081

(-0.29) (-0.81) (0.90) (1.52) (3.82) (2.93) (2.11)
Valley -0.015 -0.003 0.009 0.006 -0.022 -0.047 -0.012
(-0.37) (-0.07) (0.13) (0.09) (-0.29) (-1.12) (-0.21)
Hill -0.003 -0.037

-0.032 -0.030 0.029 0.012 -0.017
(-0.16) (-2.09) (-1.33) (-1.14) (1.07) (0.43) (-0.54)
Log population density -0.014

-0.005 0.019

0.020

0.017

-0.002 0.014
(-2.04) (-0.85) (2.63) (2.61) (1.87) (-0.23) (1.32)
Receive public TV broadcast 0.029 0.053

-0.031 -0.012 -0.022 -0.082

-0.033
(1.48) (1.99) (-1.48) (-0.48) (-0.75) (-2.78) (-1.18)
Receive private TV broadcast 0.018 -0.057

-0.014 -0.017 0.040 0.069 0.061

(0.69) (-1.98) (-0.58) (-0.62) (1.16) (1.55) (1.91)
Natural disaster in last 5 years 0.006 0.010 -0.039

-0.039

-0.045

-0.040

-0.042

(0.48) (0.83) (-3.08) (-2.78) (-3.14) (-2.46) (-2.19)
Distance to subdistrict capital 0.000 0.002 -0.000 0.000 -0.000 -0.002 -0.001
(0.47) (1.61) (-0.20) (0.38) (-0.18) (-1.50) (-0.97)
Distance to district capital -0.000 0.000 0.001

0.001 0.001

0.001

0.000
(-0.50) (0.55) (1.82) (1.55) (2.61) (1.94) (0.93)
Subdistrict PCE Gini -0.426 -0.268 0.494 0.548

0.390 0.364 0.162
(-1.41) (-1.03) (1.64) (1.70) (1.13) (0.96) (0.34)
District xed eects Yes Yes Yes Yes Yes Yes Yes
N 27750 27750 27751 27751 27750 27750 27750
Adj. R
2
0.142 0.170 0.211 0.239 0.257 0.236 0.241
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
Standard errors are robust and clustered at the subdistrict level. Included variables not shown: dummy variables for religiosity, age and
education categories, sex, risk and time preference, linear spline for log PCE, married status, village religious diversity and segregation.
165
dierences are mainly between Muslims, who are the majority in the country, and
the rest.
With respect to cooperative attitudes in the community, Protestants are less will-
ing to help their neighbors compared to adherents of other religions. Meanwhile,
Buddhists are less willing to trust their neighbors to watch their children or property.
However, on average, these dierences do not seem to be driven by dierences in their
beliefs regarding the trustworthiness of neighbors or strangers.
However, in terms of discriminative trust, Muslims trust their coreligionists more
compared to adherents of other religions. They are also the most intolerant in all
tolerant measures. Meanwhile, Catholics and Buddhists are the least discriminative
with regards to both ethnicity and religion. They also tend to be among the two
most tolerant believers on most measures. Catholics are also more tolerant than
Protestants across all measures.
4.A.3 The role of the majority status
In their cross-country analysis, Guiso et al. (2003) found that adherents of the ma-
jority religion tend to be more intolerant. I examine whether this phenomenon exists
within countries and explains the inter-religion dierences in attitudes. To explore
this question, I include an indicator variable of whether the respondent adheres to
the majority religion in the village. In this sample, only 3 Catholics and 6 Buddhists
live in a village where their respective religion is the majority religion. Our discussion
will therefore focus only on the other three religions.
The right halves of Tables 4.9 to 4.11 present the religion results with the village-
majority status variables included. Protestants exhibit an even less willingness to
help neighbors, and minority Protestants nd their community less safe compared to
166
Table 4.9: Inter-religion Differences in Community Cohesion
Willing
to help
Village is safe [. . . ] Trust nbr. to watch Willing
to help
Village is safe [. . . ] Trust nbr. to watch
generally at night kid(s) house generally at night kid(s) house
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Catholic -0.027 0.019 0.046

0.015 -0.035 -0.058 -0.019 0.027 0.045 -0.027
(-1.50) (0.77) (1.98) (0.38) (-1.19) (-1.57) (-0.47) (0.51) (0.67) (-0.54)
Protestant -0.029

-0.011 0.008 0.038

0.005 -0.091

-0.097

-0.045 0.036 -0.001
(-1.72) (-0.67) (0.45) (1.69) (0.27) (-2.52) (-2.38) (-0.92) (0.53) (-0.03)
Hindu 0.001 -0.039

0.023 0.062 -0.034 -0.000 -0.059 -0.009 0.065 0.037
(0.03) (-2.17) (1.04) (1.29) (-0.63) (-0.00) (-1.53) (-0.19) (0.70) (0.58)
Buddhist 0.004 -0.070 -0.083 -0.211

-0.207

-0.046 -0.122 -0.119 -0.176 -0.206

(0.11) (-0.90) (-1.28) (-2.37) (-3.50) (-0.92) (-1.27) (-1.49) (-1.49) (-2.40)
Majority religion in village -0.035 -0.043 -0.021 0.026 0.004
(-1.03) (-1.09) (-0.47) (0.40) (0.08)
. . . Catholic 0.052 0.001 -0.059 -0.627

-0.486

(1.19) (0.02) (-0.74) (-7.33) (-1.76)
. . . Protestant 0.119

0.172

0.112

0.044 0.028
(2.04) (3.03) (1.81) (0.54) (0.48)
. . . Hindu 0.001 0.043 0.068 -0.003 -0.146
(0.01) (0.65) (0.88) (-0.02) (-1.62)
. . . Buddhist 0.242

0.178

0.196

-0.140 0.040
(2.42) (1.73) (2.12) (-0.96) (0.56)
Constant 3.084

3.120

2.936

2.805

3.013

3.111

3.151

2.950

2.772

3.007

(23.27) (27.19) (26.45) (15.69) (24.11) (22.90) (26.68) (24.95) (14.45) (22.22)
Province xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
P-val of joint test on:
Religions 0.372 0.132 0.267 0.054 0.004 0.025 0.057 0.062 0.225 0.014
Majority status & int. 0.032 0.007 0.019 0.312 0.021
Religions majority status 0.018 0.004 0.112 0.361 0.032
N 27751 27748 27746 20912 27749 27751 27748 27746 20912 27749
Adj. R
2
0.057 0.051 0.054 0.069 0.034 0.057 0.052 0.055 0.069 0.034
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
Standard errors are robust and clustered at the community level. Included variables not shown: dummy variables for religiosity, age and education categories, sex,
risk and time preference, married status, linear spline for log PCE, urban/rural status, log population density, dummy variables for topography, whether village
experienced natural disaster in the last ve years, receptions of public and local television signals, distance from subdistricts and districts. \Muslim" is the omitted
category.
167
Table 4.10: Inter-religion Differences in Trust Beliefs
Trust [. . . ] to return lost wallet Trust [. . . ] more Trust [. . . ] to return lost wallet Trust [. . . ] more
neighbors strangers police corelgn. coethnics neighbors strangers police corelgn. coethnics
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Catholic -0.020 0.046 -0.051 -0.305

-0.065

-0.096 -0.042 -0.171

-0.254

-0.046
(-0.33) (1.03) (-0.66) (-7.55) (-1.85) (-0.63) (-0.52) (-1.65) (-3.33) (-0.66)
Protestant -0.073 -0.006 -0.018 -0.195

-0.033 -0.195 -0.121 -0.169

-0.175

-0.091
(-1.23) (-0.19) (-0.59) (-4.94) (-0.86) (-1.47) (-1.61) (-1.86) (-2.75) (-1.30)
Hindu 0.028 -0.005 0.084 -0.149

-0.033 -0.033 -0.133

0.005 -0.169

-0.032
(0.26) (-0.09) (1.34) (-3.70) (-0.95) (-0.21) (-1.69) (0.04) (-2.01) (-0.39)
Buddhist -0.051 -0.067 -0.133 -0.298

-0.175

-0.133 -0.124 -0.249

-0.206

-0.139
(-0.44) (-0.80) (-1.42) (-3.32) (-2.71) (-0.83) (-1.11) (-1.88) (-2.03) (-1.52)
Majority religion in village -0.078 -0.096 -0.128 0.058 0.016
(-0.59) (-1.35) (-1.60) (0.91) (0.23)
. . . Catholic 0.656

-0.242

0.075 0.458

-0.296
(2.46) (-2.05) (0.33) (4.77) (-0.92)
. . . Protestant 0.217 0.160 0.218

0.026 0.181
(1.00) (1.23) (1.73) (0.23) (1.56)
. . . Hindu 0.123 0.250

0.148 0.059 0.014
(0.31) (1.73) (0.83) (0.38) (0.10)
. . . Buddhist 0.190 -0.307

0.096 -0.465

-0.223

(0.93) (-2.67) (0.57) (-2.82) (-1.72)
Constant 3.154

1.059

3.161

3.082

3.334

3.218

1.143

3.275

3.021

3.304

(12.68) (5.31) (10.30) (18.84) (18.02) (11.85) (5.47) (10.38) (17.15) (16.90)
Province xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
P-val of joint test on:
Religions 0.725 0.745 0.373 0.000 0.038 0.262 0.144 0.189 0.018 0.414
Majority status & int. 0.828 0.000 0.539 0.003 0.023
Religions majority status 0.715 0.000 0.394 0.013 0.043
N 27178 26278 25690 27750 27750 27178 26278 25690 27750 27750
Adj. R
2
0.074 0.041 0.062 0.108 0.135 0.074 0.041 0.062 0.109 0.136
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
Standard errors are robust and clustered at the community level. Included variables not shown: dummy variables for religiosity, age and education categories, sex,
risk and time preference, married status, linear spline for log PCE, urban/rural status, log population density, dummy variables for topography, whether village
experienced natural disaster in the last ve years, receptions of public and local television signals, distance from subdistricts and districts. \Muslim" is the omitted
category.
168
Table 4.11: Inter-religion Differences in Tolerance
Tolerate non-corlgn living in [. . . ] Tolerate non-corlgn to [. . . ] Tolerate non-corlgn living in [. . . ] Tolerate non-corlgn to [. . . ]
village neighbor house marry reltv. bld h. wrshp village neighbor house marry reltv. bld h. wrshp
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Catholic 0.129

0.160

0.421

0.776

0.557

0.046 0.033 0.270

0.583

0.286

(5.50) (6.30) (13.07) (16.37) (15.06) (1.02) (0.67) (3.19) (6.53) (2.85)
Protestant 0.117

0.159

0.411

0.686

0.459

0.013 0.007 0.238

0.422

0.216

(5.95) (6.49) (11.26) (15.20) (9.53) (0.32) (0.16) (2.86) (5.22) (2.17)
Hindu 0.119

0.160

0.319

0.704

0.210

0.096 0.059 0.175 0.499

0.155
(3.23) (3.78) (5.20) (8.52) (1.88) (1.42) (0.80) (1.42) (3.35) (1.34)
Buddhist 0.175

0.225

0.398

0.987

0.563

0.087 0.091 0.240

0.761

0.306

(3.93) (4.48) (5.84) (11.13) (7.79) (1.50) (1.43) (2.25) (6.39) (2.71)
Majority religion in village -0.089

-0.136

-0.161

-0.210

-0.288

(-2.01) (-2.83) (-2.00) (-2.59) (-2.77)
. . . Catholic 0.009 0.046 0.007 -0.233 0.007
(0.10) (0.47) (0.04) (-0.67) (0.05)
. . . Protestant 0.154

0.209

0.225

0.399

0.239

(2.23) (2.74) (1.91) (3.04) (1.69)
. . . Hindu 0.035 0.189

0.273

0.398

0.057
(0.35) (1.98) (1.67) (1.71) (0.30)
. . . Buddhist 0.164

0.211

0.235 0.526

0.172
(1.76) (2.22) (1.61) (3.86) (1.24)
Constant 1.954

1.808

1.787

1.820

1.661

2.034

1.931

1.934

2.004

1.939

(8.64) (7.91) (6.15) (7.04) (6.10) (8.84) (8.27) (6.44) (7.22) (6.59)
Province xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
P-val of joint test on:
Religions 0.000 0.000 0.000 0.000 0.000 0.278 0.462 0.039 0.000 0.023
Majority status & int. 0.091 0.053 0.349 0.002 0.000
Religions majority status 0.081 0.031 0.231 0.001 0.225
N 27751 27751 27750 27750 27750 27751 27751 27750 27750 27750
Adj. R
2
0.172 0.200 0.221 0.205 0.202 0.172 0.200 0.222 0.206 0.203
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
Standard errors are robust and clustered at the community level. Included variables not shown: dummy variables for religiosity, age and education categories, sex, risk and time preference,
married status, linear spline for log PCE, urban/rural status, log population density, dummy variables for topography, whether village experienced natural disaster in the last ve years,
receptions of public and local television signals, distance from subdistricts and districts. \Muslim" is the omitted category.
169
their minority-Muslim counterpart. Minority Christians and Buddhists also tend to
be more distrustful of the police. However, minority Hindus exhibit more trusting
behaviors than minority Muslims. In contrast, as minorities, the gaps in religio-
discriminative trust between Muslims and adherents of other religions, except Hindus,
tend to be smaller.
Majority status in the village does not appear to signicantly aect community or
discriminative trust. However, majority status is negatively correlated with all aspects
of tolerance, and the magnitude of the negative coecient is largest on tolorance of
non-coreligionists' house of worship. This suggests that among the tolerance issues,
this issue may be the most political. Meanwhile, majority Protestants and Hindus
tend to be more tolerant on most measures, except on the issue of non-coreligionists'
house of worship.
4.A.4 Religiosity and community heterogeneity: Muslim v.
non-Muslim
Tables 4.12-4.14 further decompose the analysis and separately examine dierences
between Muslims and non-Muslims. Most of the results in the pooled regressions are
replicated in the Muslim subset. Among Muslims, the associations between religiosity
and measures of community cohesion as well as trust beliefs (except that of strangers)
tend to be weaker in more diverse villages and subdisctricts. For these outcomes,
subdistrict-level segregations appear to play a more important role. The associations
between religiosity and the willingness to entrust neighbors with one's children and
house are stronger in more segregated subdistricts.
With regards to in-group preferences and tolerance, given the across-the-board
eects of both village and subdistrict diversities on the associations between religiosity
170
Table 4.12: Community Compositions, Religiosity and Community Cohesion by Muslim/Non-Muslim
Muslim Non-Muslim
Willing
to help
Village is safe [. . . ] Trust nbr. to watch Willing
to help
Village is safe [. . . ] Trust nbr. to watch
generally at night kid(s) house generally at night kid(s) house
A. Village Heterogeneity (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Religious/ very religious 0.009 0.025

0.007 0.010 0.002 0.034 0.006 0.004 0.069 -0.035
(1.32) (4.53) (1.10) (0.83) (0.21) (1.27) (0.21) (0.13) (1.14) (-1.00)
. . . vilage diversity
y
-0.085

0.006 0.015 -0.062 -0.072

0.051 0.075 -0.019 -0.142 0.053
(-3.02) (0.17) (0.47) (-1.03) (-1.95) (0.87) (1.08) (-0.25) (-1.02) (0.79)
. . . village segregation
y
0.371

0.062 0.089 -0.079 0.174 0.130 -0.368 0.159 -0.571 0.059
(3.86) (0.51) (0.71) (-0.33) (1.15) (0.58) (-1.21) (0.59) (-1.39) (0.23)
Very religious 0.163

0.180

0.090

0.041

0.045

0.156

0.170

0.126

0.132

0.117

(9.71) (9.96) (5.19) (1.80) (2.58) (6.88) (4.68) (3.76) (2.34) (3.17)
P-val of joint test of:
Village interactions 0.001 0.560 0.240 0.141 0.126 0.319 0.424 0.827 0.068 0.538
N 25249 25246 25244 19090 25247 3038 3038 3038 2214 3038
Adj. R
2
0.081 0.088 0.091 0.100 0.051 0.117 0.113 0.067 0.111 0.086
B. Subdistrict Heterogeneity (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Religious/ very religious 0.009 0.025

0.007 0.013 0.002 0.031 0.003 0.001 0.059 -0.024
(1.36) (4.44) (1.13) (1.07) (0.28) (1.18) (0.09) (0.03) (0.93) (-0.76)
. . . subdistrict diversity
y
-0.026 0.007 -0.016 -0.112

-0.074

0.143

0.105 0.075 -0.235 0.087
(-0.84) (0.23) (-0.52) (-2.15) (-2.33) (2.79) (1.13) (0.84) (-1.30) (1.12)
. . . subdistrict segregation
y
0.215

0.143 0.133 0.406

0.311

-0.553

-0.589

-0.556

0.332 -0.496
(1.76) (1.19) (0.93) (2.58) (2.37) (-2.79) (-1.94) (-1.73) (0.85) (-1.60)
Very religious 0.163

0.180

0.086

0.041

0.047

0.155

0.168

0.125

0.131

0.116

(9.80) (10.04) (4.99) (1.77) (2.78) (6.87) (4.55) (3.70) (2.35) (3.16)
P-val of joint test of:
Subdistrict interactions 0.200 0.257 0.645 0.030 0.034 0.004 0.154 0.202 0.429 0.278
N 25514 25511 25509 19291 25512 3047 3047 3047 2220 3047
Adj. R
2
0.080 0.087 0.091 0.100 0.051 0.118 0.113 0.067 0.109 0.086
Community xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
y
De-meaned diversity/segregation variables. Standard errors are robust and clustered at the community level. Included variables not shown: sex, dummy
variables for age and education categories, risk and time preference, married status, linear spline for log PCE, and a constant.
171
Table 4.13: Community Compositions, Religiosity and Trust Beliefs by Muslim/Non-Muslim
Muslim Non-Muslim
Trust [. . . ] to return lost wallet Trust [. . . ] more Trust [. . . ] to return lost wallet Trust [. . . ] more
neighbors strangers police corelgn. coethnics neighbors strangers police corelgn. coethnics
A. Village Heterogeneity (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Religious/ very religious 0.108

0.024

0.141

0.082

0.066

0.103 0.091 0.082 0.144

0.132

(6.26) (1.76) (6.75) (7.48) (5.82) (1.13) (1.08) (0.79) (3.70) (3.11)
. . . vilage diversity
y
-0.177

-0.060 -0.047 -0.012 -0.070 -0.125 -0.248 0.022 -0.161

-0.119
(-2.25) (-0.95) (-0.58) (-0.15) (-0.98) (-0.73) (-1.38) (0.11) (-1.89) (-1.22)
. . . village segregation
y
0.502 0.141 -0.015 0.100 -0.097 -0.111 0.369 0.349 -0.081 -0.284
(1.59) (0.60) (-0.05) (0.24) (-0.25) (-0.25) (0.73) (0.63) (-0.30) (-1.10)
Very religious 0.125

-0.039 0.185

0.179

0.070

-0.039 -0.101 0.094 0.074

0.051
(3.65) (-1.48) (5.96) (8.90) (3.19) (-0.67) (-1.62) (1.43) (1.88) (1.19)
P-val of joint test of:
Village interactions 0.080 0.624 0.637 0.961 0.036 0.553 0.389 0.669 0.020 0.076
N 24728 23854 23356 25248 25248 2967 2923 2840 3038 3038
Adj. R
2
0.102 0.063 0.080 0.135 0.176 0.122 0.043 0.107 0.187 0.181
B. Subdistrict Heterogeneity (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Religious/ very religious 0.108

0.026

0.140

0.084

0.065

0.053 0.081 0.145

0.145

0.124

(6.32) (1.92) (6.72) (7.64) (5.80) (0.56) (1.06) (1.72) (3.73) (3.20)
. . . subdistrict diversity
y
-0.156

-0.087 -0.066 -0.012 -0.087

-0.067 -0.227 -0.064 -0.149

-0.050
(-2.20) (-1.49) (-0.95) (-0.24) (-2.12) (-0.35) (-1.20) (-0.44) (-1.69) (-0.61)
. . . subdistrict segregation
y
0.411 0.053 0.014 0.125 0.091 0.721 0.439 -0.487 -0.174 -0.693

(1.20) (0.24) (0.05) (0.46) (0.41) (1.16) (0.79) (-0.62) (-0.46) (-1.97)
Very religious 0.123

-0.041 0.181

0.179

0.071

-0.037 -0.100 0.094 0.076

0.053
(3.60) (-1.56) (5.88) (8.99) (3.29) (-0.64) (-1.60) (1.46) (1.99) (1.26)
P-val of joint test of:
Subdistrict interactions 0.090 0.224 0.538 0.898 0.066 0.496 0.471 0.635 0.112 0.035
N 24991 24116 23617 25513 25513 2976 2932 2849 3047 3047
Adj. R
2
0.101 0.063 0.080 0.136 0.175 0.121 0.042 0.107 0.185 0.180
Community xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
y
De-meaned diversity/segregation variables. Standard errors are robust and clustered at the community level. Included variables not shown: sex, dummy
variables for age and education categories, risk and time preference, married status, linear spline for log PCE, and a constant.
172
Table 4.14: Community Compositions, Religiosity and Tolerance by Muslim/Non-Muslim
Muslim Non-Muslim
Tolerate non-corlgn living in [. . . ] Tolerate non-corlgn to [. . . ] Tolerate non-corlgn living in [. . . ] Tolerate non-corlgn to [. . . ]
village neighbor house marry reltv. bld h. wrshp village neighbor house marry reltv. bld h. wrshp
A. Village Heterogeneity (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Religious/ very religious -0.063

-0.085

-0.127

-0.138

-0.109

0.016 0.007 -0.016 -0.029 -0.001
(-6.08) (-8.26) (-10.62) (-9.45) (-8.39) (0.72) (0.34) (-0.48) (-0.55) (-0.02)
. . . vilage diversity
y
0.132

0.169

0.156

0.015 0.190

0.026 0.040 0.110 0.072 0.125
(2.85) (3.79) (2.63) (0.27) (2.32) (0.50) (0.71) (1.54) (0.59) (1.25)
. . . village segregation
y
-0.187 -0.278 -0.226 0.192 -0.356 0.121 0.174 -0.405

-0.267 -0.176
(-0.80) (-1.54) (-1.28) (1.11) (-1.31) (0.76) (1.09) (-2.23) (-0.58) (-0.60)
Very religious -0.034

-0.032 -0.128

-0.176

-0.108

0.101

0.069

0.065

-0.020 -0.082
(-1.80) (-1.56) (-5.86) (-6.94) (-4.35) (3.58) (2.78) (1.82) (-0.44) (-1.49)
P-val of joint test of:
Village interactions 0.001 0.000 0.019 0.230 0.049 0.462 0.158 0.080 0.793 0.436
N 25249 25249 25248 25248 25248 3038 3038 3038 3038 3038
Adj. R
2
0.219 0.246 0.238 0.148 0.220 0.084 0.078 0.074 0.108 0.169
B. Subdistrict Heterogeneity (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Religious/ very religious -0.064

-0.087

-0.129

-0.138

-0.110

0.025 0.013 -0.010 -0.017 -0.006
(-6.07) (-8.48) (-10.80) (-9.58) (-8.49) (1.12) (0.57) (-0.29) (-0.33) (-0.14)
. . . subdistrict diversity
y
0.152

0.185

0.126

0.017 0.150

0.025 0.070 0.124 -0.005 0.183

(4.21) (4.67) (2.37) (0.31) (2.09) (0.47) (1.25) (1.41) (-0.03) (1.85)
. . . subdistrict segregation
y
-0.097 -0.205 -0.131 0.288

-0.270 -0.089 -0.196 -0.713

0.020 -0.561
(-0.54) (-1.13) (-0.65) (1.69) (-1.19) (-0.47) (-0.92) (-1.73) (0.03) (-1.41)
Very religious -0.034

-0.033 -0.127

-0.168

-0.106

0.101

0.068

0.064

-0.020 -0.086
(-1.86) (-1.63) (-5.89) (-6.64) (-4.33) (3.55) (2.76) (1.79) (-0.42) (-1.58)
P-val of joint test of:
Subdistrict interactions 0.000 0.000 0.050 0.077 0.113 0.878 0.439 0.201 0.999 0.179
N 25514 25514 25513 25513 25513 3047 3047 3047 3047 3047
Adj. R
2
0.220 0.248 0.239 0.148 0.225 0.083 0.077 0.076 0.108 0.169
Community xed eects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
y
De-meaned diversity/segregation variables. Standard errors are robust and clustered at the community level. Included variables not shown: sex, dummy variables for age
and education categories, risk and time preference, married status, linear spline for log PCE, and a constant.
173
and outcomes, I examine outcomes whose associations with religiosity are not aected
by community diversities among Muslims. Two outcomes appear to be \immune" to
the eect of diversity among Muslims: trust of coreligionists and tolerance of inter-
faith marriage.
In contrast, I nd many of the eects of community diversity on the association
between religiosity and outcomes to be absent among non-Muslims. Village and
subdistrict diversities do not aect the association between religiosity and almost
all measures of community cohesion, except for the willingness to entrust neighbors
to watch one's children. In contrast to results from the Muslim subset, village and
subdistrict diversities are associated with a weaker association between religiosity and
trust of coreligionists and coethnics { although the coecient on the religiosity
village diversity interaction is not signicant for trust of coethnics. They are not
associated with the association between religiosity and measures of tolerance, except
for tolerance of non-coreligionists' house of worship.
Meanwhile, among non-Muslims, subdistrict segregation weakens the links be-
tween religiosity and community cohesive behaviors. The associations between re-
ligiosity and helpfulness as well as their willingness to entrust neighbors to watch
their house are weaker among non-Muslims living in more segregated subdistricts.
Similarly, the association between religiosity and the sense of safety for non-Muslims
are weaker in more segregated subdistricts. Village segregation does not aect the
association between religiosity and any tolerance measure among non-Muslims; how-
ever, subdistrict segregation is associated with a weaker link between religiosity and
tolerance in having a non-coreligionist stay in one's home.
174
4.A.5 Community heterogeneity and the very religious
Tables 4.15 and 4.16 present results of the regressions with a specications similar
to (4.4), but with interactions between community heterogeneity and the \very reli-
gious", instead of the \religious", dummy variable.
Table 4.15: Diversity, Segregation, and Community Cohesion & Trust
Beliefs
Willing
to help
Village is safe [. . . ] Trust nbr. to watch Trust [. . . ] to return lost wallet
generally at night kid(s) house neighbors strangers police
A. Village Heterogeneity (1) (2) (3) (4) (5) (6) (7) (8)
Religious/ very religious 0.013

0.025

0.007 0.010 -0.001 0.105

0.024

0.140

(1.88) (4.62) (1.16) (0.86) (-0.18) (6.12) (1.78) (6.92)
Very religious 0.163

0.177

0.097

0.058

0.059

0.092

-0.053

0.171

(11.99) (11.56) (6.36) (2.59) (3.65) (3.08) (-2.14) (6.05)
. . . vilage diversity
y
-0.080 -0.119

-0.072 -0.073 -0.041 -0.154 -0.069 0.094
(-1.31) (-1.88) (-1.07) (-0.69) (-0.50) (-1.10) (-0.57) (0.76)
. . . village segregation
y
-0.097 -0.010 0.097 0.201 0.310 1.123

0.659 -0.030
(-0.37) (-0.04) (0.30) (0.67) (1.10) (1.88) (1.42) (-0.09)
P-val of joint test of:
Village interactions
N 28287 28284 28282 21304 28285 27695 26777 26196
Adj. R
2
0.085 0.094 0.090 0.101 0.059 0.106 0.063 0.084
B. Subdistrict Heterogeneity (1) (2) (3) (4) (5) (6) (7) (8)
Religious/ very religious 0.013

0.025

0.007 0.013 -0.001 0.105

0.025

0.139

(1.93) (4.56) (1.17) (1.10) (-0.13) (6.13) (1.87) (6.87)
Very religious 0.163

0.176

0.094

0.057

0.061

0.093

-0.051

0.168

(11.91) (11.41) (6.15) (2.56) (3.93) (3.09) (-2.08) (6.03)
. . . subdistrict diversity
y
-0.020 -0.082 -0.041 -0.116 -0.040 -0.087 0.048 0.077
(-0.40) (-1.34) (-0.70) (-1.03) (-0.58) (-0.62) (0.41) (0.66)
. . . subdistrict segregation
y
-0.477

-0.123 -0.184 0.279 -0.090 0.426 0.068 -0.828
(-2.12) (-0.47) (-0.69) (0.80) (-0.36) (0.48) (0.15) (-1.44)
P-val of joint test of:
Subdistrict interactions
N 28561 28558 28556 21511 28559 27967 27048 26466
Adj. R
2
0.084 0.092 0.090 0.102 0.058 0.105 0.062 0.084
Community xed eects Yes Yes Yes Yes Yes Yes Yes Yes
Religion xed eects Yes Yes Yes Yes Yes Yes Yes Yes
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
y
De-meaned diversity/segregation variables. Standard errors are robust and clustered at the community level. Included variables not shown:
sex, dummy variables for age and education categories, risk and time preference, married status, linear spline for log PCE, and a constant.
175
Table 4.16: Diversity, Segregation, and Discriminative Trust & Tolerance
Trust [. . . ] more Tolerate non-corlgn living in [. . . ] Tolerate non-corlgn to [. . . ]
corelgn coethnics village neighbor house marry rltv. bld h. wrshp
A. Village Heterogeneity (1) (2) (3) (4) (5) (6) (7)
Religious/ very religious 0.083

0.066

-0.059

-0.078

-0.121

-0.130

-0.101

(7.90) (6.09) (-5.95) (-8.05) (-10.78) (-9.32) (-8.29)
Very religious 0.158

0.065

-0.007 -0.013 -0.091

-0.145

-0.102

(8.24) (3.31) (-0.39) (-0.73) (-4.26) (-5.92) (-4.43)
. . . vilage diversity
y
0.063 -0.092 0.100 0.102 0.140 -0.120 0.054
(0.62) (-1.05) (1.53) (1.44) (1.61) (-1.14) (0.64)
. . . village segregation
y
-0.155 0.080 -0.224 0.026 0.051 0.617

0.214
(-0.41) (0.33) (-0.86) (0.09) (0.19) (2.15) (0.93)
P-val of joint test of:
Village interactions
N 28286 28286 28287 28287 28286 28286 28286
Adj. R
2
0.155 0.180 0.220 0.248 0.259 0.247 0.256
B. Subdistrict Heterogeneity (1) (2) (3) (4) (5) (6) (7)
Religious/ very religious 0.085

0.066

-0.058

-0.080

-0.122

-0.130

-0.101

(8.12) (6.05) (-5.92) (-8.28) (-10.94) (-9.43) (-8.37)
Very religious 0.157

0.066

-0.008 -0.014 -0.090

-0.139

-0.100

(8.59) (3.51) (-0.48) (-0.77) (-4.28) (-5.77) (-4.49)
. . . subdistrict diversity
y
0.050 -0.047 0.080 0.121

0.106 -0.141 0.075
(0.53) (-0.51) (1.18) (1.86) (1.26) (-1.48) (0.88)
. . . subdistrict segregation
y
-0.997

-0.692

0.103 -0.009 0.121 1.069

0.438
(-1.94) (-1.89) (0.34) (-0.03) (0.36) (3.06) (0.98)
P-val of joint test of:
Subdistrict interactions
N 28560 28560 28561 28561 28560 28560 28560
Adj. R
2
0.155 0.179 0.221 0.250 0.261 0.246 0.261
Community xed eects Yes Yes Yes Yes Yes Yes Yes
Religion xed eects Yes Yes Yes Yes Yes Yes Yes
t statistics in parentheses.

p< 0:1,

p< 0:05,

p< 0:01
y
De-meaned diversity/segregation variables. Standard errors are robust and clustered at the community level. Included variables not
shown: sex, dummy variables for age and education categories, risk and time preference, married status, linear spline for log PCE, and a
constant.
176
Appendix 4.B Data appendix
4.B.1 Data sources
IFLS
IFLS is a publicly-available longitudinal, socio-economic household survey based on a
sample representing 83% of the Indonesian population living in 13 out of 26 provinces
in 1993. Four full-sample waves of the survey (IFLS1-IFLS4) have been conducted
in 1993, 1997, 2000, and late 2007.
11
Subsequent IFLS's placed a signicant eort
in tracking down original IFLS1 households as well as their split-os, resulting in
a low attrition rate over the course of four waves.
12
In IFLS4, for original IFLS1
households, all members were interviewed irrespective of whether they were there
in IFLS1. Meanwhile, for split-o households, only original IFLS1 respondents, and
their spouses and children were interviewed.
2000 Indonesian Population Census
The 2000 population census collected a basic set of information on individuals (e.g.,
sex, age, highest level of education, religion, and ethnicity) and their households in
all of Indonesia. In principle, data were supposed to be collected over the whole
population; however, due to local con
icts in some regions during the post-transition
Indonesia, the numbers for these areas were estimated (Suryadinata et al., 2003, p.
xxiv). Since none of the IFLS communities are located in these high-con
ict areas,
11
In 1998, an additional survey interviewing 25% of the sample, known as IFLS2+, was conducted
to measure the impact of the economic crisis.
12
Among all individuals living in IFLS1 households, 81.7% were recontacted in IFLS4, while
among the main respondents of IFLS1, 87% were recontacted in IFLS4. The dierence in the
recontact rates are due to the fact that, unlike in subsequent waves, not all members of the households
in IFLS1 were interviewed. See Strauss et al. (2009) for details.
177
this issue does not aect the numbers used in this analysis. This chapter uses the
individuals' religion variable to construct community religious heterogeneity variables.
2000 Indonesia Poverty Map
The 2000 Indonesia Poverty Map was developed by Indonesia's statistical agency
BPS-Statistics using the census data based on the initiatory work of Suryahadi et al.
(2003). To develop the map, researchers imputed the per-capita expenditure of each
household in the population by applying observed correlations between household
characteristics and per-capita expenditure from a survey that contains both infor-
mation onto the 2000 population census data (which only has the former). These
imputed data can then be used to construct aggregate expenditures of communi-
ties, as well as expenditure distributions of in communities, such as the expenditure
subdistrict Gini used in this chapter.
Village Potential Statistics (Podes)
Podes is a census of villages in Indonesia, and contains information on the charac-
teristics of each of Indonesia's villages. I use Podes 2005 { which is the most recent
Podes dataset prior to IFLS4 { to obtain topographical, demographical and other
village characteristics used as the community control variables.
4.B.2 Merging the datasets
Table 4.17 summarizes the data source for each of the variables used for the analyses
in Chapters 3 and 4. IFLS4 contains the main outcome variables, as well as indi-
vidual and household variables. The total number of adults { dened as individuals
aged 15 years old and above { in the IFLS4 sample is 29,054. To focus on the ve
178
Table 4.17: Variables by data sources
Data source Variables
IFLS HH
y
All outcome variables, religiosity, sex, age, years of education,
risk and time preference, married status, education history,
monthly PCE
IFLS Community
y
Natural disaster in the previous 5 years.
Population Census 2000
z
Village, subdistrict religious segregation and diversity.
Poverty Map 2000 Subdistrict PCE gini coecient.
Podes 2005 Urban/rural status, topography (plain, coast, valley,
hill/mountainous), population density, receptions of TV and
radio broadcasts, distance to district and subdistrict capitals.
Notes:
y
Except for education history, which was constructed by merging using previous waves, all variables come
from IFLS4.
z
These variables are author's own calculation.
main world religions, 17 observations were dropped either because they refused to
answer (10 observations), listed \other" as their religion (5 observations) or are Con-
fucians (2 observations). Therefore, 29,037 individual observations are available for
the individual religiosity analysis of Chapter 3.
The rst step was to combine the IFLS4 household and community datasets. For
the community variable of interest, which is recent experience of natural disaster,
168 observations were missing from the IFLS4 community dataset. These missing
observations come from 51 communities outside of the original IFLS1 communities.
Next, I combined the IFLS dataset with the population census and the poverty
map. I therefore merged IFLS4 villages' BPS location codes with those of the 2000
population census.
13
Conveniently, IFLS4 includes a 2000 BPS village-level codes,
which makes the merging between the IFLS4 and the 2000 Census as well as the 2000
Poverty Map a relatively painless process. Nonetheless, these merges still produced
an additional 986 observations with no village-level heterogeneity variables.
14
13
IFLS village location codes are not available in its public data and need to be requested from
the IFLS support team.
14
The merges between IFLS4, the census, and the poverty map produced a total of 988 unmatched
observations, two of which overlapped with unmatched set from the merge between IFLS4 household
and community datasets. Between IFLS4 and the census, 649 observations were unmatched, while
between IFLS and the poverty map, 592 individual observations were unmatched; 259 of these
unmatched observations overlapped.
179
Finally, I merged the dataset with Podes 2005 using a dictionary that was con-
structed to match the 2000 and 2005 BPS village codes. This nal process resulted in
an additional 82 unmatched observations.
15
Hence, for analyses requiring community
regressors, the dataset has 27,844 observations with complete community variables.
15
Between IFLS4 and Podes 2005, 312 observations were unmatched, but 230 observations were
in the set of previous unmatched observations.
180
Appendix 4.C A note on segregation measures
Massey and Denton (1988) propose ve dimensions of segregation: evenness (dissim-
ilarity), exposure (isolation), concentration (the amount of physical space occupied
by the minority group), clustering (the extent to which minority neighborhoods abut
one another), and centralization (proximity to the center of the city). Of these dimen-
sions, the last three require information of the physical space that is not available in
the Indonesian Census data. Hence, the summary focuses mainly on indices looking
at the rst two.
4.C.1 Criteria for evaluation
Frankel and Volij (2011) examine axiomatically the theoretical properties of eight
common multi-group (instead of a two-group) ordinal segregation measures. To eval-
uate these measures, they rst consider the following six properties to evaluate these
measures (all explanations are based on their example of analyzing segregation in
schools within a district):
16
1. Scale Invariance. Scale Invariance states that the scale of a district does
not matter: if the number of students in each ethnic group in each school is
multiplied by the same positive factor, segregation is unaected.
2. Symmetry. A segregation ordering satises Symmetry if it is invariant to a
renaming of the ethnic groups (e.g., if blacks are renamed to whites and vice
versa).
16
The authors also discuss the issue of the decomposability properties of these indices, which I
do not summarize here.
181
3. Independence. Independence states that if the students in a subset of schools
are reallocated within that subset, then segregation in the whole district rises
if and only if it rises within the subset.
4. School Division Property (SDP). The School Division Property states that
splitting a school into two schools (a) cannot lower segregation in the district,
and (b) leaves segregation unchanged if the ethnic distributions of the two
resulting schools are identical.
5. Composition Invariance. Composition Invariance states that the segregation
of a district does not change if the number of students in a given ethnic group
is multiplied by the same positive constant throughout the district (e.g., if the
number of blacks in every school rises by ten percent).
6. Group Division Property (GDP). The Group Division Property states that
the segregation of a district does not change if an existing ethnic group is sub-
divided into two groups that have the same distribution across schools (e.g., if
whites are divided by gender and white boys have the same distribution across
schools as white girls).
4.C.2 Indices
There are many segregation indices in the literature { Massey and Denton (1988)
examined 20 dierent segregation indices. Frankel and Volij (2011) examined a subset
that are used often in the literature (i.e., the dissimilarity and isolation indices), as
well as some lesser used ones. A similar set of popular measures for multi-group
segregation was also analyzed by Reardon and Firebaugh (2002).
17
The measures are
17
Two indices included by Reardon and Firebaugh (2002) but not included here is the squared
coecient of variation index, and the relative diversity index.
182
as follows:
1. Atkinson Index (A). Based on the Atkinson inequality index, the formula for
the Atkinson Index is:
A
w
= 1
X
n2N
Y
r2R
(t
n
r
)
wr
: (4.6)
wheret
n
r
=
T
r
n
Tr
(T denotes the population size) andw
r
is an arbitrary and xed
non-negative weight for each religion that sums up to one.
The drawback of the Atkinson index is that it is sensitive to zeroes in the case
of three or more social/religious groups. It is therefore not designed to compare
communities with dierent number of nonempty groups. It is possible to modify
the index to address the problem, but by doing so, makes the index violate GDP.
2. Mutual Information Index (M). This is the index used in this chapter,
which takes the dierence between the entropy at the community level and the
weighted-average of the entropy at each community's sub-communities:
M
i
=h
i

X
n2N

n
h
n
: (4.7)
where the entropy for community or subcommunity n, h
n
is calculated as:
h
n
=
R
X
r
s
nr
:ln(
1
s
nr
) (4.8)
Frankel and Volij show that this index satises 5 of the 6 above properties for
multi-group analysis, except for Composition Invariance. However, Coleman et
al. (1982) argue that composition invariance is not necessary when segregation
is interpreted in terms of the frequency of inter-group relations.
183
3. Entropy Index (H). Entropy Index is essentially a Mutual Information Index
normalized to between 0 and 1: For each community i, H
i
is calculated by
dividing the Mutual Information Index at the community level with the entropy
of the community's religious distribution:
H
i
= 1
P
n2N

n
h
n
h
i
: (4.9)
In addition to violating the Composition Invariance,H also violates GDP: With
this index, splitting a religion into two sub-religions with identical distribution
across sub-communities changes the community's segregation ordering.
4. Dissimilarity Index (D). Both the Dissimilarity Index and the Gini Index be-
low are measures of evenness. Evenness captures the dierential distribution of
dierent social groups among sub-communities in the community. Segregation
is minimized if subcommunities in the community have similar distributions of
social groups (and, hence, a similar distribution of social groups with the com-
munity overall). For a multi-group dissimilarity index, let P denote the religious
distribution in a community, and letI(P ) =
P
r2R
P
r
(1P
r
). The multi-group
dissimilarity index is therefore:
D
i
=
1
2I(P )
X
r2R
P
r
X
n2N

n
jr
n
r
1j: (4.10)
where r
n
r
is the disproportionality ratio of religion r in subcommunity n.
18
D
takes on a value between zero and one. By multiplying by I(P ), we obtain the
unnormalized version of the index,D
0
which equals to the minimum proportion
18
That is, r
n
r
=
N
r
n
:N
Nr:Nn
where N
r
is the population with religion r in the community, N
n
is the
population in sub-community n and N is the total population.
184
of individuals that would need to move between sub-communities, for xed
sub-communities, in order to completely integrate the community.
Of the six properties, the multi-group dissimilarity index, D violates Indepen-
dence and GDP, and maintains Composition Invariance only when the index is
used to analyze two social groups. However, at least for the two-group case,
Massey and Denton (1988) proposed the use of this indicator to maintain histor-
ical comparability given that past studies of US data often utilized this measure.
5. Gini Index (G). The multi-group Gini index measures the area between the
Lorenz curve and the 45 degree line. Similar to the Dissimilarity Index above,
G violates Independence and GDP, and maintains Composition Invariance only
when the index is used to analyze two social groups.
6. Normalized Exposure Index (NE). This index captures the exposure aspect
of segregation and in the two-group setting, it is known as the Isolation Index.
This index captures the degree to which members of a particular social group
is exposed to other members of their social group, correcting for the fact that
groups forming a larger share of the population have naturally higher exposure
rates. In the two-group case (following the formulation of Cutler et al. (1999)),
this index for a community with N subcommunities (indexed by n) can be
constructed as:
I =
P
n2N

Trn
Tr

Trn
Tn

Tr
T
min(1;
Tr
T
smallest
)
Tr
T
(4.11)
where T is the population size and like before, r indexes the social group (but
in a case of two groups, an insider and an outsider), and T
smallest
denotes the
population of the smallest sub-community in the community.
185
Meanwhile, following James (1986), the multi-group NE/Isolation Index is:
NE
i
=
X
n2N
X
r2R

n
P
r
P
r
1P
r
(r
n
r
1)
2
(4.12)
Frankel and Volij show that this multi-group NE violates GDP and is not Com-
position Invariant. With more than two groups, NE also violates the Indepen-
dence property, which requires that reallocation within a subset of subcommu-
nities only increases segregation if segregation increases within this subset of
communities.
7. Frankel and Volij (2011) also discuss the Clotfelter's index (C
K
) and Card
and Rothstein index(CR). Both of these indices are calculated for a xed
set of groups.
Overall, Frankel and Volij nd that the Mutual Information and the Atkinson
Index to be the most well-behaved indicators. The Atkinson Index, however, does not
handle cases where communities have dierent numbers of social groups very well. In
addition, the Mutual Information index can be nicely decomposed into between-group
and within-group measures. These ndings support a similar (but not axiomatic)
analysis of multi-group segregation indicators by Reardon and Firebaugh (2002), who
also recommended the use of the Mutual Information index for multi-group analysis
given its mathematical properties.
In addition, Echenique and Fryer (2007) propose a measure based on social net-
works and interactions. However, their measure requires the availability of network
structure information.
186
4.C.3 Usage in the literature
The indices that are most often used in the literature, in particular for two-group
analysis, are the dissimilarity index and the isolation index (Echenique and Fryer,
2007; Iceland and Scopilliti, 2008). For instance, in their series of papers on seg-
regation, Cutler, Glaeser and Vigdor typically use both indices (see, e.g., Cutler et
al., 1999, 2008a,b). However, they utilized the indices in the two-group setting, e.g.
black vs. others (Cutler et al., 1999) or immigrant from a particular country of origin
vs. others (including immigrants from other countries) (Cutler et al., 2008a). For
instance, in the case of the analysis of immigrants, the unit of analysis is thus a
country-of-origin group residing in one specic city or MSA, which they dened as an
immigrant community. In the case of religion, this is tantamount to dening segrega-
tion for each religion in each community, instead of having a community indicator of
segregation. In the case of the isolation index, to aggregate to one community indica-
tor, James (1986) proposes a weighted average of the isolation index across religions.
However, this aggregation violates the Independence axiom (Reardon and Firebaugh,
2002; Frankel and Volij, 2011).
In the case of Indonesia, Barron et al. (2009) implemented the Entropy Index
(which is the Mutual Information index divided by the entropy at the community
level) to capture ethnic/religious clustering within the district.
187
Chapter 5
Conclusion
The three essays here present empirical evidence on how economic incentives and
social identities may shape social cooperation. Chapter 2 presents results from an
experimental test a social network formation model that anonymize agents and, in
doing so, isolate their focus on the economic incentives of cooperation. Meanwhile,
Chapters 3 and 4 suggest that religious identities may have important roles in shaping
individual cooperative attitudes. This chapter will conclude with a brief review the
ndings so far and discuss their limitations, possible implications, as well future
research directions.
5.1 Social networks, cooperation, economic incentives
The network formation experiment provided empirical evidence on how social net-
works are formed in a controlled environment where the economic benets and costs
are fully known. Unsurprisingly, we nd that these benets are important in shap-
ing these networks. More specically, we nd that on average, myopic rationality in
linking decisions { i.e., where decisions to form, maintain, or break a bilateral link
depends solely the immediate net benet on the link in question { approximates in-
dividual behavior rather well. As suggested in Table 2.6, a signicant majority of the
linking decisions in our experiment did re
ect myopic rational behavior.
Nonetheless, under certain conditions, people make decisions that deviate from
188
myopic rationality, and these deviations often led the dynamic process of network
formation to nal network outcomes that stray from the theoretical predictions. As
shown in Table 2.2, this often leads to a dynamic process that does not end with the
predicted, unique PWS network. Moreover, with multiple PWS networks with dier-
ent individual payos where the Pareto-superior PWS network is unreachable under
perfectly myopic-rational behavior, individuals appear able to strategically deviate
from myopic rationality to reach this Pareto-superior network.
On the theory front, understanding when these deviations occur and their implica-
tions on predicted outcomes is, therefore, important. Our data suggest the tendency
towards fewer deviations from myopic rationality as marginal losses increase and as
matches get closer to the end { all strongly suggestive that players optimize subject to
imperfect choice, imperfect foresight and/or imperfect information processing capac-
ity. These ndings should motivate further research to incorporate these behavioral
imperfections into existing theoretical models.
Meanwhile, on the experimental front, ecological validity is a concern. Indeed, we
feel that our cost and benet representation of adding and removing links captures
the essence of social networks in an excessively abstract way. Real-world applica-
tions of social networks are rarely anonymous; moreover, evidence presented in the
previous two chapters suggests that social identities can be important for coopera-
tion. This suggests two ways forward. First, the use of laboratory studies in the
eld or laboratory studies that exploit social technologies (facebook, twitter, second
life, etc.) would add a more realistic dimension to the network formation problem.
Second, manipulations of the level of anonymity in an otherwise identical experiment
can oer insights into the relative importance of identities over economic incentives in
network formation. Either modication can be implemented without compromising
the controlled environment of the laboratory.
189
5.2 Religious identities and social cooperation
On the other hand, evidence from observational data on Indonesia suggest that reli-
gious identities matter for cooperation. In particular, I nd strong and robust pos-
itive correlations between religiosity and cooperative attitudes toward members of
one's community, but not to strangers. Religiosity is also associated with a stronger
trust of coreligionists and coethnics, as well as intolerance of non-coreligionists. For
these measures, the strengths of their associations with individual religiosity are much
stronger than those with gender, education, or per-capita expenditure. These ndings
provide evidence of the parochial-altruistic nature of religion in Indonesia. In a more
benign form, it is present in all religions. However, as evidenced in the negative links
between religiosity and all tolerance measures for Indonesian Muslims (Table 3.10),
this manifestation of parochial altruism tends to be strongest among these adherents
of the country's majority religion.
At the community level, consistent with the empirical ndings from the ethnic
fragmentation literature, we nd a negative correlation between religious diversity
and generalized trust. At the same time, in support of the inter-group contact hy-
pothesis, diversity tends to be positively correlated with tolerance. People living in
segregated communities, in contrast, tend to have more trust, but are less tolerant.
The individual-level evidence suggests the presence discriminatory preference. How-
ever, our examination of the interactions between religiosity and community diversity
and segregation suggest that social interactions may also matter: In more diverse
communities, the associations between religiosity, and community cohesiveness mea-
sures as well as intolerance tend to be weaker; they tend to be stronger in more
segregated communities.
There are a number of directions for further research. As discussed in the chap-
190
ters, some issues of endogeneity remain to be addressed. In particular, there is the
issue of residential sorting in the community-level analysis. Although the data used
to calculate religious diversity and segregation were collected in 2000 { eight years
prior to the outcomes { these variables may not be free of endogeneity if the resi-
dential sorting process (that may be in
uenced by the attitudes studied here) has
reached its equilibrium prior to 2000. Further research should work on nding a bet-
ter identication strategy { perhaps, from a credible instrument { to allow for causal
inference.
1
Second, a natural follow-up to this study is to look at the impact of religious iden-
tities on a number of aggregate economic outcomes, particularly those are in
uenced
by inter-group cooperation. In addition, the literature on religion as a club suggest
the value, cost, as well as potential economic implications of the parochial nature
of religion (e.g., see Iannaccone, 1992; Berman, 2000; Chen, 2010). This literature
predicts an association between religious identities and a number of outcomes, to wit,
fertility decisions, labor market outcomes, and abilities to cope with economic shocks.
Finally, recent developments in local governance in Indonesia may allow for a
more careful analysis of the eects of religious identities on various outcomes. The
fall of the secular, authoritarian, and centralized government in 1998 has allowed
a more important role for religion in the public space. With the implementation
of decentralization in 2001, many regional governments have started implementing
policies with distinctly religious
avor (Bush, 2008).
2
These policy changes provide
an opportunity to study in detail whether such high-level policy changes have an
eect on religious identities, and to the extent that they do, whether these changes
1
One possibility for this is to perhaps use a natural experiment from government a program that
exogenously move groups of people from one place to another, such as the Indonesian transmigration
program.
2
As of 2008, Bush (2008) counted 78 district regulations in 52 districts/municipalities that are
religion-based.
191
aect intra- and inter-group cooperation in the diverse society that is Indonesia.
192
Chapter bibliography
Berman, Eli, \Sect, Subsidy, and Sacrice: An Economist's View of Ultra-Orthodox
Jews," Quarterly Journal of Economics, August 2000, 115, 905{953.
Bush, Robin, \Regional `Sharia' Regulations in Indonesia: Anomaly or Symptom?,"
in Greg Fealy and Sally White, eds., Expressing Islam: Religious Life and Politics
in Indonesia, Singapore: Institute of Southeast Asian Studies, 2008.
Chen, Daniel L., \Club Goods and Group Identity: Evidence from Islamic Resur-
gence during the Indonesian Financial Crisis," Journal of Political Economy, April
2010, 118 (2), 300{354.
Iannaccone, Laurence R., \Sacrice and Stigma: Reducing Free-riding in Cults,
Communes, and Other Collectives," The Journal of Political Economy, April 1992,
100 (2), 271{291.
193
Comprehensive bibliography
Ai, Chunrong and Edward C. Norton, \Interaction terms in logit and probit
models," Economics Letters, July 2003, 80, 123{129.
Akerlof, George A. and Rachel E. Kranton, \Economics and Identity," Quar-
terly Journal of Economics, 2000, 115 (3), 715{753.
Alesina, Alberto and Ekaterina Zhuravskaya, \Segregation and the Quality of
Government in a Cross Section of Countries," American Economic Review, August
2011, 101 (5), 1872{1911.
and Eliana La Ferrara, \Participation in Heterogeneous Communities," Quar-
terly Journal of Economics, 2000, 115 (3), 847{904.
and , \Who trusts others?," Journal of Public Economics, August 2002, 85 (2),
207{234.
Allport, Gordon W., The nature of prejudice, Reading Mass.: Addison-Wesley
Pub. Co., 1979 [1954].
and Bernard M. Kramer, \Some roots of prejudice," The Journal of general
psychology, 1946, 22, 9.
Altonji, J. G, T. E Elder, and C. R Taber, \Selection on observed and un-
observed variables: Assessing the eectiveness of Catholic schools," Journal of
Political Economy, 2005, 113 (1), 151{184.
194
Anderson, Lisa, Jennifer Mellor, and Jerey Milyo, \Did the Devil Make
Them Do It? The Eects of Religion in Public Goods and Trust Games," Kyklos,
2010, 63 (2), 163{175.
Arrow, Kenneth J., \What Has Economics to Say about Racial Discrimination?,"
The Journal of Economic Perspectives, April 1998, 12 (2), 91{100.
Bala, Venkatesh and Sanjeev Goyal, \A Noncooperative Model of Network For-
mation," Econometrica, September 2000, 68 (5), 1181{1229.
Barron, Patrick, Kai Kaiser, and Menno Pradhan, \Understanding Variations
in Local Con
ict: Evidence and Implications from Indonesia," World Development,
March 2009, 37 (3), 698{713.
Batson, Charles, Patricia Schoenrade, and W. Larry Ventis, Religion and
the Individual : A Social-Psychological Perspective, New York: Oxford Univ. Press,
1993.
Bellows, John and Edward Miguel, \War and Local Collective Action in Sierra
Leone," Working Paper December 2008.
Benjamin, Daniel J., James J. Choi, and Georey W. Fisher, \Religious
Identity and Economic Behavior," NBER Working Paper 15925 April 2010.
Berg, Joyce, John Dickhaut, and Kevin McCabe, \Trust, Reciprocity, and
Social History," Games and Economic Behavior, July 1995, 10 (1), 122{142.
Berman, Eli, \Sect, Subsidy, and Sacrice: An Economist's View of Ultra-Orthodox
Jews," Quarterly Journal of Economics, August 2000, 115, 905{953.
Bernhard, Helen, Urs Fischbacher, and Ernst Fehr, \Parochial altruism in
humans," Nature, August 2006, 442 (7105), 912{915.
195
Berninghaus, Siegfried K., Karl-Martin Ehrhart, and Marion Ott, \A net-
work experiment in continuous time: The in
uence of link costs," Experimental
Economics, September 2006, 9 (3), 237{251.
, , , and Bodo Vogt, \Evolution of networks|an experimental analysis,"
Journal of Evolutionary Economics, January 2007, 17 (3), 317{347.
Bertrand, Jacques, Nationalism and ethnic con
ict in Indonesia, Cambridge, UK;
New York: Cambridge University Press, 2004.
Bowles, Samuel and Herbert Gintis, A cooperative species: human reciprocity
and its evolution, Princeton: Princeton University Press, 2011.
Bramoull e, Y. and R. Kranton, \Risk-sharing networks," Journal of Economic
Behavior & Organization, 2007, 64 (3-4), 275{294.
Burger, Martijn J. and Vincent Buskens, \Social context and network forma-
tion: An experimental study," Social Networks, January 2009, 31 (1), 63{75.
Bush, Robin, \Regional `Sharia' Regulations in Indonesia: Anomaly or Symptom?,"
in Greg Fealy and Sally White, eds., Expressing Islam: Religious Life and Politics
in Indonesia, Singapore: Institute of Southeast Asian Studies, 2008.
Callander, S. and C. R. Plott, \Principles of network development and evolution:
An experimental study," Journal of Public Economics, 2005, 89 (8), 1469{1495.
Camerer, Colin and Teck Hua Ho, \Experience-weighted Attraction Learning in
Normal Form Games," Econometrica, 1999, 67 (4), 827{874.
Carter, Michael R. and Marco Castillo, \Trustworthiness and Social Capital
in South Africa: Analysis of Actual Living Standards Data and Artifactual Field
196
Experiments," Economic Development and Cultural Change, July 2011, 59 (4),
695{722.
Chaney, Eric, \Tolerance, Religious Competition and the Rise and Fall of Muslim
Science," 2008.
Charness, Gary, Francesco Feri, Miguel A. Mel endez-Jim enez, and
Matthias Sutter, \Equilibrium Selection in Experimental Games on Networks,"
Working Paper February 2012.
Chen, Daniel L., \Club Goods and Group Identity: Evidence from Islamic Resur-
gence during the Indonesian Financial Crisis," Journal of Political Economy, April
2010, 118 (2), 300{354.
Choi, Jung-Kyoo and Samuel Bowles, \The Coevolution of Parochial Altruism
and War," Science, October 2007, 318 (5850), 636 {640.
Clingingsmith, David, Asim Ijaz Khwaja, and Michael Kremer, \Estimating
the Impact of The Hajj: Religion and Tolerance in Islam's Global Gathering," The
Quarterly Journal of Economics, 2009, 124 (3), 1133 {1170.
Coleman, James, Thomas Hoer, and Sally Kilgore, \Achievement and Seg-
regation in Secondary Schools: A Further Look at Public and Private School Dif-
ferences," Sociology of Education, April 1982, 55 (2/3), 162{182.
Costa, Dora L. and Matthew E. Kahn, \Civic Engagement and Community
Heterogeneity: An Economist's Perspective," Perspectives on Politics, 2003, 1 (01),
103{111.
Croson, Rachel and Uri Gneezy, \Gender Dierences in Preferences," Journal
of Economic Literature, May 2009, 47, 448{474.
197
Cutler, David M., Edward L. Glaeser, and Jacob L. Vigdor, \The Rise and
Decline of the American Ghetto," Journal of Political Economy, June 1999, 107
(3), 455.
, , and , \Is the Melting Pot Still Hot? Explaining the Resurgence of Im-
migrant Segregation," Review of Economics and Statistics, August 2008, 90 (3),
478{497.
, , and , \When are ghettos bad? Lessons from immigrant segregation in the
United States," Journal of Urban Economics, May 2008, 63 (3), 759{774.
Deaton, A. S, \Aging, religion, and health," NBER Working Paper 15271, National
Bureau of Economic Research 2009.
Di Cagno, Daniela and Emanuela Sciubba, \The Determinants of Individual
Behaviour in Network Formation: Some Experimental Evidence," in Mohammed
Abdellaoui, John D. Hey, W. Leinfellner, G. Eberlein, and S. H. Tijs, eds., Advances
in Decision Making Under Risk and Uncertainty, Vol. 42 of Theory and Decision
Library, Springer Berlin Heidelberg, 2008, pp. 219{241.
Eagly, Alice H. and Maureen Crowley, \Gender and helping behavior: A meta-
analytic review of the social psychological literature.," Psychological Bulletin, 1986,
100 (3), 283{308.
Echenique, F. and R. G. Fryer, \A Measure of Segregation Based on Social
Interactions," The Quarterly Journal of Economics, May 2007, 122 (2), 441{485.
Fafchamps, M. and S. Lund, \Risk-sharing networks in rural Philippines," Journal
of Development Economics, 2003, 71 (2), 261{288.
198
Fafchamps, Marcel, Market institutions in Sub-Saharan Africa : theory and evi-
dence, Cambridge Mass.: MIT Press, 2004.
Falk, Armin and Michael Kosfeld, \It's all about connections: Evidence on
network formation," Working paper, Institute for Empirical Research in Economics,
University of Zurich March 2003.
Fatas, Enrique, Miguel Mel endez-Jim enez, and Hector Solaz, \An experi-
mental analysis of team production in networks," Experimental Economics, 2010,
13 (4), 399{411.
Fehr, Ernst, \On The Economics and Biology of Trust," Journal of the European
Economic Association, April 2009, 7, 235{266.
Fraley, Chris and Adrian E. Raftery, \Model-Based Clustering, Discriminant
Analysis, and Density Estimation," Journal of the American Statistical Association,
June 2002, 97 (458), 611{631.
and , \MCLUST Version 3 for R: Normal Mixture Modeling and Model-Based
Clustering," Technical Report 504, Department of Statistics, University of Wash-
ington December 2009.
Frankel, David M. and Oscar Volij, \Measuring school segregation," Journal of
Economic Theory, January 2011, 146 (1), 1{38.
Glaeser, Edward L., David I. Laibson, and Bruce Sacerdote, \An Economic
Approach to Social Capital," The Economic Journal, 2002, 112 (483), F437{F458.
, , Jose A. Scheinkman, and Christine L. Soutter, \Measuring Trust,"
Quarterly Journal of Economics, August 2000, 115 (3), 811{846.
199
Goeree, Jacob K., Arno Riedl, and Alja z Ule, \In search of stars: Network for-
mation among heterogeneous agents," Games and Economic Behavior, November
2009, 67 (2), 445{466.
Goyal, Sanjeev and Sumit Joshi, \Unequal connections," International Journal
of Game Theory, 2006, 34 (3), 319{349.
Granovetter, Mark, \The Impact of Social Structure on Economic Outcomes,"
Journal of Economic Perspectives, January 2005, 19 (1), 33{50.
Greene, William, \The behaviour of the maximum likelihood estimator of limited
dependent variable models in the presence of xed eects," Econometrics Journal,
2004, 7 (1), 98{119.
Guiso, Luigi, Paola Sapienza, and Luigi Zingales, \People's opium? Religion
and economic attitudes," Journal of Monetary Economics, January 2003, 50 (1),
225{282.
, , and , \Civic Capital as the Missing Link," in Jess Benhabib, Alberto
Bisin, and Matthew Jackson, eds., Handbook of social economics., Vol. Volume 1,
North-Holland, 2011, pp. 417{480.
Habyarimana, James, Macartan Humphreys, Daniel N. Posner, and
Jeremy M. Weinstein, \Why Does Ethnic Diversity Undermine Public Goods
Provision?," American Political Science Review, November 2007, 101 (04).
Hall, Deborah L., David C. Matz, and Wendy Wood, \Why Don't We Practice
What We Preach? A Meta-Analytic Review of Religious Racism," Personality and
Social Psychology Review, February 2010, 14 (1), 126 {139.
200
Hamoudi, A., \Risk Preferences in Households and Families," Technical Report,
UCLA working paper 2006.
Herings, P. Jean-Jacques, Ana Mauleon, and Vincent Vannetelbosch, \Far-
sightedly stable networks," Games and Economic Behavior, November 2009, 67 (2),
526{541.
Iannaccone, Laurence R., \Sacrice and Stigma: Reducing Free-riding in Cults,
Communes, and Other Collectives," The Journal of Political Economy, April 1992,
100 (2), 271{291.
Iceland, John and Melissa. Scopilliti, \Immigrant Residential Segregation in
U.S. Metropolitan Areas, 1990{2000," Demography, 2008, 45 (1), 79{94.
Jackson, M and A Wolinsky, \A Strategic Model of Social and Economic Net-
works," Journal of Economic Theory, October 1996, 71 (1), 44{74.
Jackson, M. O, \A Survey of Network Formation Models: Stability and Eciency,"
in Gabrielle Demange and Myrna Wooders, eds., Group formation in economics :
networks, clubs and coalitions, Cambridge UK, New York: Cambridge University
Press, 2005, pp. 11{57.
Jackson, Matthew O. and Alison Watts, \The Evolution of Social and Economic
Networks," Journal of Economic Theory, October 2002, 106 (2), 265{295.
James, F. J., \A New Generalized "Exposure-Based" Segregation Index: Demon-
stration in Denver and Houston," Sociological Methods & Research, February 1986,
14 (3), 301{316.
201
Johnson, Megan K., Wade C. Rowatt, and Jordan LaBou, \Priming Chris-
tian Religious Concepts Increases Racial Prejudice," Social Psychological and Per-
sonality Science, April 2010, 1 (2), 119 {126.
Kirchsteiger, Georg, Marco Mantovani, Ana Mauleon, and Vincent Van-
netelbosch, \Myopic or farsighted? An experiment on network formation," Dis-
cussion Paper 8263, Centre for Economic Policy Research, London February 2011.
Knack, Stephen and Philip Keefer, \Does Social Capital Have An Economic
Payo? A Cross-Country Investigation," Quarterly Journal of Economics, 1997,
112 (4), 1251{1288.
Kranton, Rachel E and Deborah F Minehart, \A Theory of Buyer-Seller Net-
works," American Economic Review, June 2001, 91 (3), 485{508.
La Porta, Rafael, Florencio Lopez de Silanes, Andrei Shleifer, and
Robert W. Vishny, \Trust in Large Organizations," The American Economic
Review, May 1997, 87 (2), 333{338.
Landes, David, The Wealth and Poverty of Nations : Why Some Are So Rich and
Some So Poor, New York Ny: W.W. Norton, 1998.
Lehrer, Evelyn L, Religion, economics, and demography : the eects of religion on
education, work, and the family, London: Routledge, 2009.
Levy, Gilat and Ronny Razin, \Religious Beliefs, Religious Participation, and
Cooperation," American Economic Journal: Microeconomics, August 2012, 4 (3),
121{151.
Luttmer, Erzo F. P., \Group Loyalty and the Taste for Redistribution," Journal
of Political Economy, 2001, 109 (3), 500{528.
202
Maluccio, John, Lawrence Haddad, and Julian May, \Social capital and
household welfare in South Africa, 1993-98," Journal of Development Studies, Au-
gust 2000, 36 (6), 54{81.
Massey, D. S. and N. A. Denton, \The Dimensions of Residential Segregation,"
Social Forces, December 1988, 67 (2), 281{315.
McCauley, John F., \Distinguishing Religious and Ethnic Politics: Evidence from
a Field Experiment in West Africa," 2009.
McKelvey, Richard and Thomas Palfrey, \Quantal Response Equilibria for
Normal Form Games," Games and Economic Behavior, July 1995, 10 (1), 6{38.
Miguel, E and M Gugerty, \Ethnic diversity, social sanctions, and public goods
in Kenya," Journal of Public Economics, December 2005, 89 (11-12), 2325{2368.
Mujani, Saiful, \Religious Democrats: Democratic Culture and Muslim Political
Participation in Post-Suharto Indonesia." PhD dissertation, Ohio State University
2004.
Munshi, Kaivan, \Information Networks in Dynamic Agrarian Economies," in
T. Paul Schultz and John A. Strauss, eds., Handbook of development economics,
Vol. 4 of Handbook of economics, Elsevier, 2007, pp. 3085{3113.
Nannestad, P., \What have we learned about generalized trust, if anything?," Annu.
Rev. Polit. Sci., 2008, 11, 413{436.
Narayan, Deepa and Lant Pritchett, \Cents and Sociability: Household In-
come and Social Capital in Rural Tanzania," Economic Development and Cultural
Change, July 1999, 47 (4), 871{897.
203
Neusner, Jacob and Bruce Chilton, Altruism in world religions, Washington,
D.C.: Georgetown University Press, 2005.
Norenzayan, Ara and Azim F. Shari, \The Origin and Evolution of Religious
Prosociality," Science, October 2008, 322 (5898), 58 {62.
Nunn, Nathan and Leonard Wantchekon, \The Slave Trade and the Origins of
Mistrust in Africa," American Economic Review, December 2011, 101 (7), 3221{
3252.
Olken, Benjamin A, \Do Television and Radio Destroy Social Capital? Evi-
dence from Indonesian Villages," American Economic Journal: Applied Economics,
September 2009, 1 (4), 1{33.
Pantz, Katinka, The strategic formation of social networks : experimental evidence,
Aachen: Shaker, 2006.
Parra, Juan Carlos, \The eects of religion on social cooperation: Results from a
eld experiment in Ghana," 2011.
Pettigrew, Thomas F. and Linda R. Tropp, \A Meta-Analytic Test of Inter-
group Contact Theory.," Journal of Personality and Social Psychology, 2006, 90
(5), 751{783.
Putnam, Robert, Robert Leonardi, and Raaella Y. Nanetti, Making Democ-
racy Work : Civic Traditions in Modern Italy, Princeton N.J.: Princeton University
Press, 1993.
Reardon, Sean F. and Glenn Firebaugh, \Measures of Multigroup Segregation,"
Sociological Methodology, 2002, 32 (1), 33{67.
204
, John T. Yun, and Tamela Mcnulty Eitle, \The Changing Structure of School
Segregation: Measurement and Evidence of Multiracial Metropolitan-Area School
Segregation, 1989-1995," Demography, 2000, 37 (3), 351{364.
Rothwell, Jonathan T., \Trust in Diverse, Integrated, Cities: A Revisionist Per-
spective," SSRN Electronic Journal, 2010.
Rue, Bradley J. and Richard Sosis, \Cooperation and the in-group-out-group
bias: A eld test on Israeli kibbutz members and city residents," Journal of Eco-
nomic Behavior & Organization, June 2006, 60 (2), 147{163.
Schechter, L., \Traditional trust measurement and the risk confound: An experi-
ment in rural Paraguay," Journal of Economic Behavior & Organization, 2007, 62
(2), 272{292.
Sheng, Shuyang, \Identication and Estimation of Network Formation Games,"
Job market paper, University of Southern California December 2012.
Sosis, Richard and Bradley J. Rue, \Ideology, religion, and the evolution
of cooperation: Field experiments in Israeli Kibbutzim," Research in Economic
Anthropology, 2004, 23, 89{117.
Strauss, John, Firman Witoelar, Bondan Sikoki, and Anna Wattie, \The
Fourth Wave of the Indonesia Family Life Survey: Overview and Field Report,
Volume 1," RAND Working Paper WR-675/1-NIA/NICHD, RAND April 2009.
Suryadinata, Leo, Evi Arin Nurvidya, and Aris Ananta, Indonesia's popula-
tion : ethnicity and religion in a changing political landscape, Singapore: Institute
of Southeast Asian Studies, 2003.
205
Suryahadi, Asep, Chris Elbers, and Menno Pradhan, Developing a Poverty
map for Indonesia: an initiatory work in three provinces, SMERU Research Insti-
tute and BPS, 2003.
Tan, Jonathan H.W. and Claudia Vogel, \Religion and trust: An experimental
study," Journal of Economic Psychology, December 2008, 29 (6), 832{848.
Theil, Henri and Anthony J. Finizza, \A note on the measurement of racial
integration of schools by means of informational concepts," The Journal of Math-
ematical Sociology, July 1971, 1 (2), 187{193.
Udry, Christopher, \Risk and Insurance in a Rural Credit Market: An Empirical
Investigation in Northern Nigeria," The Review of Economic Studies, July 1994, 61
(3), 495{526.
Uslaner, Eric M., \Segregation, mistrust and minorities," Ethnicities, December
2010, 10 (4), 415{434.
and Richard S. Conley, \Civic engagement and particularized trust," American
Politics Research, 2003, 31 (4), 331{360.
Vigdor, Jacob L., \Community Composition and Collective Action: Analyzing
Initial Mail Response to the 2000 Census," Review of Economics and Statistics,
2004, 86 (1), 303{312.
206

Abstract (if available)

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Essays on the empirics of risk and time preferences in Indonesia

PDF

Essays on development and health economics: social media and education policy

PDF

Essays on health and aging with focus on the spillover of human capital

PDF

Essays in political economy and mechanism design

PDF

Essays on economics of education and private tutoring

PDF

Essays on development and health economics

PDF

A structural econometric analysis of network and social interaction models

PDF

Three essays on economics of early life health in developing countries

PDF

Selection and impacts of early life events on later life outcomes

PDF

Essays on education: from Indonesia to Los Angeles, CA

PDF

The EITC, labor supply, and child development

PDF

Essays on education and institutions in developing countries

PDF

Essays on health insurance programs and policies

PDF

Essays on labor and development economics

PDF

Essays on the new economic history of the Middle East

PDF

Essays in development and experimental economics

PDF

Essays on wellbeing disparities in the United States and their social determinants

PDF

Essays on human capital accumulation -- health and education

PDF

Three essays on supply chain networks and R&D investments

PDF

Essays on development economics and adolescent behavior

Asset Metadata

Creator Gaduh, Arya B. (author)

Core Title Three essays on cooperation, social interactions, and religion

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Publication Date 05/14/2013

Defense Date 04/17/2013

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag economic experiment,Indonesia,OAI-PMH Harvest,Religion,segregation,social capital,social networks

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Strauss, John A. (committee chair), Carrillo, Juan D. (committee member), Nugent, Jeffrey B. (committee member), Painter, Gary Dean (committee member)

Creator Email abgaduh@gmail.com,gaduh@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c3-255969

Unique identifier UC11292621

Identifier etd-GaduhAryaB-1692.pdf (filename),usctheses-c3-255969 (legacy record id)

Legacy Identifier etd-GaduhAryaB-1692.pdf

Dmrecord 255969

Document Type Dissertation

Format application/pdf (imt)

Rights Gaduh, Arya B.

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA