Upload
vandang
View
213
Download
0
Embed Size (px)
Citation preview
Friendship Networks and Political Opinions: A
Natural Experiment among Future French Politicians
Yann Algan∗ Nicolo Dalvit† Quoc-Anh Do‡ Yves Zenou§
March 19, 2018
Abstract
This paper shows how friendship shapes political opinion in newly-formed social networks in a college that produces most of France’s toppoliticians. We make use of a unique natural experiment that randomlyassigns new freshmen into “integration groups”. Pairs of students in thesame integration group are more likely to become friends. This same-group membership thus serves as instrumental variable to estimate theeffect of friendship in dyadic regressions. We find strong, robust effectsof friendship on differences in beliefs after six months: becoming friendsreduces the difference by half a point on a ten-point scale of politicalopinions. It works mostly by keeping friends from diverging, rather thanpulling their opinions closer. The effect becomes insignificant for second-degree friends. The friendship effect completely dominates the effect ofbelonging to the same study group throughout the freshman year. Thefindings highlight the importance of analyzing elicited friendship datainstead of using peer groups.
Keywords: Political beliefs, friendship effect, social networks, conver-gence, learning, natural experiment.
JEL classification codes: C93, D72, Z13.
∗Sciences Po and CEPR. Email: [email protected]†Sciences Po, France. Email: [email protected]‡Harvard University, Sciences Po, and CEPR. Email: [email protected]§Monash University, Australia, IFN, and CEPR. Email: [email protected] are grateful to Alexis Le Chapelain for his contribution at an early stage of the project. We
also thank valuable comments from Alberto Alesina, Yann Bramoulle, Jing Cai, Ruben Durante, RubenEnikolopov, Matt Gentzkow, Ben Golub, Sanjeev Goyal, Emeric Henry, Emir Kamenica, Ethan Kaplan,Eliana La Ferrara, Horacio Larreguy, Alexandre Mas, Friederike Mengel, Markus Mobius, Gautam Rao,Imran Rasul, Adam Szeidl, Andrei Shleifer, Noam Yuchtman, and from seminar and conference par-
1
1 Introduction
What do Emmanuel Macron, Francois Hollande, Nicolas Sarkozy, Jacques Chirac, Fran-
cois Mitterrand, and Georges Pompidou, six among the seven French presidents since
Charles de Gaulle, have in common? What is the smallest common denominator be-
tween twelve of the previous Prime Ministers of the French Republic?1 The short answer
is Sciences Po: the elite French School of Public Affairs that has been training all the top
French politicians since World War II. All the different generations of French political
leaders have met for the first time on Sciences Po’s benches, and most of them have
formed their more solid friendship and network relationship by that time. This has led
to harsh criticism against friendship connection and homophily among the leaders ruling
France.
This anecdotal example raises a broader research question on the role of friends and
social networks in beliefs formation. How does social networks influence political opin-
ions? What determines the persistence of disagreement among friends, and what are
the mechanisms behind it? While there has been extensive research on how leaders and
group of people may influence voting behavior,2 or on the impact of media exposure and
persuasion on turn-out and political opinions,3 the role of social networks and friends on
political opinion has received less attention. This is what we investigate in this paper.
In a seminal theoretical paper, Golub and Jackson (2012) analyze how the speed of
convergence of agents’ beliefs depends on network structure in a model with homophily
to explain the persistence of segregation and persistence in disagreement. But this study
has not led to empirical identification, due to the lack of observational studies where
the social network formation is exogenous and not plagued by sorting and reflection
problem. In a seminal non-experimental study, Lazarfeld et al. (1944) found that US
voters in the 1940 Presidential election were more influenced by friends than mass media.
ticipants at the University of Chicago, the Barcelona Summer Forum, the Social Impacts of EconomicPolicies and Change conference in Heidelberg, the AEA Meeting, the Information Transmission in Net-works conference at Harvard, and Aix-Marseille School of Economics. We are grateful for Sri Srikandan’sassistance with the survey website. Algan acknowledges support from the European Research CouncilGrant 647870. Do acknowledges support from the French National Research Agency’s (ANR) “Investisse-ments d’Avenir” grants ANR-11-LABX-0091 (LIEPP) and ANR-11-IDEX-0005-02, and the Banque deFrance - Sciences Po partnership. The remaining errors are ours.
1Edouard Philippe, Dominique de Villepin, Lionel Jospin, Alain Juppe, Edouard Balladur, MichelRocard, Laurent Fabius, Raymond Barre, Jacques Chaban-Delmas, Maurice Couve de Merville, andMichel Debre, in addition to Jacques Chirac and Georges Pompidou.
2See e.g. Della Vigna and Gentzkow (2010), Gabel and Scheve (2007), Carlsson et al. (2015).3See e.g. Della Vigna and Kaplan (2007), Gentzkow (2006), Gentzkow, Shapiro and Sinkinson (2011),
Kendall, Nannicini and Trebbi (2015), Gerber, Karlan and Bergan (2009).
2
More recent empirical papers exploit online networks such as Facebook to measure social
contagion of messages (Bond et al., 2012) on political mobilization and turn-out, but
without identifying the impact of friends’ political beliefs.
This paper addresses this issue by exploiting the near ideal natural experiment of
Sciences Po to provide causal evidence of the role of friends in political opinion and
behaviors. While most of the top political leaders meet each other and start their political
enrollment at Sciences Po, all students need to spend a week together in August each
year before starting to studying in September. Because this week’s aim is to facilitate
the integration of the students into Sciences Po, it is called the integration week. The key
feature of the integration week is that students are randomly allocated to groups based
on the alphabetic order of their last names.
We exploit this random assignment policy by performing a survey for the first-year stu-
dent who have entered Sciences Po in early September 2013, and ask incentive-compatible
questions to elicit their social networks (as, for example, in Leider et al., 2009, 2010).
Students were asked to provide a list of friends and were rewarded if friends provide cross-
validated answers to a few questions. We also surveyed students’ political opinions over
time, and behaviors in latest national and local elections, and students’ opinions on key
policy issues such as immigration policy. We then examine the diffusion and convergence
of political opinions within social networks by using the panel of all pairs of students and
by estimating dyadic regression of differences in beliefs on friendship.
Naturally the main traditional concern in such regression is reverse causality and cor-
relation in unobservables due to homophily: people have proclivities to link to others
similar to themselves, both on observed and unobserved characteristics that could influ-
ence their behavior (Manski, 1993; McPherson, Smith-Lovin and Cook, 2001; Currarini,
Jackson and Pin, 2009). As a result, the OLS estimate of the role of friendship in beliefs
formation might be upward biased since people tend to be similar. They tend to form
friendship relationships with individuals who originally share their political opinions. We
address this concern by using an instrument for friendship formation based on the initial
exogenous allocation of Sciences Po students into small groups during the integration
week.
The group membership during the integration week is thus by far the first place
within the institution where first year students interact together. By exploiting our 2014
survey, we find a very strong first-stage relationship between same integration group
membership and friendship: being in the same integration group increases the chance of
a pair of students to become friends by 16 percentage points. Common integration group
3
membership alone explains about 15% of the variation in pairwise undirected friendships
across dyads, whereas alternative control variables could explain only 1.6% more.
We next estimate the causal impact of friends on differences in opinions by using inte-
gration group membership as an instrument, interpretable as the local average treatment
effect (LATE) among compliers, i.e., among those who become friend only due to the
group assignment. A pair’s friendship reduces differences in political opinions 6 months
into the first college year by half a point on a scale from 1 to 10, equivalent to a quarter of
the mean difference, and a third of its standard deviation, among all pairs of students. It
is considerably larger than the OLS estimate, suggesting that complying pairs are more
susceptible to the friendship effect than others.
Besides, we show that convergence in political opinion happens only among friends,
while the effect would be insignificant if we were focusing on pairs of non-friends belonging
to the same group. In other words, the friendship effect completely dominates the effect
of belonging to the same study group throughout the freshman year. This finding is
important with respect to the literature on peer effects since what really matters in
social learning or imitation is not peers, but friends. Moreover, friends not only converge
in political beliefs, but also in behaviors : pairs of friends significantly converge in their
associative lives.
We further find that friendship contributes to narrow friends’ opinion gap mostly by
reducing the incidence of divergence (when two opinions move apart), rather than by
increasing convergence (when two opinions move together). Friendship also increases
the chance of co-movement of two opinions. It effectively helps reduce the instances of
extremist political views, without pushing for conformism among students.
The results are driven by pairs of students that started out with very similar political
opinions. They suggest that students who befriend in part because of political similarity
would continue to develop along this dimension, including joining the same associations
with political motivations, and end up remaining close to each other’s political views,
whereas other pairs who befriend based on other dimensions do not follow this mechanism,
and do not experience a friendship effect on differences in opinions..
Finally, we find that network strength and structure matter to the degree of friendship
effect, a fact that can be useful to all models of social learning with an average belief
updating process a la DeGroot (DeGroot, 1974; DeMarzo, Vayanos and Zwiebel, 2003;
Golub, and Jackson, 2010, 2012). Notably, the effect is strong among self-declared strong
friendships, but still remains among weak friendships, and up to social distance 2 (between
friends of friends). Also, network central students are a lot less likely to be influenceable
4
than others.
The rest of the paper unfolds as follows. Section 2 relates our paper to the literature on
peers, social networks and social learning. Section 3 describes the study’s context. Section
4 details our empirical strategy, the timing and design of our surveys, and discusses the
collected data. Section 5 provides all our empirical results, including the first stage
(section 5.1), the main results on political opinions (section 5.2), results on associative
behaviors (section 5.3), the mechanism at work (section 5.5), heterogeneous effects and
the role of network structure (section 5.5), and robustness checks (section 5.6). Finally,
section 6 concludes.
2 Related Literature
Although our paper provides a novel analysis of the role of social networks on political
belief formation, there has been an extensive literature on the impact of peers and social
networks on other beliefs and behaviors. We discuss below the main identification issues
raised in this literature and how we address them in this paper.
2.1 Peer Effects
There is an important literature on peer effects. This literature looks at the causal impact
of peers on different outcomes and mainly uses field experiments (see e.g. Angrist and
Lavy, 1999; Sacerdote, 2001; Zimmerman, 2003).4
For example, Boisjoly et al. (2006) show that the racial composition of freshman
housing assignments can have a long run impact on student attitudes. They show that,
if a student is randomly assigned to a black roommate, he/she is somewhat more likely
to support Affirmative Action in admissions and societal income redistribution.
Other interesting peer-effect studies include that of Sacerdote (2001) and Carrell et
al. (2009) who analyze specific contexts in which first-year roommates (or hallmates or
squadron mates in the case of military academies) are randomly assigned by the housing
office. This creates exogenous variation in one’s peer group, which is then used to ask how
much peers matter, which peers matter, and for what outcomes. They find strong peer
effects in education. A recent paper by Carrell et al. (2013) examines squadron mates
in the case of military academies. They manipulate the groups by putting together low-
ability and high-ability incoming cadets at the US Air Force Academy. They show that
4For an overview of this literature, see Sacerdote (2011, 2014).
5
performance for the lower-ability students fell relative to lower-ability students in the
randomly assigned control group.
Compared to this literature, we study network rather than peer effects and show that
the friendship effect (social network) completely dominates the effect of belonging to
the same study group (peer effects) throughout the freshman year. This highlights the
importance of analyzing elicited friendship data instead of using peer groups. We also
believe that this is one of the few papers that study peer effects over time. Exceptions
include Zax and Rees (2002), Bifulco et al. (2011), Patacchni et al. (2016), Lavy and Sand
(2016). However, these studies look at individual behavior over time and mainly focus on
the long-run effects of friendship on the individual human capital accumulation. In our
paper, we look at the “long-term” effect of friendship on the dyadic behavior of friends
in terms of political opinion, which allows us to study the convergence or divergence of
these opinions over time.
2.2 Empirical Aspects of Networks
A growing empirical literature has documented the effects of social networks on behavior
(see e.g. Advani and Malde, 2018; Bramoulle et al., 2016; De Paula, 2017; Jackson, 2008,
2011, 2014; Jackson et al., 2017; Jackson and Zenou, 2015; Blume et al., 2011; Ioannides,
2012; Topa and Zenou, 2015, for overviews). Since social networks are so prevalent
in economic settings, modeling these networks is essential in order to understand how
network structure affects behavior. However, it is very difficult to cleanly test theoretical
predictions using data, since there are many confounding features in the environment.
The estimation of network effects is indeed complicated by several issues: reflection
problem, common shocks and endogenous network formation. It is well-known that when
estimating peer effects using a linear-in-means model (i.e. when regressing individual
activity level on the average activity level in the neighborhood/among the peers), the
endogenous and contextual effects cannot be separately identified due to the reflection
problem, first formulated by Manski (1993). With explicit social network data, as we
have here, this problem is eluded (see, e.g. Bramoulle et al., 2009; Lee and Liu, 2010;
Calvo-Armengol et al., 2009; Lin, 2010; Liu et al., 2014). Indeed, the reflection problem
arises in linear-in-means models because individuals interact in groups - individuals are
affected by all individuals belonging to their group and by nobody outside the group. In
the case of social networks, instead, since the reference group is individual specific, this is
not true because peer groups are overlapping. Formally, social effects are identified (i.e.
6
no reflection problem) if at least two individuals in the same network have different links
(Bramoulle et al., 2009) is generally satisfied in any real-world network.
Another concern is the fact that students might be exposed to common shocks that
drive the convergence of beliefs, irrespective of their social interaction. One of the most
plausible candidates is a teacher fixed effects. Students who have been randomly assigned
to a group are also exposed to the same teachers. We run various placebo tests in this
paper to rule out the direct impact of the group on beliefs formation. In particular,
we show that the main driving force in the convergence process is the structure of the
network within the group, and in particular direct paths between friends.
The other (and often most) important issue is the potential endogeneity of the net-
works. Indeed, as stated above, people have proclivities to link to others similar to
themselves (homophily), both on observed characteristics and unobserved characteristics
that could influence their behavior. By failing to account for similarities in (unobserved)
characteristics, similar behaviors might be mistakenly attributed to interactive effects
when they are actually due to underlying characteristics or exposure to common stimuli.
As discussed by Goldsmith-Pinkham and Imbens (2013), Graham (2015), Jackson et al.
(2015), there is no simple solution to this problem. One can explicitly model the network
formation process and structurally estimate it (see, e.g. Mele, 2017; Goldsmith-Pinkham
and Imbens, 2013; Chandrasekhar and Jackson, 2014; Graham, 2017; Badev, 2018), or
use instrumental variables (Bifulco et al., 2011; Bramoulle et al., 2009; Calvo-Armengol
et al., 2009; Patacchini and Zenou, 2017).
Another way out is to use controlled experiments. In the field of networks, this has
been implemented by either (i) fully controlling the network of relationships in the labora-
tory (Choi et al., 2012; Kearns et al., 2009) or (ii) assigning subjects in the field positions
in a network through which they must communicate (Centola, 2010, 2011; Goeree et al.,
2010; Babcock and Hartman, 2010; Cai et al., 2015). In the latter papers, the network is
still endogenous and not randomized. What is randomized is the intervention. For exam-
ple, in Babcock and Hartman (2010), the network is self-reported and thus not random,
but subjects (students at UC Santa Barbara) have different fractions of their friends who
are being randomly treated (free access to exercise at the university gym).
In our paper, we use a natural experiment where agents are randomly allocated to
the network. We believe this is one of the first papers that uses such a strategy in the
context of networks. It gives a ”clean” identification strategy that allows testing the
effect of network position on educational outcomes.
7
2.3 Social Learning
There is an interesting literature on learning in social settings.5 From a theoretical
viewpoint, two basic approaches have been considered. One is a Bayesian approach,
where agents update their beliefs based on either communication or observation of other
agents’ actions over time. This approach provides a nice benchmark for what happens
with full rationality (see e.g. Bala and Goyal, 1998, 2001; Acemoglu et al., 2011). Another
approach is more mechanical, where agents repeatedly process the information from their
neighbors according to fixed rules (see e.g. DeGroot, 1974; DeMarzo et al., 2003; Golub,
and Jackson, 2010, 2012).
From an empirical viewpoint, the social learning literature has provided several major
insights. First, careful observational studies, natural experiments, and field experiments
have established the importance of social learning in many domains, such as in the intro-
duction of new technologies and innovations (e.g. Foster and Rosenzweig, 1995; Bandiera
and Rasul, 2006; Henkel and Maurer, 2010; Conley and Udry; 2010) and labor markets
(Granovetter, 1974; Munshi, 2003; Topa, 2001; Bayer et al., 2008; Damm, 2009; Beaman,
2012). Second, some studies carefully measure social networks and estimate the relative
effect of geographic neighbors, direct friends, and second-order friends. The evidence on
this question is mixed, with some papers finding an almost equal influence of second-order
neighbors (Kremer and Miguel, 2007) and others finding no effect (Rao et al., 2007)6 or
significant decay (Patacchini and Zenou, 2012; Mobius et al., 2015).
However, most of these studies focus on outcomes, with the adoption of new tech-
nologies or product, or employment outcomes. But they do not distinguish outcomes
from beliefs, whether the process goes through the transmission of information or just
imitation. This is what we do in our paper.
3 Sciences Po background and organization
This section provides a description of the context of the natural experiment at Sciences
Po, including its role in French politics and its organization of the curriculum.
Since its foundation, Sciences Po has always been strongly involved in the training of
politicians and high level civil servants. The university (initially called “Ecole libre des
Sciences Politiques”) was founded in 1872, after the defeat of France against Prussia,
5For overviews, see Goyal (2011), Mobius and Rosenblat (2016) and Golub and Sadler (2016).6They find that there is important positive social learning from direct friends but not second-order
friends about the benefits of flu vaccination in a study of under-graduate students at a private university.
8
and its explicit aim was to provide a modern training to the French elite, akin to the one
received in Germany. Although being a fully private institution until 1945, it soon gained
a de facto monopoly for the preparation of examinations enabling to enter the highest
level of the French administration. In particular, 80% to 90% of the students entering
the Ecole Nationale d’Administration (the most prestigious public exam in France) each
year have been trained originally at Sciences Po (Rouban, 2014a). The university has
also strong links with French politicians. Between 12 to 15% of French MPs elected in
the last decades graduated from Sciences Po (Rouban, 2011), as well as more than fifteen
percent of the mayors of cities above 30,000 inhabitants (Rouban, 2014b). Sciences Po
students are also extremely present in the executive branch, with many ministers coming
from Sciences Po.
While not all Sciences Po students want to become politician or civil servant, politics
is much more important for them than for students from other universities or Grandes
ecoles. One tenth of the students are member of a political party, a very large proportion
compared to their age group. About one quarter of the students choose a master in
public affair, which is the traditional path of access to French administration. Therefore,
Sciences Po plays the central role in the training of the French political and administrative
elite. As most Grandes ecoles students, Sciences Po students are very different from the
students enrolled in public universities. On average they are academically stronger, and
come from a much wealthier background.
To change its reputation of an elite school nurturing networking and homophily, Sci-
ences Po has implemented various innovative pedagogical reforms in the recent years,
such as a special admission procedure to examine and recruit students from poor suburbs
for around 20% of the total number of admissions, and a widening of recruitment outside
Paris from different French regions and abroad. One of the most important innovations
implemented in 2007 by the former dean, Richard Descoings, was to force first-year stu-
dents from different socio-economic background to mix with each other. To achieve this
goal, students are exogenously allocated to different classes at school entry.
The first year of Sciences Po bachelor is made of a large curriculum core, with six
introductory courses in microeconomics, macroeconomics, history, sociology, political sci-
ence, and constitutional law. Each of these courses consists in one weekly lecture, and
compulsory weekly tutorials. The tutorials last two hours, gather about twenty students
each, and enable to review the materials seen in the lecture and to do exercises (such
as oral presentation, for instance). It involves a lot of work in small groups in order to
do collective assignments. Since the number of students enrolled in the lectures is large
9
(about 800), the tutorials are where most social interactions across students take place.
In order to foster interactions, the administration chose few years ago to link tutorials
together into packs. Therefore, students are assigned the same classmates across all the
subjects. This was implemented because the administration felt that many students were
isolated because of the large size of the program. It further reinforced the role of tutorial
as creating social links.
Other sources of social interactions are associations (close to one hundred, including
political parties and student unions), sports, and the monthly parties organized by the
student office. However, the interactions in the tutorials are much stronger, with at least
three meetings a week, and many opportunities of working together through group work.
As we show in the next section, tutorials are responsible for the majority of friendship
link formation.
4 Empirical design, methodology, and measurement
4.1 Empirical strategy
Our empirical design focuses on the dyadic relationship among all pairs of students
(i, j), between a measure of pairwise difference DYij and a measure of friendship Linkij,
DYij = g(Linkij,Xij, ηij). Undirected friendship Linkij is defined as an indicator of
whether i or j names the other as friend in their answers. We focus on the undirected
network of friendship, thus use all symmetric dyadic variables.7 DYij is the absolute
difference between i and j of the variable Y , such as political opinion on a scale from 1 to
10. Xij and ηij represent respectively observable dyadic covariates and the unobservable
idiosyncratic residual. Xij includes all observable pairwise variables representing com-
monness and differences across predetermined dimensions: the pre-Sciences Po difference
in political opinions DY 0ij (surveyed from a retrospective question), common gender, com-
mon nationality, common academic program, common admission type (essentially regular
admission or Affirmative Action admission), common graduation with honor from high
school, common district (French departement) of high school, common professions of
parents, common current residence’s ZIP code, dummies for being both female, for being
both French, for being both French with double nationality, and the difference in tuition
7We use the OR network, similar to Leider et al. 2009 and many other papers that survey friendships.The results remain robust to using the AND network.
10
fees that proxies for the difference in parents’ income.8 As will be shown, this list of
observables has very limited explanatory power on friendship.
We are generally interested in the average causal effect of friendship on pairwise
differences of opinion, as βL = E[DYij|Linkij = 1,Xij]−E[DYij|Linkij = 0,Xij] (i.e., the
“partial effect” of Linkij on f(·)). A negative (positive) βL means that friendship makes
people’s opinions closer (further apart) over the observed period of 6 months from August
2013 to March 2014, and vice versa. It is identified under the conditional independence
assumption that conditional on observables, friendship is assigned exogenously, in which
case a simple OLS regression would suffice to estimate β. However, this conditional
independence assumption is likely violated because of homophily based on unobserved
characteristics.
The homophily bias occurs when there is an unobserved certain dimension U along
which individuals’ similarity Uij correlates with the formation of friendship links Linkij,
and influences the outcome DYij = g(Linkij,Xij, Uij, ηij). Denote friendship formation
as Linkij = f(Xij, Uij, εij), then the homophily bias due to the omission of Uij is ∂g∂U×
Corr(Linkij, Uij). It is larger when Uij is more important to the outcome, and when it
is more predictive of link formation. In our context, it likely pushes the OLS estimate
away from zero. While we can measure the importance of the homophily bias due to the
potential omission of the observables Xij, it is difficult to gauge its magnitude, especially
when observables only explain a small share of the variation in Linkij. In effect, this
remains a thorny issue in all estimations of effects of links on networks, and one that, to
our best knowledge, has not been addressed in the empirical literature using an exogenous
source of variation.
In addition to the predetermined covariates, we also include TGij, a dyadic indicator
whether two students are members of the same tutorial group or not. Since membership
in a specific tutorial group is essentially arbitrary, a linear regression can estimate the
causal average effect of being in the same tutorial group on the outcome, namely βTG =
E[DYij|TGij = 1, Linkij,Xij] − E[DYij|TGij = 0, Linkij,Xij] (the partial effect of TGij
on f(·)). It is then possible to compare the different effects of tutorial group membership
and friendship on the difference in opinions.
8We create those dyadic control variables based on the list of predetermined variables that we couldcollect from administrative data, and on our assessment of a priori importance of those conditions in theformation of friendship among Sciences Po students. Unfortunately, there is no better information onthe precise household income, as the administrative data have a very high rate of missing observationsfor this question. The inclusion or exclusion of any dyadic control variable does not make any noticeablequalitative or quantitative difference to our results.
11
Our key methodological innovation consists of treating the homophily bias by in-
strumenting Linkij by the indicator IGij of whether i and j participated in the same
integration group (IG) at the beginning of the year (August 2013). We argue that this
instrument satisfies all the LATE conditions (Imbens and Angrist, 1994).
First, in section 5.1, we will test the instrument’s relevance, namely βIG = E[Linkij|IGij =
1,Xij] − E[Linkij|IGij = 0,Xij] 6= 0 (with Linkij = f(IGij,Xij, εij)). This first stage
condition is satisfied if the integration week is a strong enough catalyst to form lasting
friendships among students.
Second, this instrument’s exogeneity is based on the mechanism of assignment into
integration groups by alphabetical order of the family name, arguably independent from
individual characteristics that matter to the formation of links. We will further test the
claim of exogeneity in a balance test in section 4.5.
Third, the instrument IGij arguably satisfies the exclusion restriction. The integration
week was exclusively meant to facilitate students’ familiarization and socialization with
their new peers and new environment in Paris, without any academically- or politically-
related activities. The integration groups are dissolved after that week, and does not
relate to any other academic or extra-curricular activities afterwards. (No subsequent
academic or extra-curricular activities among Sciences Po students are organized based
on alphabetical order.) Thus, it should have no meaningful channel to affect the formation
and adjustment of individual opinions six months later, which guarantees the exclusion
restriction of the instrument.
Fourth, it is natural to make the monotonicity assumption that being in the same IG
always (weakly) increases the incidence of friendship formation for any pair of potential
friends, such that f(IGij = 1,Xij, εij) ≥ f(IGij = 0,Xij, εij) ∀(i, j). 9
Taken together, those four assumptions guarantee a causal LATE interpretation of
our estimate (Imbens and Angrist, 1994). That is, our IV estimate can be interpreted as
the average causal effect of friendship on the “compliers” pairs of students, namely those
who would have become friends thanks to being in the same IG group in the integration
week. Since this is a condition that characterizes a rather strongly-complying group
of student pairs (for instance, pairs that only become friends after weeks or months of
acquaintance are not included), we remain cautious in generalizing our estimates to all
9Its violation would mean the rather improbable event of a “defier” pair that would have becomefriends had they not met in the same IG, but would not have become friends because they met earlyin the same IG. Even without the monotonicity assumption, de Chaisemartin (2017) shows that, undera much weaker condition, one could still interpret the IV estimator as the Average Treatment Effectamong a subgroup of compliers.
12
possible pairs of Sciences Po students. However, in Imbens’ (2010) spirit of “better LATE
than nothing”, we argue that the correct estimation of the LATE in our context already
lays strong ground for further research on transmission of beliefs among students.
4.1.1 Clustered standard errors
In a dyadic setting, each individual is repeated in her pairs with all other students,
resulting in natural correlations between the residual terms of pairs sharing an individual.
Furthermore, there can naturally occur common shocks within the same group, such as
teacher’s biases, that could drive all group members’ opinions. While those shocks are
uncorrelated to our instrument, and cannot bias our IV estimates, they produce clustered
standard errors, and must be taken care of in order to obtain correct standard errors and
confidence intervals.
Throughout the paper, we choose to correct for potential clustered standard errors by
a two-way group clustering strategy. That is, we allow for arbitrary correlations in the
idiosyncratic component ηij between any pair of observations that overlap in a group.10
We make sure results are robust to different types of clustering correction.
4.2 Timing of events
We conduct our first and major survey in March 2014, and a follow up survey in July
2015. The first survey in March 2014 gave strong participation incentives, with 25 mini
iPads to win, and thus recorded 526 out of the 800 first year cohort who completed the
survey. The second survey in July 2015 was much less well-funded, and could only attract
300 participants. Overall, there are 235 students who have completed in both surveys.
The paper makes most use of the first survey, while the second only serves in robustness
checks.
10Cameron and Miller (2014) discusses Fafchamps and Gubert’s (2007) method to fully account for allpossible correlations between all dyads that overlap with a group or share an individual. Unfortunately inthis case, Cameron, Gelbach, and Miller’s (2011) decomposition of the the sandwich formula for standarderrors (used for a fast calculation of the two-way clustering correction) becomes intractable. The onlypossible implementation is to undertake the full calculation of Fafchamps and Gubert’s formula, whichrequires an excessive amount of computing memory and time, given our relatively large sample size.Therefore, throughout our paper, we choose to implement a simplified version of this method, in whichwe allow for non-zero correlations between any residual terms ηij and ηi′,j′ such that either i and i′
belong to the same group, or j and j′ belong to the same group, or both (thus ignoring the possiblesame-group memberships of i and j′, or of i′ and j. We have also fully implemented Fafchamps andGubert’s (2007) formula in a few benchmark regressions, and found similar and better levels of standarderrors and p-value.
13
The first survey in March 2014 collects information to construct the network of friends
among first-year students. They are also asked about their contemporaneous political
opinions in March 2014, as well as their recalled political opinion from August 2013
before they join Sciences Po. Those questions are asked again in the second survey in
July 2015, about both contemporaneous opinions and recalled opinions from March 2014
(thus it is possible to compare actual versus recalled opinions). The timing of our surveys
is summarized below:
(1) August 2013. The activities are concentrated in the integration week, a week that
precedes the official starting date at Sciences Po, meant to help newly-admitted students
socialize with each other and familiarize with the environment in Paris. Students are
divided into groups of around 20 by alphabetical order based on their last name. Most
students participate in this integration week.
(2) September 2013. First-year students start the scholar year with the same set of
mandatory core courses in all of Sciences Po’s five disciplines. Every week, each course
is composed of a full-cohort lecture, followed by an additional section in a small tutorial
group of around 20. The process of tutorial group allocation is rather arbitrary and
exogenous.
(3) March 2014: We survey all the 800 first-year students on their social networks and
their current political opinions as well as their political opinions in August 2013 before
entering Sciences Po. 526 out of the 800 first-year students completed the survey.
(4) July 2015: We survey the now-second-year students. In particular, we ask their cur-
rent political opinions and their political opinions in March 2014. 300 students completed
in the survey.
4.3 Survey design
Most of our analysis is based on first-year students who enter Sciences Po Paris starting in
September 2013. In March 2014, we run an internet-based survey among all 800 first-year
students, and ask incentive-compatible questions to elicit social networks (as in Leider et
al., 2009 and 2010). We also survey their cultural values and political opinions.
In empirical studies on networks, a large rate of non-response can produce unpre-
dictable biases in the estimates (Chandrasekhar and Lewis 2011). We thus offer strong
material incentives in the first survey in the form of a lottery for twenty-five mini iPads at
14
a monetary value of approximately 300 Euros each, so each student has an average prob-
ability of about 6% to win. Eventually, 68.4% (547 out of 800) of the students answered
to at least some question in the survey, and 65.6% (526 out of 800) completed the whole
survey. This is about the same level of participation as the best-participated studies of
social networks of students, such as Leider et al. (2009) or Goree et al. (2011). It is
well above the standard participation rate of around 20% found in studies using online
surveys (Cantoni et al. 2017).
In order to incentivize truthful answers, we design the elicitation of friendships as
a coordination game, similarly to Leider et al. 2009. Not only do we ask students to
name a list of friends of up to 10 names, but we also ask how they meet each of them,
how much time they spend together, and in which activities, and how strong do they
evaluate their relationships. We announce in the survey that their answers would be
cross-checked with those of the other students, and that if both answers match, they
would gain points , later converted into an additional probability of winning the gift. We
do not disclose the exact mechanism, in order to avoid that some students engaged into
strategic behavior and try to actively coordinate with other people. The survey is carried
out during a vacation week, which limit the possibility for the students to interact with
each other and to complete the survey together. We also require that they complete the
whole questionnaire in order to be included in the lottery. We present in Appendix the
details of the questions and procedures.
The second part of the survey was devoted to questions about political opinion and
values. A stated above, in the first wave, we asked them their political opinion on March
2014, and the summer before admission (August 2013) while, in the second wave, we
again asked them their political opinion on July 2015, and more than one year before
(March 2014). These questions were asked by using a Likert scale from 1 to 10 (1 being
extreme left and 10 extreme right). We also ask if they were a member of political party
(today, or in the past) and questions on related political opinions such as attitudes toward
immigrants, by using question taken from the World Value Survey.
4.4 Data description
As mentioned above, we consider the (symmetric) OR network in which two students
are linked if at least one nominates the other. The reciprocal rate among answers is
about 50%. The probabilities of a well-matched answer in terms of the context of the
first meeting between the two friends, of the amount of time spent every week, of the
15
type of activities mostly spent together, and of the self-evaluated strength of friendship
are respectively 76%, 52%, 46%, and 52%. If answers are completely made up and
randomized, the probability of matching on any of those dimensions would be rather low,
given that respondents have many choices for each answer, especially in the question on
the context of their first meeting. We therefore interpret those numbers as the survey
answers are indeed very reliable, especially for the purpose of picking up friendships.
Table 1 Panel A reports the main statistics on the number of friends and the social
network structure. The average and maximum number of nominated friends per student
is 8.8 and 21, respectively, with a very high variance. Moreover, there seems to be some
small world properties with a very small average path length (3.7) and a relatively small
diameter (9). The clustering coefficient is also relatively high, which means that roughly
25 percent of students have friends of friends who are friends. In terms of network
position, the mean eigenvector centrality is relatively low (0.0361).
Panel B shows the descriptive statistics of the friendship dyadic measures. We dis-
tinguish between the full sample (column (1)) of all students who have participated and
the benchmark sample (column (2)) that corresponds to the benchmark regression (the
two samples differ slightly because of certain missing values). By nature, the share of
measured friendship links is relatively small at 1.6%, and that of second and third order
indirect links are larger at 9.3% and 38%, respectively. The dyadic same group variables
are of similar magnitudes, at an average of 1.6% for same integration group, and 2.3%
for same tutorial groups. The friendships are partitioned rather evenly across different
levels of friendship strength, especially from 2 (ordinary friends) to 4 (very close friends).
We also observe that there is little difference between the full sample and the benchmark
sample.
Panel C shows the descriptive statistics about the political opinion and values of
Sciences Po students. We also look at political opinions in the first and the second
surveys (current and recalled). The average political opinion of Sciences Po students is
left of, and close to the center, and does not change much in the different surveys (whether
it is recalled or current). Observe, however, that, even though the mean of the political
opinion hardly changes before and after the entry in Sciences Po, there is a reduction of
the variance of about 15 percent. And the entrance in Sciences Po is associated with a
large increase in the enrollment in a political party.
[Insert Table 1 here]
16
Figure 1 shows the distribution of declared political opinion as of March 2014 (orange)
and of (recalled) political opinion in August 2013 (green). Political opinion as of August
2013 clearly sees a mode at center-left positions (3 and 4), which disappears in March
2014. In general, over the seven months covered by the survey, the distribution of political
opinion clearly shifts to the right and, with a lot of students moving from center-left to
center opinions (5), an increase in center-right positions (6 and 7) and a decrease in
rightist opinion.
[Insert Figure 1 here]
Panel E displays other individual descriptive statistics of independent variables.
Panels D and F present dyadic relationships. On average, within a couple of friends,
the students reduce their difference in political opinion from 2.211 (August 2013) to 1.932
(March 2014). This is also true in 2015 and even in 2014 as recalled in 2015.
Panel G presents some results on the declared place and time when couples of friends
first met each other, differentiating between (1) meeting during the integration week, (2)
meeting during the integration week and being in the same integration group and (3)
meeting before Sciences Po and being in the same integration group. Column (1) shows
that around 24 percent of the couples of friends in our baseline sample met during the
summer integration week. This gives us confidence on the relevance of our instrumental
variable. Column (2) shows that around 94 percent of the couples of friends that were
in the same integration group (August 2013) declared to have first met each other in
that occasion, supporting the idea that summer integration groups indeed creates an
opportunity to meet new friends. Finally, column (3) shows that only 0.2 percent of the
couples of individuals that attended the same integration group (i.e. 2 couples) knew
each other before coming to Sciences Po. It is thus not a concern that students have
known each other before Sciences Po.
4.5 Exogenous assignment mechanisms and balance test
Our instrumental variable strategy depends fundamentally on the claim that the assign-
ment into integration groups is exogenous. While the basis of the assignment is simply
an alphabetical order by students’ family name, we need to make sure that alphabet-
ically close family names do not carry other information that could stack up students
with similar backgrounds in the same groups. First, we show in Table 2 the results of
a balance test of exogeneity in a linear regression of IGij on all observable pairwise co-
17
variates.11 In comparison with the mean and standard deviation of the dyadic variable
Linkij (respectively 0.017 and 0.13, as shown in Table 1.B), all coefficients in Table 2 are
very small and mostly insignificant. Among the few significant coefficients, it is natural
to find the variable common program, which specifies mostly dual-degree programs joint
with other universities, in which students simultaneously take courses at both Sciences
Po and the other institution, therefore are grouped together since the integration week.
We make sure to control for this covariate throughout the paper. (Removing all stu-
dents from double-degree programs leaves all estimates practically unaltered.) Finally,
another significant coefficient is on same high school. However, the sign is opposite to
that predicted by homophily, and we take it as significant by chance.
[Insert Table 2 here]
Second, we run robustness checks by dropping all last names starting with each letter,
or dropping all students with a specific non-French nationality, given the potential concern
that the family names of certain ethnic origin may cluster by certain letter (such as
Chinese family names starting with Z or W, or Vietnamese names starting with N.) We
also run robustness checks in which we exclude all family names starting with “de”, which
might correspond to an aristocratic family name. The results shown in the Appendix
confirm that our results are strongly robust to all those concerns.
In our empirical design, we also use the dyadic variable whether two students are
members of the same tutorial group, in order to compare the friendship effect with the
effect of exposure in a common group through 6 months. It is thus worth arguing that
the assignment into tutorial groups is also quite exogenous.
Before the academic year starts in September 2013, students have a short time window
to enroll in different tutorial groups. Each group is fixed for all compulsory courses they
take during the first year and a half (three courses each semester). Several features of the
enrollment process make extremely unlikely that students choose tutorial on purpose, for
instance for schedule reason, for being assigned a given instructor, or to be with their
friends. The enrollment slots come in batches every five minutes (to avoid over-burdening
the system), and in practice each student only has enough time to pick a single tutorial
11Standard errors are clustered two-way among observations sharing either student i or student j. Thisclustering approach is less conservative than the standard two-way clustering by i’s and j’s groups carriedout throughout the paper, so we expect to have higher power in detecting observables that significantlycorrelate with IG. See Bruhn and McKenzie (2009) for a discussion on common practices of balancetests in field experiments.
18
group based on a single discipline at a specific schedule, without information on the
teacher apart from the name, and without information on the schedules of the other
tutorials in other disciplines that are associated with this tutorial group. The possibility
of cherrypicking teachers or coordinating to get in the same group is therefore very limited.
5 Empirical results
5.1 First stage results on friendship formation
We first establish the relevance of the instrumental variable IGij, i.e., the causal effect of
participating in the same integration group in August 2013 on forming and maintaining a
lasting friendship 6 months later. Table 3 presents the regression of Linkij on IGij, with
and without observable covariates Xij, yielding an estimate of βIG = E[Linkij|IGij =
1,Xij]−E[Linkij|IGij = 0,Xij] in the coefficient of IGij of around 16%. The F statistic
in all columns clearly shows that the instrument is strong.
[Insert Table 3 here]
It is remarkable that this coefficient is far stronger than any other coefficient on any
observable predetermined characteristics. It shows that “chance exposure” to other stu-
dents during the first week of a student’s college life has an effect on friendship formation
several orders of magnitude larger than that of most predetermined characteristics that
are typically observed or obtained from administrative records.
The result can be further interpreted as evidence of the first week’s role as a “window of
opportunity” for friendship formation. It is rather natural to think that friendships tend
to form at the beginning of college, during a week meant to facilitate socialization and
familiarization with a completely new environment, in groups composed of same-cohort
students. What is perhaps more striking is that this window of opportunity results in
lasting friendships. Its effect on friendship is about half as large as the effect of common
membership of tutorial groups, which has already lasted for 6 months up to the survey.
Table 3 also shows that homophily plays a statistically significant role in friendship
formation, notably along the dimensions of political opinions, gender, background and
origin (such as departement and region of high school, or admission category), interest
(the type of high school major), and family income (proxied by the level of tuition).
However, its role is rather limited, seen in the very small size of all coefficients on prede-
termined characteristics, and the very small increase in R-squared of 0.004 from column
19
2 to column 3 when all those covariates are introduced. In contrast, the two variables of
same group memberships, IGij and TGij, could explain almost 20% of the variation in
friendship formation. The inclusion of predetermined covariates does not alter much the
coefficients of IGij and TGij.
Overall, the results in Table 3 clearly confirms the relevance of the instrumental
variable IGij. They also highlight the importance of “chance encounters” among similar
individuals during a window of opportunity in formation of friendship links, and a much
weaker role of homophily based on predetermined observable characteristics. This finding
can be useful in calibrating network-formation models such as Jackson and Rogers (2007).
5.2 Impact of friendship on opinion differences
Table 4 present the OLS and IV results on the effect of friendship on differences in
political opinions, using the empirical strategy described in Section 4.1. Panel A first
shows a set of linear OLS regression results with different sets of controls. The partial
correlation between friendship and difference in political opinion ranges from 0.10 to 0.13
(in absolute value) across specifications, roughly 6% of the mean difference 1.932, and 8%
of the standard deviation 1.467 (see Table 1 Panel D), and statistically significant at 5%:
On average and after controlling for covariates, pairs of friends have lower differences in
political opinion. The inclusion of predetermined dyadic control variables in column (2)
reduces this coefficient from 0.13 to 0.10, suggesting that there is some homophily bias
due to those covariates. Tutorial group fixed effects (one for each individual in a dyad) do
not alter much the coefficient on friendship (columns (4) and (6)), suggesting that there
is no sorting by group, as expected based on the tutorial group formation. Furthermore,
the coefficient does not seem to be significantly different between friendships within and
across tutorial groups (as seen on the coefficient of the interaction term in columns (5)
and (6)).
Finally, column (7) shows that the tutorial group variable alone has a rather weak,
negative, and statistically significant coefficient on differences in opinion. This is the
typical peer effect in the literature that averages among pairs of friends and non-friends:
the results may be driven entirely by strong effects among pairs of friends (as students
in the same tutorial group are more likely to become friends.) The study of such average
effect thus hides important heterogeneity among different types of relationship, as also
argued by Carrell et al. (2013).
20
[Insert Table 4 Panel A here]
Table 4 Panel B presents the IV results and thus the causal LATE of friendship on
differences in political opinion, while controlling for initial differences in political opinion
and other predetermined characteristics. Columns (1) to (6) carry similar sets of controls
as their counterparts in Panel A. The estimates of the LATE vary from 0.49 to 0.57 (in
absolute value), all statistically significant at 5%. They imply that the causal effect of
friendship on opinion differences among compliers, over the course of the first 6 months
at Sciences Po, is about half a point on a 1-to-10 scale, about a quarter of the mean
difference, and a third of the standard deviation of the differences. Friendship seems to
have a considerably larger effect for those who are also members of the same tutorial
group, with the additional effect ranging from 0.24 in column (5) to 0.47 in column (6),
although this additional effect is not statistically significant.
It is remarkable that after taking into account the effect of friendship, being in the
same tutorial group no longer helps reduce differences in opinion. This observation,
combined with column (7) in Panel A, again stresses the fact that the usual common-
group peer effect is merely an average among the effects of pairs of friends and non-friends,
and shadows the truly important effect among pairs of friends alone. It strengthens
the importance of measuring friendship and separating the effect of friendship from the
common-group peer effect.
[Insert Table 4 Panel B here]
The IV estimates are much larger in absolute value than the OLS estimates. This
difference is not only due to a correction of the homophily bias (which alone would have
predicted a smaller IV estimate), but also the LATE interpretation. In our context,
compliers are the ones that befriend very easily if exposed to each other during the
integration week. We may also expect that they have strong influence on each other,
which could result in a much stronger effect than the average effect among all pairs.
When this phenomenon dominates the homophily bias, the IV coefficient will be larger
than the OLS coefficient. To study this mechanism, we will detail the heterogeneity of
the impact of friendship in the next section.
21
5.3 Friendship effect on political and associative activities
It is important to link students’ beliefs to their behaviors. In this context, the most
natural behavior following changes in political opinions is participation in political orga-
nizations, including students’ associations with certain political inclinations, and polit-
ical parties. Table 5 shows results on the effect of friendship on the difference between
students’ decisions to enroll in those organizations. Unfortunately, we do not observe
the intensity of participation, and could only consider the extensive margin. Another
shortcoming is that formal political party enrollment is still very rare among first-year
students.12
[Insert Table 5 here]
In Panel A, while we do not see a significant effect of friendship in the probability of
joining any association (column (1)), students do follow their friends in joining exactly
the same association. The effect is about −0.18 among students who do join some orga-
nization (column (2)), and −0.08 in the full sample (column (3)). On the other hand,
columns (4) and (5) show no significant effect on enrollment in political parties. 13
Table 5 Panel B focuses on different subsets of students associations, to understand
friends’ motivation in joining the same association. The effect of friendship is concen-
trated on associations that have a political inclination and those related to one’s identity.
Column (2) groups associations that are strongly motivated by political parties and po-
litical activities14 Column (3) includes all associations with some political motivation. In
column (4), we consider all associations pertaining to certain aspect of a student’s iden-
tity, such as nationality, region of origin, religion, and ethnicity. The estimate is strongly
positive, but not significant, which could reflect the large variation in the activities and
the appeal of those associations. Other types of associations such as related to sports
(column (1)) or policy issues (column (5)) do not exhibit a friendship effect.
In columns (6) and (7), we attempt to understand whether the motivation of joining
the same political association has to do with spending time together among friends, or
inspiring each other into a common area of interest (such as raising interest in political
12Also, they are too young to have voted in political elections.13The outcome variable in this Panel is the dyadic difference, so a negative estimate means friendship
makes friends closer in their behaviors.14They include the famous student union UNEF, with a rather radical agenda even though it is officially
declared as apolitical. It organizes most of the students’ loudest manifestations on campus. While astudent at Sciences Po, the former French president Francois Hollande was also president of UNEF.
22
debates). The almost zero coefficients in those columns show that students are no more
interest in joining any political association other than the one that their friends are in.
Friendship thus has a very targeted effect on participation in associations.
Taken together, these results are important in showing an effect of friendship on
actual behaviors, beyond the one on self-reported political opinion. Interestingly, while
friendship leads friends to further exposure within an association, in this context it only
does so for politically-motivated associations.
5.4 Mechanisms
In this section, we investigate the mechanism driving our main result by considering the
different types of dyadic changes in opinions between August 2013 and March 2014. First,
we classify the changes in a pair’s opinions into three major categories: Opinions that
move towards each other (convergence), opinions that move away from each other (diver-
gence), and opinions that move in the same direction (co-movement). Those categories
include “weak” cases, in which only one of the two opinions changes. Thus they are not
mutually exclusive: the case of one fixed opinion and the other moving away would be
branded as both divergence and co-movement. To further separate those cases, we also
focus on strong convergence and divergence categories, in which both opinions change.
Table 6 shows estimates of the effect of friendship on the incidence of each category, using
the IV strategy set out in section 4.1.
[Insert Table 6 here]
The estimates of the effect of friendship on the incidence of both fixed opinions,
weak convergence, single convergence (specifically the case when only one opinion moves
towards the other), and strong convergence are not statistically significant, and of rather
modest magnitude (shown in columns (1) to (4)). On the other hand, we find strong
evidence that friendship reduces the incidence of divergence, and particularly of strong
divergence when both friends move apart (columns (5) to (7)). Friendship reduces the
chance of divergence by 12 to 17 percentage points, a large magnitude compared with
the probability of divergence. We also find evidence of friendship inducing students to
move in the same direction (column (8)), and especially doing so while reducing their
differences (column (9)).15
15Table 6’s approach focuses on the extensive margin of each category of movement, thus alleviatesthe concern of recall bias due to the retrospective nature of students’ answers on political opinion before
23
The lack of result on convergence shows that friendship does not necessarily bring
conformism or create “echo chambers”, in terms of compressing the diversity of opinions.
Yet the strong effect of discouragement on divergence suggests that friends do keep each
other from deviating away. At least, if they do move their opinions, they would more
likely move in the same direction.
One natural corollary of this observation is that friendship reduces the incidence of
students with extremist views. Table 7 tests this prediction by applying our IV strategy
to study the effect of friendship on the existence of at least one student with an extreme
political opinion, defined as an answer either in {1, 2} (extreme left) or {9, 10} (extreme
right).16 In general we find a very strong negative estimate of the friendship effect on the
incidence of extremism after 6 months at Sciences Po, among the pairs that started out
as two moderates (column (1)), a moderate and an extremist (column (3)), or with at
least one extremist (column (4)). Yet, perhaps due to the relatively small proportion of
extremists makes, those estimates are rather imprecise, and not statistically significant at
conventional levels. Further investigation into pairs that likely have stronger friendship
effects, namely those whose opinions do not differ by more than 1 point (as will become
clear in the next Table), reveals a friendship effect that is statistically significant at 10%
(column (2)). This evidence hints that friendship may reduce extremism.
[Insert Table 7 here]
The difference between friendship effects on divergence and convergence suggests that
the main driver of the effect may be pairs that have the most room for divergence and
the least room for convergence, i.e., those whose initial opinions are very close. Follow-
ing this direction, Table 8 investigates the friendship effects on samples conditional on
initial differences in opinions. Overall, the effect is stronger for pairs with smaller initial
differences, as shown in columns (1) to (6), and by the interaction term between initial
differences and friendship shown in column (1). The lack of statistical significance of this
coefficient in column (1) can also reflect the nonlinear nature of the relationship, as well
as the large amount of noise among pairs with large initial differences.
joining Sciences Po. Indeed, even if those answers can be biased towards students’ opinions, this biasdoes not affect results on the weak categories, such as friendship’s reduction effect on divergence. Wewill return to this issue in section 5.6.
16As usually found in French universities, there exist few extreme right views. The number of extremeleft views is much larger.
24
[Insert Table 8 here]
Taken together, the findings from Tables 6-8 are consistent with the following mech-
anism. Students who form new friendships because of their encounter in an integration
group may do so according to one or a few areas in which they share common interest.
One important area can be politics: certain pairs of students form friendships because of
their similarity in political opinions (i.e., homophily by political opinion). Others, how-
ever, may form friendships along other dimensions, and can become friends despite large
differences in political opinion. Over time, interactions between pairs are shaped along
the lines of common interests, so friends with similar political opinions tend to often dis-
cuss politics, and even join the same politically-inclined associations, while friends with
different political opinions may strengthen their relationship through other dimensions
on which they are more similar. In consequence, friends who started out with similar
political opinions continue to influence each other’s political view, especially in keeping
each other from diverging, while friends who started out with large differences in opinions
do not exert much influence on each other.
This mechanism echoes Golub and Jackson’s (2012) analysis on homophily and the
speed of convergence in beliefs, but with the introduction of an endogenous selection of the
dimension of interaction based on homophilous preferences. Since our newly discovered
empirical facts imply a rather nonlinear mechanism of diffusion of beliefs, notably in the
asymmetry between converging and diverging, it would be interesting to reconsider their
results in light of those facts.
5.5 Heterogenous effects and social network structure
The theoretical and empirical literatures on non-Bayesian learning in networks typically
assume homogenous, linear effects of direct links on a node’s beliefs. Examples include
theories using average-based belief updating processes (the term coined by Golub and
Jackson (2012) for a generalized definition of DeGroot’s (1974) belief updating), or em-
pirical estimations of peer effect in networks (such as Calvo-Armengol et al., 2009, or
Bramoulle et al., 2009). It is natural to worry about the empirical validity of such an
assumption, as theoretical results on networks can depend critically on linearity and
homogeneity.
In section , we have shown the highly asymmetric nature of friendship effects between
convergence and divergence, suggesting that the linearity assumption in average-based
25
belief updating is rather unrealistic. In what follows we will further explore the hetero-
geneity of belief updating with respect to different types of relationships, and network
characteristics of friends.
Table 9 reports the results of our baseline IV specification in Table 4 for different
levels of intensity of friendship, measured on a scale of 1 to 4 based on students’ answers.
The effect of friendship is clearly increasing in the intensity of the friendship link going
from a 3.247 points reduction in political difference in the case of the sample restricted
to “very close friendships,” to a 0.575 reduction when considering any type of friendship
link. This suggests that not only does friendship make students’ political opinions more
similar, but stronger friendships also keep students closer. This effect could be inherently
due to the strength of friendship, or to the possibility that friends with similar political
opinions tend to be close friends, and based on Table 8 they tend to experience a stronger
friendship effect.
[Insert Table 9 here]
In Table 10, we further differentiate pairwise relationships in categories based on
social distance, namely the shortest-path distance between any pair of students in their
social network. Column (1) first replicates the benchmark result from Table 4B’s column
(2) highlights a special case of the causal comparison between direct friends (i.e., of
social distance 1) and others (social distance greater than 1). In column (2), we consider
the effect of a pair’s social distance, by asking what happens to a pair’s difference in
political opinion if their social distance changes from d to d+ 1. Using integration group
membership as instrument for social distance, we can interpret the estimate of the effect
of social distance on difference in opinion as a causal response averaged over all pairs that
are induced by the IV to move closer in social distance (Angrist and Pischke, 2009). The
magnitude of the effect in column (2) is a lot smaller than the corresponding benchmark
coefficient, suggesting that the effect is not always strong for all values of social distance
(now a positive coefficient means a shorter social distance leads to closer opinions.)
In this direction, we further cut the sample into subsamples to compare between pairs
of consecutive social distances: social distance 1 versus 2 in column (3), social distance 2
versus 3 in column (4), and social distance 3 versus farther distances in column (5). In
each column, we can always use the same integration group membership as instrument,
subject to a test of instrument strength. We find a strong and precisely estimation effect
on political opinion when social distance shrinks from 2 to 1, i.e., when indirect friends
26
become direct friends. The magnitude of the convergence effect of this treatment is not
too far from the benchmark result: 0.38 versus 0.54 points on a scale from 1 to 10. Beyond
direct friends, we find some evidence of a similar-sized effect of reducing social distance
from 3 to 2, although the estimate remains rather imprecise, but not any statistically
significant result above social distance 3.17
[Insert Table 10 here]
We further study whether students’ positions in the social network, notably by the
network centrality, may matter to the friendship effect on differences in political opin-
ions. This question is related to the line of thought that centrality in social networks
also conveys information about a person’s influence and charisma. If that is the case, the
friendship effect should be much larger for pairs of a highly-central student (a “star”) and
a non-central student (a “non-star”) than pairs of two non-central students or pairs of two
highly-central students. Table 11 Panel A explores this idea using the measure of eigen-
vector centrality (see Table 1, Panel A), according to which a student’s centrality measure
is a linear aggregate of his friends’ centrality. The three columns consider respectively
the subsamples of pairs of two stars (column (1)), of one star and one non-star (column
(2)), and of two non-stars (column (3)), for which a student is considered highly-central
(a star) if his centrality measure is in the highest quartile of its distribution. The friend-
ship effect is not present among pairs of stars, with a positive, statistically insignificant
estimate. Among the other two subsamples, the effect is rather strong and statistically
significant at 10%. Those results are further confirmed in Table 11 Panel B, in which we
show the friendship effect for pairs of students belonging to different quartiles of central-
ity. While there is a considerable amount of noise, the overall message is that individuals
in quartiles 1 to 3 are generally more malleable to their friends’ influence. Overall, the
evidence is consistent with the view that network stars are hardly influenceable, whereas
non-stars tend to be more influenceable. However, network stars do not necessarily wield
stronger influence than non-stars.
[Insert Table 11 here]
17This is similar to Leider et al.’s (2009) description that from a social preference perspective, indirectfriends of distance 3 are not distinguishable from strangers.
27
5.6 Robustness checks
So far, all results on political opinions use initial differences in political opinions in August
2013 that are surveyed in March 2014 (see the timing in section 4.2.) There is a concern,
usually raised in the experimental economics literature, that those answers may incor-
porate a recall error biased towards the present. We will provide additional robustness
checks regarding this concern.
In the second survey in June 2015, we collect data from a subsample of participants in
the first survey, including their present political opinion, and their recalled opinion back
in March 2014. First, we check if the recalled opinions are much far off the actual opinions
based on contemporaneous answers in March 2014. Table 12 shows the joint distribution
of both surveyed and recalled opinions for 2014. The mass is clearly concentrated on the
diagonal, with 90% of the observations not differing more than 1 point between the two
measures, implying a very strong correspondence between recalled and actual answers.
This gives us confidence on the fact that the recalled opinion expressed in March 2014
for the political opinion in August 2013 is relatively accurate.
[Insert Table 12 here]
Table 13 presents further results on students’ recall error, measured as recalled opinion
for 2014 minus actual opinion surveyed in 2014. The absolute magnitude of the recall error
has practically zero partial correlations with past and present actual political opinion, as
shown in column (1). However, in column (2) we do find evidence that the signed recall
error is strongly correlated with the change in opinions from 2014 to 2015, signifying that
recalled opinions are biased towards present opinions (as surveyed in 2015).18
[Insert Table 13 here]
How can the recall error affect our results? First, if less than 10% of answers suffer
a recall error, the resulting bias on our benchmark result would probably be minimal.
Second, since we mostly control for initial political opinions of August 2013, if this variable
is biased towards actual opinions of March 2014, it would create an attenuation bias of
our coefficient of interest towards zero. Indeed, the recalled opinion with error will tend
to absorb more variation in the outcome variable than the actual initial opinion, and by
doing so it reduces the coefficient of the other variables.
18This is reminiscent of the experimental economics literature on similar recalled errors.
28
This attenuation bias can be best illustrated in the small sample of students that
we could observe in both waves. Table 14 replicates the benchmark IV regression on
differences of opinions in 2015, using as control either the recalled opinions for 2014 (in
column (1)) or the actual opinions surveyed in 2014 (in column (2)), as well as the full
set of predetermined covariates. While the main qualitative effects remain in the same
direction as our benchmark results from Table 4B, the friendship effect is larger when the
actual opinion is used in column (2), confirming the attenuation bias in this case. Taken
together, the evidence suggests that if we have actual data on political opinions before
Sciences Po, the benchmark estimate of the friendship effect may be even larger.
Table 14 also provides additional insight into the friendship effect after two years.
The magnitude of the friendship effect is comparable to that in the benchmark regression
from Table 4B’s column (2), suggesting that most of its effect has taken place by March
2014, and that there remains little additional effect one year further.
[Insert Table 14 here]
6 Concluding remarks
In this paper, we investigate empirically how newly-formed friendships through a one-
week exposure among new students at the entry of college may shape friends’ political
beliefs. We find that within the first 6 months of college, friendship causes a substan-
tial reduction in the gap between friends’ political opinions by around 0.5 points (over a
scale from 1 to 10 in political opinions), equivalent to a quarter of the average gap and a
third of its standard deviation. Friendship also causes friends’ participation in the same
politically-inclined associations. The effect of friendship works through a discouragement
of divergence and an encouragement of co-movement of opinions, but not an encourage-
ment of convergence. Consequently, friendship tends to reduce extremist views among
students, without forcing conformism. The friendship effect is strongest among students
with high similarity in political opinions before joining Sciences Po.
The findings are consistent with an explanation that politically similar friends would
socialize more along the political dimension, including participation in the same politically-
inclined associations, which results in a strong effect of friendship on their opinions. In
contrast, politically dissimilar friends would not follow such a path, hence friendship has
little effect on their political opinions.
We find that friends’ effect on political opinion is highly heterogeneous and nonlinear.
29
It is stronger for stronger friendship, closer social distance. It is likely asymmetric between
network stars and non-stars, defined by network eigenvector centrality, in that stars a
much less likely to be influenced. Those findings will be important in shaping how we
model the process of belief formation and transmission in networks.
References
[1] Acemoglu, D., Dahleh, M.A., Lobel, I. and A. Ozdaglar (2011), “Bayesian learning
in social networks,” Review of Economic Studies 78, 1201-1236.
[2] Advani, A. and B. Malde (2018), “Credibly identifying social effects: Accounting for
network formation and measurement error,” Journal of Economic Surveys, forth-
coming.
[3] Angrist, J.D. and G.W. Imbens (1994), “Identification and Estimation of Local
Average Treatment Effects,” Econometrica 62, 467-475.
[4] Angrist, J.D. and V. Lavy (1999), “Using Maimonides’ rules to estimate the effect of
class size on scholastic achievement,” Quarterly Journal of Economics 114, 533-575.
[5] Angrist, J.D. and J.-S. Pischke (2009), Mostly Harmless Econometrics: An Empiri-
cists Companion. Princeton: Princeton University Press.
[6] Babcock, P.S. and J.L. Hartman (2010), “Networks and workouts: Treatment size
and status specific peer effects in a randomized field experiment,” NBER Working
Paper No. 16581.
[7] Badev, A. (2018), “Discrete Games in Endogenous Networks: Theory and Poli-cy,”
Unpublished manuscript, University of Pennsylvania.
[8] Bala, V. and S. Goyal (1998), “Learning from neighbors,” Review of Economic Stud-
ies 65, 595-621.
[9] Bala, V. and S. Goyal (2001), “Conformism and diversity under social learning,”
Economic Theory 17, 101-120.
[10] Bandiera, O. and I. Rasul (2006), “Social networks and technology adoption in
northern Mozambique,” Economic Journal 116, 869-902.
[11] Bayer, P., S.L. Ross and G. Topa (2008), “Place of work and place of residence:
Informal hiring networks and labor market outcomes,” Journal of Political Economy
116, 1150-1196.
30
[12] Beaman, L.A. (2012), “Social networks and the dynamics of labor market outcomes:
Evidence from refugees resettled in the U.S.,” Review of Economic Studies 79, 128-
161.
[13] Bifulco, R., Fletcher, J.M. and S.L. Ross (2011), “The effect of classmate charac-
teristics on post-secondary outcomes: Evidence from the Add Health,” American
Economic Journal: Economic Policy 3, 25-53.
[14] Blume, L.E., Brock, W.A., Durlauf, S.N. and Y.M. Ioannides (2011), “Identification
of social interactions,” In: J. Benhabib, A. Bisin, and M.O. Jackson (Eds.), Handbook
of Social Economics, Volume 1B, Amsterdam: Elsevier Science, pp. 853-964.
[15] Boisjoly, J., Duncan, G.J., Kremer, M., Levy, D.M. and J. Eccles (2006), “Empathy
or Antipathy? The Impact of Diversity,” American Economic Review 96, 1890-1905.
[16] Bond, R.M., Fariss, C.J., Jones, J. J., Kramer, A.D.I., Marlow, C., Settle, J. E. and
J.H. Fowler (2012), “A 61-million-person experiment in social influence and political
mobilization,” Nature 489, 295-298.
[17] Bramoulle, Y., Djebbari, H. and B. Fortin (2009), “Identification of peer effects
through social networks,” Journal of Econometrics 150, 41-55.
[18] Bramoulle, B.W. Rogers and A. Galeotti (2016), The Oxford Handbook of the Eco-
nomics of Networks, Oxford: Oxford University Press.
[19] Bruhn, M. and D. McKenzie (2009), “In pursuit of balance: randomization in prac-
tice in development field experiments,” American Economic Journal: Applied Eco-
nomics 1, 200–232.
[20] Cai, J., de Janvry, A. and E. Sadoulet (2015), “Social networks and the decision to
insure,” American Economic Journal: Applied Economics 7, 81-108.
[21] Calvo-Armengol, A., Patacchini, E. and Y. Zenou (2009), “Peer effects and social
networks in education,” Review of Economic Studies 76, 1239-1267.
[22] Carlson, M., Dalh, G., D.O. Rooth (2015), “Do politicians change public attitudes?”
NBER Working Paper No. 21062.
[23] Carrell, S.E., Fullerton, R.L. and J.E. West (2009), “Does your cohort matter?
Estimating peer effects in college achievement,” Journal of Labor Economics 27,
439-464.
[24] Carrell, S.E., Sacerdote B.I. and J.E. West (2013), “From natural variation to op-
timal policy? The importance of endogenous peer group formation,” Econometrica
81, 855-882.
31
[25] Centola, D. (2010), “The spread of behavior in an online social network experiment,”
Science 329, 1194-1197.
[26] Centola, D. (2011), “An experimental study of homophily in the adoption of health
behavior,” Science 334, 1269-1272.
[27] Chandrasekhar, A.G. and M.O. Jackson (2013), “Tractable and consistent random
graph models,” Unpublished manuscript, Stanford University.
[28] Chandrasekhar, A.G. and R. Lewis (2011), “Econometrics of sampled networks,”
Unpublished manuscript, Stanford University.
[29] Choi, S., Gale, D., and S. Kariv (2012), “Social learning in networks: A quantal
response equilibrium analysis of experimental data,” Review of Economic Design
16, 93-118.
[30] Conley, T.G. and C.R. Udry (2010), “Learning about a new technology: pineapple
in Ghana,” American Economic Review 100, 1-28.
[31] Currarini, S., Jackson, M.O., and P. Pin (2009), “An economic model of friend-ship:
Homophily, minorities, and segregation,” Econometrica 77, 1003-1045.
[32] Damm, A.P. (2009), “Ethnic enclaves and immigrant labor market outcomes: Quasi-
experimental evidence,” Journal of Labor Economics 27, 281-314.
[33] DeGroot, M. H. (1974), “Reaching a Consensus,” Journal of the American Statistical
Association 69, 118-121.
[34] DellaVigna, S. and E. Kaplan (2007): “The Fox News Effect: Media Bias and Vot-
ing,” Quarterly Journal of Economics 122, 1187-1234.
[35] DellaVigna, S. and M. Gentzkow (2010): “Persuasion: Empirical Evidence,” Annual
Review of Economics 2, 643-669.
[36] DeMarzo, P., D. Vayanos, and J. Zwiebel (2003), “Persuasion Bias, Social Influence,
and Uni-Dimensional Opinions,” Quarterly Journal of Economics 118, 909-968.
[37] De Paula, A. (2017), “Econometrics of network models,,” In: B. Honore, A. Pakes,
M. Piazzesi and L. Samuelson (Eds.), Advances in Economics and Econometrics:
Theory and Applications: Eleventh World Congress, Cambridge: Cambridge Uni-
versity Press, pp. 268–323.
[38] Foster, A. and M. Rosenzweig (1995), “Learning by doing and learning from others:
human capital and technical change in agriculture,” Journal of Political Economy
103, 1176-1209.
32
[39] Gabel, M. and K. Scheve (2007), “Estimating the Effect of Elite Communications
on Public Opinion Using Instrumental Variables,” American Journal of Political
Science 51,1013-1028.
[40] Gentzkow, M. (2006), “Television and Voter Turnout,” Quarterly Journal of Eco-
nomics 121, 931-972.
[41] Gentzkow, M., Shapiro, J.M. and M. Sinkinson (2011), “The Effect of Newspaper
Entry and Exit on Electoral Politics,” American Economic Review 101, 2980-3018.
[42] Gerber, A. S., D. Karlan, and D. Bergan (2009), “Does the Media Matter? A Field
Experiment Measuring the Effect of Newspapers on Voting Behavior and Political
Opinions,” American Economic Journal: Applied Economics 1, 35-52.
[43] Goldsmith-Pinkham, P. and G.W. Imbens (2013), “Social networks and the iden-
tification of peer effects,” Journal of Business and Economic Statistics 31, 253-264.
[44] Goeree, J.K., M.A. McConnell, T. Mitchell, T. Tromp, and L. Yariv (2010), “The
1/d Law of Giving,” American Economic Journal: Micro-Economics 2, 183-203.
[45] Golub, B. and M.O. Jackson (2010), “Naive learning in social networks and the
wisdom of crowds,”American Economic Journal: Microeconomics 2, 112-149.
[46] Golub, B. and M.O. Jackson (2012), “How homophily affects the speed of learn-ing
and best-response dynamics,” Quarterly Journal of Economics 127, 1287-1338.
[47] Golub, B. and E. Sadler (2016), “Learning in social networks,” In: Bramoulle, Y.,
A. Galeotti and B.W. Rogers (Eds.), The Oxford Handbook on the Economics of
Networks, Oxford: Oxford University Press.
[48] Goyal, S. (2011), “Learning in networks,” In: J. Benhabib, A. Bisin and M.O. Jack-
son (Eds.), Handbook of Social Economics, Volume 1A, Amsterdam: Elsevier Sci-
ence, pp. 679-727.
[49] Graham, B.S. (2015), “Methods of identification in social networks,” Annual Review
of Economics 7, 465–485.
[50] Graham, B.S. (2017), “An econometric model of network formation with degree
heterogeneity,” Econometrica 85, 1033–1063.
[51] Granovetter, M. (1974), Getting a Job, Cambridge, MA: Harvard University Press.
[52] Henkel, J. and S.M. Maurer (2010), “Network effects in biology R&D,” American
Economic Review, Papers and Proceedings 100, 159-164.
33
[53] Hoffmann, F. and P. Oreopoulos (2009), “Professor qualities and student achieve-
ment,” Review of Economics and Statistics 91, 83-92.
[54] Imbens, G.W. (2010), “Better LATE than nothing: Some comments on Deaton
(2009) and Heckman and Urzua (2009),” Journal of Economic Literature 48, 399–
423.
[55] Ioannides, Y.M. (2012), From Neighborhoods to Nations: The Economics of Social
Interactions, Princeton: Princeton University Press.
[56] Jackson M.O. (2008), Social and Economic Networks, Princeton, NJ: Princeton Uni-
versity Press.
[57] Jackson, M.O. (2011), “An overview of social networks and economic applications,”
In: J. Benhabib, A. Bisin and M.O. Jackson (Eds.), Handbook of Social Economics
Volume 1A, Amsterdam: Elsevier Science, pp. 511-585.
[58] Jackson, M.O. (2014), “Networks in the understanding of economic behaviors,” Jour-
nal of Economic Perspectives 28, 3-22.
[59] Jackson, M.O., Rogers, B.W. and Y. Zenou (2017), “The economic consequences of
social network structure,” Journal of Economic Literature 55, 49–95.
[60] Jackson, M.O. and Y. Zenou (2015), “Games on networks,” In: P. Young and S.
Zamir (Eds.), Handbook of Game Theory, Vol. 4, Amsterdam: Elsevier, pp. 91-157.
[61] Kearns, M.J., Judd, S., Tan, J. and J. Wortman (2009), “Behavioral experiments on
biased voting in networks,” Proceedings of the National Academy of Sciences of the
USA 106, 1347-1352.
[62] Kendall, C., T. Nannicini, and F. Trebbi (2015), “How Do Voters Respond to In-
formation? Evidence from a Randomized Campaign,” American Economic Review
105, 322–353.
[63] Kremer, M. and T. Miguel (2007), “The illusion of sustainability,” Quarterly Journal
of Economics 122, 1007-1065.
[64] Lavy, V. and E. Sand (2016), “The effect of social networks on student’s academic
and non-cognitive behavioral outcomes: Evidence from conditional random assign-
ment of friends in school,”Unpublished manuscript, University of Warwick.
[65] Lazarfeld, P., Berelson, B. and H. Gaudet (1944), People’s Choice: How the Vot-er
Makes Up His Mind in a Presidential Campaign, Third edition, New York : Columbia
University Press.
34
[66] Leider, S., Mobius, M., Rosenblat, T. and Q-A. Do (2009), “Directed Altruism
and Enforced Reciprocity in Social Networks,” Quarterly Journal of Economics 124,
1815-1851.
[67] Leider, S., Mobius, M., Rosenblat, T. and Q-A. Do (2010), “What Do We Expect
From Our Friends?” Journal of the European Economic Association 8, 120-138.
[68] Lin, X. (2010), “Identifying peer effects in student academic achievement by a spatial
autoregressive model with group unobservables,” Journal of Labor Economics 28,
825-860.
[69] Lee, L.-f. and X. Liu (2010), “Efficient GMM Estimation of High Order Spatial
Autoregressive Models with Autoregressive Disturbances,” Econometric Theory 26,
187-230.
[70] Liu, X., Patacchini, E. and Y. Zenou (2014), “Endogenous peer effects: Local aggre-
gate or local average?” Journal of Economic Behavior and Organization 103, 39-59.
[71] Manski, C.F. (1993), “Identification of endogenous effects: The reflection problem,”
Review of Economic Studies 60, 531-542.
[72] McPherson, M., Smith-Lovin, L. and J.M. Cook (2001), “Birds of a feather: Ho-
mophily in social networks,” Annual Review of Sociology 27, 415-444.
[73] Mele, A. (2017), “A structural model of segregation in social networks,” Economet-
rica 85, 825–850.
[74] Mobius, M. and T. Rosenblat (2014), “Social Learning in Economics,” Annual Re-
view of Economics 6, 827-847.
[75] Mobius, M., Szeidl, A. and T. Phan (2015), “Treasure hunt: Social learning in the
field,” Working Paper, Central European University, Budapest.
[76] Munshi, K. (2003), “Networks in the modern economy: Mexican migrants in the US
labor market,” Quarterly Journal of Economics 118, 549-599.
[77] Patacchini, E., Rainone, E. and Y. Zenou (2017), “Heterogeneous peer effects in
education,” Journal of Economic Behavior and Organization 134, 190–227.
[78] Patacchini, E. and Y. Zenou (2012), “Ethnic networks and employment out-comes,”
Regional Science and Urban Economics 42, 938-949.
[79] Patacchini, E. and Y. Zenou (2017), “Social networks and parental behavior in the
intergenerational transmission of religion,” Quantitative Economics 7, 969–995.
35
[80] Rao, N., Mobius, M. and T. Rosenblat (2007), “Social networks and vaccination
decisions,” Federal Reserve Bank of Boston, Working Paper No. 0712.
[81] Rouban, L. (2011), “Sociologie politique des deputes de la Ve Republique, 1958-
2007,” Cahiers du CEVIPOF, No. 55, Paris, CEVIPOF.
[82] Rouban, L. (2014a), La Fonction Publique en Debat, Paris: Documentation francaise.
[83] Rouban, L. (2014b), “L’elite urbaine : les maires des villes de plus de 30 000 habitants
de 1983 a 2008,” Note No. 3, Paris, CEVIPOF.
[84] Sacerdote, B. (2001), “Peer effects with random assignment: Results from Dart-
mouth roomates,” Quarterly Journal of Economics 116, 681-704.
[85] Sacerdote, B. (2011), “Peer effects in education: How might they work, how big
are they and how much do we know thus far?” In: E.A. Hanushek, S. Machin and
L. Woessmann (Eds.), Handbook of Economics of Education, Vol. 3, Amsterdam:
Elevier Science, pp. 249-277.
[86] Sacerdote, B. (2014), “Experimental and quasi-experimental analysis of peer effects:
Two steps forward,” Annual Review of Economics 6, 253-272.
[87] Topa, G. (2001), “Social interactions, local spillovers, and unemployment,” Review
of Economic Studies 68, 261-295.
[88] Topa, G. and Y. Zenou (2015), “Neighborhood versus network effects,” In: G. Du-
ranton, V. Henderson and W. Strange (Eds.), Handbook of Regional and Urban Eco-
nomics, Vol. 5, Amsterdam: Elsevier Publisher, pp. 561-624.
[89] Wasserman, S. and K. Faust (1994), Social Network Analysis. Methods and Applica-
tions, Cambridge: Cambridge University Press.
[90] Zax, J.S. and D.I. Rees (2002), “IQ, academic performance, environment, and earn-
ings,” Review of Economics and Statistics 84, 600–616.
[91] Zimmerman, D. (2003), “Peer effects in academic outcomes: Evidence from a natural
experiment,” Review of Economics and Statistics 9-23.
36
Table 1 - Panel A: ”OR” Network statistics
Mean of degree per individual 8.8625Variance of degree per individual 18.4842Median of degree per individual 10Maximum of degree per individual 21Minimum of degree per individual 0Diameter of the network 9Average path length 3.7008Overall clustering coefficient 0.241Average clustering coefficient 0.271Mean eigenvector centrality 0.0361Standard deviation of
0.0200eigenvector centrality
Notes: Summary statistics are computed on thefull sample.
Table 1 - Panel B: Dyadic Links and Groups
Variable(1) (2)
Full Sample Benchmark Sample
Mean Standard deviation Mean Standard deviation
Friendship 0.016 (0.124) 0.017 (0.13)2nd Order Links 0.093 (0.29) 0.099 (0.299)3rd Order Links 0.38 (0.485) 0.402 (0.49)Mere relationship (strength 1) 0.0014 (0.0382) 0.0017 (0.0421)Friendship link (strength 2) 0.0063 (0.0791) 0.0068 (0.082)Close friendship (strength 3) 0.0041 (0.0642) 0.0045 (0.067)Very close friendship (strength 4) 0.0035 (0.0593) 0.004 (0.0632)
Same Integration Group 0.016 (0.128) 0.018 (0.133)Same Study Group 0.023 (0.149) 0.023 (0.15)
Notes: Summary statistics (1) refer to the full dyadic sample, where full sample is defined as the setof all pairs for which both members named at least one friend or stated that they have no friends inSciences Po. Summary statistics (2) refer to the benchmark dyadic sample as described in Table 2.
Table 1 - Panel C: Dependent Variables (at the individual level)
Variable(1) (2)
Full Sample Benchmark Sample
Mean Standard deviation Mean Standard deviation
Political Opinion in 2014 (1-10) 5.044 (1.755) 5.091 (1.712)Pre-Sciences Po Political Opinion (in 2013) (1-10) 5.108 (1.958) 5.148 (1.934)Political Opinion in 2015 4.841 (1.807) 4.818 (1.746)Political Opinion in 2014 as recalled in 2015 4.907 (1.647) 4.941 (1.642)Enrolled in a Political Party in 2014 (yes / no) 0.102 (0.303) 0.121 (0..326)Enrolled in a Political Party in 2013 (yes / no) 0.067 (0.249) 0.076 (0.265)Enrolled in an Association in 2014 0.621 (0.485) 0.642 (0.48)
Notes: Summary statistics (1) refer to the full individual sample, where the full sample is made of all the individualobservations for which the variable described is not missing. Summary statistics (2) refer to the benchmark sample,where the benchmark sample is defined as the individual sample containing all the individuals that are present inour benchmark dyadic sample as described in Table 2.
Table 1 - Panel D: Dependent Variables (at the dyadic level)
Variable(1) (2)
Full Sample Benchmark Sample
Mean Standard deviation Mean Standard deviation
Difference in Political Opinion in 2014 1.932 (1.467) 1.926 (1.468)Initial difference in Political Opinion (2013) 2.211 (1.631) 2.2 (1.623)Difference in Political Opinion in 2015 2.014 (1.538) 1.94 (1.496)Difference in Political Opinion in 2014 (as recalled in 2015) 1.835 (1.424) 1.798 (1.412)diff. Enrollment in a Political Party in 2014 0.815 (0.388) 0.785 (0.411)diff. Enrollment in a Political Party in 2013 0.874 (0.332) 0.858 (0.349)diff. Enrollment in the Same Political Party in 2014 0.827 (0.379) 0.8 (0.4)diff. Enrollment in an Association 0.479 (0.4995) 0.461 (0.498)diff. Enrollment in the Same Association 0.915 (0.279) 0.903 (0.296)
Notes: Summary statistics (1) refer to the full dyadic sample, where full sample is defined as the set of all pairs for whichboth members named at least one friend or stated that they have no friends in Sciences Po. Summary statistics (2) referto the benchmark dyadic sample as described in Table 2. The prefix ”diff.” in front of certain variables signifies a dummyvariable that is equal to 1 iff the corresponding variable is the not same for both students in the pair and 0 iff it is the same.
Table 1 - Panel E: Independent Variables (at the individual level)
Variable(1) (2)
Full Sample Benchmark Sample
Mean Standard deviation Mean Standard deviation
Gender (1= Female) 0.592 (0.492) 0.583 (0.494)Honors Graduation 0.753 (0.432) 0.831 (0.375)Tuition Fees 3580 (3493) 3826 (3328)
Notes: Summary statistics (1) refer to the full individual sample, where the fullsample is made of all the individual observations for which the variable describedis not missing. Summary statistics (2) refer to the benchmark sample, where thebenchmark sample is defined as the individual sample containing all the individualsthat are present in our benchmark dyadic sample as described in Table 2.
Table 1 - Panel F: Independent Variables (at the dyadic level)
Variable(1) (2)
Full Sample Benchmark Sample
Mean Standard deviation Mean Standard deviation
Same Gender 0.522 (0.4995) 0.512 (0.4998)Same Female 0.369 (0.483) 0.339 (0.473)Same Nationality 0.928 (0.259) 0.964 (0.186)Same French 0.716 (0.451) 0.751 (0.432)Same Double French Nationality 0.774 (0.418) 0.778 (0.416)Same Admission 0.565 (0.496) 0.675 (0.468)Same Priority Admission 0.029 (0.168) 0.015 (0.122)Same Department of High School 0.052 (0.221) 0.059 (0.237)Same Region of High School 0.253 (0.435) 0.251 (0.434)Same High School Major 0.363 (0.481) 0.375 (0.484)Difference in Tuition Fees 3879 (3005) 3759 (2832)Same No Fees 0.476 (0.499) 0.614 (0.487)Same Parents’ Profession 0.422 (0.494) 0.442 (0.497)Same ZIP code 0.026 (0.160) 0.026 (0.159)Same Program 0.520 (0.4995) 0.508 (0.4999)
Notes: Summary statistics (1) refer to the full dyadic sample, where full sample is defined as theset of all pairs for which both members named at least one friend or stated that they have nofriends in Sciences Po. Summary statistics (2) refer to the benchmark dyadic sample as describedin Table 2.
Table 1 - Panel G: Friendship
Friend Met During Friend Met During Friend Met BeforeIntegration Week Integration Week Sciences Po
(1) (2) (3)
No 76.43 5.62 99.80Yes 23.57 94.38 0.20
Observations 946 178 178
(1) Couples of friends that declare to have met during the integration week.(2) Couples of friends that attended the same integration group and that declareto have met during the integration week.(3) Couples of friends that attended the same integration group and that declareto have met before entering Sciences Po.Notes: Summary statistics refer to the benchmark dyadic sample as described inTable 2.
Table 2 - Balance Test of Integration Group
Dependent Variable: Common Integration Group
Same Gender 0.000673(0.00148)
Same Female -0.00194(0.00135)
Same Nationality 0.00622(0.00449)
Same Mixed -0.00206French Nationality (0.00421)Same Admission 0.00137
(0.00116)Same Priority 0.00182Admission (0.00697)Same Department 0.00299of High School (0.00420)Same Region 0.00170of High School (0.00156)Same High -0.00203*School Major (0.00105)Diff. in -3.18e-07Tuition Fees (2.35e-07)Same No Fees -0.00138
(0.00115)Same Parents 0.00103Profession (0.00130)Same ZIP Code -0.000761
(0.00337)Same Program 0.00270**
(0.00121)
Observations 54,615R-squared 0.001F-stat 2.957
Sample: the benchmark sample includes pairs that arein the upper triangular part of the adjacency matrix,whose individuals both answered to the question ontheir friendship network and for which we have valuesfor the controls and for the variable ”Difference in Po-litical Opinion in 2014”. We drop couples that containat least one individual in the top 5 percent of the dis-tribution of time taken to name each friend (equal to81.625 seconds per friend).Notes: F-stats are for the joint significance of the vari-ables included in the model. Standard errors (in brack-ets) are clustered at the individual 1 and individual 2level.
Table 3: First Stage on Friendship Formation
Dependent Variable: Friendship
(1) (2) (3)
Same Integration Group 0.166*** 0.161*** 0.160***(0.0154) (0.0152) (0.0152)
Same Tutorial Group - 0.358*** 0.353***(0.0210) (0.0207)
Initial Diff. in Political - - -0.000932**Opinion (August 2013) (0.000395)Same Gender - - 0.0115***
(0.00155)Same Female - - -0.00794***
(0.00164)Same Nationality - - -0.00303
(0.00460)Same French - - 0.00524
(0.00387)Same Double - - -0.00203Nationality (0.00372)Same Admission - - 0.00459***
(0.00131)Same Priority - - 0.00215Admission (0.00597)Same Department of - - 0.0113***High School (0.00353)Same Region of - - 0.00247*High School (0.00136)Same High - - 0.00368***School Major (0.000914)Diff. in - - -4.37e-07*Tuition Fees (2.36e-07)Same No Fees - - 0.00146
(0.00131)Same Parents - - 0.00104Profession (0.00102)Same ZIP Code - - 0.0129***
(0.00493)Same Program - - 0.00799***
(0.00172)
Observations 54,615 54,615 54,615R-squared 0.029 0.198 0.202Group Clustering Yes Yes YesF-stat 114.2 256.2 47.22
Notes: Standard errors are clustered at the group of individual1 and at the group of individual 2 level. F-stats are for the jointsignificance of the variables included in the model. The sampleused is the benchmark sample described in the footnote to theTable 2.
Table 4 - Panel A: Friendship and Difference in Political Opinion with OLS
Dependent Variable: Difference in Political Opinion (March 2014)
(1) (2) (3) (4) (5) (6) (7)
OLS OLS OLS OLS OLS OLS OLS
Friendship -0.127*** -0.0977** -0.109** -0.119** -0.104** -0.105** -(0.0411) (0.0489) (0.0497) (0.0508) (0.0513) (0.0514)
Friendship * Same - - - - -0.0132 -0.0393 -Tutorial Group (0.110) (0.110)Same Tutorial Group - 0.0255 0.266*** 0.0285 0.278*** -0.072***
(0.0325) (0.0514) (0.0515) (0.0644) (0.0154)Initial Diff. in Political 0.527*** 0.526*** 0.526*** 0.526*** 0.526*** 0.526*** -Opinion (August 2013) (0.0257) (0.0252) (0.0252) (0.0252) (0.0252) (0.0252)
Observations 54,615 54,615 54,615 54,615 54,615 54,615 54,615IV No No No No No No NoControls No Yes Yes Yes Yes Yes NoGroup Clustering Yes Yes Yes Yes Yes Yes YesFixed Effects No No No Yes No Yes No
Notes: Standard errors are clustered at the group of individual 1 and at the group of individual 2 level. Thesample used is the benchmark sample described in the footnote to Table 2. Controls include the followingvariables: c. Gender, c. Female, c. Nationality, c. French, c. Double Nationality, c. Admission, c. PriorityAdmission, c. Honour Graduation (High School), c. District of High School, c. Region of High School, c. HighSchool Major, Diff. in Tuition Fees, c. No Fees, c. Parents Profession, c. ZIP Code, c. Program. Weak IV statreports the Kleibergen-Paap cluster-robust statistic, distributed as a Chi-squared under the null hypothesis ofweak identification.
Table 4 - Panel B: Friendship Effect with IV
Dependent Variable: Difference in Political Opinion (March 2014)
(1) (2) (3) (4) (5) (6)
IV IV IV IV IV IV
Friendship -0.489** -0.538** -0.554** -0.572** -0.531** -0.531**(0.217) (0.225) (0.234) (0.234) (0.257) (0.257)
Friendship * Same - - - - -0.236 -0.471Tutorial Group (0.479) (0.468)Same Tutorial Group - - 0.183** 0.461*** 0.262* 0.655***
(0.0837) (0.110) (0.159) (0.170)Initial Diff. in Political 0.526*** 0.526*** 0.526*** 0.526*** 0.526*** 0.525***Opinion (August 2013) (0.0258) (0.0252) (0.0252) (0.0253) (0.0252) (0.0253)
Observations 54,615 54,615 54,615 54,615 54,615 54,615IV S. Int. Gr S. Int. Gr. S. Int. Gr. S. Int. Gr. S. Int. Gr. S. Int. Gr.IV Same Group Same Group Same Group Same Group Same Group Same GroupControls No Yes Yes Yes Yes YesGroup Clustering Yes Yes Yes Yes Yes YesFixed Effects No No No Yes No YesWeakIV test stat. 114.3 113.2 109 105.5 20.08 18.07
Notes: Standard errors are clustered at the group of individual 1 and at the group of individual 2 level. The sample used isthe benchmark sample described in the footnote to Table 2. Controls include the following variables: c. Gender, c. Female, c.Nationality, c. French, c. Double Nationality, c. Admission, c. Priority Admission, c. Honour Graduation (High School), c.District of High School, c. Region of High School, c. High School Major, Diff. in Tuition Fees, c. No Fees, c. Parents Profession,c. ZIP Code, c. Program. Weak IV stat reports the Kleibergen-Paap cluster-robust statistic, distributed as a Chi-squared underthe null hypothesis of weak identification.
Table 5 - Panel A: Friendship Effect on Political and Associative Activities
Dependent Variables:d. Enrolment in d. Enrolment in d. Enrolment in d. Enrolment in d. Enrolment in the
Student Associations Same Association* Same Association a Political Party Same Political Party*
(1) (2) (3) (4) (5)
Friendship 0.0353 -0.180* -0.0789* -0.0278 0.0182(0.119) (0.0954) (0.0429) (0.0656) (0.318)
Political Enrollment - - - 0.468*** -(August 2013) (0.0735)c. Group -0.0180 0.00944 0.00949 0.0112 -0.00646
(0.0418) (0.0387) (0.0158) (0.0283) (0.0926)
Observations 47,895 24,310 54,615 52,650 780IV Yes Yes Yes Yes YesControls Yes Yes Yes Yes YesDouble Group Clust. Yes Yes Yes Yes YesWeakIV test stat. 106.9 63.72 109 101.2 5.696
* conditional on the individuals in the couple being enrolled in some association/political party.Notes: The dependent dummy variables are equal to 0 if the characteristic is the same for the two individuals in the pair and 1 if thecharacteristic is not the same. Standard errors are clustered at the group of individual 1 and at the group of individual 2 level. Thesample used is the benchmark sample described in the footnote to Table 2. Weak IV stat reports the Kleibergen-Paap cluster-robuststatistic, distributed as a Chi-squared under the null hypothesis of weak identification.
Table 5 - Panel B: Friendship Effect on Different Association Types
Dependent Variable:d. Enrolment in
Same Association Associations of Different AssociationsSame Type of Same Type
Type of Association:Sports Strictly Political or Political Related to Related to Political Political
Student Union One’s Identity Policy Issues
(1) (2) (3) (4) (5) (6) (7)
Friendship -0.00769 0.0495* 0.135* 0.141 0.00466 0.00329 -0.0119(0.0962) (0.0282) (0.0740) (0.0927) (0.0825) (0.0173) (0.0211)
Observations 54,615 24,310 24,310 24,310 24,310 24,310 24,310IV Yes Yes Yes Yes Yes Yes YesControls Yes Yes Yes Yes Yes Yes YesDouble Group Clust. Yes Yes Yes Yes Yes Yes YesWeakIV test stat. 108.8 63.77 63.77 63.77 63.77 63.77 63.77
Notes: The dependent dummy variables are equal to 0 if the characteristic is the same for the two individuals in the pair and 1 if the characteristicis not the same. Standard errors are clustered at the group of individual 1 and at the group of individual 2 level. The sample used is the benchmarksample described in the footnote to Table 2. Weak IV stat reports the Kleibergen-Paap cluster-robust statistic, distributed as a Chi-squaredunder the null hypothesis of weak identification.
Tab
le6:
Fri
end
ship
Eff
ect
on
Conve
rgen
ce,
Div
ergen
ce,
an
dC
o-m
ovem
ent
Dep
end
ent
Var
iab
les:
Sta
yin
gin
Wea
kS
ingle
Str
on
gW
eak
Sin
gle
Str
on
gM
ovin
gin
Mov
ing
inS
am
eS
ame
Pos
itio
nC
onve
rgen
ceC
onve
rgen
ceC
onve
rgen
ceD
iver
gen
ceD
iver
gen
ceD
iver
gen
ceS
am
eD
irec
tion
Conve
rgin
gD
irec
tion
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
Fri
end
ship
0.08
270.
0336
0.08
23
-0.0
493
-0.1
70*
-0.0
532
-0.1
16***
0.1
04*
0.1
30**
(0.0
699)
(0.1
00)
(0.1
01)
(0.0
882)
(0.0
871)
(0.0
784)
(0.0
325)
(0.0
561)
(0.0
554)
Init
ial
Diff
.in
Pol
itic
al-0
.005
270.
110*
**0.
0598
***
0.0
598***
-0.0
833***
-0.0
646***
-0.0
187***
-0.0
277***
-0.0
124**
Op
inio
n(A
ugu
st20
13)
(0.0
0378
)(0
.006
71)
(0.0
0323)
(0.0
0656)
(0.0
0256)
(0.0
0288)
(0.0
0200)
(0.0
0599)
(0.0
0482)
Sam
eT
uto
rial
Gro
up
-0.0
394
0.00
459
-0.0
207
0.0
256
0.0
611**
0.0
138
0.0
473***
-0.0
445**
-0.0
573**
(0.0
268)
(0.0
423)
(0.0
395)
(0.0
316)
(0.0
306)
(0.0
311)
(0.0
112)
(0.0
224)
(0.0
236)
Mea
nof
0.15
60.
456
0.28
90.1
50.2
27
0.188
0.0
38
0.1
83
0.1
45
Dep
end
ent
Var
iab
leS
tan
dar
dD
evia
tion
of0.
3628
0.49
810.
4535
0.3
571
0.4
186
0.391
0.1
919
0.3
868
0.3
524
Dep
end
ent
Var
iab
leO
bse
rvat
ion
s54
,615
46,9
0149
,636
46,9
01
54,6
15
54,6
15
54,6
15
54,6
15
54,6
15
IVY
esY
esY
esY
esY
esY
esY
esY
esY
esC
ontr
ols
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Dou
ble
Gro
up
Clu
st.
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Wea
kIV
test
stat
.10
8.8
109.
710
5.4
109.7
108.8
108.8
108.8
108.8
108.8
Note
s:SSta
ndard
erro
rsare
clust
ered
at
the
gro
up
of
indiv
idual
1and
at
the
gro
up
of
indiv
idual
2le
vel
.T
he
sam
ple
use
dis
the
ben
chm
ark
sam
ple
des
crib
edin
the
footn
ote
toT
able
2.
Wea
kIV
stat
rep
ort
sth
eK
leib
ergen
-Paap
clust
er-r
obust
stati
stic
,dis
trib
ute
das
aC
hi-
square
dunder
the
null
hyp
oth
esis
of
wea
kid
enti
fica
tion.
Table 7: Friendship Effect on Extremism
Dependent Variables: At Least One Extremist in the Couple in March 2014
Initial Conditions:Both Moderate Both Moderate and One Extremist and At Least One
Diff.P.P.≤ 1 One Moderate Extremist
(1) (2) (3) (4)
Friendship -0.0908 -0.128* -0.400 -0.305(0.0769) (0.0667) (0.313) (0.548)
Same Tutorial Group 0.0228 0.0624*** 0.198* 0.219(0.0255) (0.0159) (0.118) (0.213)
Observations 43,956 20,324 10,098 3,654IV Yes Yes Yes YesControls Yes Yes Yes YesDouble Group Clust. Yes Yes Yes YesWeakIV test stat. 102.1 61.05 24.96 9.469
Notes: Standard errors are clustered at the group of individual 1 and at the group of individual 2 level.The sample used is the benchmark sample described in the footnote to Table 2. Weak IV stat reportsthe Kleibergen-Paap cluster-robust statistic, distributed as a Chi-squared under the null hypothesis ofweak identification.
Table 8: Friendship Effect and Initial Differences in Political Opinion
Dependent Variables: Difference in Political Opinion (March 2014)
Conditions: - Diff.P.P.=0 Diff.P.P.=1 Diff.P.P.=2 Diff.P.P.≥ 3 Diff.P.P.≤ 1
(1) (2) (3) (4) (5) (6)
Friendship -0.996** -0.779*** -0.759** -0.411 -0.266 -0.767***(0.427) (0.299) (0.313) (0.491) (0.413) (0.239)
Friendship*Initial Pol. Opinion 0.206 - - - - -(0.174)
Initial Diff. in Political 0.522*** - - - 0.715*** 0.141***Opinion (August 2013) (0.0252) (0.0417) (0.0179)Same Tutorial Group 0.186** 0.290** 0.393*** 0.0604 0.0155 0.354***
(0.0832) (0.127) (0.127) (0.174) (0.129) (0.107)
Observations 54,615 7,714 14,235 11,644 21,022 21,949IV Yes Yes Yes Yes Yes YesControls Yes Yes Yes Yes Yes YesDouble Group Clust. Yes Yes Yes Yes Yes YesWeakIV test stat. 26.54 42.95 41.64 20.09 68.85 58.10
Notes: Standard errors are clustered at the group of individual 1 and at the group of individual 2 level. The sample usedis the benchmark sample described in the footnote to Table 2. Weak IV stat reports the Kleibergen-Paap cluster-robuststatistic, distributed as a Chi-squared under the null hypothesis of weak identification.
Table 9: Friendship Effect by Type of Friendship
Dependent Variable: Difference in Political Opinion (March 2014)
Condition: I - - - -
(1) (2) (3) (4) (5)
Friendship -0.565** - - - -(0.233)
Very Close Friendship - -3.247** - - -(1.396)
Close Friendship - - -1.377** - -or Stronger (0.594)Friendship Link - - - -0.743** -or Stronger (0.310)Mere Relationship - - - - -0.575**or Stronger (0.248)Initial Diff. in Political 0.526*** 0.525*** 0.525*** 0.525*** 0.526***Opinion (August 2013) (0.0253) (0.0253) (0.0253) (0.0253) (0.0253)Same Tutorial Group 0.184** 0.299** 0.252** 0.228** 0.190**
(0.0817) (0.132) (0.110) (0.0985) (0.0878)
Observations 54,556 54,615 54,615 54,615 54,615IV Yes Yes Yes Yes YesControls Yes Yes Yes Yes YesDouble Group Clust. Yes Yes Yes Yes YesWeakIV test stat. 111.9 28.26 50.64 71.52 99.89
I - Only friends that did not know each other from before entering Sciences Po.Notes: Standard errors are clustered at the group of individual 1 and at the group ofindividual 2 level. The sample used is the benchmark sample described in the foot-note to Table 2. Weak IV stat reports the Kleibergen-Paap cluster-robust statistic,distributed as a Chi-squared under the null hypothesis of weak identification.
Table 10: Effect of Reduction of Social Distance on Difference in Political Opinion
Dependent Variable: Difference in Political Opinion
Condition: - - Degree≤ 2 Degree = [2,3] Degree ≥ 3
(1) (2) (3) (4) (5)
Friendship -0.538** - - - -(0.225)
Shortest Path - 0.151** - - -(0.0616)
1st vs 2nd order only - - 0.376*** - -(0.122)
2nd vs 3rd order only - - - 0.440* -(0.238)
3rd vs more order only - - - - 4.067(7.311)
Observations 54,615 54,615 6,369 27,359 48,246IV Yes Yes Yes Yes YesControls Yes Yes Yes Yes YesDouble Group Clust. Yes Yes Yes Yes YesWeakIV test stat. 113.2 150.8 209.6 171.9 0.332A.R. Wald test F-stat. 5.96 5.96 9.00 3.41 1.33
Notes: specification (1) - benchmark sample; specification (2) - benchmark sample; specification(3) - only pairs of individuals that are at most within two degrees of distance; specification (4) -only pairs of individuals that are between two and three degrees of distance; specification (5) - onlypairs of individuals that are at three or more degrees of distance. Standard errors are clustered atthe group of individual 1 and at the group of individual 2 level. The sample used is the benchmarksample described in the footnote to Table 2. Weak IV stat reports the Kleibergen-Paap cluster-robuststatistic, distributed as a Chi-squared under the null hypothesis of weak identification.
Table 11 - Panel A: Friendship Effect and Network Centrality
Dependent Variable: Difference in Political Opinion (March 2014)
Condition: See Legend in the Notes
(1) (2) (3)
Friendship 0.141 -0.829* -0.767*(0.230) (0.456) (0.404)
Observations 3,240 20,250 31,125IV Yes Yes YesControls Yes Yes YesDouble Group Clust. Yes Yes YesWeakIV test stat. 160.9 58.91 57.02
Notes: specification (1) - two individuals in the top quartile of eigen-vector centrality; specification (2) - one individual in the top quartile ofeigenvector centrality; specification (3) - no individual in the top quartileof eigenvector centrality. Standard errors are clustered at the group ofindividual 1 and at the group of individual 2 level. The sample used isthe benchmark sample described in the footnote to Table 2. Weak IVstat reports the Kleibergen-Paap cluster-robust statistic, distributed asa Chi-squared under the null hypothesis of weak identification.
Table 11 - Panel B: Friendship Effect and Network Centrality
Dependent Variable: Difference in Political Opinion (March 2014)
Quartile of Eigenvector Centrality1st 2nd 3rd 4th
Quart
ile
1st -1.331*** -2.029 -0.538 -3.304
(0.410) (3.955) (1.106) (4.594)
2nd - -0.546 0.00464 -0.864
(1.264) (0.431) (0.805)
3rd
- - -1.013** -0.640(0.511) (0.440)
4th
- - - 0.118(0.226)
Notes: The table presents a matrix of results for different sub-samplesof couples selected based on their eigenvector centrality. Entry (1,1)in the matrix reports the coefficient on friendship in specification (2) -Table 4 for the sub-sample of couples where one both individuals arein the first quartile of eigenvector centrality. Entry (1,2) reports thesame coefficient for the sub-sample of couples where one individual isin the first quartile of eigenvector centrality and the other is in thesecond quartile and so on. Standard errors are clustered at the groupof individual 1 and at the group of individual 2 level. The sample usedis the benchmark sample described in the footnote to Table 2. Weak IVstat reports the Kleibergen-Paap cluster-robust statistic, distributed asa Chi-squared under the null hypothesis of weak identification.
Table 12: Descriptive Statistics on Recall Bias
Percentages (Actual Pol. Op. in 2014 as a reference) N
Actual (Individual) Political Opinion in 20141 2 3 4 5 6 7 8 9 10 Total
Rec
alled
P.
Op.
in2014
1 0 1 0 0 0 0 0 0 0 0 12 0 5 1 2 0 0 0 0 0 0 83 1 6 19 7 3 1 0 0 0 0 374 0 0 8 16 22 4 1 0 0 0 515 0 0 2 7 26 6 1 0 0 0 426 1 0 0 1 6 22 8 3 0 0 417 0 0 0 1 0 6 12 5 0 0 248 0 0 0 0 0 1 6 6 1 0 149 0 0 0 0 0 0 2 1 0 1 410 0 0 0 0 0 0 0 0 0 0 0
Total 2 12 30 34 57 40 30 15 1 1 222
Notes: The table reports the joint empirical distribution of actual (horizontalaxes) and recalled (vertical axes) individual political opinion in 2014. The tablemakes use of the individual linked 2014-2015 dataset. The sample includes onlythose individuals present both in the 2014 and in the 2015 survey and for whichthe variables ”actual political opinion in 2014” and ”recalled political opinionin 2014” are not missing.
Table 13: Recall Bias Regression on Individual Data
Dependent Variable:Absolute Recall Bias Recall Bias
(1) (2)
Actual (Ind.) Pol. Op. in 2015 0.000691 -(0.115)
Actual (Ind.) Pol. Op. in 2014 0.00703 -(0.137)
Diff. in Actual (Ind.) Pol. Op. - 0.574***Between 2015 and 2014 (0.0435)
Observations 220 220Double Group Clust. Yes Yes
Notes: The table makes use of the individual linked 2014-2015 dataset.Standard errors are clustered at the group level. The sample includesonly those individuals present both in the 2014 and in the 2015 surveyand for which the variables ”political opinion in 2015”, ”actual politicalopinion in 2014” and ”recalled political opinion in 2014” are not missing.We define as ”Recall Bias” the difference between political opinion in2014 as declared in the 2015 survey (i.e. recalled political opinion in2014) and political opinion in 2014 as declared in the 2014 survey (i.e.actual political opinion in 2014) and ”Absolute Recall Bias” the absolutevalue of this difference.
Table 14: Benchmark Regression on 2015 Data
Dependent Variable:Difference in Political Opinion in 2015
(1) (2)
Friendship -0.295** -0.481**(0.138) (0.192)
Recalled Diff. in Political Opinion in 2014 0.763*** -(0.0325)
(Actual) Diff. in Political Opinion in 2014 - 0.644***(0.0448)
Observations 11,836 11,836Controls Yes YesDouble Group Clust. Yes YesWeakIV test stat. 72.91 73.03
Notes: The table makes use of the dyadic linked 2014-2015 dataset on pairs of indi-viduals. Standard errors are clustered at the group of individual 1 and at the group ofindividual 2 level. The sample used is the benchmark sample described in the footnoteto Table 2, minus those observation for which either the variable ”difference in politicalopinion in 2015” or the variable ”recalled difference in political opinion in 2014” or bothare missing or for which at least one individual is not present in either the 2014 or the2015 survey or both. Weak IV stat reports the Kleibergen-Paap cluster-robust statistic,distributed as a Chi-squared under the null hypothesis of weak identification.
Figure 1: Distributions of Individual Political Opinion
A Appendix
Table A1: Description of Variables - Dyadic Data
Variable Description
Friendship1 if at least one of the two individual has named the other as one ofher friends (i.e. undirected friendship), zero otherwise.
Same Integration Group1 if the two individuals have attended the same integration group overthe 2013 summer, 0 otherwise.
Same Tutorial Group1 if the two individuals have been part of the same ”triplette” group,0 otherwise.
Difference in Political Opinion (March2014)
Absolute difference between the political opinions of the twoindividuals, as declared on a 1-10 scale.
Initial difference in Political Opinion(August 2013)
Absolute difference between the political opinions of the twoindividuals, as declared on a 1-10 scale.
Difference in Political Opinion in 2015Absolute difference between the political opinions of the twoindividuals, as declared on a 1-10 scale.
Difference in Political Opinion in 2014Absolute difference between the political opinions of the twoindividuals, as declared on a 1-10 scale.
diff. Enrollment in a Political Party in2014
0 if the two individuals are enrolled in a political party. Missing if atleast one of them did not answer to this question. 0 otherwise.
diff. Enrollment in a Political Party in2013
0 if the two individuals are enrolled in a political party. Missing if atleast one of them did not answer to this question. 0 otherwise.
diff. Enrollment in the Same PoliticalParty in 2014
0 if the two individuals are enrolled in the same political party.Missing if at least one of them did not answer to this question. 0otherwise.
diff. Enrollment in an Association0 if the two individuals are enrolled in the a student association.Missing if at least one of them did not answer to this question. 0otherwise.
diff. Enrollment in the Same Association0 if the two individuals are enrolled in the same student association.Missing if at least one of them did not answer to this question. 0otherwise.
Sum of MovementsSum of absolute movements in political opinion among the twoindividuals between August 2013 and March 2014.
S. Staying in Same Position1 if both individuals have not changed political opinion betweenAugust 2013 and March 2014
Moving in Same Direction1 if both individuals have changed their political opinion betweenAugust 2013 and March 2014 and their new political opinion havemoved in the same direction relative to their initial one, 0 otherwise.
Moving in Non-Opposing Direction1 if between August 2013 and March 2014 the two individuals havenot changed their political opinion in opposing directions (e.g. onehas moved to the right and the other to the left), 0 otherwise.
Moving in Same Converging Direction
1 if both individuals have changed their political opinion betweenAugust 2013 and March 2014, their new political opinion have movedin the same direction relative to their initial one and the difference intheir political opinions has diminished, 0 otherwise.
Strong Convergence
1 if the two individuals have different initial political positions, noneof them have moved away from and at least one of them has movedtowards the other initial political opinion relative to her own initialposition. Missing if the two individuals have the same initial politicalopinion. 0 otherwise.
Weak Convergence
1 if the two individuals have different initial political positions, noneof them have moved away from the other initial political opinionrelative to her own initial position. Missing if the two individuals havethe same initial political opinion. 0 otherwise.
Strong Divergence1 if the two individuals have both moved away from each others initialpolitical position relative to their own initial position, 0 otherwise.
Weak Divergence1 if the two individuals have not moved towards each others politicalposition relative to their own initial position, 0 otherwise.
Any of the above mentioned variables is missing whenever at least one of the two individuals in the couple did notanswer to the related question in the survey.
Table A1: Description of Variables - Dyadic Data
Variable Description
Left-Left1 if both individuals in the couple have an initial political opinion ofless than 4, 0 otherwise.
Left-Center1 if one individual has an initial political opinion of less than 4 andthe other has an initial political opinion of more than 3 and less than8, 0 otherwise.
Left-Right1 if one individual has an initial political opinion of less than 4 andthe other has an initial political opinion of more than 7, 0 otherwise.
Center-Center1 if both individuals in the couple have an initial political opinion ofmore than 3 and less than 8, 0 otherwise.
Center-Right1 if one individual has an initial political opinion of more than 3 andless than 8 and the other has an initial political opinion of more than7, 0 otherwise.
Right-Right1 if both individuals in the couple have an initial political opinion ofmore than 7, 0 otherwise.
Fr. Strength 11 1 if at least one of the two individual has named the other as one ofher friends and has stated that their friendship is at least as intense asa ”mere relationship”, 0 otherwise.
Fr. Strength 21 1 if at least one of the two individual has named the other as one ofher friends and has stated that their friendship is at least as intense asa ”friendship link”, 0 otherwise.
Fr. Strength 31 1 if at least one of the two individual has named the other as one ofher friends and has stated that their friendship is at least as intense asa ”close friendship”, 0 otherwise.
Fr. Strength 41 1 if at least one of the two individual has named the other as one ofher friends and has stated that their friendship is at least as intense asa ”very close friendship”, 0 otherwise.
Shortest Path Shortest path between the two individuals.
1st vs 2nd order only Equal to the shortest path if this is either 1 or 2, missing otherwise.
2nd vs 3rd order only Equal to the shortest path if this is either 2 or 3, missing otherwise.
3rd vs more order onlyEqual to the shortest path if this is either 3 or more, missingotherwise.
Difference in Differences in PoliticalOpinion
”Difference in Political Opinion (March 2014)” minus Difference inPolitical Opinion (August 2013)
d. Importance of Sc Po Degree0 if the two individuals have the same (qualitative) opinion on theimportance of a degree from Sciences Po as a determinant of success,1 otherwise.
d. Importance of Individual Network0 if the two individuals have the same (qualitative) opinion on theimportance of a person’s individual network as a determinant ofsuccess, 1 otherwise.
d. Importance of Individual Effort0 if the two individuals have the same (qualitative) opinion on theimportance of a person’s individual effort as a determinant of success,1 otherwise.
d. Importance of Sc Po Network0 if the two individuals have the same (qualitative) opinion on theimportance of Sciences Po network as a determinant of success, 1otherwise.
d. Importance of Family Network0 if the two individuals have the same (qualitative) opinion on theimportance of family network as a determinant of success, 1 otherwise.
Any of the above mentioned variables is missing whenever at least one of the two individuals in the couple did notanswer to the related question in the survey.
Table A1: Description of Variables - Dyadic Data
Variable Description
Same Gender 1 if the two individuals are of the same gender, 0 otherwise.
Same Female 1 if the two individuals are both female, 0 otherwise.
Same Nationality 1 if the two individuals share a common nationality, 0 otherwise.
Same French 1 if the two individuals are both french, 0 otherwise.
Same Double French Nationality1 if the two individuals are both french and they both have a secondnationality.
Same Admission1 if the two individuals have been admitted through the sameadmission procedure, 0 otherwise.
Same Priority Admission1 if the two individuals have both been admitted through the priorityadmission procedure, 0 otherwise.
Same Department of High School1 if the two individuals have completed their high school diploma inthe same french department, 0 otherwise.
Same Region of High School1 if the two individuals have completed their high school diploma inthe same french region, 0 otherwise.
Same High School Major1 if the two individuals have a high school diploma with the samemajor, 0 otherwise.
Difference in Tuition FeesAbsolute difference in tuition fees among the couple (proxy for familyincome).
Same No Fees 1 if both individuals do not pay tuition fees, 0 otherwise.
Same Parents’ Profession1 if at least one of an individual’s parents has common profession withthe parents of the other individual, 0 otherwise.
Same ZIP code 1 if the two individuals live in the same ZIP code area, 0 otherwise.
Same Program 1 if the two individuals are enrolled in the same program, 0 otherwise.
Any of the above mentioned variables is missing whenever at least one of the two individuals in the couple did notanswer to the related question in the survey.
Table A2: Convergence of Political Opinion - Alternative Specification
Dependent Variable:Difference in Differences in Political Opinion
(March 2014 - August 2013)
(1) (2) (3) (4)
IV IV IV IV
Friendship -0.841*** -0.878*** -0.904*** -0.917***(0.275) (0.282) (0.293) (0.326)
Same Tutorial Group - - 0.303*** 0.259(0.0997) (0.189)
Friendship * Same Tutorial Group 0.131(0.613)
Observations 54,615 54,615 54,615 54,615IV S. Int. Gr. S. Int. Gr. S. Int. Gr. S. Int. Gr.Controls Yes Yes Yes YesDouble Group Clust. Yes Yes Yes YesWeakIV test stat. 114.2 113.2 108.9 20.06
Notes: Standard errors are clustered at the group of individual 1 and at the group ofindividual 2 level. The sample used is the benchmark sample described in the footnote toTable 2. Weak IV stat reports the Kleibergen-Paap cluster-robust statistic, distributed as aChi-squared under the null hypothesis of weak identification.
Table A3: Determinant of Success
Dependent Variables:d. Importance of
Sc Po Degree Individual Network Individual Effort Sc Po Network Family Network
(1) (2) (3) (4) (5)
Friendship -0.0867 0.0283 0.0893 -0.0764 0.131(0.102) (0.116) (0.0548) (0.114) (0.0837)
Observations 54,285 54,285 54,285 54,285 54,285IV Yes Yes Yes Yes YesControls Yes Yes Yes Yes YesDouble Group Clust. Yes Yes Yes Yes YesWeakIV test stat. 115.9 115.9 115.9 115.9 115.9
Notes: The dependent dummy variables are equal to 0 if the characteristic is the same for the two individuals in thepair and 1 if the characteristic is not the same. Standard errors are clustered at the group of individual 1 and at thegroup of individual 2 level. The sample used is the benchmark sample described in the footnote to Table 2. Weak IVstat reports the Kleibergen-Paap cluster-robust statistic, distributed as a Chi-squared under the null hypothesis of weakidentification.
Table A4 - Panel A: Convergence Within ”Political Segment”
Dependent Variables: Difference in Political Opinion (March 2014)
Additional Controls: - - - I I I
(1) (2) (3) (4) (5) (6)
Friendship -0.490* -0.515** - -0.537** -0.559** -(0.252) (0.233) (0.270) (0.245)
Friendship * Same Tutorial Group -0.262 - - -0.231 - -(0.463) (0.489)
Initial Diff. in Political 0.486*** 0.486*** 0.487*** 0.508*** 0.508*** 0.508***Opinion (August 2013) (0.0400) (0.0400) (0.0400) (0.0626) (0.0626) (0.0626)Left - Left 0.862*** 0.863*** 0.866*** - - -
(0.191) (0.190) (0.190)Left - Center 0.621*** 0.621*** 0.622*** - - -
(0.228) (0.228) (0.228)Left -Right 1.011*** 1.012*** 1.011*** - - -
(0.309) (0.308) (0.308)Center - Center 0.483*** 0.484*** 0.486*** - - -
(0.181) (0.181) (0.180)Center - Right 0.511** 0.511** 0.511** - - -
(0.226) (0.225) (0.225)Right - Right 0.471* 0.472* 0.467* - - -
(0.251) (0.250) (0.251)Same Tutorial Group 0.256 0.168** -0.0141 0.251 0.174** -0.0239
(0.161) (0.0840) (0.0314) (0.159) (0.0873) (0.0379)Same Integration Group - - -0.0824** - - -0.0895**
(0.0357) (0.0369)
Observations 54,615 54,615 54,615 54,615 54,615 54,615IV Yes Yes Yes Yes Yes YesControls Yes Yes Yes Yes Yes YesDouble Group Clust. Yes Yes Yes Yes Yes YesWeakIV test stat. 20.10 109.3 - 20.26 110 -
I - Include a set of dummies (not shown) controlling for the political segments the two individuals belong to.The dummies are constructed in the same way as the ones in column 1, 2 and 3, but the definition of politicalsegment is finer. We segment the range of political opinions into five bins: [1,2], [3,4], [5,6], [7,8], [9,10]. Wethen construct a set of dummies in the same way as for the ones used in columns 1, 2 and 3 (see Table A1for more details).Notes: Standard errors are clustered at the group of individual 1 and at the group of individual 2 level.The sample used is the benchmark sample described in the footnote to Table 2. Weak IV stat reportsthe Kleibergen-Paap cluster-robust statistic, distributed as a Chi-squared under the null hypothesis of weakidentification.
Table A4 - Panel B: Convergence - Different Sub-samples
Dependent Variables: Difference in Political Opinion (March 2014)
Conditions: Left-Left Left-Center Left-Right Center-Center Center-Right Right-Right
(1) (2) (3) (4) (5) (6)
Friendship -2.566** -0.491 -0.466 -0.267 -0.738* -0.0124(1.191) (0.473) (1.078) (0.375) (0.406) (0.275)
Initial Diff. in Political 0.259** 0.547*** 0.789*** 0.345*** 0.579*** 0.392*Opinion (August 2013) (0.112) (0.0418) (0.133) (0.0492) (0.0559) (0.215)Same Tutorial Group 0.959** 0.180 0.0802 0.0547 0.277 0.0752
(0.398) (0.157) (0.320) (0.156) (0.215) (0.177)
Observations 2,628 15,622 3,212 22,791 9,416 946IV Yes Yes Yes Yes Yes YesControls Yes Yes Yes Yes Yes YesDouble Group Clust. Yes Yes Yes Yes Yes YesWeakIV test stat. 4.594 45.25 18.59 45.99 35.40 25.92
Notes: Standard errors are clustered at the group of individual 1 and at the group of individual 2 level. The sampleused is the benchmark sample described in the footnote to Table 2. Weak IV stat reports the Kleibergen-Paapcluster-robust statistic, distributed as a Chi-squared under the null hypothesis of weak identification.
Table A6 - Panel A: Convergence of Political and Associative Activities - Conditional on Similar Initial Opinion
Dependent Variables:d. Enrolment in d. Enrolment in d. Enrolment in d. Enrolment in d. Enrolment in the
Student Associations Same Association* Same Association a Political Party Same Political Party*
Condition: Diff. in Initial Political Opinion ≤ 1
(1) (2) (3) (4) (5)
Friendship 0.279* -0.227 -0.0786 -0.0436 -0.173(0.164) (0.156) (0.0684) (0.0985) (0.328)
Political Enrollment - - - 0.449*** -(August 2013) (0.0769)c. Group -0.138* 0.0273 -0.00142 0.0134 0.109
(0.0729) (0.0602) (0.0228) (0.0386) (0.157)
Observations 19,073 9,727 21,949 21,274 305IV Yes Yes Yes Yes YesControls Yes Yes Yes Yes YesDouble Group Clust. Yes Yes Yes Yes YesWeakIV test stat. 51.74 32.44 58.10 59.98 3.295
* conditional on the individuals in the couple being enrolled in some association/political party.Notes: The dependent dummy variables are equal to 0 if the characteristic is the same for the two individuals in the pair and 1 if thecharacteristic is not the same. Standard errors are clustered at the group of individual 1 and at the group of individual 2 level. Thesample used is the benchmark sample described in the footnote to Table 2. Weak IV stat reports the Kleibergen-Paap cluster-robuststatistic, distributed as a Chi-squared under the null hypothesis of weak identification.
Table A6 - Panel B: Convergence of Political and Associative Activities - Conditional on Different Initial Opinion
Dependent Variables:d. Enrolment in d. Enrolment in d. Enrolment in d. Enrolment in d. Enrolment in the
Student Associations Same Association* Same Association a Political Party Same Political Party*
Condition: Diff. in Initial Political Opinion > 1
(1) (2) (3) (4) (5)
Friendship -0.139 -0.140 -0.0752* -0.0362 0.682(0.139) (0.103) (0.0443) (0.0883) (0.435)
Political Enrollment - - - 0.480*** -(August 2013) (0.0749)c. Group 0.0586 -0.00256 0.0136 0.0120 -0.116
(0.0518) (0.0421) (0.0183) (0.0354) (0.150)
Observations 28,822 14,583 32,666 31,376 475IV Yes Yes Yes Yes YesControls Yes Yes Yes Yes YesDouble Group Clust. Yes Yes Yes Yes YesWeakIV test stat. 71.76 42.66 92.03 84.86 1.475
* conditional on the individuals in the couple being enrolled in some association/political party.Notes: The dependent dummy variables are equal to 0 if the characteristic is the same for the two individuals in the pair and 1 if thecharacteristic is not the same. Standard errors are clustered at the group of individual 1 and at the group of individual 2 level. Thesample used is the benchmark sample described in the footnote to Table 2. Weak IV stat reports the Kleibergen-Paap cluster-robuststatistic, distributed as a Chi-squared under the null hypothesis of weak identification.
Table A7 - Panel A: Convergence on Associative Activities - Conditional on Similar Initial Opinion
Dependent Variable:d. Enrolment in
Same Association Associations of Different AssociationsSame Type of Same Type
Type of Association:Sports Strictly Political or Political Related to Related to Political Political
Student Union One’s Identity Policy Issues
Condition: Diff. in Initial Political Opinion ≤ 1
(1) (2) (3) (4) (5) (6) (7)
Friendship -0.0758 0.138** 0.262** 0.462** 0.198 0.0432 -0.0388***(0.121) (0.0609) (0.114) (0.190) (0.176) (0.0407) (0.0119)
Observations 21,949 9,727 9,727 9,727 9,727 9,727 9,727IV Yes Yes Yes Yes Yes Yes YesControls Yes Yes Yes Yes Yes Yes YesDouble Group Clust. Yes Yes Yes Yes Yes Yes YesWeakIV test stat. 58.10 32.44 32.44 32.44 32.44 32.44 32.44
Notes: The dependent dummy variables are equal to 0 if the characteristic is the same for the two individuals in the pair and 1 if the characteristicis not the same. Standard errors are clustered at the group of individual 1 and at the group of individual 2 level. The sample used is thebenchmark sample described in the footnote to Table 2. Weak IV stat reports the Kleibergen-Paap cluster-robust statistic, distributed as aChi-squared under the null hypothesis of weak identification.
Table A7 - Panel B: Convergence on Associative Activities - Conditional on Different Initial Opinion
Dependent Variable:d. Enrolment in
Same Association Associations of Different AssociationsSame Type of Same Type
Type of Association:Sports Strictly Political or Political Related to Related to Political Political
Student Union One’s Identity Policy Issues
Condition: Diff. in Initial Political Opinion > 1
(1) (2) (3) (4) (5) (6) (7)
Friendship 0.0489 -0.0108 0.0500 -0.0727 -0.123 -0.0251*** 0.00578(0.143) (0.0298) (0.0802) (0.163) (0.146) (0.00893) (0.0355)
Observations 32,666 14,583 14,583 14,583 14,583 14,583 14,583IV Yes Yes Yes Yes Yes Yes YesControls Yes Yes Yes Yes Yes Yes YesDouble Group Clust. Yes Yes Yes Yes Yes Yes YesWeakIV test stat. 92.03 42.66 42.66 42.66 42.66 42.66 42.66
Notes: The dependent dummy variables are equal to 0 if the characteristic is the same for the two individuals in the pair and 1 if the characteristicis not the same. Standard errors are clustered at the group of individual 1 and at the group of individual 2 level. The sample used is thebenchmark sample described in the footnote to Table 2. Weak IV stat reports the Kleibergen-Paap cluster-robust statistic, distributed as aChi-squared under the null hypothesis of weak identification.
Table A8: Robustness: Excluding Names Starting with a Given Alphabet Letter
Dependent Variable: Difference in Political Opinion
Excluding: A B C D E F G H I
Friendship -0.601** -0.492* -0.387 -0.869*** -0.592*** -0.569** -0.460** -0.558** -0.614**(0.252) (0.284) (0.301) (0.283) (0.226) (0.237) (0.228) (0.241) (0.249)
Observations 50,403 40,186 46,056 45,150 53,628 50,721 48,205 49,770 53,301
Excluding: J K L M N O P Q R
Friendship -0.561*** -0.516** -0.349* -0.423* -0.547** -0.641** -0.616*** -0.573** -0.590**(0.213) (0.239) (0.205) (0.232) (0.235) (0.257) (0.230) (0.233) (0.240)
Observations 51,681 53,628 43,071 46,056 53,956 53,301 50,086 54,285 49,141
Excluding: S T U V W X Y Z -
Friendship -0.561** -0.540** -0.560** -0.488* -0.551** -0.553** -0.560** -0.547** -(0.243) (0.244) (0.243) (0.251) (0.242) (0.237) (0.244) (0.240)
Observations 48,205 50,403 54,285 52,003 54,285 54,615 53,956 54,285 -
Notes: Standard errors are clustered at the group of individual 1 and at the group of individual 2 level. The sample used isthe benchmark sample described in the footnote to Table 2. Weak IV stat reports the Kleibergen-Paap cluster-robust statistic,distributed as a Chi-squared under the null hypothesis of weak identification.
Figure A1: Distribution of Common Group Fixed Effects
Notes: The graph shows kernel density estimates of the distribution of common groupfixed effects as estimated in specification (9), Table 4.
Figure A2: Individual Political Opinion - Actual, Recalled and Over Time
Notes: The graph shows kernel density estimates of the distribution of individualpolitical opinion. The table makes use of the individual linked 2014-2015 data-set.The sample includes only those individuals present both in the 2014 and in the 2015survey and for which the variables ”political opinion in 2015”, ”actual political opinionin 2014” and ”recalled political opinion in 2014” are not missing.
Formulaire de consentement
Nous vous invitons à participer à une étude sur les attitudes et opinions des étudiants de
Sciences Po.
Ces informations sont collectées uniquement afin de réaliser un projet scientifique. Elles ne
seront pas utilisées pour des buts administratifs et lucratifs. Vos réponses resteront
strictement anonymes. Dès lors que les données auront été collectées, vos nom et prénom
seront remplacés par un code ne permettant en aucun cas de vous identifier. Les personnes
utilisant ces données n’auront donc à aucun moment accès à votre identité réelle.
Ce questionnaire est administré dans le cadre d’une recherche scientifique financée par la
commission européenne (ERC Starting Grant «Trust N° 240923 » de Yann Algan) et validé
par le comité d’éthique de la commission européenne sur la protection et l’anonymat des
données. En répondant honnêtement et exhaustivement aux questions posées, vous permettrez
de faire progresser les connaissances scientifiques.
La participation à ce questionnaire ne vous prendra qu'une vingtaine de minutes.
La participation à cette étude donne lieu à une gratification et vous offre la possibilité de
participer à d’autres études dans le futur, qui donneront lieu elles- aussi à une
gratification.
Vous n’êtes soumis à aucune obligation de participer à cette étude et, si vous changez d’avis,
vous pouvez à tout moment arrêter votre participation. Vous ne recevrez dans ce cas aucun
paiement.
J'ai pris connaissance des conditions d'accès et des modalités de l'étude et les accepte.
Labeds - SciencesPo. 2014
Introduction : votre réseau social
Attention ! Certaines questions sont proches, mais requièrent des réponses
différentes de votre part. Il est important de lire les questions de manière
attentive !
Nous allons vous demander maintenant quels étaient vos amis durant votre scolarité à
Sciences Po.
En participant à cette étude vous pouvez recevoir une gratification : 50 mini-iPad sont à
gagner !
Attention ! Répondez fidèlement à ce questionnaire : votre chance de gagner augmentera en
fonction de la véracité de vos réponses, croisées avec les réponses de vos camarades. Ce
croisement est effectué par ordinateur, les personnes que vous citerez ne pourront à aucun
moment avoir accès à vos réponses. Réciproquement, vous ne pourrez à aucun moment savoir
qui vous a cité.
Il n’y a aucune façon de déduire les réponses des autres participants à l’issue de cette enquête.
Nous vous poserons plusieurs fois des questions identiques, mais qui portent
sur des périodes différentes. Vérifiez donc bien à quelle période correspond la
question.
Nous vous demandons de répondre à ce questionnaire seul et sans parler de
vos réponses à vos amis. Ce questionnaire est en effet strictement personnel. Nous vous
demandons également de répondre à ce questionnaire en une seule fois, et de la manière la
plus spontanée possible.
Labeds - SciencesPo. 2014
Q1.a. Citez vos amis (jusqu'à 10 maximum) à la fin de la première année du
premier cycle (en juin 2010) parmi les étudiants de Sciences Po dans votre
promotion d’entrée.
Exemple : J’ai rencontré Z en septembre 2009 à SciencesPo., et je suis devenu son ami. Je
cite son nom. J’ai rencontré W en août 2010. Je suis devenu son ami, mais je ne cite pas son
nom dans cette partie.
Je n'ai aucun ami(e)s.
Q1.b. Veuillez compléter le tableau ci-dessous pour chacun de vos amis :
Si vous et votre ami vous citez mutuellement, vous recevrez tous les deux un jeton
supplémentaire. De plus, si vous choisissez des réponses compatibles à la question sur la
situation dans laquelle vous vous êtes rencontrés, et à celle concernant le temps passé
ensemble, vous recevrez tous les deux un jeton supplémentaire.
Exemple : J’ai rencontré mon ami Z en septembre 2009 dans une association de Sciences Po.,
et pendant l’année universitaire 2009-2010 nous avons passé en moyenne moins de 30
minutes ensemble chaque semaine. Je cite son nom, et il me cite aussi dans ce questionnaire.
Nous gagnons chacun un jeton. Nous choisissons les réponses « Dans une association de
Sciences Po. » et « Moins de 30 minutes ». Nous gagnons chacun un jeton de plus.
Ami
Comment
avez-vous
rencontré
cet ami ?
Pensez-vous que
cette personne
vous citera dans le
cadre de ce
questionnaire ?
Indiquez le
temps passé en
moyenne
chaque semaine
pendant l’année
scolaire 2009-
2010 avec cette
personne (en
dehors des
heures de
cours) :
Quelle activité
effectuiez vous
principalement avec
cet ami en dehors des
cours (activité
académique, loisirs et
sports, associations
syndicales ou
politiques…)?
ABDELKARIM
Fatima Ezzahra Oui Non
ABOU-
JAOUDE Marie Oui Non
Labeds - SciencesPo. 2014
NB : Ne peut-on pas mettre des noms bidon ?
Q2.a. Citez vos amis parmi les étudiants de Sciences Po (jusqu'à 10
maximum) dans votre promotion d’entrée aujourd’hui (en janvier 2014),
dans l’ordre d’importance.
Exemple : J’ai rencontré W en août 2011. Je suis devenu(e) son ami(e), et je cite son nom
dans cette partie.
Je n'ai aucun ami(e)s.
Labeds - SciencesPo. 2014
Q2.b. Veuillez compléter le tableau ci-dessous pour chacun de vos amis :
Si vous et votre ami vous citez mutuellement, vous recevrez tous les deux un jeton
supplémentaire. De plus, si vous choisissez des réponses compatibles à la question sur la
situation dans laquelle vous vous êtes rencontrés, et à celle concernant le temps passé
ensemble, vous recevrez tous les deux un jeton supplémentaire.
Exemple : J’ai rencontré mon ami Z en septembre 2011 dans une association de Sciences Po,
et pendant l’année scolaire 2012-2013 nous avons passé en moyenne moins de 30 minutes
ensemble chaque semaine. Je cite son nom, et il me cite aussi dans ce questionnaire. Nous
gagnons chacun un jeton. Nous choisissons les réponses « Dans une association de Sciences
Po » et « Moins de 30 minutes ». Nous gagnons chacun un jeton de plus.
Ami
Comment
avez-vous
rencontré
cet ami ?
Pensez-vous que
cette personne vous
citera dans le cadre
de ce
questionnaire ?
Indiquez le temps
passé en moyenne
chaque semaine
pendant l’année
scolaire 2012-2013
avec cette
personne (en
dehors des heures
de cours) :
Quelle activité effectuiez
vous principalement avec
cet ami en dehors des
cours (activité
académique, loisirs et
sports, associations
syndicales ou
politiques…)?
ADANT
Juliette Oui Non
ACHART
Isabelle Oui Non
ABESS
Karima Oui Non
Q3 : Avez-vous été membre actif d’organisations de Sciences Po. ?
Ces organisations peuvent être les associations reconnues par Sciences Po, les partis
politiques actifs à Sciences Po, les syndicats étudiants, le BDE, le BDA, l’AS...
a) Citez au plus 5 organisations dans lesquelles vous avez été le plus actif
depuis septembre 2012.
b) Citez au plus 5 organisations dans lesquelles vous avez été le plus actif
pendant la première année du premier cycle (2009-2010).
Q4.a : Aujourd’hui, comment vous situez-vous en terme d’opinion politique
sur une échelle de 1 à 10 où 1 correspond à l’extrême gauche, 5-6, au centre
gauche / centre droit, et 10 l’extrême droite ?
1 2 3 4 5 6 7 8 9 10
Q4.b : En juillet 2012, comment vous situiez-vous en terme d’opinion
politique
sur une échelle de 1 à 10 où 1 correspond à l’extrême gauche, 5-6, au centre
gauche / centre droit, et 10 l’extrême droite ?
1 2 3 4 5 6 7 8 9 10
Q4.c : A la fin de la première année du premier cycle (juin 2010), comment
vous situiez-vous en terme d’opinion politique
sur une échelle de 1 à 10 où 1 correspond à l’extrême gauche, 5-6, au centre
gauche / centre droit, et 10 l’extrême droite ?
1 2 3 4 5 6 7 8 9 10
Q4.d : En août 2009, juste avant votre entrée à Sciences Po comment vous
situiez-vous en terme d’opinion politique
sur une échelle de 1 à 10 où 1 correspond à l’extrême gauche, 5-6, au centre
gauche / centre droit, et 10 l’extrême droite ?
1 2 3 4 5 6 7 8 9 10
Q5.a Avez-vous été membre d’un parti politique dans le passé ?
Oui Non
Lesquels :
Parti 1 :
Période d’adhésion :
Parti 2 :
Période d’adhésion :
Q5.b Etes-vous membre aujourd’hui d’un parti politique ?
Oui Non
Lesquels :
Parti :
Depuis quelle date ?
Q5.c Pour qui avez-vous voté au premier tour de l’élection présidentielle de
2012 :
Q6.a. D’une manière générale, diriez-vous que l’on peut faire confiance à la
plupart des gens ou que l’on n’est jamais assez prudent quand on a affaire
aux autres ?
Q6.b Nous allons maintenant parler des gens qui viennent de pays moins
développés pour travailler ici. Qu'est-ce que le gouvernement devrait faire
selon vous ?
Q6.c. Pensez-vous que les facteurs suivants auront de l’influence sur votre
succès professionnel dans le futur :
votre diplôme
les connaissances et compétences acquises au cours de vos études
vos efforts personnels
votre réseau formé à Sciences Po
votre réseau familial
Q6.d Si vous avez fait un stage cette année, comment avez-vous eu
connaissance de ce stage :
Réseau de connaissance à Sciences Po
Réseau familial
Offre sur les sites de Sciences Po (Sciences Po Avenir, site de master)
Recherche personnelle
Quelle note sur 10 donneriez-vous à l’organisation des enseignements autour des triplettes en
1A ?
Quels en sont les avantages ?
Facilite l’intégration à Sciences Po
Se faire des amis plus facilement
Se faire des amis plus divers
Permet le travail en commun
Améliore la qualité du travail en commun
Autre : Précisez :
Fin du questionnaire
Merci de votre participation.
Labeds - SciencesPo. 2014