You are on page 1of 53

Peer Effects on Mental Health: Evidence from

Random Assignment into Classrooms*

Anwen Zhang†

This version: 30th November 2018


[Click for the latest draft]

Abstract
Adolescent mental health has wide-ranging and long-lasting socio-economic
consequences. Existing evidence based on observational data of friendship net-
works points to a positive correlation between an individual’s mental health and
her peers’, but concerns remain whether this link is causal. I study whether
there exist peers effects on mental health in the classroom, by exploiting vari-
ations in peer composition generated by assignment rules in junior secondary
schools in China, where students are randomly or evenly grouped into class-
rooms. In general I find no evidence of overall peer effects, and weak evidence of
context-specific peer effects when taking into account heterogeneity and nonlin-
earity. From a policy perspective, the general weakness of peer effects suggests
that group-based interventions at the classroom level probably would not gener-
ate large positive externality.

JEL Codes: I12, I21


Keywords: peer effects, mental health, classroom assignment

* I am grateful to seminar participants at LSE for helpful comments. This research uses data from
the China Education Panel Survey (CEPS). I thank the National Survey Research Center at Renmin
University of China for making the data available.

London School of Economics and Political Science; Email: a.zhang6@lse.ac.uk.
1 Introduction
Mental illness imposes a substantial cost to individuals and the society. It is the leading
cause of disability (Whiteford et al., 2013). Estimates put economic costs of mental
disorders at $2.5 trillion in 2010 (Bloom et al., 2012; Trautmann et al., 2016), compa-
rable to the size of the GDP of France. Adolescence is a key period for the onset of
mental disorders. Half of all mental health conditions are developed by the age of 14.
Worldwide 10 to 20% of children and adolescents suffer from a disabling mental illness
in any given year (MHF, 2015; WHO, 2003; WHO, 2001). Ill mental health in adoles-
cence can have wide-ranging and long-lasting consequences. It is linked with a broad
spectrum of short- and long-term socio-economic outcomes, from low educational at-
tainment (Cornaglia et al., 2015; Currie and Stabile, 2009; Fletcher, 2008, 2010), to
adverse labour market outcomes (Banerjee et al., 2015; Chatterji et al., 2007; Fletcher,
2013, 2014) and more crimes (Anderson et al., 2015). At school age, the classroom
context is one of the major social environments that shape individual behaviours, and
a key area for education policy and school-based health policy. In this paper, I study
the role of classroom peers in determining adolescent mental health.
Measuring peer effects carries significant policy implications. Peer effects on men-
tal health indicate an externality in health production, which generates prices that do
not reflect the true social benefits and costs of interventions and healthcare, leading
to inefficiencies in the allocation of such resources. If mental disorders are socially
contagious, this provides potential economic justification for health interventions, as
the benefits of these interventions will also spill over from targeted individuals to their
peers without incurring additional costs. A clean measurement of the magnitude of
peer effects is thus crucial for the cost-benefit analysis of such interventions.
There is a great need for rigorous evidence to inform health policy in school set-
tings. For instance, some mental health policy proposals such as peer support schemes
directly involve peers and require a clear understanding of its consequences on those
who are in need of support and their peers who provide the support. There is an
abundance of evidence on peer effects on educational outcomes dating back decades;1
in comparison, research on peers effects on health behaviours and outcomes is still de-
1
See Epple and Romano (2011) and Sacerdote (2011) for reviews.

1
veloping.2 A scant literature examines peer effects on various specific aspects of men-
tal health and wellbeing, including depression and mood (Hill et al., 2015; Rosenquist
et al., 2011), suicide (Mueller and Abrutyn, 2015), happiness (Fowler and Christakis,
2008a), mental health service use (Golberstein et al., 2016), and overall mental health
(Eisenberg et al., 2013).
Much of this strand of literature focuses on friendship networks (Hill et al., 2015;
Rosenquist et al., 2011; Mueller and Abrutyn, 2015; Fowler and Christakis, 2008a),
with two notable exceptions (Eisenberg et al., 2013; Golberstein et al., 2016) which
exploit natural experiments of college roommate assignments. Studies based on friend-
ship networks in general report findings that imply large peer effects. For instance,
Rosenquist et al. (2011) find that having a depressed friend increases the probability
of oneself being depressed by 118%. While these studies acknowledge the limitations
that individuals are not randomized into social networks, it remains unclear how large
a bias this creates. Besides endogenous friendship formation, unobservable common
environments facing a group of friends are another concern. Much debate is devoted to
how importance these concerns are and how to address these issues methodologically.3
Furthermore, from a policy perspective, it is unclear how intervention programs can
be designed at friendship levels. Rather, it is more common for policies to manipu-
late group compositions, or target certain groups or subgroups which can be clearly
defined. Understanding the interactions within these groups or subgroups is a pre-
requisite for such policies. To the best of my knowledge, this paper presents the first
evidence of peer effects on mental health at the classroom level.
Why peer effects might exist in mental health formation? Existing literature on
2
Existing literature on the latter has largely focused on risky behaviours (for instance, see Clark
and Lohéac, 2007; Eisenberg et al., 2014; Gaviria and Raphael, 2001) and obesity (e.g. Christakis and
Fowler, 2007; Cohen-Cole and Fletcher, 2008a,b; Fowler and Christakis, 2008b; Trogdon et al., 2008;
Yakusheva et al., 2014).
3
A contentious round of debate over peer effects on obesity within friendship networks highlights
this point. In a study that attracted much academic and media attention, Christakis and Fowler (2007)
report finding large peer effects among friends based on longitudinal data. Using a different dataset and
following Christakis and Fowler’s (2007) methodology, Cohen-Cole and Fletcher (2008b) successfully
replicate Christakis and Fowler’s (2007) results. However, Cohen-Cole and Fletcher (2008b) argue that
the effects become statistically indistinguishable from zero once common environmental confounders
and self-selection of friends are controlled for. In a related study, Cohen-Cole and Fletcher (2008a) fur-
ther argue that failure to adequately control for environmental confounders leads to finding implausible
peer effects on acne, headaches, and height among friends.

2
child development and psychology proposes a number of theories.4 First, negative
affect might get reinforced through problem-sharing. Co-rumination, or excessively
discussing and revisiting problems with a friend, while bringing friends closer, leads to
increased depression and anxiety, which in turn leads to more co-rumination (Rose,
2002; Rose et al., 2007). A second process relates to empathizing. Placing oneself in
someone else’s shoes could lead to experiencing same stress and emotions (Hatfield
et al., 1993), or feeling compelled to offer support (Joiner and Katz, 2006). Third,
social comparison, by benchmarking oneself against others, may cause stress or relief,
depending on one’s relative position in a social network (Exline and Lobel, 1997).
In light of these potential transmission mechanisms, the objective of this paper is
to identify the peer effects on mental health production at the classroom level, in the
setting of junior secondary schools in China. Classrooms are a social environment of
particular relevance to the Chinese educational setting, where most students stay in the
same classroom throughout the years in each stage of schooling. I exploit classroom
assignment rules where many schools randomly or evenly group a new cohort into
classrooms. Peer effects on mental health are identified based on within-school-grade
variation in peer composition at the beginning of the school stage, combined with
longitudinal data and rich information of teacher characteristics.
Under this empirical framework, I find little evidence of peer effects in overall
mental health based on linear-in-means models, and some evidence of heterogeneity
and nonlinearity for specific contexts. I show that the weak peer effects in overall
mental health is not due to weak data or improper research design; further, it is con-
sistent across a number of overall mental health measures. Allowing for heterogeneity
and nonlinearity, I find some evidence of small to modest peer effects for and within
certain socio-economic subgroups; further, there is also evidence that peer effects are
nonlinear and take a shallow reverse U-shape, with larger effects coming from peers
in the second quintile of the mental health distribution. The findings are consistent
with a small literature on peer effects in mental health that exploits other natural
experiments, and deviate from some observational studies that report large peer cor-
relations in mental health. From a policy perspective, the general weakness of peer
effects suggests that group-based interventions at the classroom level probably would
not generate large positive externality.
4
Eisenberg et al. (2013) offers a more thorough review.

3
2 Institutional Background

2.1 The education system in China


China’s education system is composed of three levels: primary, secondary, and post-
secondary. Typically at age 6, children enter primary education which lasts six years.
Secondary education is split into two stages, junior and senior. Junior secondary edu-
cation, also known as middle school education, are split into three years, from Grade 7
to Grade 9 (typically age 12 to 14). Primary and junior secondary education constitute
the nine-year compulsory education in China, which is provided free of charge.5 Af-
ter junior secondary school, students can elect to drop out or enter senior secondary
school, or commonly known as high school, which lasts for another three years (typ-
ically age 15 to 17).
Junior secondary school students are taught the same subjects and do not take
elective courses. Chinese, Maths, and English are three courses that are considered
“core subjects” which receive higher weights in the curriculum.

2.2 School choice and classroom assignment


In principle, junior secondary schools are required by law to accept any student finish-
ing primary school in their respective catchment areas.6 Local education authorities
are not allowed to adopt ability tracking across schools. In practice, over-subscribed
schools do set their own admission criteria, often based on conducting admission tests.
It is also widespread practice to obtain a school place by paying hefty fees (Carman
and Zhang, 2012).
At the beginning of junior secondary education, the school assigns news students
into classrooms. By law, ability tracking is not allowed during the compulsory ed-
ucation stage, but implementation varies across areas and schools. The majority of
schools assign classrooms randomly or on a balanced basis, to ensure comparable aca-
demic abilities across classes at the beginning of the school stage.
Classrooms play a key role in forming a school student’s academic and social net-
5
Free education is achieved on a national scale since 2008.
6
In contrast, it is commonplace for senior secondary schools to set admission criteria based on
admission test scores (Ding and Lehrer, 2007).

4
work, as classrooms are typically fixed for the whole three years of junior secondary
education. There are exceptions, however, where schools may reassign classrooms on
a yearly basis. Students stay in the same classroom, and teachers of different subjects
rotate in delivering lessons of their subject to the whole class. Each classroom has a
designated teacher, banzhuren (homeroom teacher), who is responsible for the overall
performance and welfare of the whole class, providing pastoral care to the students,
and communicating with the parents, on top of teaching their own academic subject.

2.3 Mental health support in schools


At the national level, there is no systematic education about mental health or manda-
tory requirement of providing counselling services to students in schools. Teachers
are usually the first line of institutional support in schools for student with emotional
and behavioral problems, with homeroom teachers taking the major responsibilities.
As non-professionals, they may receive some mental health education training funded
by local education authorities. In our data, roughly a quarter of the teachers have
received some mental health education training.
Schools are advised to include mental health education as part of the curriculum
and set up counselling offices where resources permit, according to the guidelines of
the Ministry of Education (MoE, 2002, 2012). In reality, there are large regional
disparities, particularly between urban and rural areas, in terms of the provision of
school-based mental health education and services (Zhang and Du, 2018). Even in
the schools with professional counsellors, quality and utilization of services vary sub-
stantially across schools. Of all the schools surveyed in our data, 12% report having
no counselling offices, 55% report having counselling offices but facilities need im-
provement, and 33% report having counselling offices that are running well. Use of
school-based and non-school-based mental health services is low among students with
perceived needs for professional support (Wu et al., 2012), suggesting students may
rely on informal support to deal with mental health issues.

5
3 Data

3.1 Data source


This paper uses data drawn from the China Education Panel Survey (CEPS). The
CEPS is a large-scale, nationally-representative, school-based longitudinal survey of
junior secondary school students in China. It is the first such dataset available in
China.
Two waves of the CEPS data have been made available. Wave 1 was conducted
during the 2013-14 academic year, covering around 20,000 students in 438 classes at 112
schools in 28 county-level areas. Following a stratified multi-stage sampling approach,
the CEPS first selected the 112 schools with probabilities proportional to size (PPS).
The next sampling stage involves randomly choosing two classes from Grade 7 and
two from Grade 9 in each sampled school. In small schools with only one or two
classes in Grade 7 or 9, all students in that grade were surveyed. Lastly, all students
present on the survey date in the selected classrooms were sampled, of which 10,279
were in the 7th grade. The second wave took place in the 2014/15 academic year,
following the 10,279 students who were in the 7th grade in the previous wave into the
8th grade. During the follow-up wave, 9,449 students were re-interviewed, 91.9% of
the baseline 7th grade sample.7
The CEPS collected information on the students and their family and school envi-
ronment, including socio-demographic information, academic performance, physical
and mental health status, attitudes and relationships, through questionnaires adminis-
tered to the students, parents, teachers, and school administrators.

3.2 Classroom assignment


During Wave 1 of CEPS, school administrators were asked “what are the standards
for assigning classes for new students in your school”,8 to which the responses include
“entry exam test scores” (20%), “students’ places of origin” (1%), “random or even
7
Out of the 830 students who were not successfully followed in Wave 2, 71% had transferred to a
different school, 15% had dropped out of school, 5% took authorized absence on the day, and the rest
for other reasons.
8
Out of the 109 school administrators who answered this question, 61% were principals, 28% vice
principals, 8% deans, and the rest in other administrative roles.

6
assignment” (85%, or 93 schools), and “other” (5%). Even assignment is a rule where
N students are assigned into k classes (k < N ) in an S fashion based on ranks of test
scores: the first k students are assigned into the k classes in a progressive order by
rank, the next k students in a reverse order, and so on. Based on this information, 183
classes from 93 schools were randomly or evenly formed. At Grade 8, a small number
of schools regrouped part or all of its classes. As a result, students from 30 of the 183
classes had been reassigned to new classes by the time of Wave 2.
The empirical analysis is mainly based on students that were randomly/evenly
grouped into classrooms at Grade 7, and stayed in the same classroom at Grade 8. In
total, there are 153 such classes from 80 schools in the data. Potentially the reassign-
ment rules could also be exploited, as they generate another useful source of variation
in peer composition, provided the reassignment mechanisms can be explicitly con-
trolled for; however, CEPS does not collect information on the new peers when reas-
signment takes place, thus data limitation renders this exploration impossible. Classes
with non-random assignment at Grade 7 can also contribute to identification, as long
as the assignment mechanisms are fully controlled for. This is made difficult by the
lack of details and pre-assignment test scores from the data, but test scores at Grade
7 should serve as a reasonable proxy. This subsample will used for robustness checks
later on.
Keeping the classes that were randomly or evenly grouped at Grade 7 and stayed
grouped at Grade 8 leaves a balanced two-period panel data of 6,381 students. Drop-
ping observations with missing values in key variables reduces the data by 6.5% to
5,970 students observed in both waves. The mean class size of this sample is 39; in
comparison, this figure is 46 in full Wave I sample. Judging by these numbers, 15%
of the students from a given class are not observed due to missing data from sample
attrition and invalid responses.

3.3 Mental health measures


Both waves of CEPS provide measures of mental health, with some differences across
waves, as summarized in Table 1. In Wave 1, adolescents are asked, “how often did you
have the feelings below in the last seven days”, with respect to five symptoms: “blues”,
“depressed”, “unhappy”, “not enjoying life”, and “sad”. The responses are rated on a

7
1–5 scale, with 1 = “always” and 5 = “never”.9 At wave 2, adolescents are asked again,
on a 1–5 scale, “how often did you have the feelings below in the last seven days”, but
with respect to ten symptoms: “blues”, “too depressed to focus”, “unhappy”, “not en-
joying life”, “having no passion”, “nervous”, “excessive worry”, “something bad will
happen”, and “too energetic to concentrate”. Out of these ten questions, four ques-
tions are the same as in Wave 1 (“blues”, “unhappy”, “not enjoying life”, and “sad”),
one similar (“depressed” vs“too depressed to focus”), and the remaining five are newly-
added measures at Wave 2. These ten measures cover various symptoms of depression,
anxiety, and attention deficit hyperactivity disorder (ADHD), but are closely corre-
lated with each other (See Appendix Tables A2). All questions exhibit correlation
coefficients at or above 0.42, with the exception of “too energetic to concentrate”,
whose correlation coefficients with other nine questions range between 0.27–0.37.
For brevity of analysis and presentation, I conduct principal component analysis
(PCA) to reduce multivariate data to fewer dimensions. Based on PCA, I extract
one latent principal component score (PCS) for each wave as the baseline measures
of mental health based on all available mental health questions in each wave. The
extracted PCS based on five available variables on mental health in Wave 1 is termed
MH5PCS; the PCS based on ten variables from Wave 2 is terms as MH10PCS. More
details are available in Appendix A.2. A number of robustness checks are conducted
to assess sensitivity of the results using various measures of mental health: aggregated
raw score based on simple summations over all ten questions (MH10), aggregated raw
score based on simple summations over nine questions (MH9, excluding “too energetic
to concentrate”), aggregated raw score based on the five common or similar questions
(MH5), aggregated raw score based on the four common questions (MH4), and their
respective PCS versions (MH10PCS, MH9PCS, MH5PCS, MH4PCS).
All of the measures mentioned above at the individual level are normalized to
have zero mean and unity standard deviation (SD), such that point estimates can be
interpreted based on SDs. Peer mental health is constructed as the mean of normalized
student mental health within each classroom excluding the student herself. A number
of sub-class peer mental health measures are also constructed in a leave-one-out fashion
based on some socio-economic classification criterion, to assess heterogeneity in peer
9
Original dataset has these coded as 1 = “never” and 5 = “always”. I reverse code the responses so
that higher scores represent good mental health, and lower scores poor mental health.

8
effects. For instance, along gender lines, I consider the roles of same-sex class peers
and opposite-sex class peers. In addition, the class share of peers with various levels of
mental health (for instance, proportion of class peers that fall in the bottom quintile
of the mental health distribution) is used as an extra measure of peer mental health
in the exploration of nonlinear relationships. Variable definitions and descriptions
of other variables are available in Appendix A.1. Summary statistics are available in
Appendix Table A3.

4 Empirical Strategy

4.1 Linear-in-means model


We start with a linear-in-means model of peer effects. Let i denote the student, and −i
her peer group. Student i’s mental health Hi , is described in the following equation:

Hi = α + βH −i + Xi θ + εi (1)

where Xi is a vector of individual characteristics and family ground information, and


εi the error term. [α, β, θ] are parameters to be estimated, where β represents peer
effects.
As highlighted in Manski (1993), there are three main empirical difficulties in es-
timating peer effects. First, individuals sort into groups. Peer groups are rarely not
randomly formed; rather, individuals who are already alike tend to sort into similar
social groups. This is more apparent in friendship networks (McPherson et al., 2001).
Unless sorting into peer groups can be explicitly modeled, this self-selection process
presents an upward bias in the estimation of peer effects. To address this issue, I rely
on natural experiments of classroom assignments to generate exogenous variation in
peer composition. For the main part of analysis, I examine schools that randomly or
evenly assign students to classrooms, and make use of the within-school-grade vari-
ation generated by these assignment rules for identification. These assignment rules
ensure that peer composition at the beginning of the school stage is random, thus cre-
ating an ideal source of exogenous variation for the identification of peer effects. As a
robustness check, I also include non-random assignment schools in part of the analy-

9
sis, and control for a proxy of pre-assignment test scores to account for the fact that
these school often use pre-assignment test scores as an assignment criterion.
Next, individuals in the same social groups are faced with a common environment.
Students in the same school or class are faced with the same institutional environment.
For instance, some schools are equipped with better facilities of on-campus counseling
services, which can be correlated with both the individual’s and her peers’ mental
health. Further, even when classroom assignment is random, teacher assignment and
teacher inputs may not be. A teacher with better awareness of mental health issues
might improve the mental health level of the whole class. Therefore, these group-level
characteristics and shocks that are not captured by the empirical model will present as
another source of upward bias. I take three steps to address this issue. First, I include
school-grade-year fixed effects in the model, so that all commons shocks at this level
are purged. Note that in implementation this is reduced to school fixed effects, as only
one school grade is observed and only outcomes at Wave 2 are examined. Second, at
the classroom level, a rich set of teacher characteristics are included in the estimation
models, to capture common shocks taking place at the classroom level. Third, at the
individual level, I include lagged mental health to proxy for historical inputs of mental
health.
Lastly, there is a simultaneity issue. In the presence of peer effects, when an in-
dividual affects her peers, the peers also affect the individual. Causality can go both
ways as an individual is also their peers’ peer; by affecting peers, individuals eventu-
ally affecting themselves. Using contemporaneous measures for both individual and
peer outcomes will make it impossible to judge the direction of causality. To break
simultaneity, I sidestep this issue by using lagged measures for peer mental health.
These adjustments lead to a liner-in-means model specified as below, applied to the
sample of students who are randomly or evenly assigned into classrooms:

Hicst = αst + βH −ics,t−1 + γHics,t−1 + Xicst δ + TEcst θ + εicst , (2)

where Hicst denotes the mental health of student i in class c at school s during year t,
αst school-year fixed effects, H −ics,t−1 the lagged mental health of student i’s classroom
peers at Grade 7, T E a vector of teacher characteristics, and ε idiosyncratic shocks.
This equation is essentially the reduced form of an instrumental variable (IV)

10
model, where the contemporaneous peer outcome H −icst is instrumented for with
lagged peer outcome H −ics,t−1 , represented by the following equation system:

0
H −icst = αst + β 0 H −ics,t−1 + γ 0 Hics,t−1 + Xicst δ 0 + T Ecst θ 0 + ε0icst , (3)
00
Hicst = αst + β 00 H −icst + γ 00 Hics,t−1 + Xicst δ 00 + T Ecst θ 00 + ε00icst , (4)

where the prime superscript 0 indicate parameters of the first stage equation, and 00
the second stage, in the language of 2-stage least squares (2SLS). The identification
assumptions of these two models are essentially the same: lagged peer mental health
H −ics,t−1 is conditionally mean independent of individual unobservables εicst (or ε00icst
in the IV case), i.e.
E(εicst |H −ics,t−1 , Ω) = E(εicst |Ω),

where Ω = (αst , Hics,t−1 , Xicst , T Ecst ). The difference between model (2) and (4) lies
in the interpretation of the peer effects: β represents the lagged peers effects whereas
β 00 represents the contemporaneous peer effects. In this paper, I follow a reduced-form
specification as in equation (2).
Peer effects can derive from two sources: endogenous effects deriving from peer
outcomes and exogenous effects deriving from peer background characteristics (Man-
ski, 1993). Separately identifying these two effects can be very difficult, as peer char-
acteristics also affect peer outcomes. Separating these two channels would require two
sources of exogenous variation. Due to the difficulty of finding two exogenous varia-
tions at the same time, the peer effects literature has largely stayed silent on this issue
(Sacerdote, 2011). In this paper I do not attempt to separately identify endogenous
and exogenous peer effects, as I rely on one source of exogenous variation.

4.2 Nonlinear models


One limitation with the linear-in-means model is that it masks potential heterogenous
effects, which may depend on an individual’s own characteristics, her position in the
group, and her peers’ position in the group. Interactions may only occur at certain
subgroups of the classroom, or the strength of interactions may be different across
subgroups. While it remains policy relevant to define peer group at the classroom
level, it is more beneficial to understand how interactions between various subgroups

11
might lead to different consequences.
To relax the linear assumption, I try a number of approaches. First, I start with
assessing heterogeneity across socio-economic subgroups, by including an interaction
of a socio-economic indicator and classroom peer mental health in the model. Second,
I distinguish classroom peers of the same socio-economic status from those of different
socio-economic status, to assess heterogeneous peers effects from these two sub-class
groups. Third, I estimate a series of quantile regressions, to allow for heterogeneity
peer effects to vary based on one’s position in the mental health distribution. Fourth,
I follow an full-interaction approach (Hoxby and Weingarth, 2005), to allow for het-
erogenous effects based on a student’s position in the mental health distribution, and
her peers’ position in the distribution. Specifically, I divide the whole mental health
distribution at Wave I into five quintile groups, and include all interaction terms be-
tween an indicator of the student’s quintile group, and the class share of her peers who
fall in a certain quintile group, specified in the following equation:

5 X
5
X k
Hicst = αst + j
βjk I−ics,t−1 · I −ics,t−1 + γHics,t−1 + Xicst δ + T Ecst θ + εicst , (5)
j=1 k=1

where I j is an indicator function for a student’s mental health being in the jth quintile
k
of the mental health distribution, and I is the mean of such indicator for the kth
quintile for her classmates. In this specification, βjk represents the effects of peers who
are in the kth quintile of the mental health distribution at Grade 7, on the students
who are in the ith quintile of the distribution at Grade 7.

5 Results: Linear-in-Means Models

5.1 Balancing test


Before going into the main results, I conduct a series of balancing tests to check that
classes are indeed randomly assigned in the selected sample. Specifically, I regress a
series of pre-determined characteristics on peer mental health (MH5PCS) at Grade 7,
to check whether they are balanced across students who are faced with different com-
positions of peers. These results are presented in Table 2. Panel A presents the results

12
for raw correlations between a student’s background characteristics and her peers’ av-
erage mental health at Grade 7, without any other controls. A number of background
characteristics are shown to be correlated with peer mental health, including cogni-
tive skills, whether living with both parents, and household financial situation. These
correlations hightlight the need for taking into account self-selection and common
environment issues when studying peers effects based on observational data. Panel B
runs the regressions conditional on school fixed effects. The coefficients are reduced
in size or become less significant. Out of the ten background variables, only gender
remains significantly correlated with peer mental health. This suggests that peer gen-
der composition may be important for mental health, and will be taken into account
in later analysis. Other than this, these pre-determined characteristics seem to be well
balanced across students who are faced with various levels of peer mental health.

5.2 Baseline results


Table 3 presents the linear-in-means results. The baseline outcome variable is MH10PCS
at Grade 8. Column (1) presents a model where only the student’s observable back-
ground information is controlled for. This specification finds a strong and significant
correlation between a student’s Grade 8 mental health and her class peers’ Grade 7
mental health, with a point estimate of 0.27. Column (2) proceeds by including school
fixed effects in the model. The change in the peer effects coefficient is substantial. The
point estimate is reduced to 0.02, and insignificant at 0.10 level. This large reduction
in the peer effects estimates demonstrates that common environment at the school
level is the major confounder in the estimation of peer effects in our setting. Column
(3) further controls for the student’s lagged mental health, which serves as a proxy for
observable historical inputs of mental health. Although mental health shows a strong
tendency of autocorrelation over time, this does not affect the peer effects estimate by
much. The coefficient on lagged peer mental health is now at −0.01 and remains indis-
tinguishable from zero. Lastly, column (4) further controls for teacher characteristics.
This addresses a concern that teacher assignment may not be random or balanced even
classroom assignment is. The peer effects estimate changes very little, with point es-
timate now becoming −0.02 and staying insignificant. Despite the little contribution
to estimation, this does not necessarily suggest teachers do not contribute to student

13
mental health. It is likely school fixed effects absorb much of the teacher effects, due
to teacher selection into schools. A second reason could be that there is not enough
variation in teacher characteristics in the data, as there are at most two classrooms
observed in each school.
Taken together, comparison across models suggests that common environment is
by far the most important confounder in our setting. Failing to take this into account
will produce a large upward bias in the estimation of peer effects on mental health.
Once this is accounted for, I do not find evidence of peer effects from the average
classmate.

5.3 Robustness checks


A main concern with the CEPS data is its inconsistent measurement of mental health
across the two waves. The contemporaneous mental health measure is based on 10
questions at Wave 2, whereas the lagged mental health measure is based on 5 questions
at Wave 1. While the two measurements seem to largely pickup depression and anxiety
symptoms, it could be argued that mental health might be measured with error in
either or both waves. To address this issue, I conduct a battery of robustness checks
based on various measurements of mental health.
Table 4 presents the results for these tests. For measuring individual mental health
at Grade 8, a number of variables are used, including MH10PCS, MH10, MH9PCS,
MH9, MH5PCS, and MH5. For measuring individual and peer lagged mental health at
Grade 7, alternative variables MH5, MH4PCS, and MH4 are employed in addition to
the baseline measure MH5PCS. In a total of 12 combinations, the peer effects estimates
are quantitatively similar, ranging between −0.03 and 0.03, and none is significantly
different from zero. Same conclusion can be reached for the coefficient on lagged own
mental health.
Another check for robustness considers including schools with non-random class-
room assignment at Grade 7. Most of these schools arrange classrooms based on some
pre-assignment test scores. Although these are not available in the data, Grade 7 test
scores should serve as a reasonable, albeit noisy, proxy. Thus I include Grade 7 test
scores as an additional control variable, extend the analysis to include non-randomly

14
assigned classes that are not reassigned at Grade 8.10 Results are presented in Table 5, in
a similar manner as the presentation for baseline results as in Table 3. Quantitatively
the results are similar as the baseline results.
These robustness checks point to no evidence of linear peer effects on overall men-
tal health at the classroom level. Beyond non-existence of peer effects, a number of
alternative explanations are consistent with these findings. First, it could be that the
data and research design do not have the power to detect any peer effects. Second, it’s
likely that the linear-in-means models masks heterogeneous and nonlinear relation-
ships of peer-to-peer effects. In the following analyses, I conduct checks to test these
explanations.

5.4 Placebo tests


Previous results finding no evidence of linear-in-means peer effects could be a conse-
quence of weak data and improper research design. For instance, it might be the case
that insignificant results are due to small variation in peer composition after control-
ling for school fixed effects. In light of these concerns, I run a series of placebo tests
using alternative outcomes, to check if potential peer effects can be detected by the
data and model.
I run the same analysis based on the baseline linear-in-means model, except that
the key variables are academic outcomes instead of mental health. Specifically, I test
whether there are peers effects on cognitive skills, and test scores in Chinese, Maths,
and English subjects.11 Cognitive skills are assessed by the survey collectors based on
a standardized test. Test scores are reported for the mid-term exams. Table 6 show
that, based on the same dataset and research design, linear peer effects can be detected
on cognitive skills, Chinese, and Maths. Peer effects on cognitive skills and Chinese
test scores are significant at 0.05 levels or higher, and the size of peer effects are around
0.23. Peer effects on Maths scores are smaller at 0.13, marginally significant at 0.10
level. No peer effects are found for English scores.
These results lend support to the empirical strategy, in that it is capable of detect-
ing peer effects based on linear-in-means models. The evidence of no detected effects
10
Reassigned classes at Grade 8 are not fully observed in Wave 2, thus excluded from the analysis.
11
Cognitive skills are included as a control variable in the mental health specifications, but dropped
from the covariate set in this part of the analysis.

15
on overall mental health suggests that the average classmate does not affect student
mental health. From a policy perspective, group-based mental health interventions
are unlikely to generate multiplier effects.

5.5 Separate aspects of mental health


While the analysis above points to little evidence of peer effects on overall mental
health, I conduct further analysis based on separate aspects of mental health to under-
stand if there is peer spillovers in specific settings. Results for these tests are available in
Table 7, which presents the estimates based on individual symptoms of mental health:
feeling blues, depressed, unhappy, not enjoying life, and sad. These measures are nois-
ier in nature, as they are measured on a shorter scale thus have less variation. The
point estimates for peers effects on these measures are larger and come with larger
standard errors. One measure, feeling depressed, shows marginally significant peer ef-
fects at 0.10 level. The point estimate is 0.23, large compared to that of overall mental
health in previous subsection, but only modest compared to some reports of previous
correlation studies.
Broadly, these findings are consistent with existing literature in this area that ex-
ploit nature experiments. Two existing studies that exploit college roommate assign-
ment rules as a source of exogenous variation in peer composition find overall weak
evidence of social contagion of mental health at the roommate level, but small to mod-
est peer effects in specific measures and settings (Golberstein et al., 2016; Eisenberg
et al., 2013). This offers an interesting contrast to research that studies peer effects
in friendships networks, most of which are in the medical literature and often find
sizable peer effects. While friends and classroom/roommate peers are distinct social
groups, the studies based on natural experiments, this paper included, do highlight
the importance of dealing with confounding unobservable common environments in
such research.

6 Heterogeneity and Nonlinearity


While the linear-in-means offers a straightforward interpretation, it probably is not
the most insightful or interesting model. The model specification masks potential

16
heterogeneity and nonlinearity. Indeed, allowing for peers effects to vary across differ-
ent dimensions can reconcile seemingly contradictory results based on linear-in-means
models (Sacerdote, 2011). In order to gain a fuller picture, this section examines the
heterogeneity and nonlinearity omitted from the linear-in-means model.

6.1 Heterogeneity by student background


First, I take a simple approach by allowing for peer effects to vary based on observable
student background. I add the interaction term between lagged classroom peer mental
health and one of eight indicators of socio-economic background separately: female,
age, ethnic minority, hukou status, financial status, mother’s education, father’s educa-
tion, and alcoholic father. As presented in panel A of Table 8, there is scant evidence
that peer effects vary by a student’s socio-economic background. Out of the eight
interacted specification, only one finds a significant interaction term. Peer effects for
students with a college-educated mother are marginally significantly larger than for
those with a less-educated mother, although the level term (lagged peer mental health)
and interaction term are not jointly significant.
Panel B takes a different approach to assess potential heterogeneous effects by socio-
economic status. Sorting along socio-economic lines suggests that social interactions
take place more often between peers who are alike, thus it is possible that peer effects
may be stronger within a subgroup of classroom peers who share similar character-
istics. In light of this hypothesis, I divide the classroom peers into two subgroups
for each students based on one of the eight socio-economic indicators: one subgroup
sharing the same socio-economic characteristics with the student, and the other who
don’t. For instance, along gender lines, I divide the peer group into same-sex peers
and opposite-sex peers. Again evidence of heterogeneous peer effects is scant along
these lines. The coefficients of the interaction term is only marginally significant in
one out of the eight specifications. Peers who share the status of having a college-
educated father tend to have larger effects. None of the coefficients on cross-group
peers is statistically significant.
Further evidence on heterogeneous peer effects on separate aspects on mental
health along socio-economic lines is presented in Appendix Tables A4–A8. These re-
sults provide supportive evidence that peers effects vary along socio-economics lines,

17
although they can be quite specific to measurement and context. There is some het-
erogeneity for feeling depressed based on father’s drinking status, for not enjoying life
on ethnic minority status and father’s education, and for feeling unhappy based on
father’s education. There is also evidence suggesting significant positive peer effects
among same-subgroup peers based on sex (for feeling depressed), financial situation
(for feeling unhappy), parental education background (for feeling depressed and not
enjoying life), and father’s drinking status (for feeling unhappy).
Taken together, there is some context-specific evidence suggesting heterogeneous
peer effects by socio-economic background, and peer effects tend to be stronger within
sub-class groups that share similar socio-economic characteristics. A caveat for these
findings is warranted: although class-level peer composition is random, these results
might reflect the possibility that students sort into these subgroups and thus estimates
are subject to some bias due to unobservable common environment at the subgroup
level.

6.2 Nonlinearity by distributional position


Next I allow for peer effects to vary depending on the student’s own position in the
distribution of mental health. In so doing, I run a series of quantile regressions, to
assess the effects of average class peer mental health for students at different quantiles
from 0.1 to 0.9. Results are presented in Table 9. Point estimates suggest some quan-
tiles (0.2, 0.3, 0.5, 0.7, and 0.9) are associated with larger effects than the rest, but
there lacks a clear pattern. Moreover, none of these quantile regression estimates are
statistically significant.
I then consider allowing for peer effects to vary depending on the peers’ position
in the distribution of mental health. Specifically, I estimates the effects of raising the
class shares of peers who were in various quintiles of the mental health distribution
at Grade 7. Table 10 presents results from models that estimate peers effects based
on various combinations of these shares. A consistent pattern across these models is
that the class share of quintile 2 peers (peers who fall into the second quintile of the
mental health distribution at Grade 7) tend to generate significantly positive effects,
compared to the peers in other positions of the mental health distribution. Estimates
on the other class share variables are in general smaller in terms of magnitude and

18
insignificant. This finding seems to suggest that peer effects take an inverse U-shape.
In term of magnitude, raising the class share of quintile 2 peers by 10 percentage points,
generate peers effects that are 0.07–0.08 SD larger, relative to increasing the share of
other quintile peers.
A last check on nonlinearity allows for peer effects to differ by both the student’s
position in the mental health distribution and her peers’ position, as specified in equa-
tion (5). The peer effects coefficients estimated from this model are plotted in Figure 1.
Each dot represents the effects of raising by 10 percentage points the class share of peers
falling into a quintile group in the mental health distribution, on the students who fall
into a certain quintile group. Each lines traces the effects of raising class shares of peers
falling into various quintile groups, for a given quintile group of students. Note that
while relative vertical position of each dot is meaningful, the position of each line
relative to zero is not. In general the point estimates are noisy and should be treated
with caution. That said, we can observe some patterns of nonlinear effects. The most
noticeable pattern is that the plots confirm the shallow reversed U-shape relationship
previously found. Peers in the second quintile group generate large effects, more so
for students in the top two quintiles.
Taken together, the exploration of heterogeneity and nonlinearity in peer effects
on mental health finds some evidence of positive peer effects for and within some
certain subgroups. First, on socio-economic dimensions, effects tend to be stronger
for some subgroups, particularly those students with a better-educated parent. This is
mainly driven by stronger interactions within these subgroups than within the whole
class. Second, across the distribution of peer mental health, those in the second quin-
tile group tend to generate positive effects for all students, more so for students in the
top two quintiles.

7 Conclusion
Clean estimation of peer effects is key for the design of cost-effective policy interven-
tions. This paper provides novel evidence on peers effects on mental health in the
classroom, by exploiting variation in peer composition generated by random or even
assignment of new students into classes, in the setting of junior secondary schools in
China. I find little evidence of peer effects in overall mental health, and some evidence

19
of heterogeneity and nonlinearity for specific aspects of mental health and sub-class
level social contexts. I show that the weak peer effects in overall mental health is not
due to weak data or improper research design; further, it is consistent across a number
of overall mental health measures. Going beyond linear-in-means models, I find peer
effects are heterogeneous and nonlinear. There is evidence of small to modest peer
effects for and within certain socio-economic subgroups; further, across the distribu-
tion of peer mental health, effects are nonlinear and take a shallow reverse U-shape,
with larger effects coming from peers in the second quintile of the mental health dis-
tribution.
Methodologically, the results reveal that the common environment issue can be a
major obstacle in finding unbiased peer effects. The findings are consistent with a small
literature on peer effects in mental health that exploits other natural experiments,
and deviate from some observational studies that report large peer correlations in
mental health but are more correlational in nature. This research also highlights the
importance of addressing self-selection issues in peer effects research, particularly for
naturally-occurring social networks. Without too much generalization, I present an
empirical framework that can be applied to study peer effects at the classroom level
based on exogenous variation in peer composition, in the context of schools in China.
Random or even assignment of new students into classrooms is a common feature of
the Chinese school system, which can be exploited to study peer effects on a range of
health and educational outcomes.
The results in this paper suggests that peer effects might be specific to certain as-
pects of mental health and social contexts. Weak evidence points to peer effects on
one aspect of mental health, feeling depressed, but not others such as feeling unable
to enjoy life, or more general measures of mental health. Stronger peer effects within
certain socio-economics subgroups in the class suggest that peer effects are more likely
to be generated in social contexts where interactions are more intensive. Future work
would benefit from richer data on detailed measures of mental health and social in-
teractions within the classroom, to better understand the underlying mechanisms of
social spillovers of mental health.
A key motivation for peer effects studies is to inform policy. From this perspective,
the general weakness of overall peer effects suggests that group-based interventions at
the classroom level probably would not generate large positive externality. Targeting

20
some socio-economic subgroups may see additional benefits due to likely multiplier
effects within these small groups, but it’s unlikely these will spill over to other socio-
economic subgroups. Instead, it would be more efficient to follow an individual-based
approach to target those in need of mental health support.

21
References
Anderson, D. M., Cesur, R., and Tekin, E. (2015). Youth depression and future crim-
inal behavior. Economic Inquiry, 53(1):294–317.
Banerjee, S., Chatterji, P., and Lahiri, K. (2015). Effects of psychiatric disorders on la-
bor market outcomes: A latent variable approach using multiple clinical indicators.
Health Economics.
Bloom, D. E., Cafiero, E., Jané-Llopis, E., Abrahams-Gessel, S., Bloom, L. R.,
Fathima, S., Feigl, A. B., Gaziano, T., Hamandi, A., and Mowafi, M. (2012). The
global economic burden of noncommunicable diseases. Report, World Economic
Forum, Geneva.
Carman, K. G. and Zhang, L. (2012). Classroom peer effects and academic achieve-
ment: Evidence from a Chinese middle school. China Economic Review, 23(2):223–
237.
Chatterji, P., Alegría, M., Lu, M., and Takeuchi, D. (2007). Psychiatric disorders and
labor market outcomes: Evidence from the National Latino and Asian American
Study. Health Economics, 16(10):1069–1090.
Christakis, N. A. and Fowler, J. H. (2007). The spread of obesity in a large social
network over 32 years. New England Journal of Medicine, 357(4):370–379.
Clark, A. E. and Lohéac, Y. (2007). “It wasn’t me, it was them!” Social influence in
risky behavior by adolescents. Journal of Health Economics, 26(4):763–784.
Cohen-Cole, E. and Fletcher, J. M. (2008a). Detecting implausible social network
effects in acne, height, and headaches: Longitudinal analysis. BMJ, 337:a2533.
Cohen-Cole, E. and Fletcher, J. M. (2008b). Is obesity contagious? Social networks
vs. environmental factors in the obesity epidemic. Journal of Health Economics,
27(5):1382–1387.
Cornaglia, F., Crivellaro, E., and McNally, S. (2015). Mental health and education
decisions. Labour Economics, 33:1–12.
Currie, J. and Stabile, M. (2009). Mental health in childhood and human capital. In
Gruber, J., editor, The Problems of Disadvantaged Youth: An Economic Perspective,
National Bureau of Economic Research Conference Report, pages 115–148. Univer-
sity of Chicago Press, Chicago and London.
Ding, W. and Lehrer, S. F. (2007). Do peers affect student achievement in China’s
secondary schools? The Review of Economics and Statistics, 89(2):300–312.

22
Eisenberg, D., Golberstein, E., and Whitlock, J. L. (2014). Peer effects on risky be-
haviors: New evidence from college roommate assignments. Journal of Health Eco-
nomics, 33:126–138.

Eisenberg, D., Golberstein, E., Whitlock, J. L., and Downs, M. F. (2013). Social
contagion of mental health: Evidence from college roommates. Health Economics,
22(8):965–986.

Epple, D. and Romano, R. E. (2011). Peer effects in education: A survey of the theory
and evidence. In Benhabib, J., Bisin, A., and Jackson, M. O., editors, Handbook of
Social Economics, volume 1, pages 1053–1163. North-Holland.

Exline, J. J. and Lobel, M. (1997). Views of the self and affiliation choices: A social
comparison perspective. Basic and Applied Social Psychology, 19(2):243–259.

Fletcher, J. (2013). Adolescent depression and adult labor market outcomes. Southern
Economic Journal, 80(1):26–49.

Fletcher, J. M. (2008). Adolescent depression: Diagnosis, treatment, and educational


attainment. Health Economics, 17(11):1215–1235.

Fletcher, J. M. (2010). Adolescent depression and educational attainment: Results


using sibling fixed effects. Health Economics, 19(7):855–871.

Fletcher, J. M. (2014). The effects of childhood ADHD on adult labor market out-
comes. Health Economics, 23(2):159–181.

Fowler, J. H. and Christakis, N. A. (2008a). Dynamic spread of happiness in a large


social network: Longitudinal analysis over 20 years in the Framingham Heart Study.
BMJ, 337:a2338.

Fowler, J. H. and Christakis, N. A. (2008b). Estimating peer effects on health in social


networks: A response to Cohen-Cole and Fletcher; and Trogdon, Nonnemaker, and
Pais. Journal of Health Economics, 27(5):1400–1405.

Gaviria, A. and Raphael, S. (2001). School-based peer effects and juvenile behavior.
The Review of Economics and Statistics, 83(2):257–268.

Golberstein, E., Eisenberg, D., and Downs, M. F. (2016). Spillover effects in health
service use: Evidence from mental health care using first-year college housing as-
signments. Health Economics, 25(1):40–55.

Hatfield, E., Cacioppo, J. T., and Rapson, R. L. (1993). Emotional contagion. Current
Directions in Psychological Science, 2(3):96–100.

23
Hill, E. M., Griffiths, F. E., and House, T. (2015). Spreading of healthy mood in
adolescent social networks. Proceedings. Biological Sciences, 282(1813):20151180.

Hoxby, C. M. and Weingarth, G. (2005). Taking Race Out of the Equation: School
Reassignment and the Structure of Peer Effects. Mimeo.

Joiner, T. E. and Katz, J. (2006). Contagion of depressive symptoms and mood: Meta-
analytic review and explanations from cognitive, behavioral, and interpersonal view-
points. Clinical Psychology: Science and Practice, 6(2):149–164.

Manski, C. F. (1993). Identification of Endogenous Social Effects: The Reflection


Problem. The Review of Economic Studies, 60(3):531–542.

McPherson, M., Smith-Lovin, L., and Cook, J. M. (2001). Birds of a Feather: Ho-
mophily in Social Networks. Annual Review of Sociology, 27(1):415–444.

MHF (2015). Fundamental facts about mental health 2015. Report, Mental Health
Foundation (MHF).

MoE (2002). Zhongxiaoxue xinli jiankang jiaoyu zhidao gangyao [Guidelines for men-
tal health education in primary and secondary schools]. Policy document, Ministry
of Education of China (MoE). [In Chinese].

MoE (2012). Zhongxiaoxue xinli jiankang jiaoyu zhidao gangyao (2012 nian xiuding)
[Guidelines for mental health education in primary and secondary schools (revised
2012)]. Policy document, Ministry of Education of China (MoE). [In Chinese].

Mueller, A. S. and Abrutyn, S. (2015). Suicidal disclosures among friends: Using social
network data to understand suicide contagion. Journal of Health and Social Behavior,
56(1):131–148.

Rose, A. J. (2002). Co-rumination in the friendships of girls and boys. Child Devel-
opment, 73(6):1830–1843.

Rose, A. J., Carlson, W., and Waller, E. M. (2007). Prospective associations of co-
rumination with friendship and emotional adjustment: Considering the socioemo-
tional trade-offs of co-rumination. Developmental Psychology, 43(4):1019–1031.

Rosenquist, J. N., Fowler, J. H., and Christakis, N. A. (2011). Social network deter-
minants of depression. Molecular Psychiatry, 16(3):273–281.

Sacerdote, B. (2011). Peer effects in education: How might they work, how big are
they and how much do we know thus far? In Hanushek, E. A., Machin, S., and
Woessmann, L., editors, Handbook of the Economics of Education, volume 3, pages
249–277. Elsevier.

24
Trautmann, S., Rehm, J., and Wittchen, H.-U. (2016). The economic costs of mental
disorders. EMBO Reports, 17(9):1245–1249.

Trogdon, J. G., Nonnemaker, J., and Pais, J. (2008). Peer effects in adolescent over-
weight. Journal of Health Economics, 27(5):1388–1399.

Whiteford, H. A., Degenhardt, L., Rehm, J., Baxter, A. J., Ferrari, A. J., Erskine,
H. E., Charlson, F. J., Norman, R. E., Flaxman, A. D., Johns, N., Burstein, R.,
Murray, C. J., and Vos, T. (2013). Global burden of disease attributable to mental
and substance use disorders: Findings from the Global Burden of Disease Study
2010. The Lancet, 382(9904):1575–1586.

WHO (2001). Mental Health: New Understanding, New Hope. The world health report
2001. World Health Organization (WHO), Geneva. OCLC: 249045544.

WHO (2003). Caring for Children and Adolescents with Mental Disorders: Setting WHO
Directions. World Health Organization (WHO), Geneva. OCLC: 666919063.

Wu, P., Li, L.-P., Jin, J., Yuan, X.-H., Liu, X., Fan, B., Fuller, C., Yao-Gui, L., and
Hoven, C. W. (2012). Need for mental health services and service use among high
school students in China. Psychiatric Services, 63(10):1026–1031.

Yakusheva, O., Kapinos, K. A., and Eisenberg, D. (2014). Estimating heterogeneous


and hierarchical peer effects on body weight using roommate assignments as a nat-
ural experiment. Journal of Human Resources, 49(1):234–261.

Zhang, H. and Du, Y. (2018). Chinese school curricula should include mental health
education. The Lancet Psychiatry, 5(8):e18.

25
Table 1: Mental health measures in CEPS: Waves 1 & 2

Question: How often do you have


the feelings below in the last seven days?
Wave 1 Wave 2
Same questions
blues blues
unhappy unhappy
not enjoying life not enjoying life
sad sad
Similar question
depressed too depressed to focus
Additional questions
N/A having no passion
N/A nervous
N/A excessive worry
N/A something bad will happen
N/A too energetic to concentrate
Number of
questions 5 10
Notes This table describes the mental health questions available in Waves
1 and 2 of the China Education Panel Survey (CEPS). Responses to
each question are “never”, “seldom”, “sometimes”, “often”, and “always”.
Questions marked “N/A” are not available in Wave 1.

26
Table 2: Balancing test

Dep. var. = predetermined characteristics at Wave 1


(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Female Minority Age Cognitive Local Parents Poor Mother Father Father
skills hukou present household college+ college+ alcoholic

Panel A: Without school fixed effects


Lagged peer MH5PCS 0.039 0.192 −0.371 0.499∗∗ 0.066 0.171∗∗ −0.198∗∗ 0.111 0.106 −0.008
(0.022) (0.104) (0.282) (0.165) (0.073) (0.057) (0.063) (0.069) (0.070) (0.019)
N 5,970 5,970 5,970 5,970 5,970 5,970 5,970 5,970 5,970 5,970

Panel B: With school fixed effects


Lagged peer MH5PCS 0.060∗ 0.035 −0.278 0.108 0.035 −0.028 0.019 0.008 0.025 −0.011
(0.027) (0.029) (0.175) (0.183) (0.038) (0.039) (0.029) (0.037) (0.030) (0.025)
N 5,970 5,970 5,970 5,970 5,970 5,970 5,970 5,970 5,970 5,970
Notes Clustered standard errors at the school level are in parentheses. ,
∗∗∗ ∗∗
, and ∗ denote statistical significance at 0.001, 0.01, and 0.05 levels respectively.

27
Table 3: Peer effects on mental health

Dep. var. = MH10PCS at Grade 8


(1) (2) (3) (4)
Control Add control Add control Add control
for own for school for lagged for teacher
characteristics fixed effects own mental health characteristics

Lagged peer MH5PCS 0.270∗∗∗ 0.015 −0.011 −0.018


(0.060) (0.112) (0.099) (0.127)
Lagged MH5PCS 0.410∗∗∗ 0.409∗∗∗
(0.015) (0.015)

Own char. Yes Yes Yes Yes


School FE No Yes Yes Yes
Teacher char. No No No Yes
N 5,970 5,970 5,970 5,970
Notes Standardized coefficients are reported. Clustered standard errors at the school level are in parentheses.
, , and ∗ denote statistical significance at 0.001, 0.01, and 0.05 levels respectively.
∗∗∗ ∗∗

28
Table 4: Robustness checks using different mental health measures
Dep. var. = mental health at Grade 8
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
MH10PCS MH10PCS MH10 MH10 MH9PCS MH9PCS MH9 MH9 MH5PCS MH5PCS MH5 MH5

Lagged peer MH5PCS −0.017 −0.010 0.024


(0.126) (0.126) (0.125)
Lagged peer MH4PCS −0.025 −0.017 0.015
(0.124) (0.124) (0.124)
Lagged peer MH5 −0.023 −0.011 0.022
(0.126) (0.125) (0.125)
Lagged peer MH4 −0.029 −0.018 0.014
(0.124) (0.123) (0.123)
Lagged MH5PCS 0.409∗∗∗ 0.412∗∗∗ 0.410∗∗∗
(0.015) (0.015) (0.015)
Lagged MH4PCS 0.404∗∗∗ 0.407∗∗∗ 0.406∗∗∗
(0.015) (0.015) (0.015)
Lagged MH5 0.407∗∗∗ 0.413∗∗∗ 0.411∗∗∗
(0.015) (0.015) (0.015)
Lagged MH4 0.402∗∗∗ 0.408∗∗∗ 0.406∗∗∗
(0.015) (0.015) (0.015)

Own char. Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,970 5,970 5,970 5,970 5,970 5,970 5,970 5,970 5,970 5,970 5,970
Notes Standardized coefficients are reported. Clustered standard errors at the school level are in parentheses. ,
∗∗∗ ∗∗
, and ∗ denote statistical significance at 0.001, 0.01, and 0.05 levels respectively.

29
Table 5: Robustness check based on a large sample including non-random
assignment school

Dep. var. = MH10PCS at Grade 8


(1) (2) (3) (4)
Control Add control Add control Add control
for own for school for lagged for teacher
characteristics fixed effects own mental health characteristics

Lagged peer MH5PCS 0.249∗∗∗ −0.040 −0.016 −0.015


(0.057) (0.113) (0.095) (0.101)
Lagged MH5PCS 0.408∗∗∗ 0.407∗∗∗
(0.014) (0.014)

Own char. Yes Yes Yes Yes


School FE No Yes Yes Yes
Teacher char. No No No Yes
N 7,210 7,210 7,210 7,210
Notes Standardized coefficients are reported. Clustered standard errors at the school level are in parentheses.
, , and ∗ denote statistical significance at 0.001, 0.01, and 0.05 levels respectively.
∗∗∗ ∗∗

30
Table 6: Peer effects on cognitive skills and test scores

Dep. var. =
(1) (2) (3) (4)
Cognitive Chinese Maths English
skills scores scores scores

Lagged peer 0.236∗∗


cognitive skills (0.091)
Lagged cognitive 0.403∗∗∗
skills (0.020)
Lagged peer Chinese 0.233∗∗∗
test score (0.069)
Lagged Chinese test 0.686∗∗∗
score (0.029)
Lagged peer Maths 0.125∗
test score (0.072)
Lagged Maths test 0.710∗∗∗
score (0.039)
Lagged peer English 0.017
test score (0.054)
Lagged English test 0.845∗∗∗
score (0.031)

Own char. Yes Yes Yes Yes


School FE Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes
N 5,970 5,812 5,815 5,807
Notes Standardized coefficients are reported. Clustered standard errors at the
school level are in parentheses. ∗∗∗ , ∗∗ , and ∗ denote statistical significance at
0.001, 0.01, and 0.05 levels respectively.

31
Table 7: Peer effects on separate aspects of mental health

Dep. var. = separate aspect of mental health at Grade 8


(1) (2) (3) (4) (5)
Blues Depressed Unhappy Enjoy Sad

Lagged peer MH1 0.179


Blues (0.142)
Lagged peer MH1 0.227∗
Depressed (0.125)
Lagged peer MH1 0.172
Unhappy (0.118)
Lagged peer MH1 0.079
not enjoying life (0.114)
Lagged peer MH1 −0.074
Sad (0.105)
Lagged MH1 0.322∗∗∗
Blues (0.014)
Lagged MH1 0.271∗∗∗
Depressed (0.013)
Lagged MH1 0.319∗∗∗
Unhappy (0.014)
Lagged MH1 0.349∗∗∗
not enjoying life (0.016)
Lagged MH1 0.246∗∗∗
Sad (0.012)

Own char. Yes Yes Yes Yes Yes


School FE Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes
N 5,970 5,970 5,970 5,970 5,970
Notes Standardized coefficients are reported. Clustered standard errors at the school level
are in parentheses. ∗∗∗ , ∗∗ , and ∗ denote statistical significance at 0.001, 0.01, and 0.05 levels
respectively.

32
Table 8: Heterogeneity by socio-economic status

Dep. var. = mental health


(1) (2) (3) (4) (5) (6) (7) (8)
Female Age Minority Local Rich Mother Father Father
hukou household college+ college+ alcoholic

Panel A: Lagged class peer mental health interacted with indicator


Lagged peer MH5PCS 0.025 −0.025 0.129 −0.012 −0.013 −0.069 −0.047 −0.062
(0.135) (0.978) (0.206) (0.144) (0.124) (0.130) (0.130) (0.240)
Interaction term −0.090 −0.001 −0.160 −0.010 −0.098 0.221∗ 0.123 0.046
(0.085) (0.067) (0.187) (0.110) (0.132) (0.120) (0.112) (0.173)
Lagged MH5PCS 0.409∗∗∗ 0.405∗∗∗ 0.408∗∗∗ 0.409∗∗∗ 0.410∗∗∗ 0.409∗∗∗ 0.409∗∗∗ 0.414∗∗∗
(0.015) (0.015) (0.015) (0.015) (0.015) (0.015) (0.015) (0.015)
Own char. Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,831 5,957 5,952 5,956 5,956 5,956 5,810
Panel B: Lagged sub-class peer mental health
Lagged same-subgroup 0.025 −0.006 −0.090 −0.020 0.023 0.120 0.133∗ −0.031
peer MH5PCS (0.077) (0.007) (0.063) (0.055) (0.063) (0.072) (0.072) (0.038)
Lagged cross-subgroup −0.018 −0.043 0.004 −0.015 −0.012 −0.076 −0.083 −0.009
peer MH5PCS (0.070) (0.059) (0.066) (0.064) (0.066) (0.064) (0.061) (0.061)
Lagged MH5PCS 0.408∗∗∗ 0.409∗∗∗ 0.412∗∗∗ 0.410∗∗∗ 0.408∗∗∗ 0.403∗∗∗ 0.403∗∗∗ 0.417∗∗∗
(0.015) (0.016) (0.015) (0.015) (0.015) (0.015) (0.015) (0.015)
Own char. Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,739 5,920 5,938 5,938 5,932 5,932 5,774
Notes Standardized coefficients are reported. Clustered standard errors at the school level are in parentheses. ,
∗∗∗ ∗∗
, and ∗
denote
statistical significance at 0.001, 0.01, and 0.05 levels respectively.

33
Table 9: Quantile peer effects on mental health

Dep. var. = mental health, quantile =


0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Lagged peer MH5PCS −0.050 0.120 0.072 0.014 0.106 −0.036 0.099 −0.011 0.092
(0.264) (0.172) (0.130) (0.100) (0.104) (0.084) (0.075) (0.072) (0.084)
Lagged MH5PCS 0.539∗∗∗ 0.523∗∗∗ 0.498∗∗∗ 0.472∗∗∗ 0.473∗∗∗ 0.455∗∗∗ 0.417∗∗∗ 0.339∗∗∗ 0.204∗∗∗
(0.029) (0.018) (0.018) (0.017) (0.017) (0.017) (0.018) (0.018) (0.024)

Own char. Yes Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,970 5,970 5,970 5,970 5,970 5,970 5,970 5,970
Notes Standardized coefficients are reported. Clustered standard errors at the school level are in parentheses. ,
∗∗∗ ∗∗
, and ∗ denote statistical
significance at 0.001, 0.01, and 0.05 levels respectively.

34
Table 10: Effects of raising class shares of various mental health levels on mental
health

Dep. var. = MH10PCS at Grade 8


(1) (2) (3) (4) (5) (6) (7) (8)

Lagged peer MH5PCS: 0.018 −0.046 −0.056 −0.086


quintile 1 (0.343) (0.290) (0.306) (0.409)
Lagged peer MH5PCS: 0.752∗∗ 0.741∗∗ 0.718∗∗ 0.804∗
quintile 2 (0.305) (0.301) (0.360) (0.430)
Lagged peer MH5PCS: −0.053 −0.081 −0.360 0.005
quintile 3 (0.275) (0.401) (0.301) (0.323)
Lagged peer MH5PCS: −0.043 −0.210 −0.276 0.043
quintile 4 (0.372) (0.245) (0.269) (0.321)
Lagged peer MH5PCS: −0.325 −0.311 −0.357 0.086
quintile 5 (0.364) (0.346) (0.341) (0.409)
Lagged MH5PCS 0.409∗∗∗ 0.409∗∗∗ 0.410∗∗∗ 0.410∗∗∗ 0.408∗∗∗ 0.408∗∗∗ 0.408∗∗∗ 0.410∗∗∗
(0.015) (0.015) (0.015) (0.015) (0.015) (0.015) (0.015) (0.015)

Own char. Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,970 5,970 5,970 5,970 5,970 5,970 5,970
Notes Standardized coefficients are reported. Clustered standard errors at the school level are in parentheses. ,
∗∗∗ ∗∗
, and ∗ denote
statistical significance at 0.001, 0.01, and 0.05 levels respectively.

35
.1
Peer effects coefficient
-.05 0
-.1 .05

1 2 3 4 5
Class share of peers in lagged MH5PCS quintile

Own lagged MH5PCS is bottom quintile


Own lagged MH5PCS is 2nd quintile
Own lagged MH5PCS is 3rd quintile
Own lagged MH5PCS is 4th quintile
Own lagged MH5PCS is top quintile

Figure 1: Peers effects by the student’s and her peers’ position in the mental health
distribution

36
A Appendix

A.1 Variable description


Beyond the key variables described in the main text, a rich set of control variables are
included in the analysis. Adolescent characteristics and family background variables
include cognitive scores at Wave 1, gender, age, ethnic minority, local hukou,12 rural
hukou, mother’s education level, father’s education level, and whether father often
gets drunk. Teacher characteristics include teacher’s gender, teaching experience, ed-
ucation level, and mental health training. These characteristics are included for four
teachers of each class (homeroom teacher, Chinese teacher, maths teacher, and English
teacher). In addition, homeroom teacher’s years of experience as a homeroom teacher
is also controlled for. Definitions of all the variables used in the analysis are provided
in Table A1.

A.2 Principal component analysis


Principal component analysis (PCA) is a data reduction method that tries to identify
the subspace in which the data approximately lies. It can be used to find the principal
directions in which the data varies through elucidating the covariance structure of a set
of variables. The principal components are found by calculating the eigenvectors and
eigenvalues of the data covariance matrix to transform it into diagonal vectors. The
largest eigenvalue corresponds to the eigenvector that explains the most variation, the
second largest eigenvalue is associated with the eigenvector that explains the second
highest variation, and so on. A rule of thumb for retaining the principal components
that are associated with a eigenvalue that is higher than 1. Figure plots the eigenvalues
for the PCA based on MH5 in Wave 1 and MH10 in Wave 2. PCA based MH10 finds
only one eigenvalue higher than 1, so one principal component is retained; PCA based
on MH10 finds two eigenvalues higher than 1, but the second largest eigenvalue is very
close to 1 at 1.02, so only the first principal component is retained for analysis.
12
Hukou is China’s household registration system. Having a local hukou gives the registree entitle-
ment to local social welfare, such as a place at a local public school.

37
A.3 Additional tables and figures
Additional tables and figures are listed below.

Table A1: Variable definitions

Variable(s)
Definition

Mental health variables


MH10PCS
Definition: Principal component score (PCS) extracted from 10 mental health ques-
tions, normalized to zero mean and unity standard deviation (SD).
MH10
Definition: Aggregated score based on summation over 10 mental health questions,
normalized to zero mean and unity SD.
MH9PCS
Definition: Principal component score (PCS) extracted from 9 mental health ques-
tions excluding “too energetic to concentrate”, normalized to zero mean and unity
standard deviation (SD).
MH9
Definition: Aggregated score based on summation over 9 mental health questions
excluding “too energetic to concentrate”, normalized to zero mean and unity SD.
MH5PCS
Definition: PCS extracted from 5 mental health questions that are the same or
similar in both waves of CEPS, normalized to zero mean and unity SD.
MH5
Definition: PCS extracted from 5 mental health questions that are the same or
similar in both waves of CEPS, normalized to zero mean and unity SD.
MH4PCS
Definition: PCS extracted from 4 mental health questions that are the same in both
waves of CEPS, normalized to zero mean and unity SD.
Continued on next page . . .

38
. . . continued from previous page
Variable
Definition

MH4PCS
Definition: PCS extracted from 5 mental health questions that are the same in both
waves of CEPS, normalized to zero mean and unity SD.
Aspect-specific mental health
Definition: Ten variables created based on 10 separate mental health questions, nor-
malized to zero mean and unity SD.

Peer mental health variables


Peer MH5PCS
Definition: Average of MH5PCS within the class excluding the respondent. Mea-
sured at Wave 1.
Peer MH5
Definition: Average of MH5 within the class excluding the respondent. Measured
at Wave 1.
Peer MH4PCS
Definition: Average of MH4PCS within the class excluding the respondent. Mea-
sured at Wave 1.
Peer MH4
Definition: Average of MH4 within the class excluding the respondent. Measured
at Wave 1.

Adolescent characteristics and family background


Cognitive skills
Definition: Cognitive score based on standardized test administered by CEPS sur-
vey collectors, measured at Wave 1.
Female
Definition: =1 if female, = 0 otherwise.
Continued on next page . . .

39
. . . continued from previous page
Variable
Definition

Minority
Definition: =1 if ethnic minority (i.e. not Han Chinese), = 0 otherwise.
Age
Definition: Years of age.
Local hukou
Definition: =1 if hukou (China’s household management system, tied to access to
local social welfare) is registered locally, =0 otherwise.
Rural hukou
Definition: =1 if hukou is registered in rural areas, =0 otherwise.
Mother’s education
Definition: Four binary variables created based on mother’s education levels: pri-
mary education, junior secondary education, senior secondary education, tertiary
education.
Father’s education
Definition: Four binary variables created based on father’s education levels: pri-
mary education, junior secondary education, senior secondary education, tertiary
education.
Financial situation
Definition: Three binary variables created based on financial situation of the re-
spondent: poor, medium, or rich.
Father alcoholic
Definition: =1 if having an alcoholic father, = 0 otherwise.

Teacher characteristics
Headroom experience
Definition: Years of experience as headroom teacher.a
Continued on next page . . .

40
. . . continued from previous page
Variable
Definition

Teacher female
Definition: =1 if female teacher, = 0 otherwise. Four binary variables created for
homeroom teacher, Chinese teacher, Maths teacher, and English teacher separately.
Teacher experience
Definition: Years of teaching experience of the teacher. Four variables created for
homeroom teacher, Chinese teacher, Maths teacher, and English teacher separately.
Teacher education
Definition: Teacher’s education level: college education, adult university education,
regular university education and above. Three binary variables created for each
teacher (homeroom teacher, Chinese teacher, Maths teacher, and English teacher),
12 in total.
Teacher mental health training
Definition: Teacher’s level of mental health training: none, level A, and level
B. Three binary variables created for each teacher (homeroom teacher, Chinese
teacher, Maths teacher, and English teacher), 12 in total.

Notes
This table details the definitions of variables used in the analysis.

41
Table A2: Correlation matrix of mental health questions

Blues Depressed Unhappy Enjoy Sad Passionless Nervous Worry Happen Concentrate

Wave 1
Blues 1.000
Depressed 0.698∗∗∗ 1.000
Unhappy 0.717∗∗∗ 0.675∗∗∗ 1.000
Enjoy 0.600∗∗∗ 0.613∗∗∗ 0.693∗∗∗ 1.000
Sad 0.585∗∗∗ 0.645∗∗∗ 0.673∗∗∗ 0.705∗∗∗ 1.000

Wave 2
Blues 1.000
Depressed 0.698∗∗∗ 1.000
Unhappy 0.717∗∗∗ 0.675∗∗∗ 1.000
Enjoy 0.600∗∗∗ 0.613∗∗∗ 0.693∗∗∗ 1.000
Sad 0.585∗∗∗ 0.645∗∗∗ 0.673∗∗∗ 0.705∗∗∗ 1.000
Passionless 0.585∗∗∗ 0.645∗∗∗ 0.673∗∗∗ 0.705∗∗∗ 1.000 1.000
Nervous 0.488∗∗∗ 0.453∗∗∗ 0.491∗∗∗ 0.427∗∗∗ 0.483∗∗∗ 0.483∗∗∗ 1.000
Worry 0.523∗∗∗ 0.503∗∗∗ 0.526∗∗∗ 0.516∗∗∗ 0.502∗∗∗ 0.502∗∗∗ 0.658∗∗∗ 1.000
Happen 0.496∗∗∗ 0.483∗∗∗ 0.503∗∗∗ 0.471∗∗∗ 0.487∗∗∗ 0.487∗∗∗ 0.552∗∗∗ 0.622∗∗∗ 1.000
Concentrate 0.266∗∗∗ 0.374∗∗∗ 0.309∗∗∗ 0.327∗∗∗ 0.353∗∗∗ 0.353∗∗∗ 0.316∗∗∗ 0.327∗∗∗ 0.372∗∗∗ 1.000
Notes This table reports the Pearson correlation coefficients between mental health questions in CEPS Waves 1 & 2. Each cell reports the correlation
coefficient for the corresponding row variable and column variable. ∗∗∗ , ∗∗ , and ∗ denote statistical significance at 0.001, 0.01, and 0.05 levels respectively.
Depressed: full wording is "too depressed to focus" in Wave 2 questionnaire.
Enjoy: "not enjoying life".
Passionless: "having no passion".
Worry: "excessive worry".
Happen: "something bad will happen".
Concentrate: "too energetic to concentrate in class".

42
Scree plots of eigenvalues
MH5 questions MH10 questions
6

6
4

4
Eigenvalue

Eigenvalue
2

2
0

1 2 3 4 5 0 2 4 6 8 10
Question Question

Figure A1: Scree plot of eigenvalues based on the principal component analysis

43
Table A3: Summary statistics

Mean SD Min. Max

Student’s contemporaneous mental health


MH10PCS 0.000 1.000 −3.3 1.4
MH10 0.000 1.000 −3.4 1.4
MH9PCS 0.000 1.000 −3.2 1.3
MH9 0.000 1.000 −3.2 1.4
MH5PCS 0.000 1.000 −3.0 1.3
MH5 0.000 1.000 −3.1 1.3
MH4PCS 0.000 1.000 −3.0 1.2
MH4 0.000 1.000 −3.0 1.2
Lagged own mental health
Lagged MH5PCS 0.000 1.000 −3.7 1.2
Lagged MH5 0.000 1.000 −3.7 1.2
Lagged MH4PCS 0.000 1.000 −3.6 1.2
Lagged MH4 0.000 1.000 −3.7 1.2
Lagged peer mental health
Lagged peer MH5PCS 0.000 0.291 −0.9 0.7
Lagged peer MH5 0.000 0.291 −0.9 0.7
Lagged peer MH4PCS 0.000 0.291 −0.8 0.7
Lagged peer MH4 0.000 0.291 −0.9 0.7
Student characteristics and family background
Cognitive skills 0.375 0.806 −3.1 2.1
Female 0.485 0.500 0.0 1.0
Minority 0.908 0.290 0.0 1.0
Missing: Minority 0.002 0.047 0.0 1.0
Age 13.629 2.271 0.0 19.0
Missing: Age 0.023 0.151 0.0 1.0
Local student 0.794 0.404 0.0 1.0
Missing: Local student 0.003 0.055 0.0 1.0

Continued on next page . . .

44
. . . continued from previous page

Mean SD Min. Max

Rural hukou 0.456 0.498 0.0 1.0


Missing: Rural hukou 0.007 0.086 0.0 1.0
Living with both parents 0.822 0.382 0.0 1.0
Living with one parent 0.095 0.293 0.0 1.0
Living with neither parent 0.075 0.264 0.0 1.0
Mother’s education: primary 0.204 0.403 0.0 1.0
Mother’s education: lower secondary 0.385 0.487 0.0 1.0
Mother’s education: upper secondary 0.239 0.427 0.0 1.0
Mother’s education: tertiary 0.169 0.375 0.0 1.0
Missing: Mother’s education 0.002 0.048 0.0 1.0
Father’s education 2.533 0.961 0.0 4.0
Father’s education: primary 0.130 0.337 0.0 1.0
Father’s education: lower secondary 0.399 0.490 0.0 1.0
Father’s education: upper secondary 0.269 0.443 0.0 1.0
Father’s education: tertiary 0.200 0.400 0.0 1.0
Missing: Father’s education 0.002 0.048 0.0 1.0
Financial situation: poor 0.165 0.372 0.0 1.0
Financial situation: middle 0.765 0.424 0.0 1.0
Financial situation: rich 0.068 0.251 0.0 1.0
Missing: Financial situation 0.002 0.048 0.0 1.0
Father alcoholic 1.047 0.313 0.0 2.0
Missing: Father alcoholic 0.027 0.162 0.0 1.0
Teacher characteristics
Homeroom teacher: female 0.674 0.469 0.0 1.0
Homeroom teacher: teaching experience 15.894 8.460 0.0 39.0
Missing: Homeroom teacher: teaching experience 0.023 0.149 0.0 1.0
Homeroom teacher: homeroom experience 13.032 8.029 0.0 38.0
Missing: Homeroom teacher: homeroom experience 0.007 0.085 0.0 1.0
Homeroom teacher: college education 0.109 0.312 0.0 1.0
Homeroom teacher: university (adult) education 0.380 0.485 0.0 1.0
Homeroom teacher: university (regular) and above 0.502 0.500 0.0 1.0

Continued on next page . . .

45
. . . continued from previous page

Mean SD Min. Max

Missing: Homeroom teacher: education level 0.009 0.096 0.0 1.0


Homeroom teacher mental health training: Level A 0.108 0.311 0.0 1.0
Homeroom teacher mental health training: Level B 0.098 0.298 0.0 1.0
Missing: Homeroom teacher mental health training 0.077 0.266 0.0 1.0
Chinese teacher: female 0.763 0.425 0.0 1.0
Missing: Chinese teacher: female 0.016 0.126 0.0 1.0
Chinese teacher: teaching experience 14.788 8.553 0.0 40.0
Missing: Chinese teacher: teaching experience 0.043 0.202 0.0 1.0
Chinese teacher: college education 0.086 0.281 0.0 1.0
Chinese teacher: university (adult) education 0.394 0.489 0.0 1.0
Chinese teacher: university (regular) and above 0.508 0.500 0.0 1.0
Missing: Chinese teacher: education level 0.012 0.109 0.0 1.0
Chinese teacher mental health training: Level A 0.098 0.297 0.0 1.0
Chinese teacher mental health training: Level B 0.138 0.345 0.0 1.0
Missing: Chinese teacher mental health training 0.070 0.255 0.0 1.0
English teacher: female 0.854 0.353 0.0 1.0
Missing: English teacher: female 0.012 0.108 0.0 1.0
English teacher: teaching experience 16.175 8.666 0.0 34.0
Missing: English teacher: teaching experience 0.020 0.140 0.0 1.0
English teacher: college education 0.122 0.327 0.0 1.0
English teacher: university (adult) education 0.431 0.495 0.0 1.0
English teacher: university (regular) and above 0.434 0.496 0.0 1.0
Missing: English teacher: education level 0.014 0.117 0.0 1.0
English teacher mental health training: Level A 0.119 0.324 0.0 1.0
English teacher mental health training: Level B 0.125 0.331 0.0 1.0
Missing: English teacher mental health training 0.089 0.285 0.0 1.0
Maths teacher: female 0.586 0.493 0.0 1.0
Missing: Maths teacher: female 0.004 0.066 0.0 1.0
Maths teacher: teaching experience 17.023 8.889 0.0 39.0
Missing: Maths teacher: teaching experience 0.022 0.146 0.0 1.0
Maths teacher: college education 0.105 0.307 0.0 1.0

Continued on next page . . .

46
. . . continued from previous page

Mean SD Min. Max

Maths teacher: university (adult) education 0.354 0.478 0.0 1.0


Maths teacher: university (regular) and above 0.541 0.498 0.0 1.0
Maths teacher mental health training: Level A 0.119 0.324 0.0 1.0
Maths teacher mental health training: Level B 0.118 0.323 0.0 1.0
Missing: Maths teacher mental health training 0.061 0.239 0.0 1.0

N 5,970

Source: Author’s calculation based on the CEPS, Waves 1 & 2.

47
Table A4: Heterogeneous effects on feeling blues by socio-economic status

Dep. var. = mental health, interaction with


(1) (2) (3) (4) (5) (6) (7) (8)
Female Age Minority Local Rich Mother Father Father
hukou household college+ college+ alcoholic

Panel A: Lagged class peer mental health interacted with indicator


Lagged peer MH1 0.132 −0.765 0.216 0.215 0.184 0.165 0.153 −0.005
Blues (0.151) (1.381) (0.242) (0.178) (0.145) (0.151) (0.149) (0.245)
Interaction term 0.093 0.067 −0.041 −0.061 −0.047 0.062 0.130 0.172
(0.128) (0.098) (0.209) (0.144) (0.151) (0.140) (0.137) (0.221)
Lagged MH1 Blues 0.321∗∗∗ 0.322∗∗∗ 0.322∗∗∗ 0.322∗∗∗ 0.323∗∗∗ 0.322∗∗∗ 0.322∗∗∗ 0.325∗∗∗
(0.014) (0.014) (0.014) (0.014) (0.014) (0.014) (0.014) (0.014)
Own char. Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,831 5,957 5,952 5,956 5,956 5,956 5,810
Panel B: Lagged sub-class peer mental health

Lagged same-subgroup 0.159∗ 0.004 0.012 −0.007 −0.074 0.141∗ 0.124∗ −0.007
peer MH1 blues (0.081) (0.009) (0.054) (0.062) (0.064) (0.073) (0.069) (0.043)
Lagged cross-subgroup 0.064 −0.025 −0.009 0.001 0.028 −0.048 −0.043 −0.028
peer MH1 blues (0.080) (0.064) (0.069) (0.067) (0.069) (0.068) (0.067) (0.061)
Lagged MH1 Blues 0.314∗∗∗ 0.320∗∗∗ 0.319∗∗∗ 0.319∗∗∗ 0.322∗∗∗ 0.313∗∗∗ 0.314∗∗∗ 0.322∗∗∗
(0.013) (0.014) (0.014) (0.014) (0.013) (0.014) (0.014) (0.013)
Own char. Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,739 5,920 5,938 5,938 5,932 5,932 5,774
Notes Standardized coefficients are reported. Clustered standard errors at the school level are in parentheses. ,
∗∗∗ ∗∗
, and ∗
denote
statistical significance at 0.001, 0.01, and 0.05 levels respectively.

48
Table A5: Heterogeneous effects on feeling depressed by socio-economic status

Dep. var. = mental health, interaction with


(1) (2) (3) (4) (5) (6) (7) (8)
Female Age Minority Local Rich Mother Father Father
hukou household college+ college+ alcoholic

Panel A: Lagged class peer mental health interacted with indicator


Lagged peer MH1 0.235∗ 0.077 0.021 0.227 0.236∗ 0.183 0.217∗ −0.218
Depressed (0.135) (1.241) (0.220) (0.176) (0.128) (0.131) (0.129) (0.267)
Interaction term −0.016 0.008 0.223 −0.018 −0.088 0.204 0.005 0.410∗
(0.122) (0.088) (0.203) (0.162) (0.176) (0.163) (0.137) (0.214)
Lagged MH1 0.271∗∗∗ 0.271∗∗∗ 0.271∗∗∗ 0.271∗∗∗ 0.271∗∗∗ 0.271∗∗∗ 0.271∗∗∗ 0.275∗∗∗
Depressed (0.013) (0.014) (0.013) (0.013) (0.013) (0.013) (0.013) (0.014)
Own char. Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,831 5,957 5,952 5,956 5,956 5,956 5,810
Panel B: Lagged sub-class peer mental health

Lagged same-subgroup 0.171∗∗ −0.002 −0.045 0.105 −0.020 0.116∗∗ 0.132∗∗∗ 0.021
peer MH1 depressed (0.067) (0.008) (0.056) (0.064) (0.071) (0.052) (0.047) (0.047)
Lagged cross-subgroup 0.059 −0.073 −0.017 −0.067 0.022 −0.073 −0.079 −0.019
peer MH1 depressed (0.078) (0.071) (0.076) (0.076) (0.076) (0.071) (0.071) (0.071)
Lagged MH1 Depressed 0.262∗∗∗ 0.270∗∗∗ 0.268∗∗∗ 0.263∗∗∗ 0.265∗∗∗ 0.261∗∗∗ 0.261∗∗∗ 0.270∗∗∗
(0.013) (0.014) (0.014) (0.014) (0.013) (0.014) (0.014) (0.014)
Own char. Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,739 5,920 5,938 5,938 5,932 5,932 5,774
Notes Standardized coefficients are reported. Clustered standard errors at the school level are in parentheses. ,
∗∗∗ ∗∗
, and ∗ denote statisti-
cal significance at 0.001, 0.01, and 0.05 levels respectively.

49
Table A6: Heterogeneous effects on not enjoying life by socio-economic status

Dep. var. = mental health


(1) (2) (3) (4) (5) (6) (7) (8)
Female Age Minority Local Rich Mother Father Father
hukou household college+ college+ alcholic

Panel A: Lagged class peer mental health interacted with indicator


Lagged peer MH1 0.096 −0.610 0.535∗∗∗ 0.091 0.084 0.057 0.040 0.364
not enjoying life (0.116) (1.012) (0.187) (0.121) (0.110) (0.119) (0.121) (0.248)
Interaction term −0.035 0.048 −0.485∗∗∗ −0.010 −0.073 0.185 0.274∗ −0.250
(0.090) (0.070) (0.165) (0.089) (0.181) (0.151) (0.156) (0.191)
Lagged MH1 0.349∗∗∗ 0.346∗∗∗ 0.349∗∗∗ 0.349∗∗∗ 0.349∗∗∗ 0.348∗∗∗ 0.348∗∗∗ 0.356∗∗∗
not enjoying life (0.016) (0.016) (0.016) (0.016) (0.015) (0.015) (0.016) (0.016)
Own char. Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,831 5,957 5,952 5,956 5,956 5,956 5,810
Panel B: Lagged sub-class peer mental health

Lagged same-subgroup 0.085 0.000 −0.054 0.085 0.088 0.048 0.111 0.086∗
peer MH1 enjoy (0.072) (0.008) (0.064) (0.060) (0.067) (0.085) (0.067) (0.050)
Lagged cross-subgroup −0.006 −0.039 −0.035 −0.057 −0.053 −0.052 −0.073 −0.032
peer MH1 enjoy (0.070) (0.069) (0.074) (0.072) (0.071) (0.076) (0.070) (0.068)
Lagged MH1 not 0.345∗∗∗ 0.346∗∗∗ 0.348∗∗∗ 0.343∗∗∗ 0.343∗∗∗ 0.343∗∗∗ 0.343∗∗∗ 0.352∗∗∗
enjoying life (0.016) (0.016) (0.016) (0.016) (0.016) (0.016) (0.016) (0.017)
Own char. Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,739 5,920 5,938 5,938 5,932 5,932 5,774
Notes Standardized coefficients are reported. Clustered standard errors at the school level are in parentheses. ,
∗∗∗ ∗∗
, and ∗
denote
statistical significance at 0.001, 0.01, and 0.05 levels respectively.

50
Table A7: Heterogeneous effects on feeling unhappy by socio-economic status

Dep. var. = mental health, interaction with


(1) (2) (3) (4) (5) (6) (7) (8)
Female Age Minority Local Rich Mother Father Father
hukou household college+ college+ alcoholic

Panel A: Lagged class peer mental health interacted with indicator


Lagged peer MH1 0.132 −0.297 0.359 0.242 0.166 0.130 0.110 0.213
Unhappy (0.132) (1.010) (0.217) (0.161) (0.116) (0.120) (0.122) (0.203)
Interaction term 0.082 0.031 −0.201 −0.094 −0.011 0.187 0.279∗∗ −0.037
(0.105) (0.070) (0.206) (0.143) (0.136) (0.132) (0.117) (0.143)
Lagged MH1 0.319∗∗∗ 0.317∗∗∗ 0.319∗∗∗ 0.320∗∗∗ 0.320∗∗∗ 0.320∗∗∗ 0.320∗∗∗ 0.325∗∗∗
Unhappy (0.014) (0.015) (0.014) (0.014) (0.014) (0.014) (0.014) (0.015)
Own char. Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,831 5,957 5,952 5,956 5,956 5,956 5,810
Panel B: Lagged sub-class peer mental health

Lagged same-subgroup 0.122 −0.004 0.014 −0.078 0.147∗∗∗ 0.126∗ 0.196∗∗ −0.025
peer MH1 unhappy (0.075) (0.007) (0.066) (0.051) (0.051) (0.070) (0.090) (0.043)
Lagged cross-subgroup 0.078 0.005 0.028 0.059 −0.015 −0.011 −0.038 0.047
peer MH1 unhappy (0.065) (0.055) (0.063) (0.058) (0.064) (0.060) (0.056) (0.056)
Lagged MH1 0.313∗∗∗ 0.317∗∗∗ 0.318∗∗∗ 0.322∗∗∗ 0.309∗∗∗ 0.313∗∗∗ 0.312∗∗∗ 0.325∗∗∗
Unhappy (0.014) (0.015) (0.014) (0.014) (0.014) (0.014) (0.015) (0.015)
Own char. Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,739 5,920 5,938 5,938 5,932 5,932 5,774
Notes Standardized coefficients are reported. Clustered standard errors at the school level are in parentheses. ,
∗∗∗ ∗∗
, and ∗
denote
statistical significance at 0.001, 0.01, and 0.05 levels respectively.

51
Table A8: Heterogeneous effects on feeling sad by socio-economic status

Dep. var. = mental health


(1) (2) (3) (4) (5) (6) (7) (8)
Female Age Minority Local Rich Mother Father Father
hukou household college+ college+ alcoholic

Panel A: Lagged class peer mental health interacted with indicator


Lagged peer MH1 −0.067 0.394 −0.027 −0.031 −0.079 −0.107 −0.078 −0.010
sad (0.112) (0.896) (0.196) (0.148) (0.105) (0.103) (0.107) (0.260)
Interaction term −0.016 −0.036 −0.054 −0.057 −0.019 0.149 0.007 −0.046
(0.087) (0.061) (0.211) (0.119) (0.136) (0.173) (0.152) (0.194)
Lagged MH1 0.246∗∗∗ 0.242∗∗∗ 0.245∗∗∗ 0.247∗∗∗ 0.247∗∗∗ 0.246∗∗∗ 0.246∗∗∗ 0.246∗∗∗
sad (0.012) (0.012) (0.012) (0.012) (0.012) (0.012) (0.012) (0.013)
Own char. Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,831 5,957 5,952 5,956 5,956 5,956 5,810
Panel B: Lagged sub-class peer mental health

Lagged same-subgroup −0.033 −0.006 −0.081 −0.047 0.043 0.089 0.139 −0.033
peer MH1 sad (0.067) (0.007) (0.063) (0.063) (0.067) (0.085) (0.090) (0.044)
Lagged cross-subgroup −0.019 −0.058 0.010 0.006 −0.016 −0.057 −0.073 0.018
peer MH1 sad (0.067) (0.065) (0.069) (0.072) (0.072) (0.074) (0.077) (0.066)
Lagged MH1 0.248∗∗∗ 0.244∗∗∗ 0.248∗∗∗ 0.251∗∗∗ 0.246∗∗∗ 0.241∗∗∗ 0.241∗∗∗ 0.249∗∗∗
sad (0.013) (0.013) (0.012) (0.013) (0.013) (0.013) (0.013) (0.013)
Own char. Yes Yes Yes Yes Yes Yes Yes Yes
School FE Yes Yes Yes Yes Yes Yes Yes Yes
Teacher char. Yes Yes Yes Yes Yes Yes Yes Yes
N 5,970 5,739 5,920 5,938 5,938 5,932 5,932 5,774
Notes Standardized coefficients are reported. Clustered standard errors at the school level are in parentheses. ,
∗∗∗ ∗∗
, and ∗
denote
statistical significance at 0.001, 0.01, and 0.05 levels respectively.

52

You might also like