Social Science & Medicine 119 (2014) 81e87

Contents lists available at ScienceDirect

Social Science & Medicine

journal homepage: www.elsevier.com/locate/socscimed

Short report

Social capital and adolescent smoking in schools and communities: A

cross-classified multilevel analysis
Bart De Clercq a, *, Timo-Kolja Pfoertner b, Frank J. Elgar c, Anne Hublet a, Lea Maes a
a
Ghent University e Faculty of Medicine and Health Sciences, Department of Public Health, Academical Hospital, K3-4, De Pintelaan, 185,
B-9000 Gent, Belgium
b
Institute of Medical Sociology, Martin Luther University, Halle (Saale), Germany
c
Institute for Health and Social Policy and Douglas Institute, McGill University, Montreal, Quebec, Canada

a r t i c l e i n f o a b s t r a c t

Article history: We sought to determine whether social capital at the individual-, school- and community-level can
Received 8 April 2014 explain variance in adolescent smoking and accounts for social inequalities in smoking. We collected
Received in revised form data as part of the 2005/6 Health Behavior in School-aged Children survey, a nationally representative
12 August 2014
survey of the health and well-being of high school pupils in Belgium (Flanders). Social capital was
Accepted 14 August 2014
Available online 15 August 2014
assessed by structural and cognitive components of family social capital, a four-factor school social
capital scale and a cognitive community social capital scale. We fitted non-hierarchical multilevel models
to the data, with 8453 adolescents nested within a cross-classification of 167 schools and 570 com-
Keywords:
Adolescence
munities. Significant variation in adolescent regular smoking was found between schools, but not be-
Smoking tween communities. Only structural family social capital and cognitive school social capital variables
Social capital negatively related to regular smoking. No interactions between socio-economic status and social capital
Inequalities variables were found. Our findings suggest that previously observed community-level associations with
Cross-classified multilevel models adolescent smoking may be a consequence of unmeasured confounding. Distinguishing nested contexts
of social capital is important because their associations with smoking differ.
© 2014 Elsevier Ltd. All rights reserved.

1. Introduction
Tobacco smoking, the largest single cause of avoidable death in
the EU (European Commission, 2013), inversely relates to socio-
economic status in adult (Winkleby et al., 1992) and adolescent
populations (Blane et al., 1996). Prevention interventions are
needed at an early stage because such health-risk behaviors are
typically established during adolescence and young adulthood
(Centers for Disease Control and Prevention, 2008). The last twenty
years have witnessed an explosion of interest in place effects on
health (Diez Roux, 1998; Macintyre et al., 2002). This research has
found between-school and between-neighborhood variance in
adolescent smoking (Aveyard et al., 2005; Henderson et al., 2008;
Kelly et al., 2011; Lovato et al., 2010). However, differences in the
associations of social contextual factors with the prevalence of
smoking remain unclear (Galea et al., 2004). The goal of this study
was to explore the social context of adolescent smoking.
* Corresponding author.
E-mail address: b.declercq@ugent.be (B. De Clercq).
The association between social capital and adult smoking is well
documented (Chuang and Chuang, 2008) including the “miniatur-
ization of community” phenomenon (Lindstrom, 2003), but evi-
dence in adolescent populations remains sparse. Few studies have
examined a link between both individual- (Curran, 2007;
Lundborg, 2005; Morgan and Haglund, 2009) or contextual social
capital (Aslund and Nilsson, 2013; Henderson et al., 2008;
Takakura, 2011; Thorlindsson et al., 2012) and adolescent smok-
ing, and more specifically in the context of families (Curran, 2007),
schools (Henderson et al., 2008; Takakura, 2011), and communities
(Aslund and Nilsson, 2013; Thorlindsson et al., 2012). The definition
and measurement of social capital are important to understand (a)
the heterogeneous findings across studies, and (b) the differential
impact of various components of social capital on health. Social
capital has been broadly defined as the resources that individuals
access through their social networks (Kawachi and Berkman, 2001).
A first distinction in the conceptualization and measurement of
social capital is commonly drawn between “structural” and
“cognitive” components of social capital (Harpham et al., 2002). The
structural component reflects behavioral manifestations of
network connections: network social capital refers to network

http://dx.doi.org/10.1016/j.socscimed.2014.08.018
0277-9536/© 2014 Elsevier Ltd. All rights reserved.
82 B. De Clercq et al. / Social Science & Medicine 119 (2014) 81e87
membership, family social capital refers to parentechild and
parenteschool interactions, school social capital consists of
childeschool interactions, and childeneighborhood interactions
are labeled as community social capital. Within these networks lie
subjective aspects of social capital, such as perceptions of trust and
reciprocity. These subjective elements of social capital are referred
to as cognitive components. Almost orthogonal to the distinction
between structural and cognitive social capital, most conceptuali-
zations can be decomposed into “horizontal” and “vertical” com-
ponents (Islam et al., 2006). Horizontal social capital reflects ties
that exist among individuals or groups of equals or near-equals, and
vertical (linking) social capital refers to interactions across explicit,
formal or institutionalized power or authority gradients in society
such as relationships between pupils and teachers (Szreter and
Woolcock, 2004). Additional distinctions have been drawn within
horizontal social capital, namely “bonding” social capital (strong
ties within homogenous groups such as family members and
friends) and “bridging” social capital (weaker ties within hetero-
geneous groups such as neighbors and school personnel, and
formal or informal social participation) (Putnam, 2000). A third
distinction concerns the level of analysis e whether social capital is
treated as an individual attribute or as a collective resource
(Kawachi, 2006). The proliferation of multilevel techniques in
public health research is consistent with the notion that social
capital influences health across multiple nested contexts (Duncan
et al., 1996, 1998). Smoking among adolescents, for instance,
might not only be shaped by where they live, but also be an
interface of simultaneous influences of the family, school, and
community contexts (De Clercq et al., 2012; Morrow, 1999). A
limitation of traditional multilevel models is the requirement that
pupils are nested within schools, and schools in their turn, are
nested within communities. This rigid classification rarely corre-
sponds to actual populations, in which pupils from several neigh-
borhoods attend the same school or schools serve multiple
communities. Ignoring the cross-classification within a population
causes underestimation of the standard error of estimates. To our
knowledge, no previous research has examined the multilevel as-
sociation between multiple levels of social capital and smoking
allowing for a non-nested data structure.
Influenced by theory on income inequality and social network
integration (Wilkinson,1996), prior research has proposed the concept
of social capital as an explanatory pathway in social inequalities in
adult (Kawachi et al., 1997) and adolescent (Waterson et al., 2004)
health. Regarding inequalities in adolescent smoking, only two studies
have examined the role of individual (Evans and Kutcher, 2011) and
contextual (Thorlindsson et al., 2012) social capital. These studies
found that community social capital buffered the negative relation
between socio-economic status and smoking. Previous research also
found that high levels of neighborhood or community social capital
might weaken (flatten) the social gradient in adolescent health
(Aminzadeh et al., 2013; De Clercq et al., 2012; Vyncke et al., 2013).
The objectives of this study were to (1) estimate the amount of
variance in adolescent smoking that can be attributed to individual-
, school- and community-level factors, (2) evaluate the importance
of social capital at the individual-, school- and community-level for
explaining variance in smoking, and (3) investigate whether dif-
ferences in individual-, school- or community-level social capital
account for social inequalities in adolescent smoking.
2. Methods
2.1. Study participants
The sample consisted of 8453 adolescents nested within a cross-
classification of 167 schools and 570 communities (postal code
areas). The data were collected in the Flemish 2005e2006 Health
Behavior in School-aged Children (HBSC) survey. Self-completion
questionnaires were administered in school classrooms with re-
quirements in terms of sampling, questionnaire items and survey
administration being set out in a standardized research protocol
(for more details of the international used methodology and survey
design, see Currie et al. (2009) and Roberts et al. (2007)). In Flan-
ders, pupils from the 1st year (12 year) to the 6th year (18 year)
secondary school were questioned. A random sample of schools
was drawn from the official school list of Flanders, keeping into
account the size of the school (schools were weighted by the
number of pupils in the school so that every pupil has the same
chance to be in the sample). In total 336 schools were asked to
participate and 167 agreed (49.7%). Within the schools, 2e3 classes
were selected to participate keeping into account the distribution
of gender and the different educational levels in Flanders (general,
technical and vocational). The response on pupil level was 98.2%.
Passive informed consent was asked to the parents. The study was
approved by the ethics review committee of the University Hospital
of Ghent (project EC UZG2005/383).
2.2. Measures
2.2.1. Smoking
Smoking frequency was measured with the following question:
“How often do you smoke tobacco at present?” (1 ¼ every day,
2 ¼ at least once a week but not every day, 3 ¼ less than once a
week, 4 ¼ never). Adolescents who smoked at least once a week
were considered regular smokers (Richter and Leppin, 2007).
2.2.2. Parental and peer smoking
Many studies show that parental smoking and peer smoking are
important predictors of adolescent smoking (Kelly et al., 2011). This
was measured with the following question: “Does one of the
following people smoke?”, with separate response options for fa-
ther, mother, and best friend (1 ¼ every day, 2 ¼ occasionally,
3 ¼ never, 4 ¼ don't know, 5 ¼ don't have or see this person).
Persons who smoked at least occasionally were considered as
smokers.
2.2.3. Socio-economic covariates
The Family Affluence Scale (FAS) is a composite indicator of self-
reported socio-economic status comprising four items that address
family assets or conditions that indicate material wealth (Currie
et al., 2008): ‘Does your family own a car, van or truck? (0 ¼ no;
1 ¼ yes one; 2 ¼ yes two or more); Do you have your own bedroom
for yourself? (0 ¼ no; 1 ¼ yes); During the past 12 months, how
many times did you travel away on holiday with your family?
(0 ¼ not at all, 1 ¼ once, 2 ¼ twice, 3 ¼ more than twice); How many
computers does your family own?’ (0 ¼ none, 1 ¼ one, 2 ¼ two,
3 ¼ more than two). Responses are summed on a 1 to 10 scale with
higher scores indicating greater affluence. In addition, adolescents'
current education was also used as an indicator of socio-economic
status (coded as 1 ¼ general, 2 ¼ technical, 3 ¼ vocational) (Richter
and Leppin, 2007).
2.2.4. Structural social capital
Network social capital was measured by participation in clubs or
organizations: “Are you involved in any of these kinds of clubs or
organizations?” A sum score was calculated from the following
response categories: sports club, voluntary service, political orga-
nization, cultural organization, religious group, youth club, other
club (0 ¼ no, 1 ¼ yes) (De Clercq et al., 2012). Family social capital
was measured with two separate questions: “How often do you
have breakfast with your mother or father?”; “How often do you
 B. De Clercq et al. / Social Science & Medicine 119 (2014) 81e87 83

have dinner with your mother or father?” (each coded from Table 1
1 ¼ never to 6 ¼ every day). According to Coleman (Coleman, 1990), Rotated pattern matrix of cognitive social capital items in 8453 adolescents, the
Flemish Health Behavior in School-Aged Children (HBSC) survey, 2005e2006.
communication between family members is critically important in
the formation of basic rules and norms. Family items Factor

1
2.2.5. Cognitive social capital
When I have problems in school, my 0.80
We selected 14 survey questions reflecting cognitive school parents are willing to help me
social capital. Because some of these items could reflect both hor- My parent are willing to come to 0.72
izontal and vertical components of social capital an exploratory school to talk with the teachers
My parents encourage me to do well at 0.76
factor analysis was conducted using Mplus 7.11 to determine a
school
parsimonious set of factors in the data (Muthe
n and Muthe
n,
My parents are interested in what is 0.76
1998e2012). An oblique rotation of factors was performed allow- happening with me in school
ing factors to covary. The plot of eigenvalues suggested a four-factor My parents are willing to help me with 0.63
solution (Table 1): Horizontal (three items, related to fellow pupils), my homework
Eigenvalue 3.14
Vertical (four items, related to teachers), Trust (three items,
Factor determinacy 0.93
reflecting an aspect of trust within the school), and Participation
School items 1 2 3 4
(four items, reflecting involvement in decision making processes
within the school) (each coded from 1 ¼ strongly disagree to Horizontal Trust Vertical Participation
5 ¼ strongly agree). Values below 0.05 on the Root Mean Square My classmates like to be together 0.58 0.02 0.02 0.09
Error of Approximation (RMSEA) and a value of 0.90 or greater on Most classmates are friendly and 0.84 0.00 0.00 0.02
the Comparative Fit Index (CFI) were considered as indicative of a helpful
My classmates accept me as I am 0.65 0.05 0.03 0.03
good fit. With the criteria, the 14 item four-factor model showed
Pupils are involved in organizing 0.28 0.00 0.05 0.33
good fit to the data (c2(df¼41) ¼ 598.34, P < 0.01; CFI ¼ 0.99; school activities
RMSEA ¼ 0.03). Good internal consistency was found in each Pupils are involved in making school 0.03 0.02 0.03 0.52
subscale (Table 1). Alternative factor solutions were considered regulations
(results not shown) and evaluated on the basis of their efficiency Pupils have a say in which kind of 0.02 0.13 0.05 0.56
activities they do
using Akaike's Information Criterion (AIC) and the more conser- Pupils have a say in how class time is 0.00 0.02 0.03 0.63
vative Bayesian information criterion (BIC) (Burnham and used
Anderson, 2002). Compared to other factor solutions, the four- Our school is a nice place 0.00 0.68 0.11 0.08
factor model showed the lowest AIC and BIC values and therefore I feel home at school 0.00 0.90 0.07 0.01
I feel safe at school 0.05 0.60 0.09 0.01
best fit to the data. Family social capital consisted of a 5-item scale.
The teachers treat us fairly 0.02 0.03 0.72 0.02
Example items are “My parent are willing to come to school to talk If necessary, I get extra help 0.00 0.02 0.69 0.01
with the teachers; My parents encourage me to do well at school” The teachers are interested in who I 0.02 0.02 0.62 0.11
(coded from 1 ¼ strongly disagree to 5 ¼ strongly agree). As in am
previous work (De Clercq et al., 2012), community social capital is Most teachers are friendly 0.00 0.09 0.69 0.04
Eigenvalue 4.63 1.57 1.38 1.07
measured using a 5-item scale. Example items are “You can trust Factor determinacy 0.90 0.92 0.90 0.81
people around here; I could ask for help or a favor from neighbours”
Community items 1
(coded from 1 ¼ strongly disagree to 5 ¼ strongly agree). Both
scales are one-dimensional and showed high factor determinacy People say ‘hello’ and often stop to talk 0.63
supporting good internal reliability. Because all social capital in- to each other in the street
It is safe for younger children to play 0.58
dicators were originally measured on the basis of individual re- outside during the day
sponses, they automatically represent the individual dimension of You can trust people around here 0.73
social capital. The collective dimension, on the contrary, was There are good places to spend your 0.54
measured on the basis of aggregated individual indicators or scales free time
I could ask for help or a favor from 0.55
(school- or neighborhood-level mean scores).
neighbors
Eigenvalue 2.39
2.3. Statistical analysis Factor determinacy 0.87

Note. Coefficients represent factor loadings.

Given the complexity of cross-classified models, it is advisable to
carry out preliminary analysis using standard multilevel models,
leaving out each of the cross-classified factors in turn (Hox, 2010).
We first used a 2-level random intercept model, with adolescents as
level-1 units and communities as level-2 units. In a second model,
schools were used as the level-2 units. We found statistically sig-
nificant variance for each of these models, so we used the cross-
classified model to assess whether these variances at each level
were independent. Additionally, a 3-level model was fitted to
compare the overall efficiency of the hierarchical multilevel models
(2-level and 3-level) versus the non-hierarchical model (cross-
classified). Then we proceeded to sequentially estimate a series of
logit models. First, a variance partition coefficient (VPC) was
calculated to estimate the proportion of the variance of the
outcome between schools and communities (Goldstein et al.,
2002). Model 2 included socio-demographic variables (gender
and age). Model 3 was the same as model 2, but with additional
Bold coefficients represent factor loadings. Italics refer to information criteria
(Eigenvalue and Factor determinancy).
socio-economic variables (family affluence and adolescent educa-
tion). Model 4 accounted for parental and peer smoking behavior.
In a next step, all individual-level social capital variables were
entered simultaneously (model 5). This model estimated the
contribution of individual-level structural (network and family)
and cognitive (network, family, school and community) social
capital to smoking frequency. In order to distinguish between
purely contextual social capital associations with adolescent
smoking and compositional associations (Diez Roux, 1998), further
models consecutively entered contextual-level social capital vari-
ables on the condition that significant contextual-level variability
was observed. Finally, (cross-level) interactions with socio-
economic status indicators were computed for all social capital
variables.
84 B. De Clercq et al. / Social Science & Medicine 119 (2014) 81e87

We applied Bayesian inference to estimate the parameters, as showed an overestimation of the community-level variance
maximum likelihood based methods (Goldstein and Rasbash, 1996) (s2m0 ¼ 0:271 ¼ 0.271, P < 0.01 compared to s2m0 ¼ 0:033, not sig-
are inefficient in the context of cross-classified models (Browne, nificant in the cross-classified model) and a rather similar school-
2012). We used Markov Chain Monte Carlo (MCMC) simulation level variance (s2v0 ¼ 1:519¼1.519, P < 0.01 compared to
procedures in MLwiN 2.26 (Rasbash et al., 2009). Little is known s2v0 ¼ 1:526, P < 0.01 in the cross-classified model). Results from
about the model parameters a priori, and so our prior distributions the 3-level model showed an underestimation of the school-level
are required to express this lack of knowledge. We used Metropolis variance (s2v0 ¼ 1:387, P < 0.01 compared to s2v0 ¼ 1:526, P < 0.01
Hastings sampling with non-informative improper uniform priors in the cross-classified model) and a correct estimation of the
for the fixed effects and weakly informative uniform priors (derived community-level variance (s2m0 ¼ 0:032, not significant compared
from the iterative generalized least squares [IGLS] algorithm to s2m0 ¼ 0:033, not significant in the cross-classified model).
(Goldstein and Rasbash, 1996)) for the between-school and Overall, MCMC diagnostics from the cross-classified model showed
between-community variances. The Raftery-Lewis diagnostic was that keeping in the neighborhood cluster-set was inefficient.
used to monitor the length of the MCMC chain required for Therefore, further modeling was reduced to the usual 2-level
convergence after a burn-in of 5000 simulations. Variance com- model.
ponents estimates are reported with their standard errors (SEs) and Table 4 presents the regression coefficient estimates from the
p-values. Model coefficient estimates were converted into odds sequential multilevel models. Variance partition calculations indi-
ratios (ORs) with 95% credible intervals (CIs). We used the Deviance cated that about one third of the variability in regular smoking was
Information Criterion (DIC) to test the improvement of fit for each attributable to contextual factors, and thus school related differ-
model (Browne, 2012). A formal description of the cross- ences (r ¼ 32%) since between community differences were
classification procedure can be found elsewhere (Rasbash and negligible. Controlling for socio-demographic differences (model
Goldstein, 1994). 2), model 3 showed no association between family affluence and
regular smoking. However, strong educational differences in reg-
3. Results ular smoking were found: compared to pupils in general education
both technical (OR ¼ 1.737; 95% CI ¼ 1.411, 2.138) and vocational
Table 2 shows the descriptive characteristics of the sample. (OR ¼ 3.651; 95% CI ¼ 2.920, 4.565) pupils reported higher odds for
Respondents (n ¼ 8453) were secondary school pupils between 12 regular smoking. The school-level variance (i.e. the random inter-
and 18 years old (Mean ¼ 15 years) and both boys and girls were cept) was reduced by 59% after adjusting for individual socio-
proportionally distributed in the overall sample. About one third of demographic characteristics (model 2) and additionally by
their fathers, mothers, and peers were smokers, and 17.7% were another 50% after adjusting for individual socio-economic variables
regular smokers themselves. Table 3 presents the variance com- (model 3), which means that individual socio-demographic and
ponents estimates from the different multilevel models. The cross- socio-economic parameters accounted for 80% of the variability
classified model shows that the between school variance in regular between schools. As expected, both parental and peer smoking
smoking was different from 0 at the 0.01 significance level were strongly and positively related to adolescent regular smoking
ðs2v0 ¼ 1:526Þ and the between community variance was not sig- and explained about two third of the remaining school variance
nificant ðs2m0 ¼ 0:033Þ. Results from the 2-level models clearly (model 4). These variables acted as covariates in the subsequent
models to control for compositional differences. Model 5 included
all individual-level social capital variables rendering the between-
Table 2 school variance nonsignificant. Only structural family social capi-
Descriptive characteristics of 8453 adolescents, the Flemish Health Behavior in
tal and cognitive school social capital variables (vertical, horizontal
School-Aged Children (HBSC) survey, 2005e2006.
and trust dimension) significantly related to regular smoking.
Socio-demographics Vertical school social capital inversely related to regular smoking
Gender
(OR ¼ 0.764; 95% CI ¼ 0.678, 0.861), while horizontal school social
Boy % 49.2
Girl % 50.8 capital positively related to regular smoking (OR ¼ 1.189; 95%
Age mean (SD), range ¼ 12e18 15.2 (1.90) CI ¼ 1.076, 1.314). Since the between-school variance turned
Socio-economics nonsignificant after controlling for model 5, no further models
Family affluence mean (SD), range ¼ 1e10 6.55 (1.76) including school-level variables were developed. Finally, in-
Education
General % 50.5
teractions with socio-economic status indicators were tested for all
Technical % 28.4 social capital variables, but none were significant (results not
Vocational % 21.1 shown).
Parental & peer smoking
Father smokes % 33.2
Mother smokes % 27.5
Best friend smokes % 33.1
Structural social capital
Network social capital mean (SD), range ¼ 0e7 1.12 (0.98) Table 3
Family social capital Variance components estimates for the multilevel models in 8453 adolescents, the
Breakfast with parents mean (SD), range ¼ 1e6 3.57 (1.78) Flemish Health Behavior in School-Aged Children (HBSC) survey, 2005e2006.
Dinner with parents mean (SD), range ¼ 1e6 5.16 (1.25)
Cognitive social capital Non-hierarchical Hierarchical models
Family social capital mean (SD), range ¼ 1e5 4.16 (0.72) model
School social capital
Cross-classified 2-Level 3-Level
Horizontal mean (SD), range ¼ 1e5 3.84 (0.81)
Vertical mean (SD), range ¼ 1e5 3.54 (0.72) School-level variance s2v0 1.526 (0.236)* 1.519 (0.237)* 1.387 (0.174)
Trust mean (SD), range ¼ 1e5 3.34 (0.89) *
Participation mean (SD), range ¼ 1e5 2.81 (0.75) Neighborhood-level 0.033 (0.025)n.s. 0.271 (0.050)* 0.032
Community social capital mean (SD), range ¼ 1e5 3.75 (0.62) variance s2m0 (0.033)n.s.
Regular smoking % 17.7
Note. Figures in parentheses represent standard errors (SE).
Note. SD ¼ standard deviation. *P < 0.01.
 B. De Clercq et al. / Social Science & Medicine 119 (2014) 81e87 85

Table 4
Fixed and random parameters of the multilevel smoking model in 8453 adolescents, the Flemish Health Behavior in School-Aged Children (HBSC) survey, 2005e2006.

Model 1 Model 2 Model 3 Model 4 Model 5

b (SE) b (SE) b (SE) b (SE) b (SE)

Fixed parameters
Constant 1.884 (0.102)* 1.881 (0.083)* 2.325 (0.095)* 3.740 (0.109)* 3.728 (0.113)*
Individual-level
Socio-demographics
Girl 0.083 (0.070) 0.090 (0.073) 0.148 (0.079) 0.149 (0.081)
Age 0.436 (0.026)* 0.406 (0.025)* 0.328 (0.025)* 0.310 (0.026)*
Socio-economics
Family affluence 0.030 (0.019) 0.017 (0.022) 0.004 (0.023)
Education (ref: general)
Technical 0.552 (0.106)* 0.321 (0.106)* 0.236 (0.105)*
Vocational 1.295 (0.114)* 0.900 (0.111)* 0.811 (0.108)*
Parental & peer smoking
Father smokes 0.311 (0.078)* 0.284 (0.083)*
Mother smokes 0.752 (0.079)* 0.688 (0.082)*
Best friend smokes 2.442 (0.082)* 2.392 (0.087)*
Structural social capital
Network social capital 0.010 (0.039)
Family social capital
Breakfast with parents 0.098 (0.025)*
Dinner with parents 0.113 (0.032)*
Cognitive social capital
Family social capital 0.060 (0.053)
School social capital
Horizontal 0.173 (0.051)*
Vertical 0.269 (0.061)*
Trust 0.218 (0.054)*
Participation 0.043 (0.058)
Community social capital 0.004 (0.065)
School-level
e e e
Random parameters
s2v0 (school) 1.519 (0.237)* 0.621 (0.110)* 0.310 (0.068)* 0.106 (0.044)* 0.044 (0.031)
DIC 6913.671 6705.187 6283.572 4812.891 4420.155

Note. Figures in parentheses represent standard errors (SE); *P < 0.05.

4. Discussion minor importance for the prediction of smoking patterns is in line

The cross-classified model showed significant variation in
adolescent regular smoking between schools, but not between
communities. Traditional hierarchical two-level (Aslund and
Nilsson, 2013; Thorlindsson et al., 2012) and three-level studies
(Kelly et al., 2011) came to different conclusions since all found
significant community-level variability in adolescent smoking.
Sources of these mixed findings could be the different study
context of Flemish, Icelandic (Thorlindsson et al., 2012), Swedish
(Aslund and Nilsson, 2013) or Australian (Kelly et al., 2011) ado-
lescents, and also the size of the neighborhood cluster scale that
differed considerably between studies. However, it is plausible that
observed community-level associations with adolescent smoking
may be a consequence of unmeasured confounding. Results from
our 2-level models clearly showed an overestimation of the
community-level variance and a slight underestimation of the
school-level variance. Such evidence suggests that 2-level studies
may overestimate the contribution of the community environment
if they disregard other contexts potentially relevant to the popu-
lation and outcome under study. Results from the 3-level model
showed an underestimation of the school-level variance. Further-
more, ignoring the cross-classification effect underestimated the
standard error of the estimates. In sum, these findings support the
need for a cross-classified multilevel approach.
Differences in adolescent regular smoking behavior were mostly
explained by individual characteristics. Accordingly, previous
research has shown that interschool variance in daily smoking is
mainly attributed to differences in the composition of schools
(Aveyard et al., 2005). The finding that family affluence is only of
with other research. As in our study, Richter and Leppin (Richter
and Leppin, 2007) found that especially pupils' educational level
had a major impact on regular smoking. Furthermore, Huisman
et al. (Huisman et al., 2012) found that adolescents' social back-
ground effects on daily smoking were almost fully explained by
educational enrollment. Thus, in accordance with our finding, the
relevance of educational tracking for adolescent smoking seems to
support the homogenizing effects within schools in adolescent risk
behavior (Richter et al., 2009). On individual level, our findings are
in line with studies on the positive role of social capital for
adolescent health (McPherson et al., 2013) as well as hypotheses on
the dark side of social capital (Portes, 1998). We found that it was
not so much the emotional embeddedness (cognitive social capital)
within the family, but social relations (structural social capital) that,
according to Due et al. (Due et al., 2003), involve the ‘paren-
techildeschool triad’. In accordance with our findings on structural
family social capital, research has shown that having meals with
parents is a fundamental aspect of family live that offers the op-
portunity of socialization that include learning of healthy behaviors
(Coleman, 1990; Ja€rvelaid, 2004). Whereas horizontal ties between
pupils increased the likelihood of regular smoking, adolescents'
trust towards schools and their relation to teachers (vertical social
capital) decreased the likelihood of regular smoking. Similar
studies reported that negative perceptions about the school envi-
ronment and about teachers support act as fundamental de-
terminants of health compromising behaviors (McLellan et al.,
1999).
Especially on school level, vertical social capital represents the
overall cohesion between pupils and teachers at school and
86 B. De Clercq et al. / Social Science & Medicine 119 (2014) 81e87
teachers' opportunity to interact with pupils and to control their
behavior directly. Our final model was subjected to some additional
simulation analysis and we found a truly contextual association (i.e.
schools with higher levels of vertical social capital had lower rates
of daily smoking) from the moment we discarded school-
eneighborhood combinations that were only applicable for a
maximum of five adolescents (results not shown). The finding was
quite robust since it could be replicated in both 2-level and cross-
classified models using the standard IGLS algorithm (Goldstein
and Rasbash, 1996) as well as Bayesian procedures (Browne,
2012). Such evidence points out specific issues of cross-
classification analytic structures besides general multilevel prob-
lems such as the ecological fallacy (Robinson, 1950), the modifiable
areal unit problem (MAUP) (Openshaw, 1984), and endogeneity
(Kawachi and Subramanian, 2007). To our knowledge, no previous
study has investigated the contextual association of different forms
of social capital with adolescent smoking. Takakura (2011) only
analyses the relation between social trust and prevalence of
smoking and found inconclusive results as school-level trust was
negatively associated with smoking only for girls. Furthermore,
similar findings are only available for school-related outcomes
(Goddard, 2003) and aspects of adolescents' well-being (Ellonen
et al., 2008). Thus, further research is needed to identify support
findings on the association of different forms of school social capital
with smoking and inequalities in smoking.
In this context, no interactions between socio-economic status
and individual-, school- or community-level social capital variables
were found indicating that social capital could not explain social
inequalities in adolescent regular smoking behavior. These results
are supported by our findings to the homogenizing effects of
educational tracking on smoking, but are in contrast with tradi-
tional single-level (Evans and Kutcher, 2011) and 2-level
(Thorlindsson et al., 2012) neighborhood studies.
4.1. Strengths and limitations
Among the strengths of this study was its comprehensive
assessment of different forms of social capital on multiple levels of
analysis. The use of an appropriate statistical method that allowed
modeling and testing of the cross-classification in neighborhoods
and schools strengthened our conclusions. The adopted method-
ology gives more accurate estimates of standard errors for the
estimated coefficients and shows how random variance is distrib-
uted among different levels and how it decreases when adding
different sets of level-specific variables of interest.
Some limitations in this study warrant comment. Communities
were operationalized as postal code areas, but smaller levels of
aggregation or spatial scales using physical distances between
residents might have tapped between community variability in
adolescent smoking (Takagi, 2013). Moreover, the research design
of the study (i.e. sampling at school-level) is biased against com-
munity effects because the sample consisted of 8453 adolescents
from 167 schools and 570 communities. Such a design undermines
community-level statistical power in that there were on average 51
cases in each school but only 15 cases in each community, which
would present stronger within-school correlation than within-
community correlation. Third, the problem of endogeneity
(Kawachi and Subramanian, 2007) in the context of the present
dataset was discussed extensively elsewhere (De Clercq et al.,
2012). Also in cross-classified multilevel models the econo-
metrical partition of variance between contexts is somewhat arti-
ficial. The community where adolescents lived, is likely to
determine the supply of schools to which they can attend. The types
of schools that are available are likely to determine the types of
families that move into the local area. Thus, neighborhoods and
schools are connected in various ways that are difficult to measure.
Other analytic approaches such as fixed-effects analyses and
instrumental variable (IV) estimation are needed to address the
endogeneity of social capital (Kawachi et al., 2013). Fourth, the
substantial meaning of the cognitive family social capital scale may
be more relevant for academic achievement, rather than health,
since two items uniquely refer to school performance (Coleman,
1988). Lastly, the present study conceptualized social capital
mainly as an ecological resource. This “communitarian perspective”
shifted the focus of public health research on social capital away
from an analytical focus on social networks and ties (Moore et al.,
2005). The distinction between social capital as an ecological
resource and social networks as an individual level attribute
already runs through the early public health literature on social
capital, as researchers focused on place effects (Diez Roux, 1998;
Macintyre et al., 2002) and began to use multilevel analysis
(Duncan et al., 1996, 1998). Future studies should measure more
structural properties of social capital as for example the density of
adolescents' networks (both at school and in the neighborhood),
number of friends, parents' social network, structural components
of parenteschool interactions (e.g. parental participation in advi-
sory organization), etc. However, this was not possible in the pre-
sent study due to lack of data. Further research is also needed to
develop an integrated model that identifies the mechanisms that
underlie various forms of social capital in their impacts on
adolescent smoking.
5. Conclusion
Our findings emphasize the importance of school as a social
context for adolescent regular smoking. A common line of argu-
mentation supports this view (Takakura, 2011): schools are (i) the
places where adolescents spend most of their daytime, and thus (ii)
the main place where they interact with their friends and teachers,
which (iii) shapes both their social relations and behavioral pat-
terns, and moreover (iv) relatively stronger than larger environ-
ments such as neighborhoods. Therefore, it is plausible that school
can be an important source of social capital and a useful unit of
analysis to study social capital for pupils. Moreover, previously
observed community-level associations with adolescent smoking
may be a consequence of unmeasured confounding. Distinguishing
between types of social capital is critical because their associations
with health are different.
Funding
This work was supported by the Flemish government.
The HBSC study
HBSC is an international study carried out in collaboration with
WHO Europe. The International Coordinator of the 2005e6 survey
was Professor Candace Currie from the University of Edinburgh and
the Data Bank Manager was Dr. Oddrun Samdal from the University
of Bergen. The 2005e6 survey was conducted by Principal In-
vestigators in 41 countries. For details, see http://www.hbsc.org.
Appendix A. Supplementary data
Supplementary data related to this article can be found at http://
dx.doi.org/10.1016/j.socscimed.2014.08.018.
 B. De Clercq et al. / Social Science & Medicine 119 (2014) 81e87 87

References Kawachi, I., 2006. Commentary: social capital and health e making the connections
one step at the time. Int. J. Epidemiol. 35, 989e993.
Kawachi, I., Berkman, L.F., 2001. Social ties and mental health. J. Urban Health 78,
Aminzadeh, K., Denny, S., Utter, J., Milfont, T.L., Ameratunga, S., Teevale, T., et al.,
458e467.
2013. Neighborhood social capital and adolescent self-reported wellbeing in
Kawachi, I., Ichida, Y., Tampubolon, G., Fujiwara, T., 2013. Causal inference in social
New Zealand: a multilevel analysis. Soc. Sci. Med. 84, 13e21.
capital research. In: Kawachi, I., Takao, S., Subramanian, S.V. (Eds.), Global
Aslund, C., Nilsson, K.W., 2013. Social capital in relation to alcohol consumption,
Perspectives on Social Capital and Health. Springer, New York.
smoking, and illicit drug use among adolescents: a cross-sectional study in
Kawachi, I., Kennedy, B.P., Lochner, K., Prothow-Stith, D., 1997. Social capital, social
Sweden. Int. J. Equity Health 12.
inequality, and mortality. Am. J. Public Health 87, 1491e1498.
Aveyard, P., Markham, W.A., Lancashire, E., Almond, J., Griffiths, R., Cheng, K.K.,
Kawachi, I., Subramanian, S.V., 2007. Neighbourhood influences on health.
2005. Is inter-school variation in smoking uptake and cessation due to differ-
J. Epidemiol. Community Health 61, 3e4.
ences in pupil composition? A cohort study. Health Place 11, 55e65.
Kelly, A.B., O'Flaherty, M., Connor, J.P., Hommel, R., Toumbourou, J.W., Patton, G.C.,
Blane, D., Hart, C.L., Davey-Smith, G., Gillis, C.R., Hole, D.J., Hawthorne, V.M., 1996.
et al., 2011. The influence of parents, siblings and peers on pre- and early-teen
The association of cardiovascular risk factors with socioeconomic position
smoking: a multilevel model. Drug Alcohol Rev. 30, 381e387.
during childhood and adulthood. BMJ 313, 1434e1438.
Lindstrom, M., 2003. Social capital and the miniaturization of community among
Browne, W.J., 2012. MCMC Estimation in MLwiN, v2.26. University of Bristol: Centre
daily and intermittent smokers: a population-based study. Prev. Med. 36,
for Multilevel Modelling.
177e184.
Burnham, K.P., Anderson, D.R., 2002. Model Selection and Multimodel Inference: A
Lovato, C.Y., Zeisser, C., Campbell, H.S., Watts, A.W., Halpin, P., Thompson, M., et al.,
Practical Information-theoretic Approach, second ed. Springer, New York.
2010. Adolescent smoking: effect of school and community characteristics. Am.
Centers for Disease Control and Prevention, 2008. Youth behavior surveillance e
J. Prev. Med. 39, 507e514.
United States, 2007. MMWR Surveill. Summ. 57 (SSe4).
Lundborg, P., 2005. Social capital and substance use among Swedish adolescents e
Chuang, Y.-C., Chuang, K.-Y., 2008. Gender differences in relationships between
an explorative study. Soc. Sci. Med. 61, 1151e1158.
social capital and individual smoking and drinking behavior in Taiwan. Soc. Sci.
Macintyre, S., Ellaway, A., Cummins, S., 2002. Place effects on health: how can we
Med. 67, 1321e1330.
conceptualise, operationalise and measure them? Soc. Sci. Med. 55, 125e139.
Coleman, J.S., 1988. Social capital in the creation of human capital. Am. J. Sociol. 94,
McLellan, L., Rissel, C., Donnelly, N., Bauman, A., 1999. Health behaviour and the
95e120.
school environment in New South Wales, Australia. Soc. Sci. Med. 49, 611e619.
Coleman, J.S., 1990. Foundations of Social Theory. Harvard University Press, Cam-
McPherson, K., Kerr, S., McGee, E., Cheater, F., Morgan, A., 2013. The Role and Impact
bridge, MA.
of Social Capital on the Health and Well-being of Children and Adolescents: A
Curran, E.M., 2007. The relationship between social capital and substance use by
Systematic Review. Glasgow Centre for Population Health, Glasgow.
high school students. J. Alcohol Drug Educ. 51, 59e73.
Moore, S.H., Shiell, A., Hawe, P., Haines, V.A., 2005. The priviledging of communi-
Currie, C., Molcho, M., Boyce, W., Holstein, B., Torsheim, T., Richter, M., 2008.
tarian ideas: citation practices and the translation of social capital into public
Researching health inequalities in adolescents: the development of the Health
health research. Am. J. Public Health 95, 1330e1337.
Behaviour in School-Aged Children (HBSC) family affluence scale. Soc. Sci. Med.
Morgan, A., Haglund, B.J.A., 2009. Social capital does matter for adolescent health:
66, 1429e1436.
evidence from the English HBSC study. Health Promot. Int. 24, 363e372.
Currie, C., Nic Gabhainn, S., Godeau, E., Committee, t.I.H.N.C, 2009. The health
Morrow, V., 1999. Conceptualising social capital in relation to the well-being of
behaviour in school-aged children: WHO collaborative cross-national (HBSC)
children and young people: a critical review. Sociol. Rev. 47, 744e765.
study: origins, concept, history and development 1982e2008. Int. J. Public
Muthe
n, L.K., Muthe
n, B.O., 1998e2012. Mplus User's Guide, seventh ed. Muthe
n &
Health 54, 131e139.
Muthe
n, Los Angeles, CA.
De Clercq, B., Vyncke, V., Hublet, A., Elgar, F.J., Ravens-Sieberer, U., Currie, C., et al.,
Openshaw, S., 1984. The Modifiable Areal Unit Problem. Geobooks, Norwich.
2012. Social capital and social inequality in adolescents' health in 601 Flemish
Portes, A., 1998. Social capital: its origins and applications in modern sociology.
communities: a multilevel analysis. Soc. Sci. Med. 74, 202e210.
Annu. Rev. Sociol. 24, 1e24.
Diez Roux, A.V., 1998. Bringing context back into epidemiology: variables and fal-
Putnam, R., 2000. Bowling Alone. The Collapse and Revival of American Community.
lacies in multilevel analysis. Am. J. Public Health 88, 216e222.
Simon & Shunster, New York.
Due, P., Lynch, J., Holstein, B., Modvig, J., 2003. Socioeconomic health inequalities
Rasbash, J., Charlton, C., Browne, W.J., Healy, M., Cameron, B., 2009. MLwiN Version
among a nationally representative sample of Danish adolescents: the role of
2.1. Centre for Multilevel Modelling, University of Bristol.
different types of social relations. J. Epidemiol. Community Health 57, 692e698.
Rasbash, J., Goldstein, H., 1994. Efficient analysis of mixed hierarchical and cross-
Duncan, C., Jones, K., Moon, G., 1996. Health-related behaviour in context: a
classified random structures using a multilevel model. J. Educ. Behav. Stat. 19,
multilevel modelling approach. Soc. Sci. Med. 42, 817e830.
337e350.
Duncan, C., Jones, K., Moon, G., 1998. Context, composition and heterogeneity: using
Richter, M., Leppin, A., 2007. Trends in socio-economic differences in tobacco
multilevel models in health research. Soc. Sci. Med. 1, 97e117.
€€ €inen, J., Autio, V., 2008. Adolescent depression and school social smoking among German schoolchildren, 1994e2002. Eur. J. Public Health 17,
Ellonen, N., Ka aria
565e571.
support: a multilevel analysis of a Finnish sample. J. Community Psychol. 36,
Richter, M., Vereecken, C.A., Boyce, W., Maes, L., Gabhainn, S.N., Currie, C.E., 2009.
552e567.
Parental occupation, family affluence and adolescent behaviour in 28 countries.
European Commission (2013). http://ec.europa.eu/health-eu/my_lifestyle/tobacco/
Int. J. Public Health 54, 203e212.
index_en.htm. (accessed 10.03.14).
Roberts, C., Candace, C., Samdal, O., Currie, D., Smith, R., Maes, L., 2007. Measuring
Evans, G.W., Kutcher, R., 2011. Loosening the link between childhood poverty and
the health and health behaviours of adolescents through cross-national survey
adolescent smoking and obesity: the protective effects of social capital. Psychol.
research: recent developments in the Health Behavior in School-aged Children
Sci. 22, 3e7.
(HBSC) study. Int. J. Public Health 15, 179e186.
Galea, S., Nandi, A., Vlahov, D., 2004. The social epidemiology of substance use.
Robinson, W.S., 1950. Ecological correlations and the behaviour of individuals. Am.
Epidemiol. Rev. 26, 36e52.
Sociol. Rev. 15.
Goddard, R.D., 2003. Relational networks, social trust, and norms: a social capital
Szreter, S., Woolcock, A., 2004. Health by association? Social capital, social theory
perspective on students' chances of academic Success. Educ. Eval. Policy Anal.
and the political economy of public health. Int. J. Epidemiol. 33, 650e667.
25, 59e74.
Takagi, D., 2013. Neighborhood social capital and crime. In: Kawachi, I., Takao, S.,
Goldstein, H., Rasbash, J., 1996. Improved approximations for multilevel models
Subramanian, S.V. (Eds.), Global Perspectives on Social Capital and Health.
with binary responses. J. R. Stat. Soc. 159, 505e513.
Springer, New York, pp. 143e166.
Goldstein, H., Rasbash, J., Browne, W.J., 2002. Partitioning variation in multilevel
Takakura, M., 2011. Does social trust at school affect students' smoking and drinking
models. Underst. Stat. 1, 223e231.
behavior in Japan? Soc. Sci. Med. 72, 299e306.
Harpham, T., Grant, E., Thomas, E., 2002. Measuring social capital within health
Thorlindsson, T., Valdimarsdottir, M., Hrafn Jonsson, S., 2012. Community social
surveys: key issues. Health Policy Plan. 17, 106e111.
structure, social capital and adolescent smoking: a multi-level analysis. Health
Henderson, M., Ecob, R., Wight, D., Abraham, C., 2008. What explains between-
Place 18, 796e804.
school differences in rates of smoking? BMC Public Health 8, 218.
Vyncke, V., De Clercq, B., Stevens, V., Ieven, A., Hrafn Jonsson, S., Darias Curvo, S.,
Hox, J., 2010. Multilevel Analysis. Techniques and Applications. Routledge, New
et al., 2013. Does neighborhood social capital make a difference for levelling the
York.
health gradient in children? A literature review. BMC Public Health 13.
Huisman, C., Werfhorst, H.G., Monshouwer, K., 2012. Adolescent tobacco use in the
Waterson, T., Alperstein, G., Stewart Brown, S., 2004. Social capital: a key factor in
Netherlands: social background, education, and school organization. Youth Soc.
child health inequalities. Arch. Dis. Child. 89, 456e459.
44, 567e586.
Wilkinson, R.G. (Ed.), 1996. Unhealthy Societies. The Afflications of Inequality.
Islam, M.K., Merlo, J., Kawachi, I., Lindstrom, M., Gerdtham, U.G., 2006. Social capital
Routledge, London.
and health: does egalitarianism matter? A literature review. Int. J. Equity Health
Winkleby, M.A., Jatulis, D.E., Frank, E., Fortmann, S.P., 1992. Socioeconomic status
5 (5), 3.
€ and health: how education, income, and occupation contribute to risk factors
Jarvelaid, M., 2004. Adolescent tobacco smoking and associated psychosocial health
for cardiovascular disease. Am. J. Epidemiol. 82, 816e820.
risk factors. Scand. J. Prim. Health Care 22, 50e53.