Multiple socio-economic contexts during adolescence and health - - PDF document

1 Multiple socio-economic contexts during adolescence and health behaviors in early adulthood Nicoletta Balbo Dondena Centre, Bocconi University, Milan, Italy Nicola Barban Nuffield College, Oxford University, Oxford, UK Frank F. Furstenberg University of Pennsylvania, Pennsylvania, U.S. Abstract This study aims to investigate whether health behaviors of young adults are influenced by different socio-economic contexts, namely family, friends and school, in which they grow up during adolescence. Existing literature has shown no consensus on how socio-economic status may influence adolescent health behaviors. Moreover, most of these studies have looked at only one environment

• f an adolescent’s socio-economic context at time (e.g., either the family, school, or neighborhood).

While focusing on health behaviors during the transition to adulthood, we extend existing literature by adopting an ecological, developmental approach. We take into account that the socio- economic context in which an adolescent grows up is composed by different environments that jointly influence an individual’s development and behaviors. Therefore, we examine how the socio- economic status of family, friends and school community during adolescence may simultaneously affect smoking, drinking and marijuana use from adolescence to early adulthood.

2 Using the National Longitudinal Study of Adolescent Health in the United States, we find that the socio-economic status of an adolescent’s family, friends and school are not highly correlated. This implies that these three contexts may be differently associated with specific health behaviors. Our multilevel analyses show that adolescents are more likely to smoke and drink if they come from more disadvantaged families, although this effect does not seem to last when the individual becomes an adult. Conversely, marijuana use in young adulthood is positively associated with the socio-economic status of the family of origin and the school. We also find that those individuals who come from higher educated families but did not reach the same educational level of their parents are those more likely to smoke and make use of marijuana.

3 Introduction There is an extensive literature on socioeconomic disparities and health behaviors coming from different fields of research, such as sociology, psychology and medical sciences (e.g., Maralani, 2013; Pampel et al., 2010; Siahpush et al., 2006; Sobal and Stunkard, 1989). The importance of uncovering socioeconomic differences in health behaviors stems from the fact that poor health behaviors are associated with increased morbidity and mortality. It has been estimated that such behaviors account for around one-quarter of health disparities, that is differences in health outcomes by socioeconomic groups (Pampel et al., 2010). Existing literature mostly focuses on the association between socio-economic status (SES) and diverse health behaviors, both (i.e., SES and health behaviors) measured either during childhood, adolescence or adulthood. The association between SES and health behaviors is shown not to be equally clear across the lifespan (Chen et al., 2002; West et al., 1990; West, 1988). While, for instance, it seems to be well-recognized that low SES adults are more likely to engage in unhealthy behaviors (Adler et al., 1994; Cutler and Lleras-Muney, 2010), there is no consensus on the relationship between SES and healthy behaviors among adolescents. Sometimes a positive relationship is found, some other studies show a negative or null effect, also depending on the health behavior under study (Hanson and Chen, 2007; Luthar, 2003). Moreover, Daw et al. (2016) have shown that young adults with low parental education are not significantly more likely to engage in unhealthy behaviors than those with high parental education. Some recent research suggests that childhood and adolescence socio-economic circumstances may be important determinant for health outcomes later in life (Cohen et al., 2010; Galobardes et al., 2008; Hayward and Gorman 2004; Luo and Waite 2005; Zimmer et al., 2016). There are several theoretical frameworks explaining the link between early-life advantages or disadvantages and later-life health, the majority of which claims that the effect of early-life SES can be reversible over time (see Cohen et al., 2010 and Pudrovska and Anikputa, 2014 for a review on the topic). Building

4

• n and extending this literature that mainly look at health outcomes only, we adopt a life-course,

developmental approach (Brandt et al., 2012; Di Gessa et al., 2016) focusing on health behaviors. We aim at uncovering whether and how the socio-economic status of family, friends and school community during adolescence may affect substance use, and more specifically smoking, drinking and marijuana use, from adolescence to early adulthood, when individuals are 15, around 20 and around 30 years-old. Most of existing research on social class gradient and health behaviors during adolescence has looked at only one dimension of an adolescent’s socio-economic context at time (e.g., either the family, school, friends, or neighborhood. Hanson and Chen, 2007). However, some developmental theories claim that the socio-economic context in which a child and then an adolescent grows up is composed by different environments that jointly work on an individual’s development (Bronfenbrenner, 1979; Cook et al., 2002). The ecological approach to child development (Bronfenbrenner, 1979) specifically maintains that an individual’s development is influenced not

• nly by the family system but also by other interconnected environments, that are nested one within

the other. In this way, there of course are more proximal and distal systems, with the former ones (such as family, peers and school) more influential than the latter ones (such as cities, regions and nations). Inspired by such ecological, developmental prospective (Bronfenbrenner, 1979; Cook, 2003), we aim at analyzing the relationship between SES during adolescent and health behaviors in young adulthood while taking into account that young people are part of a multi-level social context. More specifically, we focus on the link between the SES of the most proximal - and thereby most influential – environments in which an adolescent is embedded, namely family, friends and school, and his or her smoking, drinking and marijuana use behaviors in early adulthood. Some studies focusing on physical and psychological health outcomes, such as blood pressure

• r self-esteem, have shown the importance of adopting an ecological approach that takes into

5 account the fact that SES is a multidimensional construct, measurable at multiple levels (Chen and Miller 2013; Chen and Paterson, 2006; Cohen et al., 2010; Schreier and Chen 2013). This stream of research aims at understanding which specific level of SES indicator is significantly associated with a given health outcomes, while taking into account that each context may interact one with another. Findings show that not only family SES but also neighborhood SES is relevant for psychological and physical health outcomes (Chen and Paterson, 2006; Chuang et al., 2005; Warner, 2016). Following this literature, the present paper aims at testing the relative role of different socio- economic contexts (i.e. family, friends and school) as a way to better understand the link between SES and health behaviors. We focus on health behaviors in young adulthood and not on health

• utcomes, because we are particularly interested in shedding further light on the risk and protective

factors associated with the well-documented phenomenon of the peak in substance use in the twenties (Chassin et al., 2009; Johnston et al., 2011). Previous studies have shown that substance use can lead to several problems for young adults, such as difficulties in school, in the labor market, and problems with the criminal justice system (Humensky, 2010). We are aware that the social psychology literature (Chen and Miller 2013; Cook et al., 2002; Schreier and Chen 2013) suggests that the SES in each social context works via specific mechanisms and pathways that should be measured using specific indicators, such as social capital

• r violence indicators at the neighborhood/school level or parenting indicators at the family level.

However, measuring all of the theoretical mechanisms that have been theorized as potential pathways to health would make our models too complex and not readable. Moreover, such an analysis would require a very big sample size as well as very specific and effective measures for each mechanism, and not all of them are available in secondary data. Therefore, we decided to use a unique, synthetic SES measure computed at each of the three levels under study. We believe that such a strategy may help clarifying the relative importance of each socio-economic environment for a specific health behavior. We use parental education as the SES indicator for each of the social

6 context in which an adolescent is embedded, that is family, friends and school. Of course, to measure friends and school SES, an average of the parental education of all of the members in that community is used. As previous literature on the topic shows (Patrick et al., 2012), the education indicator alone should be sufficient and good enough to capture the SES-substance use linkages. Moreover, by using the same consistent indicator in each context, we can better understand whether focusing only on the SES of one context at time, as the majority of existing literature does (Botticello, 2009; Patrick et al., 2012), may lead only to a partial understanding of the link between adolescent’s SES and young adult’s health behaviors. Put in another way, we can uncover whether we miss out part of the link between SES and health behaviors when we use a single SES indicator measured in only one social context. According to the work of Cook and colleagues (2002), the between-context correlation of family, friends and school community is however rather modest. This finding is according to us very relevant, because a loose contextual coupling may further imply that the different socio- economic environments in which an adolescent is embedded may be differently associated with his/her health behaviors later in life. Building on that, we expect to find not a high correlation between the three contexts’ SES under study, fact that would make it possible to examine the singular as well as the joint association of family’s, friends’ and school’s SES and substance use. Therefore, the empirical starting point of the present study is to test whether such a modest between-context correlation is found in our data as well, that is The National Longitudinal Study of Adolescent Health (Add Health) in the United States. Among the arguments that have been pointed out as possible causes of a positive association between SES and healthy behaviors, there are those related to higher motivations for higher-SES individuals, such as more benefits or better knowledge of potential risks, and those related to better means or more resources to reach health goals (Pampel et al., 2010). On the other hand, there are studies showing that, among adolescents and young adults, high SES might be associated with

7 higher probability of unhealthy behaviors (Huerta and Borgonovi, 2010; Humensky, 2010; Luthar 2003; Luthar and Latendresse, 2005; Patrick et al., 2012). Potential explanatory mechanisms for the latter relationship are the fact that affluent individuals are more subject to school/job performance- related stress, or experience more exposure to substance-using peers, or live in family context with less parental supervision, and more permissive parental attitudes, or alternatively have more resources to be spent in substances. Looking at the potential mechanisms highlighted by the existing literature to explain the SES- health behaviors link, it is clear that we need to have a comprehensive view of all of the socio- economic contexts in which an adolescent grows up. Socialization processes and agents outside the kinship network are especially significant in adolescence as child move out of the exclusive domain

• f family influence. School environment and peer groups become increasingly salient as a source of

learning and personal well-being (Tinsley, 1992), and sometimes they can even outweigh the influence of the family SES (West et al., 1997). We believe that an individual’s behaviors may be influenced not only by the peers’ behaviors but also by being exposed to peers’ parental environment, in which the behaviors of peers’ parents and information by them provided may play a

• role. By following adolescents throughout the transition to adulthood we can also explore whether

their own educational attainments can change or even reverse the association between parental, friends’ and school’ SES during adolescence and health behaviors at the age of 30 years old. Understanding the link between and adolescent’s SES and his/her health behaviors in adulthood is crucial in that adolescents are targets for prevention programs focused on substance use (Tobler et al., 2000). Moreover, uncovering which and how specific socio-economic contexts are associated with a certain unhealthy behavior is relevant for effective intervention to reduce health disparities. Data and measurements Data & sample

8 We make use of The National Longitudinal Study of Adolescent to Adult Health (Add Health), which is a longitudinal study of a nationally representative sample of adolescents in grades 7-12 in the United States, during the 1994-95 school year. The Add Health cohort has been followed into early adulthood using four interviews, the most recent of which was carried out in 2008, when the sample was aged 24-32 (a fifth wave will be carried out in 2016-2018). Add Health combines longitudinal survey data on respondents’ social, economic, psychological and physical well-being with contextual data on family, school, friendships, and peer groups, providing unique opportunities to study how social environments and behaviors during adolescence are linked to health and achievement outcomes in early adulthood. Add-Health is a multi-survey study composed by different modules. In Wave I, an in-school questionnaire was administered to 90,118 students attending 145 middle, junior high and high school. Each participating school was asked to complete a School Administrator questionnaire. The in-school survey collected information on students’ and their parents’ background, their friendship networks, school life and school activities, general health status and health related

• behaviors. Each school provided a roster of all students enrolled. From the rosters and the pool of

participants in the in-school survey, adolescents in grades 7 to 12 were sampled to participate in the in-home interview. The Wave I in-home interview was administered to 20,745 students, including a sample of individuals with disability, black students whose parents were college graduates, Cubans

• r second-generation Cubans, students of Puerto Ricans descents, and students of Chinese descent.

An additional “genetic sample” of twins, siblings, half-siblings and non-related household members has been included in the study. The in-home questionnaires are the core interviews of the Add- Health study and include detailed information on, among other topics, socio-demographic characteristics, friendship relations, health measurements and health behaviours. During Wave I, an additional parents’ interview was conducted. This interview provides further information about family composition and demographic and health-related knowledge about the parents or guardians

9

• f the adolescent respondent. The parents’ interview has been administered to parents of 17,670

individuals. During the 1995-96 school year (one year after Wave I), participants of the first in-home interview were interviewed again with a new in-home questionnaire. The sample in Wave II does not include the majority of 12th grade respondents, since they exceeded the grade eligibility requirement, as well as the respondents in the Wave I disabled sample. The sample size of the Wave II in-home interview is 14,738. Additional waves were conducted in 2001-2002, when the respondents were 18-26 years-old (Wave III in-home interview, sample size 15,197) and in 2008, when the respondents were 24-32 years-old (Wave IV in-home interview, sample size 15,701). Our study makes use of this complex multi-survey structure, using information from different parts of the Add Health study. Specifically, we make use of data on school membership from the Wave I in-school interview, friendship nomination from Wave I in-school and in-home interview, parental education as a measure of SES from Wave I parents’ interview, and self-reported health behaviors from Wave I, III and IV, when respondents are respectively, adolescents (12-18), young adults (18-26) and adults (24-32). We do not analyse Wave II outcomes since the interview covers the same age period of Wave I and respondents who are not attending high school (who were in grade 12 during Wave I) are excluded from the sample. We began to construct our analysis sample by selecting a baseline from the Wave I surveys with information on parental education, sex, race/ethnicity and average parental education among respondents from the same school and same friendship community. This “baseline” sample is composed by Wave I in-home respondents (N=20,745) with information on parental education (N=17,592), race/ethnicity (N=17,571), with at least one nominated friend who is part of the in- home interview (N=11,508) and whose parental education is available (N=4,536). Friendship information is derived by the in-school questionnaire where the respondent was asked to nominate

10 up to five male and five female friends from the roster of all students enrolled in the respondent’s school and in the sister school. Sister schools are sample schools in the same respondent’s

• community. For instance, if respondent’s school is a high school, sister schools is the junior high

school or middle school that sends the majority of students to the respondent’s school. Approximately 15 percent of all respondents’ friends did not attend either their school or their sister school. We restrict our analysis sample to respondents with information on parental education of at least 10 peers attending the same school. Using a network algorithm, discussed in the following method section of the article, we identified network clusters of friends attending the same school. The network groups are constructed in such a way that every respondent of Wave I in-home interview belongs uniquely to only one cluster of friends. We restrict the analysis to respondents whose friends cluster has at least three friends, but less than 50. This leads to a final “baseline” sample of 4,122 individuals. Finally, the “baseline” sample is linked to the in-home interview at Wave I, III and IV to study the effect of different SES contexts on the propensity of substance use at different stages of the individual’s lifecourse. As shown in Table 1, the final sample sizes used in the analysis are: 4,122 for the baseline sample linked to Wave I; 3,214 for the baseline sample linked to Wave III; 3,296 for the baseline sample linked to Wave IV. Independent variables: measures of SES The indicator we use as a basis to measure SES in family, friends’ group and school is education, more specifically, number of years of education. Previous research has shown (Patrick et al., 2012) that years of education are a good and reliable indicator to capture the SES-substance use linkages. Three SES measures are used, one for the family level, another on for the friends’ group level and a third one for the school level. Each measure is computed at Wave I, when individuals were adolescents between 13 and 15 years old.

11 Parental SES: The adolescent socio-economic status is measured as the maximum number of years of education (following the ISCED code) of the two parents, when both are present. If there is

• nly one parent, his/her number of years of education is used.

Friends’ SES: To measure the average socio-economic status of a respondent’s friends within school, we first had to identify the clusters of friends for each individual/respondent, using a network algorithm described in the following method section, based on an individual’s in-school friends’ nominations. We then calculate the mean of the parental SES (measured as described above) of all friends within each cluster excluding respondent’s parents. In this way, of course, the group of friends we look at is composed by in-school friends. The number of friends from schools different from the respondent’s one was too small to allow any analysis. Therefore, those friends from other schools were not included. School SES: In order to measure the school SES, we calculate the mean of the parental SES (measured as described above) of all of the students, excluding respondent’s parental education, in each of the 184 schools sampled at Wave I. Dependent variables: measures of health behaviors The present paper focuses on a specific category of health behaviors, that are substance use-related behaviors, specifically smoking, drinking and marijuana use. These behaviors are computed at Wave I, III and IV. Following existing literature (Barban, 2013), we use binary variables to measure our health behaviors under study because scales are too sparse and skewed. Smoking is measured using a binary variable that takes value 1 when the individual has smoked at least a cigarette in the last 30 days, 0 otherwise. To measure drinking behavior we decided to focus on a deviant use of alcohol, thereby looking at binge-drinking. Specifically, we use a binary variable that takes value 1 if the individual has experienced any episode of binge-drinking (if he had 5 or more drinks in a row) in

12 the last 30 days, 0 otherwise. Finally, marijuana use is measured using a dummy variable taking value 1 if the individual has used marijuana in the last 30 days, 0 otherwise. Control variables A set of control variables have been included in our models, in order to avoid any spurious association between SES measures and the health behavior under study. Specifically we control for gender, race, and age at interview. We additionally control for marital/union status for outcomes at Wave III and IV (i.e., married (reference), cohabiting or single) and completed education (measured in number of years) for the outcomes at Wave IV. We include educational attainment only for

• utcomes at Wave IV because it is unlikely that respondents have largely completed their education

before the time of the fourth interview. Of course, the last three variables have been measured in the wave at which the outcome variable is measured. Method We estimate a series of three-level multilevel (random intercept) logistic models for each health behavior under study (i.e., smoking, drinking, using marijuana) at Wave I, III and IV. The multilevel modeling allows us to account for the nested structure of the data (Snijders and Boskers, 1999), since we have individuals nested in groups of in-school friends, nested in schools. Each model includes two random intercepts, allowing for both school-specific and friend-specific random effects. Network clustering A community, in the social network terminology, is defined as a subgroup of actors who are highly connected between them and with few links to actors of other communities. In friendship social network, communities are clusters of friends who separate from each other. To identify clusters among the friendship network, we use an algorithm called walktrap community available in the igraph package in R (Pons and Latapy, 2005).

13 The general idea of the algorithm is that if random walks are performed on the network, then the walks are more likely to stay within the same community, i.e. random walks are trapped, since friends are clustered in groups. The algorithm runs short random walks (we used as a parameter random walks of 4 steps), and uses the results of these random walks to merge separate

• communities. The results obtained can be used to “separate” the network in different friendship
• clusters. Figure 1 shows an example of how friendship clusters are derived.

>>>>FIGURE 1 ROUGHLY HERE<<<< The left panel of graph presents a friendship network of 173 respondents attending a specific school at Wave I. The right panel shows how the walktrap community algorithm identifies cluster

• f friends based on their friendship ties. As it is clear from the figure, friendship clusters vary in
• size. Some groups are formed by two respondents who have mutual friendship ties and do not list

any connection with other respondents attending the same school. Since we are interested in groups

• f friends and communities of students within the school, we selected only clusters with size greater

than three. At the same time, we exclude from the analysis very large subnetwork of friends (i.e. cluster of more than 50 friends). Results Descriptive results Descriptive results provide initial interesting insights into how substance use prevalence changes with age, as well as how much SES measures at the different levels are correlated. Looking at Table 1, we can see that, in our sample, the proportion of individuals who have smoked at least once in the last 30 days is 16% when individuals are around 15 years-old. Such a share increases up to 26% during early adulthood, when individuals are around 20 years old, and it declines again, reaching a proportion of 20% when individuals are between 25 and 30 years-old.

14 >>>TABLE 1 ROUGHLY HERE<<< The same inverse U-shaped trend is observed for binge-drinking and marijuana use, for which the prevalence rates at the above mentioned, three points in time are respectively 25%, 38% and 22% for binge-drinking and 12%, 20% and 15% for marijuana use. Results of correlation analysis are reported in Table 2. Here it is shown that an individual’s parental SES modestly overlaps (correlation around 0.2) with the SES of the group of friends as well as the SES of the school during adolescence. Such a finding replicates the results of Cook et. al. (2002). However, we find a somewhat higher correlation (0.5) between the SES of the in-school group of friends and the SES of the school. This result might be driven by the fact that, in the U.S., there are several types of schools (e.g., comprehensive public schools, public magnet school, public school of choice private, private or religious schools), and each of them has a rather homogenous catchment area. Results of multilevel analyses Table 3, 4 and 5 report the results of the three-level multilevel logistic models respectively on smoking (Table 3), drinking (Table 4) and marijuana use (Table 5), estimated when the respondents were adolescents (12-18 years-old, Model 1), young adults (18-24 years-old, Model 2) and adults (24-32 years-old, Model 3). As shown in Table 3, we find that smoking behavior of individuals during adolescence and young adulthood is negatively associated with parental socio-economic status, meaning that individuals at Wave I and III have a higher probability of smoking if they come from more disadvantaged families. It is interesting to observe that this association seems to become positive, although not significant when the individual becomes adult. A similar pattern is found in the

15 relationship between parental SES and binge-drinking (Table 4), for which the association is negative and significant during adolescence, it then becomes non-significant when the individual is around 20 years-old, finally turning to be positive, although still not significant, around the age of 30 years-old. It is interesting to observe that for both, tobacco and alcohol use, the education of parents changes its role over time, stopping to be a protective factor when the individual becomes an adult. It is relevant to say that, the results found in Model 3 were consistently found with and without including the own educational level of the respondent, variable that has a correlation with the measure of family SES only equal to 0.33. Both, smoking and drinking behaviors are not affected by the socio-economic status of the school that the individual attended during adolescence, whereas the socio-economic status of the in- school group of friends seems to be slightly, negatively associated with the likelihood of smoking during adolescence. So, consistently with the role played by the family SES, a more disadvanted group of friends is associated with higher probability of smoking during adolescence. Very different results are found for marijuana use. In Table 5 we can see that the probability

• f using marijuana is affected, positively, by family and school SES only when the individual is

between 25 and 30 years-old, no significant relationship is found earlier in time. Such result means that adults coming from more advantaged families and schools are more likely to engage in marijuana use. In order to better understand the mechanisms that could explain this finding, as well as the fact that for smoking and drinking family SES is no longer a protective factor in adulthood, we engage in a further, additional analysis. We estimate the probability of smoking, drinking and using marijuana of individuals aged between 24 and 32 years-old using the same three-level multilevel logistic models, but instead of including the measure of parental SES we use a dummy indicating whether the individual has a lower educational level of his/her parents. The set of other covariates included in the models is the same that we use in all of the previous models. Table 6 reports the results of these further analyses

16 and shows that individuals who are lower educated than their parents have a higher and significant likelihood of engaging in tobacco and marijuana use than those who have the same or higher education than that of their parents. Building on the psychological theory of the “culture of affluence” (Luthar, 2003; Luthar and Latendresse, 2005), we believe that such finding could be explained by the fact that there is a selected group of individuals who come from rather privileged families (i.e., parents with medium/high level of education), who did not reach the same educational level of the parents and therefore at the age of 25-30 years-old feel achievement pressures and anxiety deriving from not being able to find the “good job” that their parents would like for them. As a consequence, they might be more likely to engage in deviant behaviors, such as smoking or using marijuana, in light of the fact that they also have the resources for doing that. Moreover, in Model 1 of Table 6, we observe that the socio-economic status of friends at Wave I is negatively correlated with the probability of smoking at Wave IV, in line with what we found in Model 1 of Table 3, where the likelihood of smoking for an adolescent was negatively associated with the average SES of his/her friends. We are very much aware that such finding may depend on the selection of in-school friends during the high-school period. Conversely, but in line with Model 3 of Table 5, the school SES is positively associated with the probability of using marijuana, which could be a signal of the fact that having attended a privileged school may increase achievement pressures, coming from high quality jobs of others, which favors distress and in turn substance use. As far as the control variables are concerned, we find for all of the three health behaviors (Table 3, 4 and 5) that the association between age and substance use is positive during adolescence, but it becomes negative starting from young adulthood onwards. Gender differences are more pronounced for drinking and marijuana use than smoking, because we observe a significantly higher probability of using alcohol and marijuana for men than women, regardless of the lifecourse stage. Conversely, ethnic differences seem to be particularly relevant for smoking, where we find that Hispanic and Black people always have a much lower probability of using

17 tobacco than Whites. Turning to drinking behavior, we find no differences between Hispanics and Whites, whereas Blacks are less likely to engage in binge-drinking at any stage in life than Whites. No significant ethnic differences are found for marijuana use, besides a higher probability for Blacks than Whites of engaging in such unhealthy behaviors during adulthood. Being married is found to always be a protective factor against all of the three substance use behaviors compared to being single or cohabiting, statutes that show a higher likelihood of engaging in any risky

• behaviors. Finally, we consistently find a negative association between an individual’s level of

education and the probability of smoking, drinking or using marijuana in adulthood, very much in line with previous research (Pampel et al., 2010) Concluding remarks This paper aimed at uncovering whether and how different contextual levels of the socio- economic status of an adolescent influence his/her substance use behaviors over time, from adolescence to early adulthood. We specifically focused on the three socio-economic contexts that are the more proximal to the individual, according to ecological theory (Bronfenbrenner, 1979), that are family of origin, group of (in-school) friends and school. For each of them we built an indicator

• f socio-economic status measured when the individual was an adolescent. We then followed such

individual until the age of around 30 years-old, observing over time his/her smoking, binge- drinking and marijuana use behavior. Building on the work of Cook et al. (2002), we assumed that each individual is embedded and grows up in multiple socio-economic contexts. Each of these contexts could separately influence such individual’s health behaviors concurrently or later in life. In line with previous research (Cook et al., 2002), we found that the parental SES is modestly correlated with friends and school SES (correlation around 0.2), whereas the latter two measures show a greater overlap, although not a high one (correlation around 0.5).

18 Our findings suggest that the relationship between SES at different contextual levels and health behaviors is very much behavior-specific, and this is in line with the mixed results found in existing literature that looks at different substance use behaviors (Hanson and Chen, 2007). Moreover, Pampel et al. (2010), in their review, while underlining general similarities in the SES- health behaviour relationship, emphasized that each behaviour has its unique dimensions in terms of pleasure it provides, and its social meanings. Such dimensions can very much change the relationship between SES and health behaviors. In our study, there seems to be a clear difference between the use of legal and illegal substances, in terms of how the use is associated with the different contextual dimensions of and individual’s SES. We found that family is the socio-economic context that matters the most in shaping health

• behaviors. Its influence varies substantially according to the type of substance use behaviour as well

as over time within the same behaviour. It is interesting to observe that, in line with previous research (Lindstrom, 2008; Patrick et al., 2012), parental SES has a significant negative effect on the probability of smoking for individuals who are adolescents and early adults. Therefore, it seems that a high parental SES works as a protective factor against smoking, although such a role is not any longer found in adulthood, when instead the relationship between parental SES and smoking seems to become positive, although not significant. A similar pattern of association is found between parental SES and drinking, for which a significant negative relationships is found only during adolescence, but no longer later in life, with again a change in the direction of the association when the individual becomes adult. However, we found that the association between the individual’s own educational level and the likelihood of both smoking and drinking in adulthood is strongly and significantly negative. Such a picture is consistent with what Pampel et al. (2010) pointed out, that SES disparities in smoking (and likely in drinking as well) are already present during teenage years, but they do strengthen after adolescence when disparities in social capital widen as a result of the own educational achievements of the individual (Glendinning et al., 1994).

19 The negative association between parental SES and smoking as well as drinking during adolescence (and young adulthood, but for smoking only), could be explained by several mechanisms that existing literature has sparsely pinpointed. First, coming from more disadvantaged families could be associated with a situation of more stress within the household, factor that in turn could favor engagement in unhealthy behaviors. Second, being instead part of a higher educated family could imply that parents less likely are an example of unhealthy behaviors (e.g., smoking). Third, higher educated parents may more likely convey to their children information on risks associated with unhealthy behaviors, making them more aware of the potential, disruptive consequences of such behaviors (Pampel et al., 2010) Conversely, and again consistently with several studies (Humensky, 2010; Patrick et al., 2012), we found a strong and positive effect of parental SES on marijuana use, but only when the individual is between 25 and 30 years-old, not earlier. Two potential explanations have been put forward: individuals of more advantaged families are, on the one hand, more subject to job performance/achievement-related stress (Luthar, 2003), on the other hand, have more monetary resources to be spent in substances (Patrick et al., 2012). However, whereas the effect of parental SES on marijuana use is positive, the own educational level of the individual is negatively associated with the likelihood of using marijuana. This element let us think that it may not be a matter of financial resources, which can more easily be used to buy marijuana. Rather, the contrasting effects of parental education and the individual’s

• wn education on the probability of using marijuana could actually suggest that those individuals

who do not reach a certain educational level may feel particularly pressured by parents’ job expectations, causing them stress that could favour such unhealthy and deviant behavior. This reasoning is supported by our findings showing that those individuals aged between 24 and 32 years-old, who have a lower educational level than the one of their parents, are substantially more

20 likely to smoke and use marijuana than those who have the same or higher educational level of their parents. Friends’ SES seems not to play a significant role in shaping substance use behaviors before, during and after the transition to adulthood. Likely, this is because peer influence mainly works via contagion effects that may not be caught by our SES measure. However, we do find a marginally significant effect of friends’ SES on the probability of smoking during adolescence, that is in line with some literature showing that friends play a crucial role in predicting smoking among adolescents (Jacobson et al. 2001) and that contagion effect tend to occur more strongly among the highly educated (Christakis and Fowler, 2008). School SES instead plays a relevant role in influencing the likelihood of using marijuana in

• adulthood. Consistently with the relationship between parental SES and marijuana use, it seems that

coming from a more advantaged school increases the probability of using that substance. We tend to interpret this finding following the same reasoning we used to explain the positive relationship between parental SES and marijuana use: growing in a school context that is privileged, and therefore likely competitive, may increase achievement pressures that might be even higher when individuals have to find a job. Such individuals may therefore feel particularly pressured and more likely to engage in substance use behaviours, such as marijuana use. We are aware that this paper suffers from some limitations. First, in analysing the relationship between SES at different contextual levels (i.e., family, friends and school) and substance use, we could identify interesting association and shed further light on potential factors and mechanisms explaining substance use, but could not infer a causal effect of SES on an individual’s health

• behaviors. That is especially problematic for friends’ SES, where the selection of friends may very

well be endogenous to the decision to adopt a certain behavior or not, that is, some factors affecting a certain behaviors may also affect friendship formation. Endogeneity should be less of an issue when we investigate the relationship between parental and school SES with substance use

21

• behaviors. However, we cannot fully rule out all potential sources of such bias. Second, Add Health

data, although they are very rich and suitable for our research goal, did not allow us to follow individuals after they reach around 30 years-old. It would have been very interesting to adopt a complete life course approach, following individuals until older ages, and we hope further research will be able to go in such direction. Finally, we ideally would have liked to look at the effect of the socio-economic status of friends during adolescence, analysing an individual’s more complete network of friends, also outside the school the individual attended. Unfortunately, we could only focus on in-school friends, and such limitation could partially explain the lack of a stronger relationship between friends’ SES and health behaviors. Despite its limitations, we believe our work contributes to existing literature on SES disparities and health behaviors by providing a rather comprehensive picture of the relationship between SES and health behaviors. We indeed adopted a developmental approach that takes into account different contextual dimensions of an adolescent’s SES and investigated how these dimensions relate with health behaviors later in life. Moreover, as suggested by Pampel et al. (2010), we aimed at comparing measures from multiple contexts of an individual’s SES across multiple health behaviors, with the goal of further uncovering the mechanisms explaining SES disparities in health behaviors.

22 References Adler, N. E., Boyce, T., Chesney, M. A., Cohen, S., Folkman, S., Kahn, R. L., et al. (1994). Socioeconomic status and health: The challenge of the gradient. American Psychologist, 49:15-24. Barban, N. (2013). Family trajectories and health: A life course perspective. European Journal of Population, 29(4):357-385. Botticello A. L. (2009). School contextual influences on the risk for adolescent alcohol

• misuse. American Journal of Community Psychology, 43(1–2):85-97.

Brandt, M., Deindl, C., and Hank, K. (2012). Tracing the origins of successful aging. The role

• f childhood conditions and social inequality in explaining later life health. Social Science &

Medicine, 74(9):1418-1425. Bronfenbrenner, U. (1979). The Ecology of Human Development: Experiments by Nature and

• Design. Cambridge, MA: Harvard University Press.

Chassin, L., Hussong, A., and Beltran, I. (2009). Adolescent substance use. In R. Lerner & L. Steinberg (Eds.), Handbook of adolescent psychology. (Vol. 1, pp. 723–763). Hoboken, NJ: Wiley. Chen, E., Matthews, K. A., and Boyce, W. T. (2002). Socioeconomic differences in children’s health: How and why do these relationships change with age? Psychological Bulletin, 128:295-329. Chen, E., and Miller, G. E. (2013). Socioeconomic status and health: Mediating and moderating factors. Annual Review of Clinical Psychology, 9:723-749. Cohen, S., Janicki-Deverts, D., Chen, E., Matthews, K.A. (2010) Childhood socioeconomic status and adult health. Annals of the New York Academy of Sciences, 1186:37-55. Chen, E., and Paterson, L. Q. (2006). Neighborhood, family, and subjective socioeconomic status: How do they relate to adolescent health? Health Psychology, 25:704-714.

23 Christakis, N.A., and Fowler, J.H. (2008). The collective dynamics of smoking in a large social network. New England Journal of Medicine, 358:2249-2258. Chuang, Y, Cubbin, C, Ahn, D, et al. (2005). Effects of neighbourhood socioeconomic status and convenience store concentration on individual level smoking. Journal of Epidemiology & Community Health, 59:568-573. Cook, T. D. (2003). The rationale for studying multiple contexts simultaneously. Addiction, 98(supp 1):151-155. Cook, T. D., Herman, M. R., Phillips, M. and Settersten, Jr., R. A. (2002). Some Ways in Which Neighborhoods, Nuclear Families, Friendship Groups, and Schools Jointly Affect Changes in Early Adolescent Development. Child Development, 73:1283-1309. Cutler, D. M., and Lleras-Muney, A. (2010). Understanding differences in health behaviors by

• education. Journal of Health Economics, 29(1):1-28.

Daw, J., Margolis, R. and Wright L. (2017). Emerging Adulthood, Emergent Health Lifestyles: Sociodemographic Determinants of Trajectories of Smoking, Binge Drinking, Obesity, and Sedentary Behavior. Journal of Health and Social Behavior, 58(2):181-197. Di Gessa, G., Glaser, K. F. & Tinker, A. M. Mar. (2016). The impact of Caring for Grandchildren on the Health of Grandparents in Europe: a lifecourse approach. Social Science & Medicine, 152:166-175. Elo, I.T. (2009). Social Class Differentials in Health and Mortality: Patterns and Explanations in Comparative Perspective. Annual Review of Sociology, 35:553-72. Galobardes, B., Lynch, J.W., and Smith, G.D. (2008). Is the association between childhood socioeconomic circumstances and cause-specific mortality established? Update of a systematic

• review. Journal of epidemiology and community health, 62(5):387–90.

Glendinning, A., Shucksmith, J., and Hendry, L. (1994). Social class and adolescent smoking

• behavior. Social Science & Medicine, 38:1449-1460.

24 Hanson, M.D. and Chen, E. (2007) Socioeconomic Status and Health Behaviors in Adolescence: A Review of the Literature. Journal of Behavioral Medicine, 30:263-285. Hayward, M.D. and Gorman, B.K. (2004). The long arm of childhood: The influence of early- life social conditions on men's mortality. Demography, 41(1): 87-107. Huerta M. C., and Borgonovi F. (2010). Education, alcohol use and abuse among young adults in Britain. Social Science & Medicine, 71:143-151. Humensky, J. L. (2010). Are adolescents with high socioeconomic status more likely to engage in alcohol and illicit drug use in early adulthood? Substance Abuse Treatment, Prevention, and Policy, 5:19. Jacobson, P.D., Lantz, P., Warner, K., Wasserman, J., Pollack, H., Ahlstrom, A. (2001). Combating Teen Smoking: Research and Policy Strategies. Ann Arbor: University Michigan Press. Johnston, L. D., O’Malley, P. M., Bachman, J. G., and Schulenberg, J. E. (2011). Monitoring the Future national survey results on drug use, 1975–2010. Volume II: College students and adults ages 19–50. Ann Arbor, MI: Institute for Social Research, The University of Michigan. Lindstrom, M. (2008). Social capital and health-related behaviors. In: Kawachi, I.; Subramanian, SV.; Kim, D., editors. Social Capital and Health. New York: Springer, p. 215-238. Luo, Y., and Waite, L.J. (2005). The impact of childhood and adult SES on physical, mental, and cognitive well-being in later life. Journal of Gerontology: Social Sciences, 60(2):93-S101. Luthar, S. S. (2003). The culture of affluence: Psychological costs of material wealth. Child Development, 74:1581-1593. Luthar S. S., and Latendresse S. J. (2005). Children of the affluent: challenges to well-being. Current Direction in Psychological Science, 14:49-53. 10.1111/j.0963-7214.2005.00333.x. Maralani, V. (2013). Educational Inequalities in Smoking: The Role of Initiation versus

• Quitting. Social Science & Medicine, 84:129-137.

25 Patrick, M. E., Wightman, P., Schoeni, R. F., & Schulenberg, J. E. (2012). Socioeconomic status and substance use among young adults: A comparison across constructs and drugs. Journal of Studies on Alcohol and Drugs, 73:772–782. Pampel, F., Krueger, P., and Denney, J. (2010). SES disparities in health behavior. Annual Review of Sociology, 36:349-370. Pons, P., and Latapy, M. (2005). Computing communities in large networks using random

• walks. In Proceedings of the 20th International Symposium on Computer and Information Sciences,

volume 3733 of Lecture Notes in Computer Science, pp. 284–293, Springer, New York. Pudrovska, T., and Anikputa, B. (2014). Early-life socioeconomic status and mortality in later life: an integration of four life-course mechanisms. The Journals of Gerontology Series B: Psychological Sciences and Social Sciences gbt122. doi:10.1093/geronb/gbt122. Schreier, H. M. C. & Chen, E. (2013). Socioeconomic status and the health of youth: A multi-level multi-domain approach to conceptualizing pathways. Psychological Bulletin, 139:606- 654. Siahpush, M, English, D, and Powles, J. (2006). The contribution of smoking to socioeconomic differentials in mortality: Results from the Melbourne Collaborative Cohort Study,

• Australia. Journal of Epidemiology & Community Health, 60:1077-1079.

Snijders, T. A. B., and Bosker, R. J. (1999). Multilevel Analysis: An Introduction to Basic and Advanced Multilevel Modeling. London etc.: Sage Publishers. Sobal, J., and Stunkard, A.J. (1989) Socioeconomic status and obesity: a review of the

• literature. Psychological Bulletin, 105:260-275.

Tinsley, B. J. (1992). Multiple Influences on the Acquisition and Socialization of Children's Health Attitudes and Behavior: An Integrative Review. Child Development, 63:1043-1069.

26 Tobler, N. S., Roona, M. R., Ochshorn, P., Marshall, D. G., Streke, A. V., and Stackpole, K.

• M. (2000). School-based adolescent drug prevention programs: 1998 meta-analysis. The Journal of

Primary Prevention, 20(4):275-336. Warner, T. D. (2015). Up in Smoke: Neighborhood Contexts of Marijuana Use from Adolescence Through Young Adulthood. Journal of Youth and Adolescence, 45(1):35-53. Wardle, J., Jarvis, M. J., Steggles, N., Sutton, S., Williamson, S., Fairmond, H. et al. (2003). Socioeconomic disparities in cancer risk behaviors in adolescence: Baseline Results from the Health and Behavior in Teenagers Study. Preventive Medicine, 36:721-730. West, P., S. Macintyre, E. Annandale, et al. (1990). Social class and health in youth: Findings from the West of Scotland Twenty-07 study. Social Science & Medicine, 30:665-673 West, P. (1988). Inequalities? Social class differentials in health in British youth. Social Science & Medicine, 27(4):291-296. West, P. (1997). Health inequalities in the early years: is there equalization in youth? Social Science & Medicine, 44(6): 833-858 White, K.M. and Preston, S.H. (1996). How many Americans are alive because of twentieth- century improvements in mortality? Population and development review, 22(3):415-29. Zimmer, Z., Hanson, H., Smith, K. (2016). Childhood socioeconomic status, adult socioeconomic status, and old-age health trajectories: Connecting early, middle, and late life. Demographic Research, 34(10):285-320.

27 Tables Table 1: Descriptive statistics of the baseline sample, sample at Wave I, at Wave III and at Wave IV Variable Mean

• Std. Dev.

Min Max Baseline sample N=4,122 Proportion Female 0.52

• Proportion White non-

Hispanic 13.61 Proportion Black 19.39 Proportion Hispanic 67.00 Age Wave I 15.97 1.62 12.50 20.00 Parental education (years) 13.95 3.18 0.00 19.00 Friends' average parental education (years) 12.22 3.14 0.00 19.00 School's average parental education (years) 11.74 1.76 7.95 15.81 Friends' group size 12.18 11.11 3.00 45.00 School size 389.90 516.83 20.00 1721.00 Wave I sample N=4,122 Proportion smoking Wave I 0.16 0.37 Proportion drinking Wave I 0.25 0.43 Proportion marijuana users Wave I 0.12 Wave III sample N=3,214 Age Wave III 22.30 1.66 18.67 26.83 Proportion single Wave III 0.19

• Proportion cohabiting Wave

III 0.54

• Proportion married Wave III

0.27

• Respondent's education in

years Wave III 13.54 1.89 8.00 22.00 Proportion smoking Wave III 0.26 0.44 0.00 1.00 Proportion drinking Wave III 0.38 0.49 0.00 1.00 Proportion marijuana users Wave III 0.20 0.40 0.00 1.00 Wave IV sample N=3,296 Age Wave IV 28.82 1.67 24.58 33.42

28 Respondent's education in years Wave IV 14.89 2.06 8.00 21.00 Proportion single Wave IV 0.38

• Proportion cohabiting Wave

IV 0.19

• Proportion married Wave IV

0.43

• Proportion smoking Wave IV

0.20 0.40 0.00 1.00 Proportion drinking Wave IV 0.22 0.42 0.00 1.00 Proportion marijuana users Wave IV 0.15 0.35 0.00 1.00 Table 2: Correlation between SES measures at different levels, Wave I Parental SES School SES Parental SES 1 School SES 0.29* 1 Friends SES 0.27* 0.54* Note: * p < .001 after Bonferroni correction

29 Table 3: Three-level multilevel logistic models on smoking at Wave I, III and IV (1) (2) (3) Smoking WI Smoking WIII Smoking WIV Constant

• 3.1230***

4.1880*** 4.4466*** (0.9120) (1.0524) (1.1738) Age 0.0172***

• 0.1369***
• 0.0321

(0.0030) (0.0355) (0.0322) Female (ref.:male)

• 0.1086
• 0.5087***
• 0.0539

(0.0944) (0.1012) (0.0965) Race (ref.: White) Hispanic

• 0.3452*
• 0.6855***
• 0.7517***

(0.1737) (0.1966) (0.1913) Black

• 1.4785***
• 1.1723***
• 0.6685***

(0.1942) (0.1711) (0.1479) Family SES

• 0.0459**
• 0.0468*

0.0218 (0.0167) (0.0182) (0.0183) School SES

• 0.0493
• 0.0071
• 0.0300

(0.0530) (0.0483) (0.0450) Friends SES

• 0.0387~
• 0.0231
• 0.0217

(0.0203) (0.0204) (0.0198) Marital status (ref: married) Single 0.6010*** 0.5507*** (0.1338) (0.1116) Cohabiting 0.5620*** 0.6198*** (0.1265) (0.1286) Education in years

• 0.3207***

(0.0255) Random part School variance 0.1928 0.1929 0.1412 (0.0789) (0.0648) (0.0533) Friends’ variance 0.5968 0.1464 0.0016 (0.1565) (0.0776) (0.0642) N observations 4121 3155 3286 Log likelihood

• 16090.5854
• 1284.4103
• 1470.9993

Standard errors in parentheses

~ p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001

30 Table 4: Three-level multilevel logistic models on drinking at Wave I, III and IV (1) (2) (3) Drinking WI Drinking WIII Drinking WIV Constant

• 4.2275***

0.1942 4.2477*** (0.7919) (0.8899) (1.1069) Age 0.0267***

• 0.0293
• 0.1230***

(0.0026) (0.0299) (0.0312) Female (ref.:male)

• 0.3682***
• 0.8442***
• 0.9631***

(0.0811) (0.0826) (0.0958) Race (ref.: White) Hispanic 0.0576

• 0.1770

0.0535 (0.1413) (0.1459) (0.1611) Black

• 0.9564***
• 0.8176***
• 0.7201***

(0.1474) (0.1337) (0.1461) Family SES

• 0.0418**
• 0.0114

0.0097 (0.0141) (0.0144) (0.0171) School SES

• 0.0520

0.0661 0.0302 (0.0464) (0.0420) (0.0400) Friends SES

• 0.0215

0.0035 0.0031 (0.0169) (0.0170) (0.0195) Single 0.8872*** 0.8979*** (0.1228) (0.1078) Cohabiting 0.4547*** 0.6451*** (0.1044) (0.1287) Education in years

• 0.0855***

(0.0244) Random part School variance 0.2116 0.1618 0.0403 (0.0743) (0.0599) (0.0478) Friends’ variance 0.3373 0.1464 0.1932 (0.1041) (0.0776) (0.0964) N observations 4094 3155 3273 Log likelihood

• 2105.2961
• 1933.9833
• 1582.8635

Standard errors in parentheses

~ p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001

31 Table 5: Three-level multilevel logistic models on marijuana use at Wave I, III and IV (1) (2) (3) Marijuana WI Marijuana WIII Marijuana WIV Constant

• 4.8994***

1.9444*

• 0.0309

(1.1960) (0.9130) (1.2013) Age 0.0140***

• 0.1616***
• 0.0448

(0.0038) (0.0320) (0.0343) Female (ref.:male)

• 0.3101*
• 0.5746***
• 0.5181***

(0.1121) (0.0936)

• 0.0448

Race (ref.: White) Hispanic 0.2692

• 0.2029
• 0.0424

(0.1883) (0.1676) (0.1859) Black 0.0502

• 0.0231

0.2971* (0.1833) (0.1323) (0.1384) Family SES 0.0910 0.0193 0.0498* (0.0192) (0.0163) (0.0198) School SES 0.00552 0.0493 0.1134** (0.0722) (0.0394) (0.0419) Friends SES

• 0.0385
• 0.0007
• 0.0014

(0.0232) (0.0188) (0.0211) Marital status (ref: married) Single 0.8466*** 1.1217*** (0.1380) (0.1346) Cohabiting 0.1832 1.3470*** (0.1253) (0.1489) Education in years

• 0.1795***

(0.0275) Random part School variance 0.5874 0.0620 0.0247 (0.1842) (0.0486) (0.0375) Friends’ variance 0.8441 0.0435 0.0605 (0.2030) (0.0760) (0.0999) N observations 4087 3171 3286 Log likelihood

• 1354.0
• 1525.2325
• 1243.5542

Standard errors in parentheses

~ p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001

32 Table 6: Three-level multilevel logistic models on smoking, drinking and marijuana use at Wave IV comparing the respondent’s educational level with parents’ educational level (1) (2) (3) Smoking WIV Drinking WIV Marijuana WIV Constant 1.2434 3.4122**

• 1.1074

(1.2026) (1.1011) (1.1659) Age

• 0.0472
• 0.1309***
• 0.0608~

(0.0331) (0.0315) (0.0338) Female (ref: male)

• 0.1955*
• 1.0081***
• 0.5934***

(0.0946) (0.0954) (0.1060) Race (ref: White) Hispanic

• 0.5948**

0.0946

• 0.0920

(0.1881) (0.1586) (0.1840) Black

• 0.5904***
• 0.7052***

0.2831* (0.1507) (0.1479) (0.1387) Marital status (ref: married) Single 0.5626*** 0.9012*** 1.1217*** (0.1102) (0.1080) (0.1340) Cohabiting 0.7347*** 0.6770*** 1.3957*** (0.1268) (0.1287) (0.1480) Lower education than parents 0.5345*** 0.0423 0.4247*** (0.0978) (0.0983) (0.1074) School SES

• 0.0749

0.0122 0.0871* (0.0500) (0.0408) (0.0407) Friends SES

• 0.0425*
• 0.0042
• 0.0114

(0.0193) (0.0193) (0.0205) Random part School variance 0.2718 0.0606 0.0258 (0.0793) (0.0510) (0.0378) Friends variance 0.0409 0.2101 0.0606 (0.0732) (0.0986) (0.1004) N Observations 3286 3273 3286 Log likelihood

• 1544.4767
• 1590.4332
• 1257.4831

Standard errors in parentheses

~ p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001

33 Figures Figure 1