Concentrated Poverty in U.S. Schools and Adolescents’ Risk of Being Overweight

Abstract

The economic segregation of U.S. schools undermines the academic performance of students, particularly students from low-income families who are often concentrated in high-poverty schools. Yet it also fuels the reproduction of inequality by harming their physical health. Integrating research on school effects with social psychological and ecological theories on how local contexts shape life course outcomes, we examined a conceptual model linking school poverty and adolescent students’ weight. Applying multilevel modeling techniques to the first wave of data (1994–1995) from the National Longitudinal Study of Adolescent to Adult Health (Add Health; n = 18,924), the results revealed that individual students’ likelihood of being overweight increased as the concentration of students from low-income families in their schools increased, net of their own background characteristics. This linkage was connected to a key contextual factor: the exposure of students in high-poverty schools to other overweight students. This exposure may partly matter because of the lower prevalence of dieting norms in such schools, although future research should continue to examine potential mechanisms.

The persistent segregation of schools has long been one of the great social justice issues of American society. Although the nature of that segregation has changed as racial and ethnic segregation has declined, the concentration of low-income students in their schools has risen. For example, in the past 20 years, the percentage of students attending schools characterized by the U.S. Department of Education as high-poverty has increased by nearly two-thirds (National Center for Education Statistics [NCES] 2013, 2019). This phenomenon, which has been driven by increased residential segregation along with the rise in private school enrollment among middle class and affluent students (Reardon and Owens 2014; Saporito 2003; Stroub and Richards 2013), is widely considered to be problematic because of its negative repercussions for the achievement of students attending high-poverty schools (Crosnoe 2009; Logan, Minca, and Adar 2012; Rumberger and Palardy 2005). The implications of concentrating poor students, however, are not just limited to the formal outcomes of schooling. As many recent studies have demonstrated (e.g., Martin et al. 2012; Miyazaki and Stack 2015; Richmond and Subramanian 2008), schools can also shape physical health indicators that reflect and predict socioeconomic disadvantages, including one’s weight status. The goal of this study is to build on this burgeoning body of social research by examining how school contextual factors connect school poverty with adolescent students’ risks of being overweight.

We do so by making three principle contributions. First, drawing insights from the sociological literature on “school effects,” which underscores how the composition of a school’s student body serves as a social context that can influence student’s outcomes beyond their own characteristics (Coleman 1961; Logan et al. 2012; Rumberger and Palardy 2005), we investigate the role of one compositional feature surmised about in other studies that has yet to be directly investigated (Miyazaki and Stack 2005; Procter et al. 2008): whether the weight composition of the student body shapes individual students’ weight status. Second, we consider the significance of the school weight composition relative to other school-level contextual factors highlighted in research connecting school poverty and adolescent overweight: the availability of school resources (e.g., physical education programs) and the stress environment (e.g., exposure to violence). Finally, we weave together insights from several subfields, including social psychology, medical sociology, and social epidemiology to develop a conceptual model, and analysis of the social mechanisms that explain how the student body weight composition shapes students’ individual risks of being overweight.

To pursue these aims, we use data from the first wave of the National Longitudinal Study of Adolescent to Adult Health (Add Health; 1994–1995)—a nationally representative sample of over 20,000 adolescents randomly sampled from secondary schools that is the only currently available nationally representative data capturing the school-level measures central to this study, including poverty, the student’s body weight composition, and the proposed social mechanisms (e.g., weight-related norms and behaviors). The results of this investigation will provide greater understanding of the links between school poverty and student weight, while speaking more broadly to concerns around school-related inequities and economic disparities in adolescent health.

School Poverty and School Effects during Adolescence

Nearly a half century ago, the “Coleman Report” pointed to the economic composition of the student body as a critical risk factor in the academic progress of students, independent of students’ own economic circumstances (Coleman 1961). Recognition of the student body composition as a dimension of the school context was important for understanding how students of various economic backgrounds, including low-, middle-, and high-income families, were negatively affected by attending high-poverty schools. Based on this insight, Coleman argued that the primary source of the “problem” of economic school segregation was not school-level disparities in funding (i.e., resources within the school), as it had been (and often remains to be) argued, but the social conditions of schools serving large numbers of students from low-income families.

Since the publication of the report, researchers have aimed to better understand these contextual factors associated with school poverty, or “school effects” as they are often termed. Although such efforts have been largely focused on explaining student academic outcomes (Crosnoe 2009; Logan et al. 2012; Rumberger and Palardy 2005), the insights of this research can also be extended into research on adolescent health outcomes, including their weight status. The importance of doing is further supported by two other bodies of literature.

First, research on the “social ecology of health,” which articulates how the socioeconomic conditions of a “place” creates a context that shapes various health outcomes (Boardman et al. 2005; Frohlich, Corin, and Potvin 2001; Kawachi and Berkman 2003), underscores the importance of considering both institutional (e.g., school resources) and social factors. For example, research on neighborhood contexts highlights the importance of social processes such as stress exposure and community connectedness in explaining the observed worse health outcomes among people living in low-income communities (Kwarteng et al. 2017; Suglia et al. 2016). Likewise, studies on family contexts have pointed to social processes such as fewer shared family meals (Fiese, Hammons, and Grigsby-Toussaint 2012), or parents’ use of food to address children’s emotional needs (Olson, Bove, and Miller 2007), in shaping the lower levels of health of poor children.

Second, several studies on adolescent health have documented the health risks of attending high poverty schools for various outcomes, including body weight (Martin et al. 2012; Miyazaki and Stack 2015; O’Malley et al. 2007; Richmond and Subramanian 2008), but also depression (Dunn, Milliren, et al. 2015; Goodman, Slap, and Huang 2003), smoking (Dunn, Richmond, et al. 2015; Suh, Shi, and Brashears 2017), and early sexual debut (Teitler and Weiss 2000). Studies on adolescent weight, however, have given less attention to examining the school-level contextual factors linked to school socioeconomic status (SES). Notably, several studies have indicated that for adolescents, body weight is shaped by the school context more so (Lee et al. 2014; Martin et al. 2012) than neighborhood or family contexts (that later of which may be more salient to younger children). Such research further underscores the importance of homing in on the role of school contexts, although we acknowledge and take steps to account for the potential influence of family-level and neighborhood-level factors as well.

As stated above, our specific aim is to investigate the salience of one particular school-level contextual factor that has been shown to rise in tandem with the percentage of students who are from low-income families: the weight composition of the student body. This contextual factor has also been linked to individual student’s weight risks (O’Malley et al. 2007; Martin et al. 2012; Miyazaki and Stack 2015). Based on these two sets of findings, scholars have suggested that the weight composition of the school serves as a contextual factor linking school SES and student’s individual weight status (Miyazaki and Stack 2015; Procter at al. 2008). To our knowledge, however, no study has directly examined this possibility, although this link has been documented in the context of neighborhoods (Boardman et al. 2005). In the following section, we describe the theoretical importance of examining this contextual factor in relation to schools.

School Social Context and Mechanisms Linking School Level and Individual Weight

As Coleman (1961) discussed (and social ecological theories on health echo; see Frohlich et al. 2001), schools represent contexts that contain a local universe of “familiar others” (defined as peers who are similar, but do not necessarily share close friendship ties; see Suh et al. 2017) who reflect and reinforce a systems of norms, values, and behaviors (Arum 2000; Dornbusch 1989). Based on this key insight, we suggest that this system of norms, values, and behaviors reflects several school-level mechanisms that may help explain why the weight composition of the student body shapes individual students’ weight.

First, because students need reference groups to evaluate themselves, and the student body, not just close friends, provides a standard of comparison (Suh et al. 2017), adolescents may be more or less likely to view themselves as being overweight depending on the normative body size at school (Evans et al. 2016; Jones 2001). Moreover, because friendship groups are likely to be more homogeneous than the student body in terms of weight (Crosnoe, Frank, and Mueller 2008; McPherson, Smith-Lovin, and Cook 2001), the presence of overweight schoolmates may lead overweight adolescent to be more self-accepting and less compelled to alter their body size (Jones 2001; Mueller 2015).

Second, students develop perceptions of what weight-related behavioral norms are common. In contexts with few overweight schoolmates, students may presume that weight control practices are frequent and use dieting, physical activity, and healthy eating (e.g., fruits, vegetables) as visible cues that they recognize norms around thinness and “fit in” (Coleman et al. 1996; Elder, Evans, and Parker 1995; Jones 2001; Milkie 1999). For example, one study found that students’ smoking behavior was influenced by their perceptions about the normativity of smoking behavior among schoolmates, regardless of how much they actually smoked (Ellickson et al. 2003). Alternatively, in contexts with more overweight students, students will have less reason to consider weight control behaviors as normative or feel pressured to manage impressions in front of other students through such practices (Salvy et al. 2011). For example, studies have found that the likelihood that students tried to lose weight was inversely associated with the proportion of overweight schoolmates (Eisenberg et al. 2005; Mueller 2015). Of course, weight control practices might not only represent the efforts of students to “fit in.” They may be learned behaviors modeled by schoolmates (Christakis and Fowler 2007). Multiple studies have found that schoolmates (again, not just close friends) influence adolescents’ consumption of unhealthy foods (Fuenekes et al. 1998; Salvy, Kieffer, and Epstein 2008), dieting (Hutchinson and Rapee 2006), and physical activity (Baker, Little, and Brownell 2003; Fitzgerald, Fitzgerald, and Aherne 2012). Thus, in contexts with more overweight students, weight control practices like dieting may go down, and other obesogenic behaviors, like consumption of calorie dense snacks (e.g., chips, cookies), may increase.

Finally, although body size is a powerful marker of status that may also motivate weight control practices (Carr and Friedman 2005; Puhl and Brownell 2001), the social penalties and stigmatization of being overweight vary across schools due to differences in norms and values. In poorer schools in which being overweight is more common, weight may be a less salient marker of status (Crosnoe 2011). Such penalties associated with school peers are also distinct from those of close friends, who tend to be more accepting of differences among each other than classmates who do not have long histories of attachment (Crosnoe et al. 2008; Giordano 2003). In this way, the more overweight students there are in a school, the lower the social risks of overweight and social controls against weight gain there will be, both of which will reduce the individual likelihood of becoming overweight.

Other Contextual Factors Linking School Poverty and Adolescent Overweight

Importantly, as social ecological research on health and research on school effects both underscore, our investigation of the weight composition of the school must not overlook other contextual factors related to school poverty and individual overweight that may correlate with and confound the associations hypothesized above. The first of these contexts is what has been termed the “institutional context.” (Jencks and Mayer 1990), which reflects school resources. For example, schools serving large numbers of students from low-income families tend to be short on funding and lack the material resources for combatting weight gain, such as physical education programs, nutritional curricula, or gymnasiums (Delva, Johnston, and O’Malley 2007; Richmond et al. 2014). They may also rely on revenue-generating ventures (e.g., soft drink machines) that undermine efforts to promote student’s healthy weight, or feed many of their students through free and reduced lunch programs that include energy-dense food items (although free lunches have improved since the time of data collection and may now be healthier than most packed lunches; Farris et al. 2014; Price, Murnan, and Moore 2006).

The second contextual explanation has to do with exposure to stressors, which as highlighted above, has also been found to be a key factor linking neighborhood poverty and resident’s greater likelihood of being overweight (Kwarteng et al. 2017). Similar to neighborhoods, schools serving larger numbers of low-income students also experience more violence, student-teacher conflict, general disorganization, less social support, lower feelings of morale, and less connectedness (Crosnoe, Johnson, and Elder 2004; Mrug, Loosier, and Windle 2008). These factors contribute to greater stress exposure in ways that may contribute to weight gain through depression and other physiological processes, low levels of physical activity, and stress-induced eating (Adam and Epel 2007; Garasky et al. 2009; Goodman et al. 2003; Luppino et al. 2010).

Study Conceptual Model and Aims

Bringing these ideas together, we propose a conceptual model of the link between school poverty and adolescent’s individual risks of being overweight, presented in Figure 1. According to this model, students attending higher poverty schools will be more likely to be exposed to school contexts serving higher concentrations of overweight students. This pathway is sometimes referred to as the compositional effect (Frohlich et al. 2001; Kawachi and Berkman 2003). Exposure to more overweight peers, however, will also be associated with students’ individual risks of being overweight, net of their own individual characteristics, including their own economic characteristics, family factors, and neighborhoods. This pathway represents the contextual effect.

Figure 1.

Conceptual model of study.

The contextual effect of school poverty comprises several factors. The first reflects school resources (Factor B), and the second reflects school stress environment (Factor C). Although these factors may represent independent pathways linking school poverty and individual student’s weight status, they are also exogenous to the weight composition of the school and potentially confounding factors that must be adjusted for in the analysis. The third contextual factor (Factor A) posits a causal pathway between exposure to schools serving larger numbers of overweight peers and a students’ individual risk of being overweight that reflects several possible school-level social mechanisms. These include (1) body size norms, (2) weight-related behaviors, and (3) social penalties. By the same token, we acknowledge the methodological challenges to establishing this causal pathway, which we discuss in more detail below.

The aims of this study are to test the basic pieces of this conceptual model. First, we test the argument that attending high-poverty schools is associated with an increase in the odds that adolescent students, net of their own poverty status, will be overweight by exposing them to more overweight peers. Second, we examine the extent to which this association is robust to the inclusion of potentially correlated contextual factors that reflect (1) school resources and (2) the stress environment. Third, we explore the school-level social mechanisms that may link exposure to more overweight peers in the school and adolescent’s risks of being overweight themselves.

In pursuing these aims, we focus on overweight status (versus obesity) because it is relevant to understanding health outcomes among a broader population of adolescents and builds upon the one study to have examined concentrated poverty (versus some other school indicator of SES) in connection to adolescent weight (see Martin and colleagues 2012). We also, however, examine the robustness of our findings to individual and school-level measures of obesity. We analyze these aims using multilevel modeling techniques, which allowed us to examine school-level sources of variability (e.g., composition of overweight students), while adjusting for the clustering of similar students within schools (C. Duncan, Jones, and Moon 1996), as well as other sources of individual and school-level sources of variability highlighted in previous research (e.g., gender, race, family structure, neighborhood quality; Bryk and Raudenbush 1992; Dunn, Milliren, et al. 2015).

Method

Data and Sample

Add Health is a nationally representative sample of adolescents first surveyed in 1994–1995 and followed into adulthood. The sample was generated using a stratified sampling design in which 132 schools were selected from a list containing the names and characteristics of all U.S. secondary schools based on their region, urbanicity, school size, school type, and racial composition. The in-school data collection was based on 90,118 students spanning Grades 7 to 12 in the 132 selected schools who were in attendance on the targeted day and agreed to participate. This in-school survey of students was supplemented with administrative reports about the school. Following these data collections, an in-home survey was conducted. The in-home survey was based on a nationally representative subsample of 20,745 students selected from the in-school data collection (mean of around 157 students surveyed per school). Because of the initial Census-like design of the in-school data collection, with appropriate weighting, (described below) individual-level measures from the in-home survey can and have been averaged across students to approximate school-level characteristics (e.g., Bearman and Bruckner 2001; Benner, Crosnoe, and Eccles 2014; Cleveland and Weibe 2003).

The analytical sample for this study included the 18,924 youth who participated in the Wave I in-school data collection and Wave I in-home collection. Although Add Health is a panel study and includes subsequent waves of data collection, Wave II, which was collected a year after Wave I, did not include the in-school component. Consequently, as other scholars have pointed out (see Martin et al. 2012, for a more detailed explanation), it is not possible to create longitudinal measures of the school-level measures of poverty or overweight, which would have been optimal. We did explore the use of lagged models (i.e., drawing on Wave I school-level measures, and Wave II individual measures net of Wave I measures), but this approach proved problematic because a large number of students changed schools. Thus, Wave I measures did not accurately capture the experiences of many students at Wave II. The one-year span between waves was also problematic for assessing weight change, which did not change for most students in Add Health across this period (the correlation across Wave I and II was >.90; see Goodman, Hinden, and Khandelwal 2000). Again, we acknowledge and further discuss the limitations of using cross-sectional data for making any robust causal claims in the “Discussion” section.

The analytical sample was based on youth who had valid individual- and school-level sampling weights, which accounted for study design effects, the nesting of students within schools, and differential attrition between in school and in home data collections (Bryk and Raudenbush 1992). All other sources of item-level missingness were accounted for through multiple imputation techniques, explained more below as well.

Measures

Individual- and school-level overweight status.

Using self-reported measures of height and weight from the Wave I in-home interview, we calculated each student’s body mass index (BMI) using the formula: (weight [kg])/(height [m])². Following the Centers for Disease Control and Prevention (CDC) protocol, overweight was identified as a BMI at or above the 85th percentile for age and gender (which was based on CDC growth charts). Although the accuracy of self-reports of general weight statuses have been called into question (Shields, Gorber, and Tremblay 2008), self-reports of actual height and weight in the Add Health have been shown to be quite reliable (Goodman et al. 2000). Taking advantage of the clustering of students in schools, we used this measure to create a continuous measure for the proportion of overweight students in each school by dividing the total number of overweight students in a school by the total number of students sampled in it.

Individual-level and school-level poverty status.

Drawing on reports of family income from the in-home parent report, household rosters from the respondent report, and information on the federal poverty thresholds from 1994, we calculated an income-to-needs ratio for each respondent’s family. Students with a family income-to-needs ratio of 1.85 or below were classified as poor (see Duncan and NICHD Early Child Care Research Network 2003 for a rationale for this cut-off). Following the same aggregation procedure described above, the school-level poverty measure was created by dividing the total number of students from poor families in a school by the total number of students sampled in it.

School resources.

To capture resources within the school, we included three school-level measures that reflected weight-related programmatic aspects of the school. The first is a measure of the proportion of students involved in athletics, which was based on students’ reports of extracurricular involvement, including clubs, teams, and organizations, as part of the in-school survey. The school-level measure was then calculated by dividing the number of students on sports teams by the total number of students sampled in the school. Second, we measured the average number of days per week students attended physical education, which was reported during the Wave I in-home interview. We calculated the school-level average using reports of students who attended the same school. Third, we used the school administrative data to create a binary indicator for whether the school had a weight loss program on site.

Stress-related factors.

To capture the stress environment of the school, we included five measures. The first three tapped the degree to which students felt safe, accepted, and supported at the school level; the final two reflected school violence and disorganization. The first, school connectedness (validated by Furlong, O’Brennan, and You 2011), was measured using five variables from the in-school survey. This scale was created by summing students’ responses to the following prompts: I feel close to people at this school; I am happy to be at this school; I feel like I am part of this school; the teachers at this school treat students fairly; I feel safe in my school. Second, five questions from the in-home interview about the degree to which students felt socially accepted, loved, close to people at their school, and a part of the school were averaged at the individual level (Crosnoe 2011), then averaged across respondents who attended the same school to create a school-level measure of the degree to which students feel they fit in at school. Third, teacher attachment was based on the average of three items taken from the in-home interview which tapped adolescents’ attitudes about their teachers (Crosnoe et al. 2004; e.g., the extent to which adolescents had trouble getting along with teachers, believed teachers treated students fairly, felt that teachers cared about them), aggregated across students in the school. For all three measures, student responses ranged from 1 to 5. Higher values indicated greater school connectedness, feelings of fitting in, and teacher attachment, and were taken to reflect environments with lower levels of stress. The fourth measure, truancy, was a school-level aggregate of students’ responses to how often they ever skipped school during the past month (0 = never, 1 = at least once). Finally, violence was a school-level aggregate of students’ responses to how many times they got into a physical fight during the past month (1 = never, 1 = at least once, 2= more than once).

Social mechanisms.

To explore mechanisms that may explain the salience of exposure to overweight peers, we incorporated several measures that mapped onto three concepts. The first is a school-level measure of norms of body size. Students responded to a question on how they perceived their body sizes. Possible response categories spanned “very under-weight” to “about the right weight” to “very overweight.” Overweight adolescents who reported that they were about the right weight or underweight (which was less than 6% of overweight students, most of whom selected the option of “slightly underweight”) were assigned a value of 1. Those overweight adolescents who thought of themselves as overweight or very overweight were given a value of 0. Based on these reports, we calculated the proportion of overweight students in the school who thought of themselves as underweight or just right (assigned a value of 1)—in other words, who did not view their body size as divergent from the “ideal” body size. Next, we captured school-level measures of weight behaviors (created by averaging the reports of students within the same school) by drawing on student reports of whether or not they had dieted in the past week in order to lose weight (1 = yes, 0 = no); how many times in the past week they had played an active sport and exercised (range: 0–7); and the frequency with which they ate energy-dense foods, such as pasta, bread, pretzels, cookies, and cakes, in the past day (0 = did not eat; 1 = ate once; 2 = ate twice or more). Third, we examined a school-level measure for social stigmatization of body size. During the in-home interviews, adolescents were asked to nominate up to 10 of their friends in their schools, meaning that respondents gave but also received friendship nominations to/from other adolescents in their schools. These reports thus identified the number of same-sex friendship nominations that each student received. To capture the social penalty for being overweight, we averaged nominations across overweight students in each school and across students in the school who were not overweight and subtracted the average number of nominations that overweight students received from the average number of nominations for non-overweight students.

Covariates.

Several covariates taken from the in-school survey were included to address heterogeneity between schools that could be associated with the focal independent and dependent variables. At the school level, these measures included school sector (private = 1, public = 0), region (dummy coded for North-east, Midwest, South, or West), location (dummy coded urban, suburban, or rural), the size of the school student body (continuous), the proportion of students living with their two biological parents, the proportion of students attending a school in which the feeder and high school were the same, and the proportion of students who identified as White (to address race/ethnic segregation). At the individual level, covariates included students’ race/ethnicity (dummy coded and categorized as non-Hispanic white, non-Hispanic black, non-Hispanic Asian, Hispanic, other/multiracial), age (continuous based on years), family structure (1 = lived with both biological parents, 0 = other family form), mother’s employment (dummy coded as not working, works 1–29 hours/week, or works 30 hours or more a week), number of household members under age 18 other than the respondent, parental education (an ordinal variable ranging from 1, less than high school, to 5, post-college degree based on the education of the parent with the highest level of attainment), and the student’s depressive symptoms, which was based on a modified Center for Epidemiologic Studies-Depression scale in which students during the in-home survey reported on the frequency in the past week that they had nine feelings (e.g., “You felt that you could not shake off the blues, even with help from your family and your friends”). Responses, which ranged from 0 (never or rarely) to 3 (most of the time or all of the time), were summed into a 27-point scale of escalating symptomatology (Perreira et al. 2005).

Because neighborhood residence is related to both family SES and characteristics of students’ schools (Adler and Stewart 2009; Brooks-Gunn et al. 1993), we also measured neighborhood-level disadvantage based on data from the parent survey of the in-home interview, including a measure for neighborhood problems—which reflected the average response to two questions about the extent to which litter/trash and drug dealers/users were in the neighborhood (0 = no problem at all, 1 = a small problem, and 2 = a big problem)— and a binary indicator for whether or not the parents reported a strong desire to move from the neighborhood (1 = yes, 0 = no). Finally, a set of binary controls tapped other individual-level factors that could influence weight, including gender and whether the student was involved in athletic activities at school (see Crosnoe et al. 2008), had any functional limitations (1 = yes, 0 = no), or had gone through early puberty (age standardized). Univariate descriptive statistics for all school-level and individual-level covariates are presented in Table 1.

Table 1.

Descriptive Statistics for School- and Individual-Level Covariates (n = 18,924).

Covariates	Mean	(Standard deviation)	Frequency (%)
School-Level Covariates
Size of student bodya
≤125			3.32
126–350			8.90
351–775			24.68
≥776			63.09
Percent whitea
Zero %			2.18
1%–66%			54.51
67%–93%			43.02
94%–100%			0.29
Private school			7.72
Region
Northeast			14.60
Midwest			23.76
South			37.06
West			24.58
Location
Urban			29.51
Suburban			53.96
Rural			16.53
High school and feeder school same			24.26
Proportion living with two bio parents	0.505	0.118
Individual-Level Covariates
Age	16.186	1.712
Female			50.91
Race/ethnicity
White			50.97
African American			21.70
Hispanic			17.07
Asian			7.32
Other/multiracial			2.85
Live with both bio parents			53.33
Mother’s Employment
Mother out of labor force			28.22
Mother works 1–29 hours/week			11.55
Mother works 30+ hours/week			60.23
Parent’s Highest Education
Less than high school			13.19
High school graduate			29.67
Some higher education			21.13
College graduate			23.51
Post-college degree-earner			12.51
Strong desire to move from neighborhood			15.27
Problems in neighborhood	0.420	0.410
Number other minor household members	2.070	1.314
Has any functional limitation			4.63
Individual stress	7.324	5.995
Early puberty (age standardized)			15.05
Athlete			56.87

^aCategorizations mirror those used by Add Health staff in the creation of the school-level weights.

Plan of Analysis

For the multivariate models, we combined logistic regression with multilevel techniques in Mplus (Muthén and Muthén 2006). Multilevel modeling addressed the sampling design of Add Health in which students were sampled within schools, and thus, observations of students within the same school were not independent. It does so by accommodating residuals at both the individual and school level, thereby adjusting for an unobserved between-school component (the variance of the school-level residuals) that could not be accounted for with traditional regression techniques and is often correlated with the dependent variable (Bryk and Raudenbush 1992). To help adjust for the complex survey design, we also used the CLUSTER feature in Mplus to incorporate and appropriately scale the individual- and school-level sample weights and specify them at their respective levels, as recommended in the Add Health documentation (Chen and Chantala 2014). To address missingness on all variables, we used Mplus’ built in multiple imputation procedure to produce 20 fully imputed data sets. Analysis of the imputed data in Mplus is essentially equivalent to analysis of unimputed data in Mplus using FIML (see Asparouhov and Muthén 2010, for more information).

The first step of our modeling sequence was to regress the individual-level odds of being overweight on the proportion of the school that was poor (Model 1) in a logistic regression framework, controlling for the full set of school- and individual-level covariates described above, including students’ own poverty status. This step established whether there was a significant association between increases in the proportion of the school that was poor and the odds that a student was overweight, net of that student’s own background characteristics. As the next step, we tested the significance of the proportion of overweight students in the school (Model 2) and assessed the extent to which the coefficient for school-level poverty changed as a result. As a third step, we examined the robustness of this association with the inclusion of measures of school resources and the stress environment (Model 3). Finally, we explored the social processes linking exposure to overweight students in the school with one’s own risk of being overweight by adding measures for less strict body size norms (Model 4), less common weight control practices/more obesogenic behaviors (Model 5), the social penalty of being overweight (Model 6), and all such factors together (Model 7). Due to concerns that our measure of stigmatization of body size was less reliable in schools with only a handful of overweight students, we restricted Models 6 and 7 of this analysis to exclude adolescents in schools with a small number of overweight students (determined to be 10 students). This restriction, which resulted in an analytical sample of 16,591, while a precautionary step, did not alter the results compared to those using the full sample.

Results

Differences among Students by Adolescent Weight Status

As a starting point, we examined differences in the characteristics and school contexts of overweight students compared to students who were not overweight (Table 2), using t tests to assess statistical significance. Compared to students who were not overweight, a larger percentage of overweight students were from poor families (29.45% of overweight students were from poor families vs. 23.99% of students that were not overweight) and attended schools with higher concentrations of poor schoolmates (overweight students attended schools where 42.1% of the student body was poor; for students that were not overweight, this figure was 37.0%). Students who were overweight also attended schools with higher proportions of overweight schoolmates (27.1% of student body was, on average, overweight) than students who were not overweight (in which 24.6% of the student body, on average, was overweight). These patterns are in line with the basic paths of our conceptual model.

Table 2.

Bivariate Associations between Adolescents’ Weight Status and the Measures of Poverty, School Contexts, and Peer Mechanisms (n = 18,924).

	Weight status
	Not overweight	Overweight	Significance
Poverty Measures
Individual-level poverty status
Poor	70.55%	76.01%	***
Not poor	29.45%	23.99%	***
School-level proportion poor	0.370 (0.326)	0.421 (0.589)	***
School Context
1. School-level proportion overweight	0.246 (0.112)	0.271 (0.208)	***
2. School Resources
Proportion in school involved in athletics	0.425 (0.315)	0.429 (0.480)	NS
Days per week students attend physical education	2.558 (1.596)	2.570 (2.845)	NS
Presence of weight loss program at school
Weight loss program	76.69%	73.42%	**
No weight loss program	26.58%	23.31%	**
3. School Stress Environment
School attachment	−0.381 (1.019)	−0.407 (1.538)	†
Fitting in at school	3.715 (0.266)	3.714 (0.501)	NS
Attachment to teachers at school	5.001 (0.312)	5.006 (0.539)	NS
Violence at school	0.801 (0.357)	0.826 (0.367)	***
Truancy at school	0.265 (0.224)	0.267 (0.193)	NS
Social Mechanisms
School-level proportion dieting	0.336 (0.093)	0.344 (0.170)	***
School-level physical activity	3.679 (0.952)	3.646 (1.580)	*
School-level energy-dense snacks	1.151 (0.133)	1.149 (0.214)	***
Social penalty for being overweight	0.475 (1.538)	0.512 (2.592)	†
School body size norms	0.071 (0.064)	0.082 (0.127)	***
Weighted percent	74.43%	25.57%
Raw ns	13,989	4,787

Note. Sample sizes vary across cells due missing information on income or the relationship measures. All models apply sampling weight. Columns represent within group means.

^†p < .10.

^*p < .05.

^**p < .01.

^***p < .001.

A second descriptive analysis explored how measures of the other contextual school-level factors specified in the conceptual model (i.e., school resources, stress environment) and the social mechanisms varied by adolescents’ weight status. Among the three measures of school resources (proportion in athletics, frequency of physical education, on-site weight loss program), we found that overweight youth were less likely to attend schools with weight loss programs than youth who were not overweight. We did not find significant variation by weight status among four of the five indicators of the stress environmental (school connectedness, fitting in, teacher attachment, and truancy) at the minimum probability level of p < .05, although we did find that overweight students were more likely to attend schools with fighting than students who were not overweight. As for the social mechanisms, we found that overweight youth attended schools in which fewer students were dieting or exercising compared to youth who were not overweight (although their schoolmates also ate energy-dense foods less frequently) and where they and their overweight schoolmates reported that their body size was “just right” more so than adolescents who were not overweight.

School Poverty and Students’ Risk of Overweight

Turning to the multivariate analysis, Table 3 presents the coefficients from the logistic regression models predicting students’ overweight status by the focal predictors of school poverty and exposure to other overweight students in the school context. These models adjusted for the full set of individual- (e.g., student’s own poverty status, family structure, neighborhood quality) and school-level covariates (e.g., school size, region), and employed multilevel modeling techniques. In subsequent steps, they also incorporated the measures of school resources and the stress environment.

Table 3.

Logistic Regression Models Predicting Overweight by Poverty and School Contexts (n = 18,924).

	Coefficient (Standard error)
	Model 1	Model 2	Model 3
Poverty Measures
Individual-Level Poverty (income-to-needs below 185%)	0.078 (0.087)	0.076 (0.087)	0.068 (0.088)
School-Level Poverty (proportion income-to-needs below 185%)	0.865*** (0.239)	0.018 (0.139)	0.141 (0.143)
School Weight Status (proportion overweight)	—	5.060*** (0.286)	4.849*** (0.312)
School Resource Factors
Proportion in school involved in athletics	—	—	0.270 (0.197)
Number of days per week students attend physical education	—	—	−0.003 (0.016)
Presence of weight loss program at school	—	—	−0.004 (0.050)
School Stress Environment Factors
School connectedness	—	—	0.045 (0.065)
Degree to which students feel they fit in at school	—	—	0.147 (0.098)
Degree to which students feel attached to teachers at school	—	—	−0.017 (0.100)
Violence at school	—	—	−0.189* (0.087)
Truancy at school	—	—	0.458* (0.162)

Note. All models are multilevel models with individual and school weights applied, control for individual and school-level covariates, and adjust for missing data. With the exception of the individual level of poverty, all variables shown above measured at the school level.

^†p < .10.

^*p < .05.

^**p < .01.

^***p < .001.

In Model 1, school-level poverty was significantly associated with students’ overweight status (β = .865; SE = .239; p < .001) above and beyond the individual-level poverty measure, which was not significantly associated with one’s risk of being overweight with the measure of school-level poverty and other covariates in the model. To provide a substantive interpretation of these coefficients, we first converted them to odds ratios. We then estimated students’ average odds of being overweight at different school poverty exposures (i.e., percentages of the school student body that is poor). These calculations revealed that a student attending a low-poverty school (determined to be a school where 20% of students came from poor families based on establish categorizations [NCES 2019]) had about an odds of being overweight of 1.19, or a 19% increased odds of being overweight, compared to a student who attended school where none of the students were from poor families. For a student attending a school in which 50% of the student body came from poor families, the odds ratio was 1.54. For students in a high poverty school (80% of students from poor families), the odds ratio was 2.00. These latter calculations convey how the odds of being overweight for an adolescent attending a high-poverty school is twice as large as for a student attending a school with no poor peers.

As the next step, in Model 2 we included the school-level measure of the proportion of overweight students. Doing so reduced the focal coefficient for school-level poverty to non-significance. Meanwhile, the association between the proportion of the student body that was overweight and individual students’ risk of being overweight was strong and significant (B = 5.060, SE = .286). Performing the same calculation as above in order to provide a substantive understanding of the coefficient revealed that a 10% increase in the proportion of the school that is overweight was associated with a 66% increase in the odds that an individual student would be overweight.

Next, we added the measures of school resources and the stress environment to assess the extent to which correlated factors explained away the significance of school-level weight composition. These results, presented in Model 3, revealed that the proportion of overweight peers in the school remained a robust predictor of the individual-level measure of overweight (β = 4.849; SE = .312; p < .001), even with such factors included in the model. In fact, none of the school resource measures was significantly associated with individual overweight status. Among the five stress environmental factors, only one was significant in the expected direction: schools with more truancy were associated with an increased risk of being overweight. Unexpectedly, school violence was significantly associated with a lower risk of being overweight.

Taken together, these results provided support for the central elements of our conceptual model. Specifically, attending a school with higher concentrations of students from low-income families was associated with an adolescent’s increased likelihood of being overweight, net of her or his own poverty status (and other relevant factors, such as neighborhood factors), and this increased risk of being overweight in higher-poverty schools was connected to students’ exposure to other overweight peers in such schools. Furthermore, it did not appear that this link was confounded by school resources or the stress environment of the school, which could vary by both the economic composition of the student body and the weight composition of the study body.

Exploring Social Mechanisms

As a final step, we sought to examine several potential social mechanisms that may explain the link between exposure to more overweight schoolmates and adolescents’ individual risk of being overweight. In particular, we focused on school norms around acceptable body size, school-level behaviors associated with body weight, and social penalties linked to being overweight. These results appear in Table 4.

Table 4.

Logistic Regression Models Predicting Overweight by Social Mechanisms.

	Coefficient (Standard error)
	Model 4	Model 5	Model 6	Model 7
Poverty Measures
Individual-Level Poverty (income-to-needs < 185%)	0.068 (0.088)	0.069 (0.088)	0.062 (0.091)	0.062 (0.091)
School-Level Poverty (proportion income-to-needs < 185%)	0.138 (0.145)	0.062 (0.162)	0.113 (0.146)	0.006 (0.171)
School Weight Status (proportion overweight)	4.796*** (0.329)	4.772*** (0.335)	4.778*** (0.310)	4.521*** (0.369)
School Resource Factors
Proportion in school involved in athletics	0.268 (0.198)	0.324 (0.217)	0.260 (0.194)	0.250 (0.219)
Number of days per week students attend physical education	−0.004 (0.016)	0.000 (0.015)	−0.010 (0.016)	−0.013 (0.016)
Presence of weight loss program at school	−0.004 (0.051)	−0.003 (0.050)	−0.007 (0.051)	−0.003 (0.051)
School Stress Environmental Factors
School connectedness	0.040 (0.067)	0.067 (0.067)	−0.017 (0.070)	−0.021 (0.080)
Degree to which students feel they fit in at school	0.149 (0.096)	0.160 (0.112)	0.168† (0.100)	0.171 (0.115)
Degree to which students feel attached to teachers at school	−0.027 (0.104)	−0.017 (0.104)	0.000 (0.105)	0.006 (0.109)
Violence at school	−0.193* (0.087)	−0.171† (0.087)	−0.220* (0.086)	−0.209* (0.086)
Truancy at school	0.462** (0.161)	0.385* (0.171)	0.442* (0.177)	0.372* (0.184)
Social Mechanisms
1. School-level body size norms	0.188 (0.492)	—	—	0.807 (0.617)
2. School-level norms of behaviors
School-level proportion dieting	—	0.004 (0.242)	—	−0.032 (0.299)
School-level physical activity	—	−0.041 (0.044)	—	−0.047 (0.044)
School-level energy-dense snacks	—	−0.266 (0.251)	—	−0.183 (0.256)
3. School-level social penalty for being overweight	—	—	0.025 (0.019)	0.028 (0.021)

Note. In Models 4 and 5, n = 18,924. In Models 5, 6, and 7, n = 16,951 All models are multilevel models with individual and school weights applied, control for individual and school-level covariates, and adjust for missing data.

^†p < .10.

^*p < .05.

^**p < .01.

^***p < .001.

Model 1, which added the measure of school body size norms to the model, did not reveal evidence that school-level perceptions of acceptable weight were associated with individual students’ risk of being overweight. In Model 2, measures of school-level behaviors including the proportions of students who dieted, students’ average weekly physical activity, and students’ average consumption of energy-dense snacks were not significantly associated with individual student’s risks of being overweight. Likewise, Model 3 did not reveal evidence that social sanctions due to body size were associated with students’ risk of being overweight. Model 4, which included the complete set of mechanisms, also did not provide evidence that the proposed mechanisms illuminated the link between the school-level measure of overweight and individual risk of overweight. This pattern of results was also the same when we removed the variable for school violence, which was significantly negatively associated with individual student’s risk of overweight and potentially operating as a suppressor variable.

Robustness Analyses

Given the lack of significance among the social mechanisms, we pursued several robustness analyses. First, we considered the potential that the mechanisms operated at a “threshold” by modeling the measure of the proportion of the school that was overweight two other ways: by adding a quadratic term to the model and by breaking the linear measure into categories that reflected quartiles (e.g., bottom 25% of the distribution, 25th–49th percentile). These models did not reveal evidence of a threshold effect. The quadratic term was insignificant, and all dummy quartile variables were statistically significant, with the magnitude of the coefficients following a linear trend.

As a second step, we reestimated the full analysis using a measure of obesity (i.e., BMI at or above the 95th percentile for age and gender based on CDC growth charts). These results mirror the ones reported in Tables 3 and 4, using both the linear term for school-level obesity and the quadratic term/dummy categories. We also explored our models subdivided by gender. Prior research highlights how boys tend to be less concerned with body size than girls (Cohane and Pope 2001), are less penalized socially for having a non-normative body size (Crosnoe et al. 2008), and are less perceptive of messages around body size (Gillen and Lefkowitz 2009). We did not find evidence, however, of significant peer-related mechanisms among the subsample of girls.

Finally, we examined the potential for more proximate and direct individual-level measures of weight-related behaviors (e.g., dieting, physical activity, and consumption of high-calorie snacks) to capture the link between school-level and individual-level weight. Thus, we extended the models in Table 3, Model 3, by adding the individual measures of weight-related behaviors. Results indicated that physical activity was associated with lower risks of being overweight (β = −.065; p < .05), although dieting was associated an increased likelihood of being overweight (β = 2.046; p < .001), and less consumption of high-calorie snacks (β = −.181; p < .001) was associated with increased risks. These latter patterns may reflect reverse causality, as overweight students are more likely to diet and reduce their caloric intake. The inclusion of these factors to the model, however, did little to alter the coefficient for school weight composition. Adding these measures to Model 7 in Table 4, however, resulted in the school-level dieting measure to become a significant negative predictor of individual students’ overweight (β = −1.979; p < .001), while the individual factors remained significant. This change in the coefficient for school-level dieting may reflect the suppression of the inverse association between the proportion of the student body dieting and individual overweight because there was less dieting in schools with more overweight students. We are cautious of providing a more substantive interpretation of this change in coefficient size, however, given our modeling approach (see Mood 2010). The full results for the robustness analyses are available upon request.

Discussion

In a national context of increasing concentrated poverty within schools (Reardon and Owens 2014), we aimed to better understand the consequences of this phenomenon for an aspect of student well-being that is not often considered an outcome of “school effects”— student weight status. Being overweight is associated with contemporaneous risks to student’s well-being, including lower self-esteem and academic achievement (Crosnoe 2011; Taras and Potts-Datema 2005), as well as longer-term risks, including adults’ risk of obesity, other morbidities, and traditional indicators of truncated mobility, including lower levels of educational attainment, lower earnings, and less labor force participation (Cawley 2004; Crosnoe 2011; Singh et al. 2008). To better understand this important health issue and its connection to school poverty, we considered a contextual feature of schools during a critical stage of both education and development, secondary schooling and adolescence, when students and their health are more vulnerable to social influences (Crosnoe 2011; Steinberg 2005), and when lifelong health behaviors are being formed and reinforced (Kelder et al. 1994).

As our results revealed, consistent with prior research (e.g., Richmond and Subramanian 2008), students who attended higher-poverty schools had an increased risk of being overweight, net of their own family poverty background. In fact, after accounting for the economic composition of the school and other confounds, the association between one’s own poverty status and one’s risk of being overweight was non-significant. Although this finding is contrary to commonly held wisdoms about elevated rates of overweight in poorer segments of the population, this non-significant result is common in the literature on adolescent health. As Martin and colleagues (2012) argued, the risks of being overweight for poor adolescents are primarily experienced at the school level because of the ways that schools powerfully shape what students do and think during adolescence. The second aim of this study was to better understand these school-level processes.

Building on prior research calling for more attention to compositional features of low-income schools (e.g., Miyazaki and Stack 2015; Procter et al. 2008), including seminal work by Jencks and Mayer (1990), who argued that doing so highlighted the “most politically salient and easily understood differences between [neighborhoods and] schools” (p.177), we hypothesized that the weight risks of concentrated poverty in schools would be channeled through student’s increased exposure to other overweight classmates. Our results supported this hypothesis. In fact, after accounting for the association between the concentration of overweight peers in school and students’ individual-level risk of overweight, the coefficient for school-level poverty became insignificant. To our knowledge, this is the first study to provide such evidence. This finding supports Coleman’s original argument about the role of school context and adolescent development and extends this argument to their physical heath; parallels research on neighborhoods, which have documented a similar link between the weight concentration of the community and individual’s risks of being obesity (Boardman et al. 2005); and is supported by work by Evans and colleagues (2016), who suggested the importance of the social network-level effect they observed was “actually characterizing the social environment endogenous to the school” (p. 29).

At the same time, we found that other school-level contextual factors viewed as part of the link between school poverty and students’ risk of overweight—namely the stress environment and school resources (e.g., Martin et al. 2012; McFarland et al. 2015)—did not appear to be salient factors, nor did adjusting for such factors alter the coefficient for the proportion of the school that was overweight. The exceptions were violence and truancy, in which exposure to schools with more truancy was associated with an increased risk of being overweight, while exposure to schools with more violence was associated with a significantly lower risk.

These findings underscored the importance of exploring the social mechanisms linking school poverty with students’ weight risks, which we did by theorizing and measuring several school-level mechanisms, including school norms around body size, the prevalence of weight related behaviors in the school, and social stigma around body size. Unfortunately, we only found evidence for one factor, and only after controlling for individual-level behaviors. Specifically, the prevalence of reported dieting was lower in schools serving more overweight (and higher proportions of low-income) peers. This lack of dieting behavior pointed to the possibility that in such schools, dieting norms were weaker or students were less concerned about losing weight, although again, we could not tease out the causal impact of school dieting prevalence given our cross-sectional approach. Indeed, as the coefficient for individual dieting indicated, overweight students are also more likely to diet than their counterparts who were not overweight. Furthermore, the efficacy of dieting as a weight control practice is unclear (and may contribute to weight gain), as were the “dieting” practices that students actually engaged in (Neumark-Sztainer et al. 2012), which we could not shed more light on due to the questionnaire wording. It is also important to recognize that dieting practices can carry serious health risks and have been associated with other unhealthy behaviors (Eisenberg et al. 2005).

We recognize that this lack of evidence for the social mechanisms is surprising, but we offer several explanations that also highlight future areas of research. First, the Add Health measures, albeit widely used in research of this kind (Mueller 2015), may be inadequate for capturing the proposed mechanism because they were prone to error (e.g., reports of food consumption), ambiguous (e.g., dieting), unidimensional (e.g., physical activity did reflect duration or intensity), or sought to capture complex processes shaping one’s cognitions about acceptable behaviors and body size norms using discrete dichotomous measures. Scholars should continue to pursue this research by exploring other methods, such as participant observation or an experiment, the former which provides nuance and depth, the latter which captures subtle social processes; time diary methods to yield more accurate information about behaviors; or more nuanced and detailed survey measures.

Second, the salience of the proposed mechanisms linking the weight composition of the student body and adolescent’s risks of overweight may vary across schools with different student populations in ways we did not explore. For example, well-documented racial differences in body size norms may diminish the power of the mechanisms to influence body weight among black adolescents (Kemper et al. 1994), while racial and ethnic minorities may face other challenges outside of the school avoiding overweight, such as paid employment (Barroso et al. 2010). Fourth, the duration of exposure to social mechanisms, which we could not measure in the absence of longitudinal school-level data, may be extremely important. As Boardman and colleagues (2005) explain, “negative health behaviors [linked to weight] are slowly disseminated through the social and cultural environment” (p. 238). Thus, duration of exposure to social factors should also be examined. Fifth, it is possible that many of the social mechanisms we proposed are more directly channeled through immediate peers, who tend be homophilous in terms of body size (De la Haye et al. 2011; McPherson, Smith-Lovin, and Cook 2001), although there is conflicting evidence of whether close friends’ behaviors influence one another’s’ weight (Cunningham et al. 2012; Eisenberg et al. 2005). Unfortunately, we could not measure such factors because Add Health did not collect data from all close peers, although future data collections could. Sixth, it is possible that the school social context is insufficient to countering students’ perceptions of well-documented weight biases and internalized stigma against overweight people, which can lead to eating disturbances in ways adolescents may not be fully aware of, avoidance of physical activity, and psychological distress (Puhl and Heuer 2009; Puhl, Moss-Racusin, and Schwartz 2007). Weigh biases may also contribute to student’s weight through other school-related pathways we did not explore, for example, through teachers’ and administrators’ attitudes and behaviors (Nutter et al. 2019). Seventh, poverty also influences weight through food insecurity, which is an issue that deserves more attention. Finally, we acknowledge an argument informed by the fundamental cause framework; that the predictive power of the school social context may diminish when broken into its component parts, which cannot be reduced into observable mechanisms (Phelan et al. 2004) in ways that explain the “effect” of school poverty.

Beyond such factors, we acknowledge other limitations of our study. First, as mentioned several times above, we cannot make causal claims about the links between school poverty, school overweight, and student’s weight. Although we took steps to account for a large number of confounds, including individual level (e.g., depression, neighborhood quality) and school-level controls (e.g., school size), along with measures of school resources and the stress environment, the links between school overweight and individual overweight could still have been driven by unmeasured factors that selected overweight students into schools with higher concentrations of overweight peers. Taking a longitudinal approach would help address this issue by capturing the associations between changes in body size and changes in school context, but we were limited to a cross-sectional approach because the second wave of Add Health lacked the in-school component and did not provide school-level measures of poverty and overweight (Martin et al. 2012). We also acknowledge that weight/height was self-reported and cannot rule out the possibility that error in these reports was connected to student poverty in ways that confounded our results. Finally, we recognize that Wave I of the Add Health was collected over 20 years ago. School contexts and their connection to school poverty may have changed since then, as have the mechanisms shaping student’s weight. A new data collection similar in design to the Add Health would be needed to assess this possibility.

In conclusion, this study helps open up the rich tradition of research on school effects to a new line of inquiry that spans research on health, education, and stratification to better understand the causes and consequences of two timely social problems: the increasing concentration of poor students in U.S. schools and high rates of overweight amongst American youth. Doing so revealed a non-educational consequence of school concentrated poverty—increased exposure to other overweight peers in the school context—which had contemporaneous and future implications for adolescent wellbeing. These results spotlight in a new way the economic, social, and public health impact of economic school segregation.