Adolescent Body Mass Index and Exposure to Peers with Overweight and Obesity: A Structural Equation Model Approach to Longitudinal Network Data

Abstract

Purpose:

Considerable evidence has shown that social networks influence a wide variety of health behaviors. This study investigates whether having friends with overweight/obesity in one's social network (network exposure) can predict changes in body mass index (BMI) throughout high school in a diverse urban population of students.

Methods:

Racially and ethnically diverse students from five high schools in Los Angeles County were surveyed at four time points throughout high school from 2010 to 2013 (N = 2091). Surveys included questions on students' social networks, demographics, and health-related information. BMI and weight categories were calculated for all students who provided height and weight information (∼50%). A latent growth curve model was used to assess the growth trajectory of BMI and the time-varying effect of network exposure to friends with overweight/obesity while controlling for demographic covariates.

Results:

Hispanic students had a significantly higher initial BMI compared with non-Hispanic students (p < 0.01). There was a significant positive slope for time on BMI growth (p < 0.01). Greater personal network exposure to friends with overweight/obesity was associated with a significant 0.65-point average increase in BMI (p < 0.05) at the first follow-up time point (T2) and a significant 0.62-point average increase in BMI (p < 0.01) at the last follow-up (T4) while controlling for covariates.

Conclusions:

Using structural equation modeling to understand the relationship between BMI and social networks, we found that increased network exposure to peers with overweight/obesity is associated with higher individual BMI, demonstrating that friendships may influence adolescents' weight status over time.

Introduction

Health behaviors developed in childhood and adolescence can persist throughout a lifetime and have long-term impacts on health and well-being.^1–4 Overweight and obesity prevalence has been steadily increasing among adults and youth in the United States.⁵ The most recent National Health and Nutrition Examination Survey (NHANES) data show increases in the prevalence of overweight and obesity among children and adolescents, with 21.2% of 12- to 19-year-olds being categorized as obese. Higher risk subgroups, such as Hispanics, have a higher prevalence of obesity compared with non-Hispanic whites, with a combined childhood and adolescent obesity rate of 25.6%.^5,6

Having overweight or obesity presents both immediate and long-term physical health problems, in addition to social difficulties for adolescents.⁷ With no indication of overweight and obesity rates decreasing, there is an imminent need to explore social factors influencing these trends and subsequent intervention strategies to address them. One approach to exploring the mechanisms behind these social processes is social network analysis.

One's weight status can be influenced by multiple factors, including social influences, such as the behaviors of family and friends. These influences may shape diet and exercise patterns in childhood and adolescence. One mechanism that may lead to homophily on weight status is social selection, where people seek out friends with a similar weight status to feel a sense of belonging.^8,9 This can lead to people with similar attributes being connected in a social network. Another possible mechanism is social contagion, when people adopt weight influencing behaviors of others in their network. The social influence of people in our personal networks may normalize the consumption of certain foods or levels of physical activity, which ultimately impacts weight status and health outcomes.

The importance of friendships and peer relationships leaves adolescents particularly susceptible to social influences that may have powerful effects on health behaviors, weight status, and health outcomes. Peer influence has been demonstrated for many health behaviors such as tobacco, alcohol, and other substance use.^10–12 Research among adolescents and young adults has shown that higher body mass index (BMI) among one's friends is correlated with one's own higher BMI.^13–15 However, this relationship may vary by gender and ethnicity.¹⁶ The social influences from one's friends can influence physical activity levels, sedentary behavior, weight status, and affect one's efforts to lose weight.^17–20

Social selection processes may contribute to more network homophily regarding weight status. Social network studies among children and adolescents have found that weight homophily among friends may be due to the exclusion of peers with overweight/obesity, as they are less likely to receive friendship nominations compared with peers of a healthy BMI and more likely to be marginalized by others.^21–25 Social selection could explain network clustering based on weight status in which students with overweight and obesity cluster together due to social exclusion from other groups.

Hispanic children and adolescents experience culturally unique risk factors for obesity²⁶ and have a higher obesity rate compared with non-Hispanic whites,^5,6 which may be associated with greater network exposure to peers with overweight/obesity. This may in turn lead to stronger social effects on weight-influencing behaviors. Research among Hispanic adults has found that they have a higher proportion of contacts with overweight/obesity in their social networks and are more likely to regard a larger body size as normal.^27,28 Based on what we know about Hispanic adults, Hispanic adolescents may also have a higher proportion of people in their network with overweight/obesity, which may be associated with increased exposure to unhealthy eating or exercise habits.

Past social network studies examining BMI status have traditionally modeled BMI as a categorical outcome through regression or actor-oriented models,^{15,16,20,21,29} and there remains a lack of research using social network analysis to examine BMI changes among Hispanic adolescent populations. The objective of this study is to use structural equation modeling (SEM) with longitudinal social network data to analyze the relationship between longitudinal changes in BMI and network exposure to peers with overweight/obesity. By using an SEM framework, we can examine the changes in BMI growth and network exposure at each wave to further understand the influence of network exposure to peers with overweight/obesity at each time point in high school. We hypothesize that:

• 1.Greater network exposure to friends with overweight and obesity is associated with more accelerated BMI growth throughout high school.
• 2.Hispanic ethnicity has a significant effect on initial BMI status and BMI slope/growth.

This study uses latent growth curve modeling (LGCM) to analyze longitudinal weight status among adolescents with their social network data as a time-varying covariate, presented as an alternative to traditional longitudinal network analyses. With BMI as an observed characteristic that may be influenced by underlying factors, a latent growth model can be used to estimate BMI trajectory while controlling for demographic factors and fluctuations in network exposure over time.

Methods

There are various approaches for analyzing longitudinal network data, including regression models,³⁰ multilevel models,³¹ stochastic actor-oriented models (SAOM or Simulation Investigation for Empirical Network Analysis [SIENA] models),³² and SEM.^33,34 When analyzing longitudinal network data, it is important to model the data with consideration of its structure related to time and nonindependence of observations. It is also important to accurately capture the relationship between the outcome (i.e., changes in BMI) and time-varying predictors (i.e., fluctuation in network exposure). LGCM was chosen over SAOM and traditional regression modeling for the following reasons.

LGCM versus SIENA

Longitudinal changes in social networks and behavior are often simultaneously estimated using dynamic network models that take an agent-based approach to model network change, known as SAOM or SIENA models. These models take into account the effects of network structure, individual level covariates, and behavior change on selection and influence processes.³² SIENA models are widely used to examine friendship networks and behavior changes.^{21,32,35–38} Compared with conventional linear modeling approaches, SIENA models appear to have many benefits. However, recent research comparing SIENA models with conventional regression models in longitudinal peer network data has found that SIENA model estimates are not more conservative than other methods.³⁰

We propose that LGCM is appropriate for our research question because traditional SIENA models require a binary or ordinal discrete outcome variable³⁹ and our outcome, BMI status, is being modeled as a continuous variable. Whereas methods for continuous outcomes in SIENA models are being developed.^40,41 SIENA models also must be fit separately for each network, and in this case would require fitting separate models for the five schools. In addition, LGCMs are robust to missingness, whereas SIENA requires imputation or other strategies to handle missing data.⁴² Using an LGCM allows for flexibility in the dependent variable and does not require fitting models separately for each network thus retaining power and allowing us to detect more granular changes in BMI status.

LGCM versus Traditional Regression

This article uses an LGCM, which falls under the SEM framework. LGCM was chosen over traditional regression because a latent variable approach allows us to estimate latent growth factors (intercept and slope) for growth trajectory of any time-varying outcome (i.e., endogenous variables) included in the model. Using this framework, individual differences (random effects) are captured in the latent variable growth factors.³³ SEM allows for the effect of time-varying predictors to be unequal at each time point, in contrast to typical regression approaches that produce marginal effects of time-varying predictors. Essentially, regression uses a univariate approach where time-varying predictors are estimated as a pooled fixed effect that is nested within individuals.³³

In our study, the time-varying effect of network exposure on BMI will be expressed as a multivariate outcome in SEM, which also allows for residual variances to differ across time.³³ The time-varying effect of network exposure on BMI will result in four effect estimates, one for each wave of data. This allows us to see if the network exposure effects vary during different years of follow-up. We will also control for gender and ethnicity to examine the effects of network exposure on BMI growth trajectories for males and Hispanics.

Sample

Racially and ethnically diverse students from five high schools in Los Angeles county were surveyed at four time points throughout high school from 2010 to 2013.⁴³ Surveys included questions on students' social networks, demographics, and health-related information.

Students participated in a baseline survey during the fall of 10th grade, a 6-month follow-up survey later that academic year in April, and two subsequent follow-up surveys in April of their 11th and 12th grades, resulting in four waves of data. Students who provided race and gender data were retained in the sample. The sample was 49.60% male and 78.62% Hispanic, yielding an analytic sample of 2091 students. The average age at baseline was 15.07 years. This study was approved by the University of Southern California Institutional Review Board.

Measures

Covariates

Demographic data on gender and race/ethnicity were collected. Students could select multiple racial or ethnic categories (Table 1). Gender (male = 1, female = 0), and ethnicity (Hispanic = 1, non-Hispanic = 0) were controlled for as covariates in the conditional LGCM.

Table 1.

Analytic Sample Descriptive Statistics (N = 2091)

Demographics	n (%)
Gender
Male	1037 (49.60)
Female	1054 (50.41)
Ethnicity/race
Hispanic	1644 (78.62)
White	1408 (85.64)
Black	48 (2.92)
Asian	79 (4.80)
American Indian	25 (1.52)
Pacific Islander	12 (0.73)
Multiracial	72 (4.38)
Non-Hispanic	447 (21.38)
White	19 (4.25)
Black	0 (0.00)
Asian	389 (87.03)
American Indian	2 (0.50)
Pacific Islander	2 (0.50)
Multiracial	35 (7.83)
Mean (SD)
Age
T1	15.07 (0.41)
T2	15.44 (0.56)
T3	16.37 (0.59)
T4	17.60 (0.52)
BMI
T1	22.65 (4.37)
T2	23.02 (5.23)
T3	23.33 (4.55)
T4	24.01 (4.94)
Network exposure
T1	0.22 (0.25)
T2	0.24 (0.27)
T3	0.26 (0.29)
T4	0.26 (0.31)

SD, standard deviation.

BMI

BMI measured for each student over four time points is the continuous outcome in the growth model. Modeling BMI continuously allows us to evaluate more gradual changes over time compared with modeling BMI categories or z-scores as our outcome. Research has also shown that BMI z-scores and percentiles have limitations with capturing change among high BMI individuals.^44–48 BMI was calculated for all students who provided height and weight information. Based on CDC child and teen percentiles and z-scores, values that were more than four standard deviations below the mean were considered data errors or outliers and removed (e.g., observations where BMI = 0). One's own BMI trajectory was modeled as a continuous outcome, and weight status categories of one's friends were used to calculate network exposure.

Friendship network nominations

Students were asked to list their seven closest friends “Please think of your seven BEST FRIENDS in [your] grade. If you don't know their names you can refer to the GRADE ROSTER. Be sure to write your friends' real names and roster ID numbers.” The network data were used to create a directed adjacency matrix, where each directed pair of students x_ij = 1, if student i nominated student j as a person they are best friends with. Friendship nominations were collected at each wave and then used to calculate network exposure for each wave.⁴⁹

Network exposure

The time-varying predictor in our model is network exposure to friends with overweight/obesity. We created this variable by combining the friendship network nominations and these nominated friends' categorical weight status. Weight categories for friends were calculated based on CDC child and teen percentiles that are adjusted for gender and age: underweight (a BMI <5th percentile), normal weight (5th–85th percentile), overweight (85th–95th percentile), and obese (≥95th percentile).

Friendship nominations were used to calculate students' exposure to friends with overweight and obesity in their networks. Exposure was calculated as the proportion of close friends who were categorized as overweight or obese according to CDC child and teen percentiles (i.e., if a participant listed five close friends, and four of those friends were categorized as overweight/obese, then network exposure at that time point would be 0.80). Network exposure was calculated at each wave.

Statistical Analysis

Univariate longitudinal regression analyses were used to determine the relationship between one's own BMI and network exposure, controlling for relevant covariates. The extent of school clustering for BMI was negligible so school was not included as an additional level in the models.

There were 20 observations with missing data on demographic covariates that were not retained in the model, as the intercept and slope of the model were estimated conditional on these covariates.³³ Therefore, only observations that had complete gender and ethnicity data were retained in the final model, yielding an analytic sample of 2091 students. An LGCM was used to assess the relationship between the growth trajectory of BMI and the time-varying effect of network exposure to friends with overweight/obesity; the intercept represents the initial status, and the slope represents the growth rate at each wave (Fig. 1). Intercept and slope are estimated for BMI, and network exposure to friends with overweight/obesity is treated as a time-varying predictor.

Figure 1.

Conditional latent growth curve modeling. Where i = intercept; s = slope; exp1–4 = network exposure to peers with overweight/obesity at waves 1–4.

First, an unconditional univariate LGCM was estimated for BMI. Next, the time-varying network exposure predictor and the time invariant predictors, gender and Hispanic ethnicity, were added to the conditional model. The variance of network exposure across waves was constrained to be equal. The varying follow-up time between waves was controlled for by including time scores for the slope growth factor; values were assigned to indicate that there were 6 months between waves 1 and 2 and 12 months between the subsequent waves. Model fit was assessed using Comparative Fit Index (CFI), Tucker Lewis Fit Index (TLI), and Root Mean Square of Error Approximation (RMSEA). Relative fit indices, including Akaike Information Criterion and Bayesian Information Criterion, were also assessed. All social network analyses were performed in RStudio (version 1.4) and LGCM was estimated using maximum likelihood estimation in Mplus (version 8). Figure 1 shows the model.

Results

Demographic characteristics and descriptive statistics are displayed in Table 1 and show that average BMI and network exposure to friends with overweight and obesity both have an overall positive trend.

Model Fit

Table 2 shows the model fit for the conditional LGCM. The final LGCM had adequate fit, with CFI and TLI both >0.90 and RMSEA close to 0.05.

Table 2.

Model Fit Indices

Fit index	Unconditional LGCM	Conditional LGCM
CFI	0.973	0.916
TLI	0.967	0.921
RMSEA	0.093	0.064
SRMR	0.137	0.118
AIC	23032	24122
BIC	23082	24280

AIC, Akaike Information Criterion; BIC, Bayesian Information Criterion; CFI, Comparative Fit Index; LGCM, latent growth curve modeling; RMSEA, Root Mean Square of Error Approximation; SRMR, Standardized Root Mean Square Residual; TLI, Tucker Lewis Fit Index.

Growth Processes

Table 3 shows the unstandardized parameter estimates from the LGCM. Hispanic students had a significantly higher initial BMI compared with non-Hispanic students (p < 0.01). There was a highly significant increase in BMI growth over time (p < 0.01). There was a significant positive association between network exposure to friends with overweight/obesity and BMI at two time points. At the first follow-up time point (T2) having greater personal network exposure to friends with overweight/obesity resulted in a significant, on average, 0.65-point greater BMI while controlling for gender and ethnicity (p < 0.05). There was also a significant association at the last follow-up (T4), where greater personal network exposure was associated with a 0.62-point average increase in BMI while controlling for covariates (p < 0.01).

Table 3.

Unstandardized Estimates of Conditional BMI Latent Growth Curve Model (N  = 2091)

Parameter	β (SE)
Means
Intercept	20.82 (0.26)
Slope	0.16 (0.04)***
Covariates predicting intercept
Male	0.53 (0.21)*
Hispanic	2.27 (0.26)***
Covariates predicting slope
Male	0.07 (0.04)^
Hispanic	0.03 (0.04)
Time-varying covariates
BMI T1 <- Exposure T1	-0.16 (0.29)
BMI T2 <- Exposure T2	0.65 (0.27)*
BMI T3 <- Exposure T3	0.24 (0.20)
BMI T4 <- Exposure T4	0.62 (0.23)**
Variance components
Intercept	16.16***
Slope	0.014
Residual T1	5.54
Residual T2	6.14
Residual T3	1.68
Residual T4	2.37
Covariance (intercept, slope)	0.056***

^p < 0.10, ^*p < 0.05, ^**p < 0.01, ^***p < 0.001.

Figure 2 provides a visual depiction of the growth in obesity and changes in friendship networks over time.

Figure 2.

Network plot of school 1 at each wave. Nodes are sized based on indegree or the number of friendship nominations received.

Discussion

The purpose of this study was to examine the relationship between network exposure to friends with overweight/obesity and BMI trajectory over time in an ethnically diverse sample of high school students. There was a positive effect for time on BMI while controlling for network exposure, gender, and Hispanic ethnicity. In this sample, findings revealed that BMI increased on average throughout high school and that network exposure to friends with overweight and obesity was positively associated with a student's own BMI, controlling for the effect of time.

Hispanic students had significantly higher initial BMI on average. These findings are consistent with past research that has demonstrated that Hispanics are at greater risk for having overweight or obesity.^6,26 There may be strong social influence effects on weight-related health behaviors for Hispanic students, although these effects may vary by gender. Research has found that Latinas share weight control behaviors and changes in weight status with others in their social networks.⁵⁰ However, our results found that males had a significantly higher initial BMI status and a marginally significant increase in BMI growth on average compared with females.

If network influences remain significant, even after controlling for the natural course of development, as we have shown in this study, then a social network intervention⁵¹ for healthy eating or exercise may be beneficial to all adolescents. Social network interventions can be targeted to include families, friends, schools, and communities. Many interventions among Hispanic families focus on obesity prevention among children, resulting in a significant lack of targeted interventions for Hispanic adolescents.⁵² Since adolescents spend a significant portion of time at school and with friends, a peer-based intervention may be effective for encouraging healthy eating and exercise behaviors among adolescents and curtailing unhealthy behaviors before adulthood.^53,54

Trends for the overall sample showed that average BMI increased over time, reflected by the significant slope growth factor. In the LGCM, there was a significant association between network exposure on BMI growth at two time points (T2 and T4). Therefore, greater network exposure to friends with overweight/obesity is associated with significant increases in BMI at those time points. Adolescent social networks may undergo changes in composition over time due to the formation and dissolution of friendships. Students' friendship networks were measured at each follow-up time point to account for the dynamic nature of networks.

Network exposure had a strong positive association with BMI at T2 and the last follow-up (T4) during senior year. Throughout high school adolescents gain more autonomy and ability to influence their dietary and physical activity behaviors. During this critical period of development, the influence of friends may also increase over time, resulting in stronger peer influence effects. In addition, this transitional period at the end of high school may be difficult or stressful for some students, perhaps leaving them more prone to unhealthy behaviors.

The relationship between network exposure and BMI growth could be interpreted in two ways. First, friendships with peers who have overweight/obesity could affect weight-related health behaviors through social influence.²⁰ Over time adolescents may adopt unhealthy diet and exercise patterns modeled by friends, gradually resulting in increased BMI. The second process is social selection for friends of a similar weight, where students who have overweight/obesity form more friendships with each other as time goes on. Past studies have found a similar mechanism for friendship selection where people of the same BMI status are more likely to share network ties.^21,22,55 In this scenario, students with a higher BMI may be selecting friends of a similar weight status, resulting in network clusters that may be more likely to experience significant BMI increases over time leading to homophily.

Social selection and social influence processes may be interacting or occurring simultaneously. As we see BMI and network exposure gradually increase across waves it is plausible that students select friends with a similar weight status, this in turn increases their network exposure to peers with overweight/obesity and subsequent weight influencing behaviors.

Strengths and Limitations

This study used LGCM to examine the relationship between BMI and network exposure to friends with overweight/obesity. Strengths of this study include the longitudinal study design that gives us a comprehensive overview of BMI and network exposure growth trajectories throughout high school, yielding more information than a cross-sectional approach. In addition, this analytic approach estimates the growth model while controlling for demographic covariates, which allows us to capture the effects of time on BMI change more accurately, compared with a traditional multilevel model that would yield a single estimate for the time-varying effect of network exposure.

SIENA models for continuous outcomes are still under development,^40,41 but the LGCM allows for flexibility in the dependent variable and does not require fitting models separately for each network. Therefore, with increased sample size and statistical power we can detect the multivariate effects of network exposure on BMI growth over time. Finally, the ethnically diverse adolescent sample provides insight about differential outcomes and increased risk for Hispanic adolescents.

The main limitation of this study is the incomplete data. Many students did not provide height and weight data for all waves, which may have led to nonsignificant or marginally significant estimates at some time points. One school did not participate in wave 3, potentially yielding nonsignificant effects at T3. However, MPlus addresses missingness of continuous data through robust maximum likelihood estimation.⁵⁶ This allows us to retain observations that are missing BMI or network exposure data at a certain wave and maximize our sample size, which could be an advantage over SIENA models that would require further strategy or data imputation on both network and behavioral variables. However, our model does not provide insight on network structural effects or influence versus selection processes to the same degree that a SIENA model would. Height and weight data were self-reported, so accuracy may vary. Generalizability of these findings may be limited to Hispanic adolescents, particularly in urban areas.

Conclusions

This study contributes to current research by using social network data and SEM to understand the relationship between BMI and network exposure to friends with overweight/obesity among a predominantly Hispanic adolescent cohort. BMI increased in the cohort over the course of the study, whereas network exposure had a significant effect on BMI, especially at the last follow-up. Hispanic students are at higher risk for having overweight/obesity, as they had higher initial BMIs. This study gives us unique insight to social networks and behavioral dynamics among Hispanic adolescents. Increased network exposure to peers with overweight/obesity is associated with higher individual BMI, demonstrating that friendships may influence adolescents' weight status through processes such as social contagion and social selection.

Funding Information

This study was supported by NIH grant number RC1AA019239 from the National Institute on Alcohol Abuse and Alcoholism; and R01DA051843 from the National Institute on Drug Abuse.

Author Disclosure Statement

The authors have no conflicts of interest to disclose.