|Home | About | Journals | Submit | Contact Us | Français|
Background The prevalence of obesity and overweight is rapidly increasing in industrialized countries, with long-term health and social consequences. There is also a strong social patterning of obesity and overweight, with a higher prevalence among women, racial/ethnic minorities and those from a lower socio-economic position (SEP). Most of the existing work in this area, however, is based on cross-sectional data or single cohort studies. No national studies to date have examined how social disparities in obesity and overweight differ by age and historical period using longitudinal data with repeated measures.
Methods We used panel data from the nationally representative Monitoring the Future Study (1986–2004) to examine social disparities in trajectories of body mass index (BMI) over adulthood (age 18–45). Self-reported height and weight were collected in this annual US survey of high-school seniors, followed biennially since 1976. Using growth curve models, we analysed BMI trajectories over adulthood by gender, race/ethnicity and lifetime SEP (measured by parents’ education and respondent's education).
Results BMI trajectories exhibit a curvilinear rate of change from age 18 to 45, but there was a strong period effect, such that weight gain was more rapid for more recent cohorts. As a result, successive cohorts become overweight (BMI > 25) at increasingly earlier points in the life course. BMI scores were also consistently higher for women, racial/ethnic minority groups and those from a lower SEP. However, BMI scores for socially advantaged groups in recent cohorts were actually higher than those for their socially disadvantaged counterparts who were born 10 years earlier.
Conclusions Results highlight the importance of social status and socio-economic resources for maintaining optimal weight. Yet, even those in advantaged social positions have experienced an increase in BMI in recent years.
The prevalence of overweight and obesity is rapidly increasing in industrialized countries,1–8 with extensive health, economic and social implications.9–17 In the United States, currently two-thirds of adults are overweight or obese.1,2,18,19 Increases in prevalence have taken place primarily since 1980 and have occurred in all age groups simultaneously,1,3,6,18,20 prompting the use of the term ‘epidemic’ to describe these trends.
There is also repeated evidence of social disparities in the prevalence of obesity and overweight. Data from national surveys paint a consistent picture where women, individuals of lower socio-economic position (SEP) and minority racial/ethnic groups have the highest rates of obesity and overweight.3,6,18,21–23 This is consistent with the argument that social disadvantage is a ‘fundamental cause’24,25 of obesity and overweight. People with more knowledge, money, power, prestige and beneficial social connections are better able to control weight gain, either through the ability to make healthy food choices26,27 (by having greater awareness of, access to, and resources to purchase healthy foods), or through greater opportunities for exercise, and safe play.28,29
However, much of the literature on social disparities in obesity is based on data from repeated cross-sectional surveys1,4,6 or from single cohort studies.3,30,31 As a result, it is difficult to determine whether patterns of weight gain over the life course represent true age effects or simply age differences in the characteristics of cross-sectional samples over time. Given the rapid secular changes that have been occurring in obesity prevalence, social disparities in trajectories of weight gain are likely to vary across historical time. The timing of birth exposes each generation or cohort to a particular window in historical time with its own unique set of social and environmental constraints and opportunities.32–34 Social patterns of weight gain are therefore inextricably linked to the social and cultural conditions of the larger society in which they are embedded across time. Longitudinal data with repeated measures are needed to identify how social disparities in obesity and overweight differ by age and historical period.
We use nationally representative panel data from the Monitoring the Future Study (MTF) in the United States to examine social disparities in adult trajectories of body mass index (BMI) from 1986 to 2004. The MTF study is ideally suited to examine the recent upsurge in obesity and overweight among Americans since the early 1980s. With 19 years of data on annual cohorts followed prospectively over adulthood (age 18–45), it is possible to simultaneously study individual weight change as well as period shifts in BMI over time. We hypothesize that BMI scores are positively associated with both age and historical time, such that BMI trajectories will become increasingly steep in more-recent time periods. We also hypothesize that BMI is inversely related to social and economic advantage, such that, within each cohort, adult BMI trajectories will be steeper for socially disadvantaged subgroups (i.e. women, racial/ethnic minorities and individuals of lower SEP).
Data are drawn from the MTF project, a nation-wide school-based survey conducted annually in the United States since 1976.35 The study design (described in more detail elsewhere)36 involves a nationally representative sample of high-school seniors (age 18) surveyed in the spring of each year (~15 000 per year). For the panel component, a 16% representative subsample (~2400 students per year) was randomly selected from each cohort for six biennial follow-ups (to the age of 30), and then at 5-year intervals to the age of 45, using self-completed mailed questionnaires. All procedures are reviewed and approved on an annual basis by the University of Michigan's Institutional Review Board (IRB) for compliance with federal guidelines for the treatment of human subjects.
To ease respondent burden in the MTF panel study, multiple forms are used, with some questions appearing on only some of the forms. Since 1986 a random one-sixth of respondents received forms asking about height and weight. Thus, the final sample available for this analysis consists of 10 956 individuals who were age 18–45 between 1986 and 2004. (Sample sizes for each high-school cohort average 345 persons with up to nine observations per person.)
Since subjects are selected in their senior year, the survey does not include high-school dropouts.35 Drug users were oversampled for participation in the panel survey by a ratio of 3:1 (and then re-weighted for analyses to reflect population estimates, as discussed below). Student response rates in their senior year average 82.7%. Retention rates in the panel respondents are highest in the first follow-up after high school (averaging 70% of the original cohort), and fall to an average of 64% in the biennial follow-ups through to age 32. Longer-term retention rates average 50% at age 35, 57% at age 40 and 60% at age 45.35 Using logistic regression analysis (with backward elimination) we modelled the probability of participation at each wave according to a broad array of baseline socio-demographic characteristics that could potentially influence study retention. Five variables consistently predicted participation at each wave: female, White, coming from a two-parent family, enrolment in an academic high school and not living in the southern region of the US at baseline. By deriving the predicted probability of participation from these five variables, each observation at each wave of data collection was weighted to account for attrition (losses to follow-up as well as survey non-response) based on the inverse probability of participation. Using this weight in our models resulted in only trivial differences in the coefficients in our results, ruling out the possibility that losses due to attrition substantially bias our results.
We capture weight gain over adulthood with the BMI (BMI = kg/m2). Height was self-reported in inches (converted to metres for these analyses), and weight in pounds (converted to kilograms). Item non-response for height and weight was small in this panel survey, ranging from <1% (0.7%) to 2.3% over the follow-up period. A BMI of 25 or higher is used to define ‘overweight’.37 BMI scores were excluded for pregnant women.
We focus on three key social indicators: (i) gender; (ii) race/ethnicity; and (iii) lifetime SEP. Female is a dummy variable coded 1 for females and 0 for males. Race/ethnicity is represented by four dummy variables: White, Black, Hispanic and other racial/ethnic group (includes Asians and Native Americans). Consistent with a life-course perspective we include SEP both in childhood and in adulthood. Childhood SEP is the maximum of the respondent's parents’ education (less than high school, high-school diploma, college degree or higher). Adult SEP is a time-varying measure of completed education at each follow-up (high-school diploma, college degree or higher). [More sophisticated coding (differentiating associate's degree, bachelor's degree or graduate degree) did not yield differences in intercepts or slopes in the growth model]. In order to capture increases in socio-economic resources afforded by marriage, we take the maximum of the respondent's and spouse's education for those who were married at each follow-up. Although SEP can also be measured by income and occupation, social disparities in obesity have typically been observed by educational levels and less consistently by income.4 As a result, we focus on educational attainment over the life course as our measure of SEP.
Growth curve models38 were used to examine trajectories of BMI over adulthood. Growth curve models belong to a general class of mixed models that take into consideration the clustering of observations within persons and also have the capacity to handle unbalanced designs (inconsistent number of observations per person).39 A two-level linear model was analysed, with multiple observations nested within persons over time (adulthood). Age was used as the indicator of time. In order to facilitate parameter interpretation, we centred age at the initial point of data collection (age 18).
The structure of this model can be expressed by equations at two levels. At level 1 (within-person model) BMI scores at time t are nested within individuals (i):
where π0i is the expected BMI score for person i at age 18 (since age is centred), π1i captures the rate of change in BMI over the life course, and rti is the within-person residual (that part of an individual's BMI at time t not predicted by age).
To incorporate period effects we also include the year of data collection as a time-varying variable, again centred on the initial year that BMI data were collected (1986). This allows us to examine period shifts in BMI across all ages. (Cohort effects integrate age and period effects by capturing variations in BMI across groups of individuals in the same high-school year as they age. We do not focus on cohort effects in this paper. Rather, we examine period differences in age patterns of BMI over adulthood.) We examine how adult trajectory slopes vary across historical time by modelling interactions between the age and year variables:
The level-1 parameters are then modelled as a function of individual characteristics (at level two). The level two (between person) submodel assumes that BMI trajectories vary across individuals, and we explicitly model these differences as follows:
Here, for example, the intercept and age slope from equation (1) are modelled as a function of gender, where β01 represents the difference in the initial BMI score for females compared with males, and β 11 captures the difference in the rate of change in BMI over adulthood for women compared with men. (We also modelled person-level differences in the slope for year, but found no effects.) The residual errors (e0i, e1i, e2i) capture random variance in the intercept and slope values across persons, which we attempt to capture through the incorporation of additional person-level variables. (Due to model limitations on the number of estimated variance components, we set the random variance for the age by year term to zero.) Substituting equations (2.1) through (2.4) into equation (1) gives us the full composite model:
We used the MIXED procedure in SAS to estimate all models using full maximum likelihood, assuming normally distributed residuals. (The distribution of the residuals showed a good approximation to normality, with little deviation from the diagonal in the normal probability plots.) All analyses were weighted to correct for the oversampling of drug users in the panel data. Nested models were compared according to three goodness-of-fit indices: (i) change in the −2 log likelihood (or deviance statistic), which follows a χ2 distribution with degrees of freedom equal to the difference in the number of parameters tested between models; (ii) change in the Bayesian Information Criterion (BIC), which makes an adjustment for model parsimony;40 and (iii) proportion of variance in BMI that is explained by a model (pseudo R2), calculated by squaring the correlation between the observed and predicted BMI values.
The individuals in this study were born between 1958 and 1985, with the majority (53%) born before 1971. Socio-demographic characteristics for the study sample at baseline (age 18) are presented in Table 1. Approximately one-quarter (26.2%) went on to receive a college degree over the follow-up period.
Table 2 presents the weighted mean BMI scores for the study sample over adulthood. Three specific high-school cohorts are illustrated. Mean BMI scores increase over adulthood, but at any given age BMI scores are higher in more recent cohorts.
The first column in Table 3 (Model A) presents the results for the unconditional growth model without any person-level covariates. The value for the intercept (22.4) represents the average BMI score at age 18. The coefficients for age express the average rate of change in BMI over adulthood. A second-order polynomial indicates that BMI scores increase steadily over young adulthood and then begin to flatten out towards midlife, as illustrated in Figure 1. After incorporating the effects of age, variation in BMI remains both within persons and between persons (variance components), and additional person-level variables may explain this variation. Note that ageing itself explains about 6.5% of the total variance in BMI scores.
The second column of Table 3 (Model B) incorporates the period effect. The positive coefficient for year indicates that BMI scores have increased linearly since 1986, at an average rate of 2.3 BMI units every decade since 1986. BMI trajectories were then allowed to vary by historical time by including interactions between age and the year of data collection (Model C in Table 3). Compared with the model including only age, the inclusion of secular time results in an improvement in model fit and accounts for about 34% of the residual variation between individual trajectories [(0.165–0.109)/0.165]. Together, ageing and historical time explain about 8% of the total variation in BMI scores.
A visual depiction of this age and period effect is presented in Figure 2. For simplicity, only three specific high-school cohorts are illustrated. Members of the earliest cohort (born in 1958) were high-school seniors in 1976. Although height and weight data are not available for this cohort until the age of 28 (in 1986), the BMI trajectory for this group exhibits a gradual increase over adulthood. However, for each successive cohort, BMI scores are higher and the early adult BMI trajectory becomes increasingly steep. As a result, successive cohorts become ‘overweight’ (BMI > 25) at increasingly earlier points in the life course.
The second and third columns of Table 4 add the effects of gender and race/ethnicity to the age and period effect model. Females and racial/ethnic minority groups are at greater risk of weight gain over the adult life course than white males (Model C). (There were no statistical interactions between any of the socio-demographic characteristics and the period effect). However, since 1986 there has been a convergence in BMI scores by racial/ethnic groups across cohorts. Figure 3 illustrates racial differences in BMI trajectories across the same three high-school cohorts depicted in Figure 2, but for simplicity we plot trajectories for women only since there are no differences in the BMI trajectories between Whites and other race/ethnic groups and between Blacks and Hispanics (Table 4, Model C), predicted BMI trajectories are only plotted for Whites and Blacks for illustrative purposes.
As illustrated in Figure 3, BMI scores are consistently higher across adulthood for racial/ethnic minority groups than for whites. However, over time the BMI trajectories for whites tend to track those for Blacks from earlier cohorts. For example, BMI scores for white women in the class of 1986 tend to approach those of black women in the 1976 cohort over the equivalent age range (age 28–36, Figure 3). Similarly, early adult BMI scores for white women in the 1996 cohort track closely those for Black women who graduated from high school 10 years earlier.
The fourth column of Table 4 (Model D) adds childhood SEP to the model. Consistent with a life-course perspective on health disparities, higher SEP in childhood is protective of weight gain both at the age of 18 and over adulthood. Individuals who have a background of lower SEP in childhood (parents with less than college degree) have a higher BMI at age 18 (initial status) and also have higher rates of annual growth in BMI over adulthood than those whose parents had a college degree. Childhood SEP accounts for only a small part of the effect of racial/ethnic disadvantage [largely for Hispanics, see Table 4 (Model D)], indicating that there are other factors associated with racial/ethnic position that carry an increased risk for higher BMI that are not simply a function of SEP at birth.
However, because of the period effect, even white males from a high SEP have been gaining weight more rapidly over adulthood in more-recent cohorts. Figure 4 illustrates BMI trajectories for white males in the three separate high-school cohorts according to childhood SEP. While BMI scores are consistently higher in each cohort amongst those with lower childhood SEP (parents with less than a high-school education), at each age BMI scores for socially advantaged groups in later cohorts are actually higher than those for their socially disadvantaged counterparts who were born 10 years earlier.
The protective effects of adult SEP over the life course are weaker than for childhood SEP (Table 4, Model E). High-school seniors who go on to receive a college degree have somewhat flatter BMI trajectories over adulthood than their classmates who did not pursue further education. However, achieved SEP in adulthood does not fully mediate the lingering effects of childhood SEP. Model E simultaneously adjusts for adult SEP and examines the residual effect of childhood SEP on BMI after accounting for the mediating effects of adult SEP. Thus, social disadvantage in childhood carries a persisting risk for increased BMI throughout the life course. (There were no statistical interaction effects between gender and adult SEP, race and adult SEP or between childhood SEP and adult SEP.) Together, these socio-demographic characteristics, along with ageing and historical time, explain over 12% of the total variation in BMI scores over adulthood (final model Table 4).
We use 19 years of nationally representative panel data for repeated cohorts of Americans to describe adult trajectories of BMI since 1986. We found that, not only do BMI scores follow a curvilinear course over adulthood, but there has also been a linear increase in adult BMI trajectories over historical time. Thus, historical conditions have differentially affected adult trajectories of weight gain since the mid-1980s. As a result, one's BMI later in adulthood may have less to do with one's age than with historical changes in social, environmental and cultural conditions that significantly influence caloric balance.41–44
We also found evidence of a ‘social patterning’45 in adult trajectories of BMI. Weight gain over adulthood is more rapid for disadvantaged social groups, including women, racial/ethnic minorities and those of lower SEP. As a result, social disparities in BMI become increasingly more pronounced over the adult life course, which has ominous implications for future health disparities among these cohorts as they age. Targeted interventions towards these groups,46–48 particularly early in the life course,49,50 remain an important and worthwhile strategy.
However, BMI scores have also been rising steadily in socially advantaged groups since 1986. Even white males in a high SEP are at an increased risk for weight gain at any given age over time. This has implications for preventive strategies. If the rise in obesity, as a public health concern, is approached from a ‘fundamental cause’ perspective,24 weight gain is assumed to be the result of restricted access to health-enhancing knowledge and resources among disadvantaged social groups. But our results indicate that another strategy may need to be considered for the prevention of obesity in the more advantaged subgroups of society. These people presumably have access to resources but are failing to use them effectively. Similar findings are beginning to appear in the literature, where social advantage does not confer the expected protective effects against weight gain in American adults.51–53 However, if these advantaged social groups begin adopting better nutrition and exercise behaviours afforded to them by their flexible social resources,25 we may observe a decline in their BMI trajectory over time, and this will only serve to magnify the socio-economic gradient in obesity and overweight in future years.
Due to the underestimation of weight in self-reported data,54 BMI scores in this sample may be higher if weight were measured directly. However, because there is no evidence of differential underreporting of weight in ethnic or racial minority groups, or by gender or SEP, the social patterns in BMI are not likely to differ markedly from those reported here. Because study subjects were selected in their senior year of high school, high-school dropouts are not represented. Although panel attrition in the MTF data is nontrivial, statistically adjusting for attrition yielded no differences in the results. Because follow-up is ongoing in this study, relatively fewer data points are available for estimation at the tail end of the growth curves (age 35–45) in these analyses.
Our purpose in this paper was to describe social disparities in BMI over adulthood as they change over historical time. The results highlight the importance of social status and socio-economic resources for maintaining optimal weight. Yet, even those in advantaged social positions, presumably with access to health-enhancing resources, have experienced an increase in BMI in more recent cohorts. Thus, promoting access to resources and opportunities for healthy eating and exercise may not be sufficient for overturning the obesity epidemic.
Robert Wood Johnson Foundation, as part of the Youth, Education, and Society Project; National Institute on Drug Abuse, a part of the National Institutes of Health; University of Michigan Center for Social Inequality Mind and Body Life Course core [National Institutes of Health/National Institutes of Child Health and Human Development Grant (R24HD04786)].
An earlier version of this paper was presented at the Second North American Congress of Epidemiology in Seattle, WA, June 21–24, 2006. We thank Jersey Liang and Jim House for helpful comments on an earlier draft.
Conflict of interest: None declared.