The authors have declared that no competing interests exist.
Conceived and designed the experiments: JPB SVS NAC AJO. Analyzed the data: JPB SVS AJO. Wrote the paper: JPB SVS NAC AJO.
We examined body mass index (BMI) across place and time to determine the pattern of BMI mean and standard deviation trajectories.
We included participants in the Framingham Heart Study (FHS) Offspring Cohort over eight waves of follow-up, from 1971 to 2008. After exclusions, the final sample size was 4569 subjects with 28,625 observations. We used multi-level models to examine population means and variation at the individual and neighborhood (census tracts) levels across time with measured BMI as the outcome, controlling for individual demographics and behaviors and neighborhood poverty. Because neighborhoods accounted for limited BMI variance, we removed this level as a source of variation in final models. We examined sex-stratified models with all subjects and models stratified by sex and baseline weight classification.
Mean BMI increased from 24.0 kg/m2 at Wave 1 to 27.7 at Wave 8 for women and from 26.6 kg/m2 to 29.0 for men. In final models, BMI variation also increased from Waves 1 to 8, with the standard deviation increasing from 4.18 kg/m2 to 6.15 for women and 3.31 kg/m2 to 4.73 for men. BMI means increased in parallel across most baseline BMI weight classifications, except for more rapid increases through middle-age for obese women followed by declines in the last wave. BMI standard deviations also increased in parallel across baseline BMI classifications for women, with greater divergence of BMI variance for obese men compared to other weight classifications.
Over nearly 40 years, BMI mean and variation increased in parallel across most baseline weight classifications in our sample. Individual-level characteristics, especially baseline BMI, were the primary factors in rising BMI. These findings have important implications not only for understanding the sources of the obesity epidemic in the United States but also for the targeting of interventions to address the epidemic.
The obesity epidemic has progressed rapidly in the United States over the last several decades. The mean body mass index (BMI) of US adults has increased from 25.7 kg/m2 to 28.7 for men and 25.1 kg/m2 to 28.7 for women from the 1960s to 2000s
Properly accounting for heterogeneity at both the individual and neighborhood levels using longitudinal data may determine true underlying patterns of population weight change over time with possible implications for interventions
Here, using data from the Framingham Heart Study (FHS) Offspring Cohort over 37 years, including a large number of individuals who moved great distances, we examined longitudinal trends in BMI between individuals and neighborhoods. The use of this cohort, linked together by common characteristics of their parents (or in-laws), enabled us to more confidently examine complex associations between BMI and social and geographic factors prone to endogeneity.
The Institutional Review Board of Harvard Medical School approved this study. The Framingham Heart Study undertook a detailed written consent process for all aspects of data collection
Our sample came from the Framingham Heart Study (FHS) Offspring Cohort, which started in 1971 and enrolled 5124 subjects who were either the children of subjects enrolled in the FHS Original Cohort or their spouses. The FHS Original Cohort included a random sample of residents of Framingham, Massachusetts, in the 1940s. Offspring Cohort subjects have been examined and surveyed up to eight times from enrollment through 2008, roughly every four years. Our final sample included all FHS Offspring Cohort subjects excluding observations with missing BMI, smoking status, alcohol intake, or census tract of residence; we also excluded subjects at any time points if they were living in a nursing home or were less than 21 years old.
For analyses, we intended to use three-level multi-level random effects models to account for BMI clustering by neighborhood and individual with an additional pure error variance term. However, contrary to our a priori hypothesis that we would find notable variance at the neighborhood level, we found that the neighborhood level contributed near zero variance in cross-sectional models for most of the eight waves (Table S1 in
The final sample size for this study included 4569 subjects with 28,625 observations over a nearly 40 year period.
Time-varying individual-level BMI was the outcome variable, objectively calculated using in person measured weight and standing height at each wave
Our models were two-level multilevel models accounting for between-individual and within-individual (or pure “error”) variance. When determining how best to account for individual-level variance, we explored several modeling strategies. In our data, the longitudinal trajectory in BMI appeared nonlinear with some evidence of nonlinear changes in the variance over time (Figure S1 in
We first generated descriptive results using SAS statistical software, Version 9.1 (Cary, North Carolina). Using MLWin Version 2.24 (Bristol, United Kingdom)
To determine how much of the unexplained individual-level variation in BMI was accounted for by baseline BMI, we subsequently fit separate models for Waves 2 through 8 with each model fit two ways: with Wave 1 BMI as a predictor and without. These models included all of the same covariates previously specified but had smaller sample sizes because we included outcomes only for Waves 2 through 8 (4569 subjects, 24,467 observations). We also included wave by age interactions to help differentiate temporal trends in BMI from aging trends.
Further, to determine whether BMI mean and standard deviation trajectories differed by baseline weight classification, we fit four models for each gender corresponding to the four categories of baseline BMI: underweight (BMI <18.5 kg/m2), normal weight (BMI 18.5 to 24.9 kg/m2), overweight (BMI 25 to 29.9 kg/m2), and obese (BMI ≥30 kg/m2). Our a priori hypothesis was that mean BMI increased more rapidly for overweight and obese participants
For all models, we used Markov Chain Monte Carlo (MCMC) analyses to generate multiple iterative samples from the joint posterior distribution of the parameters, from which parameter estimates could be constructed
The mean number of observations per subject was 6.3 with a range of 2 to 8 observations (by construction, the lower limit was 2 not 1). The mean BMI increased from 24.0 kg/m2 at Wave 1 to 27.7 at Wave 8 for women and from 26.6 kg/m2 to 29.0 for men (
Mean Across Waves | Wave 1 Mean,Wave 8 Mean | Mean Across Waves | Wave 1 Mean,Wave 8 Mean | |||
FemaleN = 2366 |
MaleN = 2203 |
|||||
26.1 | 24.0, 27.7 | 27.7 | 26.6, 29.0 | |||
52.4 | 37.3, 66.9 | 52.6 | 38.4, 67.0 | |||
44.1 | 45.6, 35.3 | 37.4 | 39.0, 26.5 | |||
51.8 | 46.9, 65.7 | 58.4 | 53.4, 73.4 | |||
4.1 | 7.5, 0 | 4.3 | 7.6, 0.1 | |||
75.4 | 86.2, 65.0 | 85.4 | 88.6, 84.0 | |||
60.0 | 53.3, 39.4 | 77.3 | 95.6, 46.1 | |||
24.8 | 43.0, 8.2 | 24.8 | 45.0, 7.5 | |||
35.8 | 16.6, 52.0 | 23.8 | 8.8, 38.0 | |||
59.0 | 77.4, 44.9 | 55.2 | 63.9, 48.4 | |||
5.2 | 6.0, 3.0 | 21.0 | 27.3, 13.6 | |||
5.4 | 6.0, 5.6 | 5.3 | 6.1, 5.4 |
The number of female subjects was 2148 in Wave 1 and 1518 in Wave 8. The number of male subjects was 2010 in Wave 1 and 1261 in Wave 8. The total number of subjects is greater than subjects in Wave 1 because some observations did not meet inclusion criteria (e.g., a subject had missing BMI in Wave 1 but available BMI in subsequent waves).
In addition to the increase in mean BMI, the
The pattern of BMI distribution also changed over time, with less skewness for both women (0.04 to 0.01) and no change for men (0.02 to 0.02), indicating a more normal distribution of BMI by Wave 8 for women. Consistent with the foregoing, kurtosis, a measure of the presence of outliers, declined quite substantially over time for women (5.62 to 1.52) with a slight increase for men (1.12 to 1.76). Overall, the distribution of BMI over time maintained a similar shape for men (slightly skewed and with thicker tails than the normal distribution) but became substantially more normal for women.
The final models included all individual-level covariates as well as neighborhood poverty (
FemaleN = 2366Obs = 15,016 | MaleN = 2203Obs = 13,609 | ||||
Variable | β | 95% Credible Interval | β | 95% Credible Interval | |
24.4 | 24.0, 24.9 |
26.4 | 26.0, 26.8 |
||
0.26 | 0.17, 0.34 |
0.31 | 0.24, 0.38 |
||
0.61 | 0.38, 0.84 |
0.02 | −0.16, 0.21 | ||
0.04 | 0.03–0.06 |
0.02 | 0.003, 0.03 |
||
Ref | Ref | ||||
−0.61 | −0.85, −0.36 |
−0.48 | −0.69, −0.27 |
||
0.37 | −0.06, 0.79 | 0.25 | −0.10, 0.60 | ||
0.47 | 0.34, 0.60 |
0.26 | 0.14, 0.37 |
||
0.12 | 0.03, 0.20 | 0.19 | 0.10, 0.28 |
||
−0.90 | −1.0, −0.77 |
−0.73 | −0.84, −0.62 |
||
Ref | Ref | ||||
0.19 | 0.10, 0.27 |
0.22 | 0.13, 0.31 |
||
0.29 | 0.10, 0.48 |
0.32 | 0.20, 0.44 |
||
0.002 | −0.01, 0.02 | −0.006 | −0.02, 0.004 | ||
4.61 | 4.46, 4.75 |
3.41 | 3.30, 3.52 |
||
1.18 | 1.12, 1.24 |
0.87 | 0.81, 0.93 |
||
3.42 | 3.22, 3.62 |
2.42 | 2.24, 2.61 |
||
1.49 | 1.46, 1.51 |
1.25 | 1.23, 1.27 |
||
59,538 | 49,167 |
95% credible interval does not cross 0.
Census tract information was unavailable for some tracts. Almost all of this missing data was from 1970 when some land areas were not yet assigned a census tract. For this analysis, we had census tract poverty data for 14,355 of the 15,016 observations among women and 12,989 of the 13,609 included observations among men. To ensure comparability across models, we included a dummy variable accounting for the availability of census tract poverty data along with a modified poverty variable (missing poverty data set to 0 rather than missing) in the final model. This did not change results for census tract poverty but did allow us to include all observations in the analyses that included this variable.
As we found for the unadjusted BMI variance, the individual-level random slopes for time and the natural log of time in fully-adjusted models revealed increasing heterogeneity in BMI across time for women and men (
In the fully adjusted models, the total unexplained variation in BMI attributed to individuals across time (individual-level standard deviation) steadily increased from 1971 to 2008 for both women and men. The error standard deviation represents the idiosyncratic pure error variance. We accounted for non-linear increases in between-individual BMI standard deviation by including a random intercept at the individual level and random slopes for time and the natural log of time.
To determine how much of the between-individual variation in BMI was accounted for by baseline BMI at Wave 1, we ran two sets of models restricted to observations from Waves 2 to 8, with baseline BMI and without, including all of the same covariates as for the prior models. In models without baseline BMI, the standard deviation in BMI at Wave 2 was 4.63 kg/m2 for women and 3.46 for men (variance 21.5 kg/m2 and 12.0). The addition of baseline BMI decreased standard deviations at Wave 2 to 1.95 kg/m2 for women and 1.42 for men (variance 3.80 kg/m2 and 2.01). The baseline BMI, thus, accounted for 82% and 83% of the between-individual variance in BMI, respectively, for women and men (data not shown in tables).
To assess the impact of baseline weight on the trajectories of BMI mean and variance, we then fit four models for each sex, stratified by baseline BMI classification – underweight, normal weight, overweight, obese - including BMI in Waves 2 to 8 as the outcome. These models also included wave by age interactions and Wave 1 BMI as predictors (Table S4 in
In these models, mean BMI among men had parallel increases across baseline weight classifications with a plateau in BMI evident by Wave 8 (
Using results from the fully-adjusted models, we plotted the BMI trajectory for women (A) and men (B) based on their weight classification at baseline (during Wave 1, 1971–1975), controlling for covariates including baseline BMI. Weight classifications were underweight (BMI <18.5 kg/m2), normal weight (18.5 to 24.9), overweight (25 to 29.9), and obese (≥30). Lines represent trajectories for the typical male or female (mean age at each wave, married, employed,>high school education, non-smoker, consuming 1–2 alcoholic drinks daily, living in a census tract at mean poverty level, with mean baseline BMI for that weight classification).
In the fully adjusted models, the individual-level standard deviation of BMI steadily increased from 1971 to 2008 for both women (A) and men (B) in all baseline weight classifications. Standard deviation increases were similar across most weight classifications with larger standard deviations for both obese women and men, and larger increases across time for obese men. We accounted for non-linear increases in between-individual BMI standard deviation by including a random intercept at the individual level and random slopes for time and the natural log of time.
Using data from the Framingham Heart Study Offspring Cohort over a nearly 40 year period, we show that factors intrinsic to individuals accounted for the overwhelming proportion of the variation in BMI over time. We also found increasing population means and variation for BMI over time. For both men and women, baseline BMI accounted for most of the unexplained individual-level variation in BMI, demonstrating that BMI reached by the late 30 s (mean age at Wave 1 was 38 years for men, 37 for women), determined BMI until their late 60 s (mean age at Wave 8 was 67 years for both men and women). The rapidity of weight gain was similar across all baseline weight classifications except for women who were obese at baseline. Obese women gained weight somewhat more rapidly than women with lower baseline BMIs until they were in their early 50 s with an abatement of this trend thereafter. BMI variation increased over time for participants in all baseline weight categories. Variation was greatest for obese female and male subjects, demonstrating a more heterogeneous population across time.
The parallel increases in weight gain across baseline weight classifications calls for a relatively uniform population-targeted strategy to decrease risk for weight gain. Further, because weight trajectories appear to be set by the late 30 s, strategies focused on children and young adults might be most effective
These results, showing increasing variation in BMI but a more uniform distribution over time, contrasts somewhat with recent data from Flegal, et al.
Finally, our analyses shed light on the possible role of neighborhood of residence in the growth of obesity over the past four decades. In contrast to prior longitudinal studies, in our study, neighborhood of residence accounted for a very small proportion of BMI variance, and neighborhood poverty was unrelated to BMI
Our study has limitations. First, we could not measure characteristics of neighborhoods where subjects work, a possible source of unmeasured confounding between BMI and neighborhood characteristics. Second, we could more effectively determine the age at which BMI trajectories are established if we had measurements prior to the 1970s. Third, our sample lacks racial diversity, an unavoidable limitation of research with the FHS Offspring Cohort. However, this limitation in generalizability also could strengthen the plausibility of our findings. All subjects had some similar characteristics because they are the offspring (or an offspring’s spouse) of the FHS Original Cohort, a random sampling of Framingham, Massachusetts, in the 1940s. One could argue that with fewer differences between individuals on observables, such as race, that it is reasonable to assume there are also fewer differences on unobservables and thus less impact from unmeasured confounding. Further, subjects were socioeconomically quite diverse. For example, in Wave 8, the mean census tract poverty for male subjects was 5.4% (SD 4.3%, Range 0.3% –31.0%). Fourth, we had a large number of census tracts in our sample, frequently with a small number of observations per tract. Our sample included participants from 2095 different census tracts over time, with a mean of 13.7 observations per tract (SD 79.8, range 1 to 1638). Multilevel models, by design, shrink the variance estimates toward the null for higher level units (tracts) with few observations and, therefore, may underestimate the ICC at the tract level in the cross-sectional models that we ran. Yet, shrunken residuals have the benefit of helping to avoid over-interpretation of random variation in the data as true neighborhood-level variation.
In sum, over nearly 40 years, BMI mean and variation increased in parallel across most baseline weight classifications in our sample. Individual-level characteristics, especially baseline BMI, were the primary factors in rising BMI. These findings have important implications not only for understanding the sources of the obesity epidemic in the United States but also for the targeting of interventions to address the epidemic.
Table S1, Cross-sectional Variance at the Neighborhood and Individual Levels, by Wave. Table S2, Unadjusted Skewness, Kurtosis and Coefficients of Variation by Wave. Table S3, Deviance Information Criteria (DIC) for Models. Table S4, Parameter Estimates from Models for Participants Who Were Underweight (BMI <18.5 kg/m2) at Baseline (1971 to 75) Followed from 1979 to 2008 to Examine BMI Trajectories, Framingham Heart Study Offspring Cohort. Table S5, Parameter Estimates from Models for Participants Who Were Normal Weight (BMI 18.5 to 24.9 kg/m2) at Baseline (1971 to 75) Followed from 1979 to 2008 to Examine BMI Trajectories, Framingham Heart Study Offspring Cohort. Table S6, Parameter Estimates from Models for Participants Who Were Overweight (BMI 25.0 to 29.9 kg/m2) at Baseline (1971 to 75) Followed from 1979 to 2008 to Examine BMI Trajectories, Framingham Heart Study Offspring Cohort. Table S7, Parameter Estimates from Models for Participants Who Were Obese (BMI ≥30 kg/m2) at Baseline (1971 to 75) Followed from 1979 to 2008 to Examine BMI Trajectories, Framingham Heart Study Offspring Cohort. Figure S1, Mean Body Mass Index for Women and Men, Framingham Heart Study Offspring Study, 1971 to 2008. Mean unadjusted BMI increased for both women and men over the course of follow-up with a more steep trajectory for women than men. Figure S2, Model for Primary Analyses Examining Body Mass Index, Framingham Heart Study Offspring Cohort Study, 1971 to 2008. We generated this screen shot from MLWin to display the model we ran for our primary analyses, demonstrating both the fixed and random effects included. This example is for our full sample of women, but the models were equivalent for men. In these models, we include a fixed and random effect for linear time (linear time from 1 to 8, based on wave of observation (time)), the natural log of time (lntime), age (a linear variable centered on its mean (age-gm)), marital status (binary: unmarried as reference, married (married_1)), education (categorical: ≤high school as reference, >high school (educat_1), missing education (educat_2)), employment status (binary: unemployed as reference, employed (employed_1)), smoking status (binary: non-smoker as reference, smoker (smokes_1)), alcohol consumption (categorical: 0 as reference, 1–2 daily (alcgrp_1), >2 daily (alcgrp_2)), census tract poverty (linear variable centered on its mean (newpov-gm)), and whether census tract poverty was available (binary: not available as reference, available (povavail_1)). We included this last variable to allow us to have equal subjects in models with census tract poverty in the model and those without it. Models that we stratified by baseline BMI classification were similar, but the outcome in these models was BMI from Waves 2 through 8 (rather than 1 through 8) and included baseline BMI and age by time interactions as additional covariates. Figure S3, Histogram of Unadjusted BMI Distribution for Subjects in Wave 1 (1971 to 1975, Diagonal Stripes) and Wave 8 (2005 to 2008, Open Bars). These histograms represent the distribution of BMI values for Wave 1 versus Wave 8, demonstrating an increase in BMI mean and variance for both women (A) and men (B).
(DOC)
We thank Rebecca Joyce and Laurie Meneades for the expert assistance required to build the data set.