Estimates of Childhood Overweight and Obesity at the Region, State, and County Levels: A Multilevel Small-Area Estimation Approach

Abstract

Abbreviations:

 
• BMI

  body mass index

 
• CI

  confidence interval

 
• CV

  cross-validated

 
• NSCH

  National Survey of Children’s Health

 
• RMSPE

  root mean squared predictive error

 
• SAE

  small-area estimation

 
• UIC

  Urban Influence Codes

Childhood overweight and obesity is a major public health problem with physical and mental health impacts, including reduced life expectancy (1–3), an association with adulthood obesity, and the onset of additional comorbidities and increased mortality (1, 4, 5). The United States has one of the highest rates of children experiencing overweight/obesity worldwide (6). Approximately 31.5% of US youth experience overweight (7), while 18.5% experience obesity, ranging from 14% among children aged 2–5 years to 21% among those aged 12–19 years (8). The most recent data available (9) indicates that obesity remains higher than the Healthy People 2020 target of 14.5% for children and adolescents aged 2–19 years (10). For the remainder of this article, we use “overweight” to indicate “at least overweight,” which includes childhood overweight and obesity.

Disparities in childhood overweight exist across geographic regions and demographic subgroups (11, 12): female persons, members of historically marginalized groups, and persons with less educational experience higher overweight rates compared with their counterparts (13–20). Multiple studies have also identified the significant impact of children’s environments on their weight status (21–24). Specifically, environmental factors such as climate, levels of green space, and safety have been associated with experiencing childhood overweight (18, 25). As a result, policy and community-level interventions to prevent childhood overweight have focused on improving environmental factors (26–28). Accessing estimates of childhood overweight across the United States is particularly important for surveillance and benchmarking efforts, policy and program development, and resource allocation toward communities with greatest needs (27, 29, 30). To understand spatial patterns of childhood overweight and to uncover the most influential determinants, current and accurate estimates of childhood overweight at the local level are necessary. However, the most recently available childhood overweight estimates are based on data from 2007 (9).

Our study aimed to bridge this gap by estimating childhood overweight rates at the county, state, and Census regional-division levels using the 2016 National Survey of Children’s Health (NSCH) and the 2016 American Community Survey. Our objectives were to: 1) ensure that the current scope of overweight is well understood in a geographic context; 2) produce updated local-level predictive estimates that can help planning and evaluation of targeted community-level interventions against overweight; and 3) accurately characterize the uncertainty in overweight prediction estimates to identify where more data are needed. We implemented model-based small-area estimation (SAE) techniques, which are established methods to obtain prevalence predictions (31, 32) by extrapolating information from more data rich locations (9, 33–35). SAE has been successfully used to study obesity (15, 24, 36–43) and many other health behaviors and disease outcomes (43–47). Also, SAE models typically use population-based survey data linked to area-level covariates (e.g., county rurality, poverty, etc.) to model individual-level outcomes (9, 30, 34, 48) and poststratification to aggregate and weight individual-level estimates to an area level (46, 49). We demonstrate how multilevel modeling with poststratification can be used to accurately estimate uncertainty and identify areas with prevalence estimates different than the regional or national average while controlling the false discovery rate.

METHODS

Data sources

The 2016 NSCH, sponsored by the Health Resources and Services Administration, is a national survey examining the physical and emotional health of US children aged 0–17 years. This survey is conducted by the US Census Bureau in the 50 states and the District of Columbia via telephone (cellphone and landline), where information is gathered from parents/guardians of selected children. Within each state, an equal number of residences are randomly selected, and within each residence, a single child is randomly selected (50). For this study, only NSCH data for children aged 10–17 years were included because the body size of younger children changes more rapidly (51), and parent-reported height and weight measurements are more reliable for older children (52). We used the NSCH data for child-level variables.

For area-level variables (county, state, and division levels), we collected publicly available data from policy institutes and governmental organizations, including the Department of Health and Human Services, Department of Agriculture, Environmental Protection Agency, US Census Bureau, and American Community Survey among others (complete list in Web Table 1, available at https://doi.org/10.1093/aje/kwab176). Potential child- and area-level variables were chosen based upon previously established (9, 53) and hypothesized factors associated with childhood overweight and with input from obesity researchers at the University of South Carolina.

Child- and area-level model covariates

NSCH child-level variables tested included race/ethnicity (Hispanic, non-Hispanic White, Non-Hispanic Black, multiracial/other), sex (male, female), age, and educational attainment of the child’s parent with the most education (less than high school, completed high school, some college, college or higher).

Area-level variables considered included Census-defined divisions (contiguous groups of states broken into New England, Middle Atlantic, East North Central, West North Central, South Atlantic, East South Central, West South Central, Mountain, and Pacific), state bullying laws, state Medicaid expansion status, state-level school wellness policy, federal/state/county school funding, county incarceration rates, county child insurance rates, county food insecurity rates, county rate of children living in households with a single parent, number of primary care/pediatric providers per 100,000 residents in the county, county adult obesity rate, county population-weighted walkability index (54), county school access and proximity status, county rate of access to exercise locations, county food desert status, and county urban-rural designation based upon 2013 Urban Influence Codes (UIC) (Web Table 1 provides a complete list). UIC codes were aggregated together to classify counties as either metropolitan (UIC codes 1 and 2; n = 1,166), micropolitan (UIC codes 3, 5, and 8; n = 641), or rural (UIC codes 4, 6, 7, 9–12; n = 1,335).

Primary outcome: childhood overweight/obesity

Children were classified as overweight/obese based on their body mass index (BMI, calculated as weight (kg)/height (m)²) values. Using the 2000 Centers for Disease Control and Prevention sex- and age-specific growth charts (55), children with a BMI value of ≥95th percentile were classified as obese, and children with a BMI value between the 85th and 95th percentiles were considered overweight. Thus, a child with a BMI value of ≥85th percentile was considered to be overweight/obese. To calculate BMI for each child, NSCH relied on parental report of a child’s height (feet/inches or meters/centimeters) and weight (pounds/ounces or kilograms/grams).

Statistical model

Poststratification

Uncertainty and significance testing

To estimate the variability and perform hypothesis testing for the estimated overweight prevalence rates, we used a Monte Carlo–based parametric bootstrapping approach (62). Here, 10,000 bootstrap samples of estimated county-level overweight rates were computed using regression coefficients and random effects randomly drawn from estimated distributions. The 95% confidence intervals were given by the 2.5th and 97.5th percentiles of the bootstrap sample and incorporated uncertainty from both the regression coefficients and random effects. For the random effects, we used empirical Bayes predictions and standard errors assuming a normal distribution. For counties with no sampled individuals (n = 839), random effects were sampled conditional on data from neighboring counties.

We applied 2 different significance testing procedures to identify counties with outlying overweight rates. In the first procedure, P values from the county-level predicted empirical Bayes random effects were used. This approach identified counties with overweight rates significantly different from what was expected given their child- and area-level predictors (the fixed effects). In the second procedure, differences between the county and national rates were calculated for each bootstrap iteration. The proportion of times the difference was lower/higher than zero (indicating that the county rate was lower/higher than the national rate) multiplied by 2 was used as the P value. This approach identified counties that were significantly different from the national average. Both procedures used the Benjamini-Hochberg correction (63) to control the false discovery rate at a significance level of |$\alpha =0.05$|⁠.

Further statistical details and results are provided in Web Appendix 1. Our analyses/maps used SAS, version 9.4 (SAS Institute Inc., Cary, North Carolina) (64); R (R Foundation for Statistical Computing, Vienna, Austria) (65); and ArcGIS Pro 2.0 (ESRI, Redlands, California) (66). Study approval was given from the University of South Carolina Institutional Review Board (Protocol #00073928) and output approved through the US Census Bureau Disclosure Review Board.

RESULTS

Multilevel SAE model selection and validation

Table 1 shows the weighted and unweighted descriptive statistics of our sample. Out of 50,112 children in the 2016 NSCH data, 26,094 were aged 10–17 and of those, 24,162 (92.6%) had complete data; those excluded were due to missing BMI (6.5%) and missing other predictor data (<1%). In this sample, 53.9% were Non-Hispanic White (weighted proportion), 51.0% were male, and the average age was 13.5 years. Most children lived in the South Atlantic division (19.3%), Pacific division (16.2%), and East North Central division (14.7%). Our final model included the predictors that best explained the current scope of overweight based on CV RMSPE. These predictors were Census division, state wellness policy, 5 county-level variables (percentage of children in a single parent household in 2015, number of primary care providers per 100,000 residents, rurality, 2015 adult obesity rate, 2016 population-weighted walkability index), the child’s age, main effects, and all 2-way interactions for child race/ethnicity, sex, and parent’s education, along with a 3-way interaction of race/ethnicity by sex by parent’s education (see Web Tables 2 and 3). Our model-based state-level overweight estimates performed well compared with state-level direct estimates from the 2016 NSCH public use file. The average difference between the estimates was 1.55 (standard deviation, 1.25) percentage points (Pearson correlation |$\rho$| = 0.89; see Web Figure 1).

Predicted childhood overweight/obesity prevalence from multilevel SAE model

Table 2 shows the predicted national and regional division overweight rates, while Figure 1 shows a map of all predicted county-level overweight rates. The national overweight rate was 30.7% (95% confidence interval (CI): 27.0%, 34.9%, Table 2). The East South Central division had the highest rate of 36.6% (95% CI: 32.3%, 41.2%), while New England had the lowest with 27.8% (95% CI: 21.6%, 34.9%; Web Figure 2). More counties in the West South Central, South Atlantic, and East South Central divisions had higher estimated rates of childhood overweight (57.5% had rates ≥33%) compared with counties in the West North Central, Mountain and Pacific divisions (21.8% had rates ≥33%). After the false discovery rate correction, 31 counties had overweight rates significantly different from the national rate (23 lower, 8 higher). The Pacific division had the most counties that were significantly lower. We also found 14 counties with overweight rates significantly higher than expected given their child- and area-level fixed effects. Most of these counties were in California (4), Florida (4), and Texas (5). Web Table 4 gives a list of outlying counties, and Web Table 5 additional state-level results.

Figure 1

Estimated proportion of children aged 10–17 years experiencing overweight or obesity, county level, United States, 2016. Overweight/obesity estimates are indicated by color gradient: lightest color, 27% or lower; the second color, 28%–32%; the third color, 33%–37%; the darkest color, 38% or higher. The lighter the color of the county, the lower the overweight/obesity prevalence is. The darker the color of the county, the higher the overweight/obesity prevalence is. For example, counties in the Southeast United States have higher overweight/obesity prevalence than counties in New England. States included in the 9 Census divisions are New England: Connecticut, Maine, Massachusetts, New Hampshire, Rhode Island, and Vermont; Middle Atlantic: New Jersey, New York, and Pennsylvania; East North Central: Illinois, Indiana, Michigan, Ohio, and Wisconsin; West North Central: Iowa, Kansas, Minnesota, Missouri, Nebraska, North Dakota, and South Dakota; South Atlantic: Delaware, Florida, Georgia, Maryland, North Carolina, South Carolina, Virginia, West Virginia, and District of Columbia; East South Central: Alabama, Kentucky, Mississippi, and Tennessee; West South Central: Arkansas, Louisiana, Oklahoma, and Texas; Mountain: Arizona, Colorado, Idaho, Montana, Nevada, New Mexico, Utah, and Wyoming; and Pacific: Alaska, California, Hawaii, Oregon, and Washington. Full black lines on the map delineate state boundaries.

Open in new tabDownload slide

Our SAE model was also able to provide insights for targeted community-level interventions against overweight by giving predictive overweight estimates for groups of interest. Table 3 shows these predicted overweight rates by various demographic groups for each regional division. Nationally and by division, female children, non-Hispanic White children, children with a college-educated parent, and children aged 15 to 17 years had lower overweight rates. Children in metropolitan counties had lower overweight rates compared with children in micropolitan and rural counties. These results are displayed in Web Figures 3–7.

To characterize the uncertainty in our overweight predicted estimates, Figure 2 displays a bivariate map of overweight rate (low, medium, high) and the estimated 95% confidence intervals (low, medium, high). “Low” or narrow confidence intervals indicated a more certain estimate. The higher child overweight rates in the South Atlantic division also had narrower confidence intervals, indicating a higher level of confidence in these estimates compared with those of the East South Central division, for instance. In the East South Central division, child overweight rates were high, but so were the confidence intervals of the estimated rates, indicating less certainty. An unexpected finding from our study is the large amount of uncertainty for Los Angeles County, California (overweight rate = 42.7%, 95% CI: 11.9%, 78.8%), given that it is a large metropolitan county. As an alternative example, Dallas County, Texas (overweight rate = 44.4%, 95% CI: 31.3%, 57.9%), had a similar sample size and overweight rate as Los Angeles but a much narrower confidence interval.

Figure 2

Bivariate map of estimated childhood overweight/obesity and confidence interval width, county level, United States, 2016. The lightest color represents counties with a low overweight/obesity estimate and narrow confidence intervals. The darkest color represents counties with a high overweight/obesity estimate and wide confidence intervals. Counties in shades of blue have a medium to high overweight/obesity estimate with narrow- to medium-width confidence intervals. Counties in shades of red have a low to medium overweight/obesity estimate with medium to wide confidence intervals. States included in the 9 Census divisions are New England: Connecticut, Maine, Massachusetts, New Hampshire, Rhode Island, and Vermont; Middle Atlantic: New Jersey, New York, and Pennsylvania; East North Central: Illinois, Indiana, Michigan, Ohio, and Wisconsin; West North Central: Iowa, Kansas, Minnesota, Missouri, Nebraska, North Dakota, and South Dakota; South Atlantic: Delaware, Florida, Georgia, Maryland, North Carolina, South Carolina, Virginia, West Virginia, and District of Columbia; East South Central: Alabama, Kentucky, Mississippi, and Tennessee; West South Central: Arkansas, Louisiana, Oklahoma, and Texas; Mountain: Arizona, Colorado, Idaho, Montana, Nevada, New Mexico, Utah, and Wyoming; and Pacific: Alaska, California, Hawaii, Oregon, and Washington. Full black lines on the map delineate state boundaries.

Open in new tabDownload slide

DISCUSSION

Childhood overweight/obesity prevalence from SAE- and inference-based models

Our study used SAE techniques to illustrate the widespread issue of childhood overweight and its geographic variation in the United States. To do so, we produced updated local-level predictive overweight estimates and characterized their uncertainty per study objectives. Our national overweight rate of 30.7% is higher than the 16.8% obesity-only rate found by Zhang et al. using 2007 data (9). Our analysis predicted elevated rates of childhood overweight in the deep South, Four Corners, and Appalachian regions, in small pockets of the Northern Plains, and along the US-Mexico border. In prevalence studies using earlier NSCH data, patterns of elevated childhood obesity had been observed in the deep South, Appalachian region, San Joaquin Valley (California), and Four Corners region, as well as small pockets in Oklahoma, Northern Plains, and along the US-Mexico border (9, 67). Although those studies applied varied approaches to producing SAEs, there were some similarities between our maps displaying childhood overweight prevalence and those of previous studies examining only childhood obesity (9, 67). Trends in childhood overweight mirror trends in adult overweight (36).

Our study’s predicted overweight prevalence by sociodemographic groups indicated that male sex, minority race/ethnicity, younger age, lower parental educational achievement, and micropolitan/rural residence were associated with higher childhood overweight rates. While our study was not designed to look at the inferential effects of sociodemographic characteristics, other studies have observed similar trends (9, 14, 67–70). For example, other studies confirmed higher prevalence rates of childhood overweight among male persons (67, 70), younger youth (9, 67), and those of minority race/ethnicity (9, 14, 67, 69, 70). Additionally, studies using National Health and Nutrition Examination Survey data also demonstrated an inverse association between the prevalence of childhood obesity and the educational level of the head of household (14, 69). Current literature also suggests that rural children have a greater risk of obesity compared with their urban counterparts (71), as demonstrated by a 2015 meta-analysis examining urban-rural differences in childhood obesity in a pooled sample of 74,168 participants aged 2–19 years old (72).

The increased uncertainty found by our study for specific county overweight estimates, such as Los Angeles, was driven by the model’s fixed effects, in particular race/ethnicity and parental education (see Web Tables 6 and 7). The main effects for Hispanic and non-Hispanic Black respondents had over 50% more variability than the main effect for non-Hispanic White respondents, and the main effects for lower parental education attainment had over 40% more variance than for higher educational levels. The demographics of Dallas County had a higher concentration of sociodemographic groups with low uncertainty compared with Los Angeles County, thus resulting in the marked difference in uncertainty.

Implications for research and practice

Our study has important implications for future research. First, this research drew upon publicly available data and utilized an SAE and poststratification approach reproducible for other important health outcomes (details in Web Appendix 1). Second, future research may use our predictions for inferential work highlighting the relationships between food environments (i.e., presence of healthy and unhealthy food outlets), sociodemographic composition, and childhood overweight at the local level. Finally, additional validation work that compares model-based estimates with prevalence values obtained through Census-based approaches where nearly 100% of the population/area has data available (e.g., aggregation of school records) is an important next step.

This research also has implications for future public heath practice. Our small area estimates of childhood overweight and associated visualizations allow for more precise identification of populations or areas in need of resources and/or programs to combat childhood overweight. Our study showed that some areas of the United States have markedly higher childhood overweight rates and/or wide confidence intervals. This information can be used to understand which geographic areas or sociodemographic groups need additional programming, resources, or policy interventions (i.e., male adolescents, rural residents, etc.) and which geographic areas need better data collection (e.g., oversampling minority groups) to improve the certainty of estimates. Having publicly available local-level estimates of overweight is particularly important for organizations creating strategic plans, grants, toolkits, and other resources to highlight the areas/groups likely to benefit from these additional resources. Knowing which geographic areas or populations have less prediction uncertainty is also important, so that survey planners can modify their sampling plans accordingly (e.g., reducing target sample size for some groups/areas) and potentially reduce cost.

Strengths and limitations

This study had many strengths that emphasize its contribution to the field. Robust statistical techniques were used to provide estimates of childhood overweight at the county and state levels that were previously unavailable. For counties with little data, SAE allowed for reliable estimates to be produced by conditioning on data from neighboring counties. Additionally, our poststratification method accounted for each area’s underlying population when producing final estimates. Moreover, estimates were presented using choropleth and bivariate maps, offering rich visualization of the scope of the childhood overweight epidemic in the United States, as well as understanding geographic uncertainty in the data. Among 10–17-year-olds in the NSCH, the major reason for exclusion was missing BMI (6.5%). Those with and without BMI data did not differ in terms of our other covariate data (data not shown). Our study findings add to a growing body of knowledge on the prevalence and determinants of childhood overweight across the United States.

An important limitation was that parent-reported height and weight were used to calculate BMI for classification of overweight and were not independently verified. BMI derived from parent-reported height and weight was found in a meta-analysis to overestimate (prevalence ratio = 1.12) the prevalence of overweight compared with using BMI derived from direct height and weight measurements (73), yet parent-reported height and weight were more accurate for estimating overweight and obesity combined than for obesity alone (73). Additionally, there is evidence demonstrating that BMI derived from parent-reported height and weight increases in accuracy with the child’s age (73, 74).

Another limitation was that different regional divisions showed overweight rates with wide confidence intervals when the division included counties with a higher prevalence of underrepresented demographic groups (e.g., overweight rates in the Pacific division have wider confidence intervals than other divisions). This had a large impact for counties where the prevalence of rare population subgroups was relatively high but had little impact for most counties, where the prevalence of these subgroups was very low. This is indicative of a need to understand and survey demographic subgroups more effectively, to obtain more precise estimates for all demographic groups. Despite of the availability of survey weights and good sample sizes, these results point to the importance of improving the survey sampling and administration process.

Conclusion

Model-based SAE was performed to update prevalence estimates of overweight among children aged 10–17 across the United States. Overweight predictions differed by geographic area as well as child sociodemographic groups, and areas with low precision were observed (i.e., greater uncertainty in regional division/counties with a high proportion of underrepresented population subgroups). This study highlights the importance of broadening survey sampling considerations to include representation of diverse population subgroups and geographic regions.

ACKNOWLEDGMENTS

Author affiliations: Department of Epidemiology and Biostatistics, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina, United States (Anja Zgodic, Jan M. Eberth, Angela D. Liese, Alexander C. McLain); Rural and Minority Health Research Center, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina, United States (Jan M. Eberth, Charity B. Breneman); Department of Health Promotion, Education, and Behavior, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina, United States (Marilyn E. Wende, Andrew T. Kaczynski); and Prevention Research Center, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina, United States (Andrew T. Kaczynski).

This study was supported by the Federal Office of Rural Health Policy (FORHP), Health Resources and Services Administration (HRSA), US Department of Health and Human Services (HHS) (cooperative agreement U1CRH30539), and by the National Institutes of Health (NIH), National Institute of General Medical Sciences (NIGMS), through the Research Center for Child Well-Being (NIGMS, grant P20GM130420).

This research was performed at a Federal Statistical Research Data Center under FSRDC Project Number 2086. All results have been reviewed to ensure that no confidential information is disclosed.

The information, conclusions, and opinions expressed in this brief are those of the authors, and no endorsement by the US Census Bureau, FORHP, HRSA, HHS, NIH, or NIGMS is intended or should be inferred.

Conflict of interest: none declared.

REFERENCES