Adult Female Overweight and Obesity Prevalence in Seven Sub-Saharan African Countries: A Baseline Sub-National Assessment of Indicator 14 Of the Global NCD Monitoring Framework

Introduction Decreasing overweight and obesity prevalence requires precise data at sub-national levels to monitor progress and initiate interventions. This study aimed to estimate baseline agestandardized overweight prevalence at the lowest administrative units among women, 18 years and older, in seven African countries. The study aims are synonymous with indicator 14 of the global non-communicable disease monitoring framework. Methods We used the most recent Demographic and Health Survey and administrative boundaries data from the GADM. Three Bayesian hierarchical models were fitted and model selection tests implemented. The age-standardized prevalence of overweight among adult women at national, first and second administrative levels were individually reported in each country in the form of maps and tables. Results Substantial variation in the age-standardized prevalence of adult female overweight was noted across several second-level administrative units. In numerous locations in Tanzania, Nigeria and Zimbabwe, more than half of the adult female population were overweight and in one location in Tanzania, over 72% of the adult female population were overweight. These estimates were Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 5 October 2020 doi:10.20944/preprints202010.0067.v1 © 2020 by the author(s). Distributed under a Creative Commons CC BY license. roughly twice the national level overweight prevalence and, in some cases, roughly 10 – 20% greater than the overweight prevalence in first-level administrative units. Conclusion The observed overweight burden in subnational administrative units suggests the presence of an epidemic tantamount to the situation in more affluent economies. African countries lack the resources to effectively handle the fallout from such epidemic, therefore motivating the need for increased urgency in adopting WHO obesity-related intervention guidelines and implementing more rigorous studies to validate the study findings. Introduction The probability of premature death (i.e. between the ages of 30 years to 70 years) from major noncommunicable diseases (NCDs) cardiovascular disease, cancers, chronic respiratory diseases and diabetes – is projected to increase in Africa and decrease elsewhere. The global obesity target agreed upon by policymakers in 2013 was to halt the rise in its prevalence within their countries by 2025, relative to the baseline in 2010. Progress towards this target will be monitored by assessment of the age-standardized prevalence of overweight and obesity in individuals aged 18 years and older, which is indicator 14 of the Global NCD monitoring framework. Reaching this target is especially important because obesity is the second largest contributor to noncommunicable disease mortality among African women. Nonetheless, the lack of granular prevalence data to enhance precise planning and action by local stakeholders, especially in the context of the double burden of malnutrition facing many African countries, could potentially interfere with efforts to achieve NCD-related and other health system goals, warranting this study. Although recent and well-designed subnational analysis of overweight and obesity prevalence with various methodologies in SSA women provide some insight into the heterogeneity present at varying administrative levels, the majority of subnational studies are limited in the number of countries studied, the administrative regions covered, and do not account for the large sampling variance introduced by small sample sizes and complex survey sampling designs. Hierarchical Bayesian (HB) models, the method applied in this paper, explicitly accounts for small sample sizes and the large sampling variance it introduces by harnessing spatial and nonspatial relationships. 787 As such, this study applies HB models to provide baseline estimates of age-standardized prevalence of overweight (≥ 25kg/m) among women, 18 years and older, at second-level administrative units in seven SSA nations for sub-national progress monitoring of global obesity targets. Methods Data Sources We used data from the Demographic and Health Surveys (DHS). The DHS are household-based sample surveys of women aged 15 – 49 years that are collected in over 90 countries. Although a description of the survey sampling methodology is available online, we provide a summarized version here. The DHS involved a multi-stage sampling process with the goal of representative estimates for health and social indicators at the national and sub-national levels. First, countries were divided into sub-national units that correspond to existing administrative units such as states or provinces. From each region, enumeration units or clusters were randomly sampled and groups Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 5 October 2020 doi:10.20944/preprints202010.0067.v1


Introduction
The probability of premature death (i.e. between the ages of 30 years to 70 years) from major noncommunicable diseases (NCDs) -cardiovascular disease, cancers, chronic respiratory diseases and diabetes -is projected to increase in Africa and decrease elsewhere. 1 The global obesity target agreed upon by policymakers in 2013 was to halt the rise in its prevalence within their countries by 2025, relative to the baseline in 2010. 2 Progress towards this target will be monitored by assessment of the age-standardized prevalence of overweight and obesity in individuals aged 18 years and older, which is indicator 14 of the Global NCD monitoring framework. 2 Reaching this target is especially important because obesity is the second largest contributor to noncommunicable disease mortality among African women. 1 Nonetheless, the lack of granular prevalence data to enhance precise planning and action by local stakeholders, especially in the context of the double burden of malnutrition facing many African countries, could potentially interfere with efforts to achieve NCD-related and other health system goals, warranting this study.
Although recent and well-designed subnational analysis of overweight and obesity prevalence with various methodologies in SSA women provide some insight into the heterogeneity present at varying administrative levels, the majority of subnational studies are limited in the number of countries studied, the administrative regions covered, and do not account for the large sampling variance introduced by small sample sizes and complex survey sampling designs. [3][4][5][6] Hierarchical Bayesian (HB) models, the method applied in this paper, explicitly accounts for small sample sizes and the large sampling variance it introduces by harnessing spatial and nonspatial relationships. 787 As such, this study applies HB models to provide baseline estimates of age-standardized prevalence of overweight (≥ 25kg/m 2 ) among women, 18 years and older, at second-level administrative units in seven SSA nations for sub-national progress monitoring of global obesity targets.

Data Sources
We used data from the Demographic and Health Surveys (DHS). The DHS are household-based sample surveys of women aged 15 -49 years that are collected in over 90 countries. Although a description of the survey sampling methodology is available online, 9 we provide a summarized version here. The DHS involved a multi-stage sampling process with the goal of representative estimates for health and social indicators at the national and sub-national levels. First, countries were divided into sub-national units that correspond to existing administrative units such as states or provinces. From each region, enumeration units or clusters were randomly sampled and groups of households were randomly selected for inclusion in the survey. Within the selected survey households, all eligible women were interviewed. 9 Anthropometric data were obtained by trained personnel who measure and record participants' weight using the SECA 874 digital scale to the nearest 0.01kg and height with the Shorr height board. 10 The DHS geographic data consists of the latitude and longitude positions collected with a positional accuracy of 15 meters or less from the center of populated areas within surveyed clusters. 11 To ensure the confidentiality of participants, random displacement of urban clusters occurred up to two kilometers while rural clusters were displaced up to five kilometers, with 1% of rural clusters displaced up to 10 kilometers. 11 However, the displacement is restricted to stay within the country and the DHS survey region. 11 In the seven countries examined in this research study, we only included data containing explicit information on the GPS displacement. This information was not available prior to 2016 for Ethiopia, 2015 for Tanzania, and 2015 for Zimbabwe. With the exception of Nigeria, displacement of survey clusters in all the included countries were restricted to the relevant second-level administrative divisions at the time of the data compilation by the DHS program.
We acquired administrative boundaries data for mapping the surveyed clusters at the subnational level from the GADM, the database of global administrative areas (https://gadm.org/data.html).

Data Compilation
For the purposes of this study, we were interested in anthropometric data from overweight or obese women, 18 years or older. This data met the data needs for computing the global NCD indicator 14 -age-standardized prevalence of overweight and obesity in individuals aged 18 years and older. 2 We excluded pregnant women from the sample due to pregnancy-related weight changes over a short time period. 12 Overweight and obesity, henceforth called overweight, was defined as a BMI of ≥25kg/m 2 . 13 The number and names of administrative units considered in this study were based on the GADM and varied by country as seen in Table 1 below. Using their longitude and latitude positions, we mapped the DHS clusters for individual countries onto administrative units locations to obtain the names of associated first and secondlevel administrative units. We linked the resulting country datasets to the DHS individual dataset using the DHS cluster variables. This final dataset was used in the statistical analysis.
Statistical Analysis Analysis was completed on an individual country basis by fitting three separate Bayesian hierarchical models with between area variation modelled as having a spatial structured with a Besag-York-Mollie (BYM) model, a BYM2 model (a scaled version of the BYM model) or as spatially unstructured with the IID model 14 using the Integrated Nested Laplace Approximation (INLA) package in R. 15 Similar to previous work, 16 in the first stage, the estimated prevalence of overweight, y, in each second-level administrative unit, i, were calculated as the ratio of total count ( ) of overweight individuals and the total population ( ) accounting for the design weights in both the estimator and variance as specified by the Horvitz-Thompson (HT) formula. Subsequently, design weighted estimates were age-standardized using the WHO World Standard Population to ensure comparability with prevalence estimates generated by the WHO. 17 = , = ∈ Further, to develop the hierarchical smoothing model, the estimated HT area level proportions were defined as the empirical logistic transform to constrain the probability to lie in (0, 1). Thus, the likelihood was taken as the asymptotic distribution In second stage, the HB models were fitted to the logit transformed area-level data to allow partial pooling of data between areas with sparse and large samples. The independent random effects models (IID model) which included spatially unstructured random effects was of the form where is the overall or national level mean log odds of overweight in the absence of random effects, ∈ , the spatially unstructured random effects are normal and independently distributed with mean, 0, and an unknown variance, . The default flat INLA hyperprior, Gamma (1, 5e-5) was placed on the precision, , of ∈ in the IID model, and hyperprior, Gamma (1, 5e-4) on the precision, , of ∈ in the BYM model. On the other hand, , the spatially structured random effects are conditional on their neighbors, and their unknown variance, . It is based on an intrinsic conditional autoregressive model (ICAR) and follows a normal distribution with a mean defined as the mean of its neighbors, ∑ ~ , and unknown variance scaled by its neighbors, .
The default flat INLA hyperprior, Gamma (1, 5e-4) was placed on the precision, , of . The variance of the spatially unstructured random effects, , captures the between area-variability in the residual log-odds, and , the variance of the between-area spatial random effects captures spatial dependence. 18 While BYM2 maintains both the spatially unstructured and spatially structured random effects, they are both standardized such that their variance is equal to one and hyperpriors placed on the precision, and the mixing parameter, ∅, which quantifies the degree of mixing between both random effects. Therefore, instead of the sum of the random effects, , being equal to its structured component, , and unstructured components, ∈ , it is reparametrized as Penalized complexity (PC) precision hyper priors, pc.prec (1, 0.01) was placed on the precision, , and PC hyperpriors, phi (0.5, 0.5) on the mixing parameter, ∅.
The INLA provides the median, 2.5th and 97.5th percentile of the posterior distribution for the logit of the female overweight prevalence estimates. The inverse logit transformation was used to generate female overweight prevalence estimates and 95% confidence intervals (CIs) for individual second-level administrative units in each of the seven countries. Model comparison and selection was undertaken with the deviance information criteria (DIC) and the Watanabe-Akaike Information Criteria (WAIC). 19,20 To assess the impact of the DHS random displacement on the locational accuracy of the DHS cluster locations for Nigeria, the maximum probability covariate (MPC) selection method was used to determine the most probable second-level administrative unit by placing a buffer, equal to the displacement radius, around each DHS cluster location and misclassification probabilities were estimated through a simulation based process, which employs a Monte Carlo integration. 21 Finally, age-standardized overweight estimates for individual countries were mapped on to their respective second-level administrative units. To enable visualization of the heterogeneity of overweight estimates at the second-level administrative, age-standardized estimates of overweight prevalence estimates at first-level administrative units were generated using the HT formula and mapped and presented side-by-side with the second-level administrative unit estimates. Since the DHS in the studied countries were representative at the first-administrative levels, the HT formula provided representative estimates.

Funding
The first author's doctoral program was partly funded by a Rotary International District 7570 Skelton Jones Scholarship. However, the funders had no involvement in the study design, data analysis and interpretation, and manuscript preparation associated with this research study. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Participants and Locations
The number of participants and locations surveyed by the DHS program differed by country (Table 2). Nigeria, followed by Benin, had the largest sample size and numbers of DHS survey/cluster locations. After excluding adolescents and pregnant women, missing BMI data in each individual country dataset remained roughly at 5% or less. Due to the small proportion of missing data, its impact was deemed negligible and it was deleted from the analysis sample. With the exception of Tanzania and Zimbabwe, a small number of DHS clusters did not have longitude and latitude positions ( Table 2). These clusters with missing longitude and latitude values did not form part of the analysis. The order of magnitude of sample sizes and cluster locations by country remained the same after exclusions.   (13.8 -27.0%), overweight prevalence within Kandi, one of its Communes (second-level administrative unit) was as high as 32.8% (28.5 -37.4%) and in another, Banikoara, overweight prevalence was as low as 11.1% (8.3 -14.6%). Equivalently, while Ethiopia had the lowest prevalence of overweight at the national level, the Southern Nations, Nationalities and Peoples, a Regional State (first-level administrative unit), which had some of the lowest proportions of overweight women in a Zone (second-level administrative unit) also contained the Zone, Burji, with the highest overweight prevalence at 35.9% (35.5 -36.2%). In Nigeria, Adamawa, a state (first-level administrative unit) with an overweight prevalence of 24.7% (13.8 -35.6%), had an LGA (second administrative level) -Jada -with overweight prevalence at 40.1% (39.5 -40.8%) and another -Demsa-with a prevalence of 7.6% (7.3 -8.0%) Moreover, several second-level administrative units had overweight prevalence that were over twice the prevalence the national overweight prevalence level in their respective nations. This represented two districts in Tanzania, 14 LGAs in Nigeria; six districts in Mozambique, and five zones in Ethiopia. A detailed list of age-standardized overweight prevalence and 95% CIs among women, 18 years and older, at second-level administrative units in the study countries are provided in the supplementary appendix. Table 3 presents the misclassification probabilities for the 889 DHS clusters included in the estimation of overweight prevalence in Nigeria's second-level administrative units or LGAs. Our results showed that approximately 1.5% of the cluster points had more than a 50% probability of being misclassified. This implies that these clusters were more likely to be located elsewhere than the current location used for the estimation. However, the impact of such misclassification on the accuracy of second-level administrative unit estimates were reasoned as potentially low given the small number of points.

Discussion
To the best of our knowledge, this is the first study to provide evidence of highly varying disparities in the age-standardized prevalence of adult female overweight and obesity across second-level administrative units in seven African countries and a baseline against which progress towards halting the rise in the prevalence of adult female overweight can be measured at the local level. In Tanzania, we noted a location in which 72% of the adult female population were overweight or obese, and several locations in Tanzania, Nigeria and Zimbabwe where more than half of the adult female population were overweight or obese, reminiscent of the situation observed in more affluent countries, which tends to have greater burdens of overweight and obesity, especially among women, 22,23 and underscoring the need to prioritize overweight and obesity-related interventions in these settings.
However, although the micro-data offered by this study empowers local actors to engage in the prevention of overweight and obesity in their communities, the complexity of the problem requires a multisectoral approach involving stakeholders at various higher administrative levels and development agencies operating nutrition programs in these countries. WHO recommends several diet and physical activity policy and program options ranging from policy measures to reduce trans-fat, saturated fat and excess calorie intake to engagement with food service providers to improve the availability, affordability, acceptability of healthier food, and promotion of healthy food in public institutions. 2 Additionally, a recent publication identified four potential entry points and provided specific recommendations for leveraging existing undernutrition programs to accommodate overweight and obesity prevention. 24 While the recently concluded Third United Nations High Meeting on NCDs is expected to renew momentum towards uptake of overweight and obesity interventions, the 2017 Non-communicable Diseases Progress Monitor indicates that none of the countries included in this study fully complies with WHO diet and physical activity policy and program recommendations. 25 Nonetheless, this is not unique to the countries in this study as a 2017 WHO global survey reported that under 20% of African countries had an operational policy, strategy or action plan for the prevention of overweight and obesity. 26 Moreover, several research gaps exists on key contextual determinants of weight gain in women such as the drivers of food choice and physical inactivity in SSA countries 24 suggesting the low priority placed on issues and hinting at a low probability of meeting global obesity targets and associated NCD targets.
By providing disaggregated analyses on overweight and obesity burden and highlighting high burden areas, the findings of this study could potentially raise the profile of this issue in national health agendas and ensure efficiency in resource allocation. Similar to the tactics employed in tobacco control in Western countries, high-burden lower administrative units could become a testing ground for interventions prior to state and national level scale up. Given the limited funding environment, a potential source of funding for such interventions include earmarked taxes on alcohol, tobacco and sugar-sweetened beverages, which remain under-utilized in several African countries. 26 Additionally, high burden overweight second-level administrative units can be ruled out as a focus of programs primarily targeting undernutrition among adult women allowing the design and uptake of double-burden interventions in such locations.
Failure to address the large burden of overweight in the identified second-level administrative units places individuals and communities at risk of morbidity and economic losses from NCDs. Cardiovascular diseases and diabetes, the top two leading causes of BMI-related mortality 23 require substantial individual and health system resources to treat and manage, which are desperately lacking in African countries.
Enhanced precision in the estimation of overweight and obesity within lower administrative units in African countries can be achieved through the collection and availability of more publicly available micro data. While the Demographic and Health Survey remains an important resource, it is not designed to be representative at lower scales, prompting the statistical models employed in this study. However, besides this obvious limitation of our study, we also assumed that the population of women at the survey location remained statistic during the period of the study. Although the World Pop project provides migration data, these are restricted to first-level administrative units and were not suitable for use in this study. 27 Moreover, our estimates could potentially have been improved through the inclusion of area-level covariate data such as indicators of socioeconomic status and level of urbanization, which have been shown to be associated with increased burden of overweight, 28,29 but this information was not publicly available. Thirdly, since administrative boundary data was obtained from the GADM, we could not verify whether differences existed between the GADM data and the status quo in individual countries. A final limitation to note in our study is that our modeling process assumed information sufficiency in some areas in order to adjust estimates in low information areas.
In conclusion, we noted variable disparities in the age-standardized prevalence of adult female overweight within second-level administrative units in seven African countries while providing sub-national baseline data to partly satisfy measurement of indicator 14 of the global NCD monitoring framework. Overweight in adult women has become a health problem of substantial magnitude in some sub-national administrative units and requires immediate prioritization. The data provided by this study will be critical in prioritizing resources to address this emerging problem and tracking progress towards meeting global obesity and NCD targets.