Mapping male circumcision for HIV prevention efforts in sub-Saharan Africa

We compiled a dataset of 38,747 geo-referenced observations from 109 household surveys conducted between 2000 and 2017 in 38 countries in sub-Saharan Africa as defined by the Global Burden of Disease (GBD) study [11]. We assembled this dataset through a review of major survey series (Demographic and Health Survey [DHS], AIDS Indicator Survey [AIS], Multiple Indicator Cluster Survey [MICS], Core Welfare Indicators Questionnaire Survey [CWIQ], Population-based HIV Impact Assessment Survey [PHIA]) and surveys tagged with the “circumcision” keyword in the Global Health Data Exchange [38]. We included all surveys that asked men about their circumcision status, were sampled from the general population, and contained subnational geographical information (Supplementary Table 1 in Additional file 1). We excluded countries in sub-Saharan Africa with no data availability from our analysis. We also excluded the island nation of São Tomé and Príncipe, despite having one data source with low sample size, due to evidence that its MC coverage patterns were substantially lower than the regional trends, which would break our implicit model assumption that areas that are close in space and time borrow information on MC prevalence (Supplementary Table 2). Survey data were not required to distinguish between medical and traditional male circumcision, and in the few surveys that included both, MC prevalence was derived from summing across both variables. Survey data were originally in two forms: survey microdata (i.e., individual-level survey responses) or survey reports (Supplementary Table 1). For surveys with available microdata, we extracted variables related to age, MC status, location, and survey weights. After subsetting the data to ages 15–49 years and excluding observations with missing information on any of these variables (4.0% of all observations), we aggregated the data by calculating the survey-weighted MC prevalence at the most precise spatial resolution available. The ideal spatial resolution was a latitude and longitude point that represented the location of the survey cluster (point-level data), but when not available, we geo-located survey microdata to the smallest geographical areal unit (termed a polygon), typically representing an administrative unit. In instances where data were matched to a polygon rather than specific GPS coordinates, these data were “resampled” proportional to the underlying population in the areal unit to mimic point data. The population data used for this analysis were obtained from WorldPop [39]. The methods for sampling point locations from areas associated with areal data are consistent with previous studies using model-based geostatistics [31, 40‐43]. Weighting by sample size, 70.7% of all data were associated with GPS coordinates, while the remaining data were associated with polygons and resampled using this approach.

Several survey reports presented MC estimates for age groups other than 15–49 years (five sources representing 1.4% of the total effective sample size; Supplementary Table 3 in Additional file 1). In order to incorporate data from these reports, we used a cross-walking model—an approach for linking disparate data sources (in this case data sources reporting for different age groups)—that used existing microdata and a linear regression to convert the prevalence in the reported age range to the standard 15–49 age range. The age cross-walk is consistent with the method previously used in the geospatial modeling of HIV [31]; this process calculated a new sample size that reflects our confidence in the estimate of MC prevalence as a function of the uncertainty in our linear model and the original sample size. The distribution of geo-located MC prevalence data is shown in Supplementary Fig. 1, and a list of sources is also located in Supplementary Table 1 in Additional file 1. We excluded three data sources in Nigeria because lower coverage contradicted several independent DHS surveys in the same country, and there was evidence that at least one of the reports had been previously flagged for data validation issues. The excluded data and the reasoning for their exclusion can be found in Supplementary Table 2. Additional methodological details on data identification and processing are also presented in section 2 in Additional file 1.

We modeled MC prevalence using a Bayesian spatially and temporally explicit generalized linear mixed effects model developed previously for an analysis of HIV prevalence and updated for this analysis (Supplementary Fig. 2 in Additional file 1) [31]. Due to computational constraints and to allow for regional differences in the temporal and spatial autocorrelation in MC prevalence, we fit separate models for four geographically contiguous regions in sub-Saharan Africa adapted from the Global Burden of Diseases, Injuries, and Risk Factors study (Supplementary Fig. 3 in Additional file 1) [1]. We excluded countries in sub-Saharan Africa with no data availability, with the exception of São Tomé and Príncipe, which was excluded despite having one data source due to evidence that its MC patterns were substantially different from the regional trends. Within our Bayesian geostatistical model, MC coverage was modeled as binomial count data with a logit link function and the linear model included an intercept, a fixed effect on year, a country-level random effect, an autocorrelated spatiotemporal random effect, and an uncorrelated error term (nugget effect). Unlike similar past analyses [40‐43], we did not use covariates for this analysis given that predictive covariates such as ethnicity, culture, or religion were not available at our model’s spatial and temporal resolution. Due to the absence of viable covariates and the limited availability of survey data in the most recent year, extrapolation of our estimates to later years in the study period was often informed by the regional time trend provided by the fixed effect on year. After fitting the model, we generated 1000 draws of MC prevalence from the approximated joint posterior distribution at each 5 × 5-km grid cell across each model region and year in the study period. We calculated point estimates for each grid cell as the mean of these draws, and we designated the 95% uncertainty intervals as the 2.5th and 97.5th percentiles of these draws.

In addition to estimates of MC prevalence on a 5 × 5-km spatial grid, we produced estimates of MC prevalence for first and second administrative level units. We constructed these estimates by calculating population-weighted averages of MC prevalence for each grid cell or fraction of a grid cell within a given first or second administrative level unit as defined by the Database of Global Administrative Areas (GADM) [44]. We list the names and the number of designated first and second administrative level units in each country in Supplementary Table 4 in Additional file 1. Furthermore, we derived estimates of the number of circumcised and uncircumcised men in each administrative unit by multiplying estimated MC prevalence or uncircumcised prevalence, respectively, in each grid cell by the corresponding WorldPop population estimates of men ages 15–49 and aggregating to administrative units. Finally, we calculated the posterior probability of exceeding 80% MC prevalence in 2017 at the 5 × 5-km level and among administrative units as the percentage of draws from the approximated posterior distribution in which MC prevalence was 80% or greater. Throughout our analysis, we qualify statements as statistically significant if the posterior probability of that statement exceeds 95%. All models were fitted using an integrated nested Laplace approximation in the R-INLA package, version 3.3.2 [45, 46].

We performed both in-sample and out-of-sample model validation using a spatiotemporal fivefold cross-validation strategy. Out-of-sample validation metrics for MC coverage indicate good model fit with spatial stratification at the second administrative level (Supplementary Table 5 in Additional file 1), including mean error (0.6 percentage-points), root-mean-square error (8.7 percentage-points), and 95% coverage of predictive intervals (97.0%). Additional results and details on data preparation, modeling, estimation, and validation can be found in Additional file 1.

Across sub-Saharan Africa, there were stark regional contrasts in estimated MC prevalence, with high levels of MC prevalence in western and central sub-Saharan Africa, low levels of MC prevalence in southern sub-Saharan Africa, and variable levels of MC prevalence in eastern sub-Saharan Africa (Figs. 1 and 2, Supplementary Figs. 4–6). Estimated MC prevalence in most countries in western and central sub-Saharan Africa was largely uniformly high; estimated coverage was greater than 85% in > 95% of second administrative level units in these countries from 2000 to 2017. Conversely, estimated MC prevalence in southern and eastern sub-Saharan African countries with lower levels of MC overall was heterogeneous in space and time. The 14 priority countries identified by WHO/UNAIDS for VMMC were among the group with lower coverage: 13 of these countries (excluding Ethiopia) were among the 15 countries with the lowest estimated national MC prevalence at the start of VMMC campaigns in 2008 (along with Burundi and Guinea-Bissau).

Within priority countries, estimated MC prevalence varied substantially among first and second administrative level units. In Kenya, with a high estimated national prevalence of 90.8% (95% uncertainty interval, 86.8–93.8%), MC prevalence varied at the first administrative level from 99.6% (98.8–99.9%) in the Makueni county to 40.9% (24.4–57.8%) in the Turkana county in 2017. In Uganda, with a more moderate national prevalence of 45.6% (38.6–52.1%), MC prevalence varied at the first administrative level from 87.4% (79.2–93.5%) in the Kasese district to 10.0% (4.4–18.4%) in the Pader district in 2017. These considerable subnational differences in prevalence were less apparent in low-prevalence settings, but still notable; for example, in Zimbabwe (national prevalence, 15.3% [11.8–19.3%]), MC prevalence varied at the first administrative level from 28.3% (16.6–41.4%) in the Bulawayo province to 9.8% (6.1–14.5%) in the Mashonaland Central province in 2017. There was a more than twofold difference in MC prevalence between the first administrative level units with the highest and lowest estimated prevalence in most (12 of 14; barring Lesotho and Eswatini) priority countries in 2017. This heterogeneity in MC prevalence increased at the second administrative level, where nearly half (6 of 14) of the VMMC priority countries had more than a fivefold difference between their second administrative level units with the lowest and highest estimated prevalence in 2017 (Fig. 3).

We estimated that a large proportion of uncircumcised men were concentrated in a small number of geographical areas. At the country level in 2017, roughly half (49.8%) of the estimated 34.5 (32.9–36.1) million uncircumcised men in sub-Saharan Africa resided in South Africa, Uganda, Zambia, and Zimbabwe, compared to 13.6% of all men in sub-Saharan Africa aged 15–49 residing in those countries. The majority (30.5 [29.2–32.0] million; 88.5% [86.2–90.3%]) of uncircumcised men in sub-Saharan Africa were concentrated in the 14 priority countries. Furthermore, within priority countries in 2017, more than half (50.3%) of all uncircumcised men were concentrated in just 123 (9.1%) second administrative level units, compared to 29.2% of all men aged 15–49 in priority countries residing in those administrative units (Fig. 4). We also observed this heterogeneous spatial distribution of uncircumcised men within each priority country. In 12 of 14 countries (excluding Eswatini and Rwanda), more than half of the country’s estimated uncircumcised men resided in only 30% or fewer second administrative level units. In eight countries, more than half of the country’s uncircumcised men lived in 20% or fewer second administrative level units.

Estimated MC prevalence changed considerably over time, and the temporal trends were marked by the onset of large-scale VMMC campaigns in priority countries from 2008 onwards (Fig. 5, Supplementary Fig. 7). From 2000 to 2008, prior to the VMMC campaigns’ inception, the estimated mean national MC prevalence stagnated in all priority countries; the model does not predict a statistically significant (posterior probability > 95%) increase or decrease in any country’s national MC prevalence over this period. At subnational levels over the same time period, we estimated an increase in MC prevalence in 60.1% (37.1–80.1%) and 58.7% (43.7–71.6%) of first and second administrative level units, respectively. However, among those administrative units with an estimated increase in MC prevalence, the median estimated increase was slight—2.0 and 2.2 percentage-points in first and second administrative level units, respectively. Furthermore, only 0.7% of our estimated increases at all first administrative level units were statistically significant (posterior probability > 95%), indicating high uncertainty in MC temporal trends from 2000 to 2008.

In contrast, and corresponding with the onset of the VMMC campaigns, the estimated number of circumcised men and MC coverage increased in priority countries from 2008 to 2017. The total number of circumcised men in priority countries rose by 21.5 (19.9–23.0) million in this period, from 42.5 (41.8–43.2) million in 2008 to 64.0 (62.5–65.3) million in 2017. Over this same period, national MC coverage increased in all priority countries, and this increase was statistically significant (posterior probability > 95%) in all countries barring Ethiopia. Within priority countries, however, the magnitude of increase varied; for example, estimated MC prevalence rose by 1.0 (− 1.7, 3.7) percentage-points in Ethiopia and 22.9 (15.2–31.3) percentage-points in Lesotho. At subnational levels, we estimated an increase in MC prevalence in 91.9% (86.0–95.8%) and 84.7% (78.1–90.0%) of all first and second administrative level units, respectively, in priority countries. Among those administrative units with an estimated increase in MC prevalence, the median estimated increase was 15.0 percentage-points and 12.2 percentage-points in first and second administrative level units, respectively. These estimated increases were also more certain; 65.1% of the estimated increases in first administrative level units were statistically significant (posterior probability > 95%).

While there was a generalized surge in estimated MC prevalence that coincided with the onset of VMMC efforts from 2008 onwards, these broad trends mask large subnational disparities. Three of the 14 priority countries—Tanzania, Mozambique, and Ethiopia—contained areas where estimated MC prevalence increased as well as areas where estimated MC prevalence decreased among first administrative level units; this was true for seven priority countries at the second administrative level. While we estimated a decrease in 121 (8.9%) of second administrative level units in priority countries after the onset of VMMC campaigns in 2008, the magnitude of these declines was small (mean across units, 2.6 [0.4–4.9] percentage-points) and the impacted areas generally had high coverage in 2008 (mean across units, 90.5% [89.3–91.6]) and therefore still had high coverage in 2017 (mean across units, 88.0% [85.8–89.9]). In certain areas, these differences were substantial. For example, in Mozambique, where the national estimated MC prevalence rose by 7.6 (1.4–13.3) percentage-points from 2008 to 2017, the difference in coverage at the second administrative level ranged from an estimated  -23.0 (-45.2,  -3.5) percentage-point change in the Muidumbe district (from 88.9% [76.7–96.1%] to 65.9% [42.5–84.2%]) to a 28.6 (5.8–49.5) percentage-point change in the Xai-Xai district (from 27.6% [16.1–41.0%] to 56.2% [35.5–74.3%]) over the same time period. Even in countries where we estimated an increase in MC coverage for all first and second administrative level units, this heterogeneity persisted. In Uganda, where the estimated national MC coverage increased by 22.9 (15.2–31.3) percentage-points from 2008 to 2017, change in estimated MC prevalence varied from a 3.6 (-4.8, 13.4) percentage-point change in Bukonzo county (from 89.8% [79.8–95.7%] to 93.4% [86.2–97.4%]) to a 40.8 (15.3–61.7) percentage-point change in Masindi county (from 19.6% [8.3–37.0%] to 58.8% [37.8–78.9%]).

Despite these increases, we find strong evidence in only three priority countries of having met the goal of 80% MC coverage by 2017 and do not find strong evidence in any country of reaching the 80% coverage target in the collection of administrative units targeted for VMMC. Figure 6 depicts the posterior probability that a given area exceeded the 80% MC coverage target in 2017; high posterior probability indicates substantial evidence that an area met this target, while low posterior probability indicates substantial evidence the area did not meet this target. There was strong evidence that Ethiopia, Kenya, and Tanzania reached the aspirational 80% coverage goal in 2017 (posterior probability > 99.9%) and strong evidence that all other priority countries failed to meet this goal (posterior probability < 0.01%). In some countries, national coverage estimates are less relevant given that VMMC implementation is limited to a regional focus; for example, VMMC implementation in Ethiopia is focused exclusively on the Gambela Region [18]. At subnational levels, the evidence for 80% MC coverage is more mixed, even in countries with strong evidence (posterior probably > 95%) to have reached the national coverage goal. In Ethiopia, 10 (12.7%) of its second administrative level units are unlikely (posterior probability < 5%) to have achieved the coverage target. Conversely, in Mozambique, where there is strong evidence that national MC prevalence did not meet the coverage goal, there is nonetheless strong evidence (posterior probability > 95%) that 33 (25.6%) districts did accomplish this goal in 2017.

This study is subject to several limitations. First, our analysis is limited by the availability of the underlying data. For our research, we constructed a large database of geo-located MC prevalence data; still, there are gaps in data coverage in both time and space (Supplementary Figs. 1 and 7 in Additional file 1). For example, across VMMC priority countries, the most recent survey used in our analysis ranged from 2013 to 2017. As a result, in priority countries with continued VMMC scale-up but without recent data, we likely underestimate MC prevalence in recent years. Further, over 4 million VMMCs were performed in priority countries in 2018 [18], but currently, there are too few available surveys to reliably extend our analysis to include MC prevalence in 2018. In many VMMC priority countries, however, there are surveys that are either recently completed or currently in the field that will improve our estimates in the future. Second, there are several factors that could impact the quality of the data used in our analysis. Survey data are subject to social desirability bias as well as non-response bias, and the response rate can vary across countries and years. We applied a complete-case analysis that did not adjust for non-response bias; future research could assess the impact of the response rate on estimates of MC prevalence. Our results also rely on self-reported circumcision status that does not distinguish between traditional and medical male circumcision. Past research has found that self-reported MC status may be subject to misreporting errors due to confusion between medical and traditional circumcision practices [57]. Traditional circumcision is often performed in a non-clinical setting, and evidence of its protective effects against HIV transmission is mixed, largely due to discrepancies in defining MC when traditionally performed [27, 29, 30, 58]. Our current analysis provides valuable information on the unmet needs of men who can be reached by VMMC given they have not undergone any form of MC. In the future, distinguishing between non-medical and medical MC would provide a more accurate description of how MC translates to reduced HIV transmission risk. Further, the geographical precision of our data is subject to some error. Most surveys that collect GPS coordinates perform random displacement to safeguard respondent confidentiality; for example, in the Demographic and Health Surveys (DHS), survey clusters are displaced by as much as 2-km for urban clusters, 5-km for rural clusters, and 10-km in 1% of rural clusters [59]. While this can affect a geostatistical model’s predictive power, past research determined that even with GPS displacement, model estimates at a 5 × 5-km resolution are still relatively accurate [60]. Third, while we have attempted to quantify uncertainty in our estimates from various sources, some sources of uncertainty could not be included in the model, such as uncertainty associated with the population estimates. Population estimates from WorldPop are also subject to error, particularly in sparsely populated areas, which may affect the accuracy of our predictions for the number of circumcised and uncircumcised men, especially at fine geographical detail. While WorldPop estimates include censuses data as inputs to their modeling framework [61], depending on timing and data accessibility, estimates may differ from the underlying census measures and may not utilize the most recent census or the most detailed tabulations. Fourth, we use GADM shapefiles to define administrative subdivisions, and differences in administrative divisions between GADM and individual country’s designation of administrative areas may affect the accuracy of results, especially in our estimates of the number of circumcised and uncircumcised men. Fifth, we were unable to include covariates in our model because none of the variables we expect to be most predictive—such as ethnicity, culture, or religion—were available at our spatial or temporal resolution. Future covariates that correlate with ethnicity, culture, or religion would help improve our model predictions in locations and time periods with sparse data coverage. Sixth, our modeling strategy “borrows strength” across space and time to inform estimates in locations and time periods with very small sample sizes and to interpolate in locations and time periods with no directly observed MC prevalence data. While we believe this is generally appropriate, there may be specific locations and times where this methodology breaks down; for example, abrupt differences in MC prevalence in neighboring locations or time periods are unlikely to be reflected in our estimates unless they are directly observed in the underlying MC data and if those data have substantial sample sizes. Finally, our estimates of MC convey only one component of a comprehensive HIV preventative package and only one of many potential drivers of the spatial distribution of the HIV epidemic. Policymakers must take care not to interpret areas of low MC coverage as definite areas of high HIV burden and vice versa; there are many other factors that contribute to the underlying HIV burden at a local scale.

There is ample opportunity to expand this analysis in the future. In addition to including more contemporaneous surveys as they become available, our analysis could also leverage spatially located non-national data sources or data from other research studies. In addition, our current analysis focuses only on one defined age group: males ages 15–49. Past research has shown important variation in the magnitude of impact and cost-effectiveness of VMMC scale-up among different age groups [46] and has indicated that focusing on younger age groups may be the most cost-effective and impactful VMMC intervention [48]. In line with this research, the WHO has called for new coverage targets to be set by age strata as opposed to uniformly set over the entire 15–49 age range [24]. New global targets aligned with the UNAIDS fast track goals aim for 90% MC coverage in men ages 10–29 [19]. While our current analysis focuses on ages 15–49 due to the availability of data and alignment with the original fast track goals, in future analyses, progress towards coverage targets should be monitored by multiple age strata at subnational levels, including younger adolescents. Finally, future analyses that combine estimates of local MC prevalence with information on local HIV incidence could further identify high-risk areas that would benefit from VMMC intervention. At present, the only estimates of HIV burden at the same geographic precision report HIV prevalence, an imperfect measure of the risk of acquiring HIV. In the future, independently constructed estimates of local HIV incidence combined with maps of MC prevalence could provide compelling evidence to identify priority administrative areas for VMMC campaigns.