Population-Level Association of Voluntary Medical Male Circumcision Scale-Up and HIV infection in Eswatini: An Ecological and Spatial Analysis
Description
Voluntary Medical Male Circumcision (VMMC) is a cornerstone biomedical intervention against heterosexual HIV transmission, yet long-term population-level evidence of its impact in high-burden settings remains limited. Traditional evaluations rely on non-spatial individual models or mathematical simulations that overlook geographic dependencies. This study provides the first empirical, population-level evaluation in Eswatini to isolate structural spatial spillover and herd-protection effects by combining serial Population-based HIV Impact Assessment (PHIA) datasets with advanced spatial lag regression models. Objectives: To evaluate the association between community-level VMMC scale-up and reductions in new HIV infections in Eswatini between 2016/17 and 2021, explicitly assessing whether localized coverage density generates spatial spillover benefits. Methods: Using a repeated cross-sectional ecological design, secondary data were compiled from two national household surveys: SHIMS 2 (2016/17) and SHIMS 3 (2021). Survey-weighted biomarker data, including recency assays and viral load metrics, were analyzed across 17,217 individuals. Multilevel mixed-effects logistic regression and spatial lag models adjusted for demographic characteristics, behavioral risk variables, and local antiretroviral treatment saturation. Findings: Annual HIV incidence declined from 1·27% per year (SE = 0·0026) in 2016/17 to 0·76% (SE = 0·0018) in 2021, coinciding with an increase in median VMMC coverage from 29% to 52%. High-coverage communities saw incidence nearly halved (1·30% to 0·60%), while low-coverage areas showed a modest decline (1·24% to 0·91%). At the individual level, circumcised males had 52% lower odds of HIV prevalence (OR = 0·48; 95% CI: 0·35–0·65; p < 0·001). Crucially, spatial lag regression confirmed a significant inverse community-level association between VMMC density and HIV distribution ($\beta$ = –0·081; z = –1·99; p = 0·046), supported by significant positive spatial autocorrelation ($\rho$ = 0·26; p = 0·010). Interpretation By deploying spatial lag models on serial national biomarker data, this study demonstrates that VMMC scale-up functions as a public health multiplier, yielding significant structural spillover protection beyond the individual recipient. These findings move epidemiological evaluation beyond uniform national rollouts, providing actionable spatial intelligence to target remaining low-coverage clusters and optimize national epidemic control.
Files
Steps to reproduce
Both SHIMS 2 and SHIMS 3 field data collection was conducted using tablet-based systems, with data transmitted securely via a virtual private network (VPN) to a centralized server and stored in a PostgreSQL database. Cleaning of interview data was performed using Statistical Analysis System SAS 9.4 (SAS Institute Inc., Cary, North Carolina, United States), while laboratory data were processed and linked to the finalized questionnaire dataset through unique specimen barcodes and study identification numbers. For this study, anonymized SHIMS 2 and SHIMS 3 datasets were obtained through formal data use agreements. A structured data extraction table was used to isolate relevant variables, including: • Primary outcomes: HIV incidence and male circumcision status. • Covariates: Age, sex, education, alcohol consumption frequency, age at first sex, ART coverage, urban/rural In both surveys, the primary sampling units (PSUs) were enumeration areas (EAs) defined by the Central Statistics Office (CSO). Stratification was based on region (Hhohho, Manzini, Shiselweni, Lubombo) and urban/rural status, resulting in eight strata. Within each stratum, EAs were selected using probability proportional to size (PPS), based on the number of households from the most recent census data, 2007, for SHIMS 2 and an updated 2019 frame for SHIMS 3. SHIMS 2 selected n = 287 EAs from a frame of n = 2,064 EAs, targeting approximately n = 5,740 households and n = 12,042 respondents aged 15 and older. SHIMS 3 selected 200 EAs from a frame of n = 2,260 EAs, targeting n = 7,000 households and achieving approximately n =11,611 respondents aged 15 and older. Within each EA, 15 to 70 households were randomly selected, with an average of 35 households per EA. This yielded an approximate mean of 58 respondents per cluster in SHIMS 3, and an approximate mean of 42 respondents per cluster in SHIMS 2. As this is a secondary analysis of nationally representative surveys, the study utilized the complete datasets from the SHIMS 2 and SHIMS 3. The use of the full available datasets was intended to maximize statistical power and ensure the precision of community-level estimates. Both surveys have a combined total of n = 23,653 respondents. The study includes all eligible SHIMS respondents with definitive HIV test results and sociodemographic data: SHIMS 2 (2016–2017), n = 12,042. Data were processed in R (v4.4.3/v4.5.2) using RStudio, with harmonization across survey years, cleaning of demographic, behavioural, and biomarker variables, and construction of analytic indicators. VLS was defined as <1,000 copies/mL among HIV positive respondents, while recent infection candidates were identified as HIV positive individuals with a positive recency assay and viral load ≥1,000 copies/mL. The final analytic dataset included 17,217 individuals with valid survey weights, strata, and cluster identifiers.
Institutions
- Foundation for Professional DevelopmentGauteng, Pretoria
- University of the WitwatersrandGauteng, Johannesburg