Population-Based HIV Impact Assessment (PHIA)²⁰ surveys were our source of data for putative risk factors and HIV test results among AGYW of ages 15–24 years, and HIV viral load assay results among all survey respondents regardless of age or sex. The PHIA surveys are multi-stage stratified household-based surveys designed to estimate HIV viral-load suppression with specified precision at the stratum level. The PHIA surveys were reviewed and approved by the CDC and the institutional review boards of each partner implementing the surveys. We combined data from 13 surveys completed between 2015–2019 (Table 1). The numbers of respondents tested for HIV ranged from 11,682–31,579 per country and totaled 280,323, including 52,171 AGYW of ages 15–24 years. Additional details, including blood-sampling procedures are provided in the survey-specific manuals²⁰. Estimated HIV prevalence among AGYW in 2023 ranged from 0.3% in Ethiopia to 7.5% in Eswatini²¹.

PEPFAR-supported prevention services are typically provided at the second subnational administrative level, herein named target subnational units (TSNUs). Eswatini, Kenya and Lesotho are exceptions, where our TSNUs were the first subnational administrative units. We located the masked centroid coordinates of the PHIA clusters within TSNUs using the 2023 UNAIDS Combined Geographic Health Boundaries file (see Data Availability). The numbers of TSNUs per country ranged from 4−178 and totaled 1,044 (Table 1; see Supplemental Digital Content 2 for the complete list of TSNUs). We also used TSNU-level estimates of the total numbers of AGYW and AGYW living with HIV from Naomi²² model output for 2023 to estimate the numbers of at-risk (HIV-negative) AGYW.

The selection of behavioral and demographic predictors of HIV infection is described in detail in Supplemental Digital Content 1. Briefly, we extracted and recoded responses to PHIA survey questions to produce the categorical predictors: urbanicity (urban vs rural residence), age range (15−19 , 20−24 years), partner HIV status (no current partner, partner HIV negative, partner status unknown, partner HIV positive), educational attainment (never attended school or unknown, primary only, secondary, some post-secondary), sexual history (never had sex, sexual debut at age >16, sexual debut at age ≤ 16 yr), employment during past year (no, yes), history of pregnancy (no, yes), and history of transactional sex (no, yes). The covariates were selected following review by subject matter experts supporting AGYW prevention programs based on alignment with eligibility criteria measures in DREAMS program global guidance. Global DREAMS individual eligibility criteria were determined based on review of existing literature of factors associated with HIV infection among AGYW²³. Our covariates span the individual (e.g., age, education, pregnancy etc.), relational- (sex partner characteristics) and community/environment (area of residence and population viral load; see below) levels.

We developed a Bayesian spatial model which jointly estimates the probability of HIV infection among individual AGYW and the TSNU-level population HIV viral load (PVL), defined as the TSNU-level posterior mean of log₁₀(VL + 1), where VL is the number of RNA copies per ml of blood plasma. VL was measured among all HIV-positive survey participants and therefore we defined VL ≡ 0 for all HIV-negative participants. Individual probabilities of HIV infection among AGYW were modeled using a logistic component, which included the individual behavioral and demographic risk covariates and PVL as predictors. PVL cannot be observed directly but was instead estimated with error and included as a classical error-prone covariate^24,25. The possibility of spatial correlation was accommodated using TSNU-level BYM2^26,27 spatial smoothing structures in the estimation of both HIV positivity among the AGYW and estimation of PVL. The BYM2 model extends the more popular BYM model²⁸ by enabling scaling which facilitates hyperprior specification²⁹.

It is important to accommodate, to the extent possible, features of the survey design in Bayesian model-based estimation and prediction from survey data^30–33. Bayesian model-based inference from survey data predicts outcomes by including the survey sampling probabilities or weights as predictors^31,34,35. Our models were fitted to pooled data from 13 probability surveys and therefore we re-scaled the sampling weights so that their sums equaled the effective sample sizes for each survey³⁶. The effective sample size for a survey was the actual sample size divided by the design effect for estimation of HIV prevalence.

We fitted model variations ignoring the re-scaled weights, and also followed Vandendijck et al. (2016)³⁴ by including the re-scaled weights in B-spline basis functions³⁷ in the predictor for HIV positivity. We chose B-splines having 3 degrees of freedom. We fitted eight possible model variations excluding and including country-level random effects cluster-level random effects and B-spline basis functions of the rescaled weights. Models were compared using the Watanabe-Akaike Information Criterion (WAIC)³⁸ and area under the receiver-operating characteristic (ROC) curve³⁹ for HIV status based on 10-fold cross-validation^40,41 from each model (see Supplemental Digital Content 1).

We began with a search for plausible interaction effects among the predictors using interaction plots. Pairs of predictors which showed evidence of non-parallel responses with HIV status were further examined using preliminary frequentist non-spatial mixed effects logistic models wherein TSNU was the sole random effect. Interactions for which P < 0.05 from the frequentist models were included, along with their main effects, in a preliminary full joint Bayesian spatial model which included country- and cluster-level random effects and B-spline basis functions of the rescaled weights. That full model was refitted sequentially eliminating unimportant interaction terms. The fixed-effect risk covariates and their interactions from the preliminary model having the smallest WAIC were retained for final fitting of the eight joint spatial models obtained by including/excluding country- and cluster-level random effects and the smoothed rescaled survey weights.

A detailed model description, including specifications of priors, hyperpriors and example model code, is given in Supplemental Digital Content 1. All computations were performed using R version 4.3.1⁴². The joint spatial model variations were fitted using the Integrated Nested Laplace Approximation (INLA) package⁴³, version INLA_2023-09-09.

Graphical analyses and preliminary frequentist models suggested the plausible existence of interaction effects between age and age of sexual debut, age group and partner HIV status, age of sexual debut and history of work during the past year, educational attainment and partner HIV status, and history of pregnancy and partner HIV status in the prediction of HIV status of AGYW (see Supplemental Digital Content 1). Therefore, we fitted preliminary full models, including those interaction effects, country- and cluster-level random effects and containing the smoothed rescaled survey weights to confirm the importance of any of those interactions. The age × partner HIV-status and education × partner HIV-status interactions were eliminated based on WAIC differences ≥ 10 between the full and reduced models.

Our final set of behavioral/demographic risk factors were age group, urban versus rural residency, educational attainment, history of pregnancy, partner HIV status, transactional sex during the past year, age category at sexual debut, employment during the past year, and age group × age at sexual debut, age group × employment, and partner HIV status × history of pregnancy interactions. Those risk factors were included in the eight joint spatial models defined by the inclusion or exclusion of the country-level and cluster-level random effects, and the B-spline basis function of the re-scaled survey weights for prediction of HIV status. The areas under the ROC curves (AUC) were similar across all models which included the country-level random effect (Table 2). Among those, models M100, M101 and M111 produced WAIC values which were equal for practical purposes. Therefore, we chose the simplest of those, M100, including the country-level random effects but excluding the cluster-level random effects and the smoothed weights, as our final model. The rescaled weights conveyed almost no information about HIV test positivity (Supplemental Digital Content Fig. 5).

Area-level PVL was—by far—the most important predictor of HIV status among the AGYW. The adjusted odds ratio (aOR) for the effect of PVL (70.6; 0.95-probability credible interval 20.7‒240.5) represents the change in logit(positivity) for each ½-unit change in PVL (Table 3). The distribution of area-level PVL estimates was slightly skewed (Supplemental Digital Content 1) and ranged from 0–0.5449 log₁₀(RNA copies·ml⁻¹ + 1). The combination of joint estimation and spatial smoothing shrunk the highly unstable survey domain estimates of PVL, which ranged from 0–1.06 log₁₀(RNA copies·ml⁻¹ + 1) units, toward the mean as expected (Supplemental Digital Content 1).

Approximately 21% of the among-TSNU variation in PVL is attributable to spatial correlation, whereas essentially all the among-TSNU variation in HIV positivity is spatially independent (Table 3). Evidently the combination of fixed- and random-effect predictors, including PVL, exhausts the possibility for among-TSNU spatial correlation in the prediction of HIV positivity in this area-level model.

Strong spatial patterns in PVL are evident at both the subcontinental and national scales. Areas of higher PVL are more prevalent within the southern African countries of Malawi, Zambia, Zimbabwe, Lesotho and Eswatini than elsewhere (Fig. 1, Top). Visualization of PVL and the numbers of HIV-negative AGYW within countries provides a basis for aligning HIV prevention interventions to the geographic areas which might produce the greatest benefits in terms of infections averted (Fig. 1, bottom). Additional country-specific maps of PVL and numbers of at-risk AGYW are provided in Supplemental Digital Content 1.

Those AGYW who had both a partner with HIV and a history of pregnancy had the next strongest association with HIV positivity (aOR 32.06; 23.67–43.43), followed by those having a partner with HIV but no past pregnancy (aOR 12.12; 7.48–19.63) (Table 3). AGYW who reported having no partner but having a history of pregnancy, and AGYW who had a partner of unknown HIV status regardless of history of pregnancy also incurred increased odds of HIV infection. There was no detectable association between history of pregnancy and HIV infection among AGYW who had HIV-negative partners.

Completion of secondary education (aOR 0.68; 0.54–0.87) and attainment of some post-secondary education (aOR 0.35; 0.24–0.51) were protective (Table 3). Having engaged in transactional sex within the past year was associated with HIV acquisition (aOR 1.27; 1.08–1.49). AGYW of ages 15–19 years who had their sexual debut at age 16 years or younger incurred increased odds of HIV acquisition (aOR 1.42; 1.10–1.84), as did those of age 20–24 years who had their sexual debut later than age 16 (aOR 2.17; 1.72–2.74) or at 16 or younger (aOR 2.94; 2.33–3.70).

The area under the subcontinental cross-validated ROC curve was 0.82 (95%CI 0.81–0.83). Country-specific ROC curves varied; their AUC values ranged from 68.0–81.7% (Supplemental Digital Content 1). The model enables statistically principled classification of newly encountered AGYW with known sensitivity and specificity (Table 4). For example, enrollment in prevention interventions of all those HIV-negative AGYW who have a probability of test positivity of at least 0.0276 would select those for whom their risk profiles would predict positivity with 80% sensitivity and 66.1% specificity. The threshold probabilities of test positivity are easily computed from responses to the risk screening questions, country and TSNU of residence, the parameter estimates and country and TSNU random effect estimates (Supplemental Digital Content 1).