Evaluating the impact of achieving cascade equality in Eswatini: a modeling study on the prevention impacts of antiretroviral therapy

Abstract

Objective:

Inequalities in the antiretroviral therapy (ART) cascade across subpopulations remain an ongoing challenge in the global HIV response. Eswatini achieved the UNAIDS 95-95-95 ART cascade targets by 2020, with differentiated programs to minimize inequalities across subpopulations, including for female sex workers (FSWs) and their clients. We sought to estimate the impacts of this achievement, through a retrospective impact evaluation of ART scale-up in Eswatini.

Design:

Drawing on population-level and FSW-specific surveys, we developed a compartmental model of heterosexual HIV transmission, and calibrated it to observed HIV prevalence, incidence, and ART cascade scale-up in Eswatini.

Methods:

We defined four counterfactual scenarios in which the population overall reached only 80-80-90 by 2020, but where FSW, clients, both, or neither were disproportionately left behind, reaching only 60-40-80. We estimated additional HIV infections by 2020 in counterfactual vs. observed scenarios, and identified epidemic conditions which maximized differences.

Results:

Compared with observed cascade scale-up in Eswatini, leaving behind neither FSW nor their clients led to median (95% confidence interval, 95% CI) 8.8 (6.3–10.9) additional infections by 2020 vs. 14.3 (10.8–18.6) if both were left behind, a 63 (31–128) increase. The impact of leaving behind FSW and/or clients was largely determined by their population sizes and HIV incidence ratio among clients vs. men overall.

Conclusion:

Inequalities in the ART cascade across subpopulations can undermine the anticipated prevention impacts of cascade scale-up. As Eswatini has shown, addressing inequalities in the ART cascade that intersect with transmission risk can maximize incidence reductions from cascade scale-up.

Introduction

Early HIV treatment via antiretroviral therapy (ART) has important health benefits for people living with HIV [1]. A secondary benefit of early ART given Undetectable = Untransmittable (U = U) is that transmission risks are mitigated in serodifferent partnerships [2]. To realize these population-level benefits, massive efforts have been underway to achieve the UNAIDS 95-95-95 ART cascade targets [3], that is, to have 95% diagnosed among people living with HIV, 95% on ART among those diagnosed, and 95% virally suppressed among those on ART. Botswana, Eswatini, Rwanda, Tanzania, and Zimbabwe have already surpassed 95-95-95 nationally [3], and achieving these targets is expected to help reduce HIV incidence towards local elimination.

Numerous transmission modeling studies have sought to estimate the prevention impacts of achieving 90-90-90+ across Sub-Saharan Africa [4,5]. Modeled populations are often stratified by risk, including key populations like female sex workers (FSWs) and their clients, to capture important epidemic dynamics related to risk heterogeneity [6]. However, these studies have generally assumed that ART cascade attainment (i.e., proportions diagnosed, treated, and virally suppressed) or progression (i.e., rates of diagnosis, treatment initiation, and viral suppression) were equal across modeled subpopulations. For example, among the studies in [5] (see also Box 1), key populations were usually assumed to have equal cascade progression with the population overall, or greater in some scenarios, but never lesser.

There are concerns that inequalities in the ART cascade could undermine the population-level prevention impacts of ART anticipated from individual-level and model-based studies [7–9]. Available data suggest that cascade attainment can be lower among marginalized subpopulations at greater risk of HIV, including key populations, younger men and women, and highly mobile populations [8,10]. These inequalities can be driven by systemic barriers to engagement in care, which intersect with individual, network, and structural determinants of HIV risk, such as economic insecurity, mobility, stigma, discrimination, and criminalization [11–14]. Moreover, cascade data may be lacking entirely for subpopulations experiencing the greatest barriers to care, that is, the lowest ART cascades likely remain unmeasured [10]. Indeed, these inequalities are also likely to grow given recent dismantling of USAID and PEPFAR key populations programs [15].

Eswatini, which has had the highest national HIV prevalence in the world, largely minimized cascade inequalities en route to 95-95-95. This equity achievement was driven by key populations programs and numerous community engagement initiatives to identify and address subpopulation-specific barriers to care [16–18]. We sought to quantity the impacts of this achievement on HIV transmission. We used a deterministic compartmental model of heterosexual HIV transmission, including FSW and their clients, calibrated to reflect observed HIV epidemic and cascade scale-up in Eswatini. We compared cumulative HIV infections and incidence over 2000–2020 in this base-case scenario with counterfactual scenarios in which cascade scale-up was slower, and where FSW and/or their clients were disproportionately left behind (Objective 1). We also sought to identify epidemic conditions (subpopulation sizes, turnover, and HIV incidence ratios) under which such inequalities could have the largest impact on transmission (Objective 2).

Methods

Here, we describe model parameterization and calibration, followed by analyses for Objectives 1 and 2.

Data sources

We used published aggregate data from four national surveys in 2006–2021 [17,19–21]; aggregate data from a 2021 FSW bio-behavioral survey [18]; and individual-level data from FSW bio-behavioral surveys in 2011 and 2014 [22,23], accessed under approval from the Scientific and Ethics Committee of the Eswatini Ministry of Health (MH/599B), and the Institutional Review Board of the Johns Hopkins Bloomberg School of Public Health (3508).

Model parameterization and calibration

The modeled population is stratified by HIV status, sex, sexual activity, and the ART cascade of care. Appendix A gives complete details of the model structure, parameterization, and calibration.

HIV

Modeled HIV natural history includes acute infection and four stages defined by CD4⁺ cell count: CD₄ > 500, 350 < CD₄ < 500, 200 < CD₄ < 350, and CD₄ < 200 (Fig. 1b). We modeled relative rates of infectiousness by stage as an approximation of viral load [24–26], and rates of HIV-attributable mortality by stage [27].

Fig. 1.

Model structure and transitions.

Low: lowest activity; Med: medium activity; LR/HR: lower/higher risk; FSW: female sex workers; Clients: of FSW; CD4: CD4⁺ T-cell count per mm³; ART: antiretroviral therapy; rates — λ: force of infection; η: HIV progression; δ: diagnosis; τ: ART initiation; σ: viral suppression; σ’: viral re-suppression; ζ: ART failure/discontinuation; not shown: turnover amongst activity groups in (a).

Risk heterogeneity

The model includes eight subpopulations: FSW and their clients at higher vs. lower risk, plus other women and men with low and medium sexual activity (0–1 and 2+ sexual partners in the past year, respectively); and four sexual partnership types: main/spousal, casual, repeat sex work, and one-off sex work (Fig. 1a). We captured risk heterogeneity through: subpopulation sizes, turnover among subpopulations, genital ulcer disease (GUD), partnership formation rates and durations, preferential mixing, sex frequency, and condom use (Table 1). We modelled increasing condom use (Figure B.10), increasing voluntary medical male circumcision (Figure B.9), and decreasing GUD prevalence over time. We did not model nonheterosexual HIV transmission nor preexposure prophylaxis (PrEP), which first became available in Eswatini in 2017 and reached 12% of FSW by 2021 [18].

Table 1.

Selected model parameters related to risk heterogeneity.

		Posterior
Parameter	Stratification	Mean	(95% CI)	Appendixa
Subpopulation size (% of total)	FSW, of women overall	3.4	(2.1, 5.3)	A.3.11.1
	Clients, of men overall	9.3	(4.0, 18.4)	A.3.11.2
	HR of FSW/clients overall	20	—	A.3.9
Duration in subpop (mean years)	FSW overall	3.8	(2.6, 5.0)	A.3.12.2
	Clients overall	8.8	(4.9, 12.7)	A.3.12.2
	HR sex work and clients	0.6	(0.3, 0.9)	A.3.12.2
Sex work clients per year	HR FSW, one-off sex work	82	(58, 104)	A.3.13.1
	LR FSW, one-off sex work	20	(15, 26)	A.3.13.1
	HR FSW, repeat sex work	48	(20, 82)	A.3.13.1
	LR FSW, repeat sex work	11	(5, 20)	A.3.13.1
Sex work visits per year	HR clients	87	(56, 126)	A.3.13.1
	LR clients	43	(27, 59)	A.3.13.1
Total partner change rate per year	Medium activity women	0.97	(0.77, 1.18)	A.3.13.2
	Medium activity men	0.72	(0.20, 1.11)	A.3.13.2
	Lowest activity men & women	0.23	(0.14, 0.33)	A.3.13.2
Any GUD p12m prevalence (%)b	HR FSW	38	(28, 52)	A.3.5.3
	LR FSW & HR clients	29	(22, 37)	A.3.5.3
	LR clients	24	(17, 33)	A.3.5.3
	Medium activity	18	(8, 29)	A.3.5.3
	Lowest activity	7	—	A.3.5.3
Relative infectiousness	Acute infection	9.3	(5.0, 15.8)	A.3.4.1
	Any GUD p12m	2.0	(1.3, 3.0)	A.3.4.5
	On ART before viral suppression	0.30	(0.07, 0.61)	A.3.7.1
	On ART after viral suppression	0	—	A.3.7.1
	Off ART / ART failure	0.65	(0.53, 0.81)	A.3.7.1
Relative susceptibility	Receptive vaginal sex	1.7	(1.2, 2.0)	A.3.4.2
	Receptive anal sex	10.0	—	A.3.4.2
	Any GUD p12m	3.6	(1.4, 6.4)	A.3.4.5
Sex acts per partnership-year	Main/spousal	86	(56, 123)	A.3.14
	Casual	136	(78, 202)	A.3.14
	One-off sex work	12	—	A.3.14
	Repeat sex work	21	(12, 32)	A.3.14
Anal sex acts (% of all acts)	Main/spousal & casual	11	(6, 22)	A.3.14
	One-off & repeat sex work	18	(7, 41)	A.3.14
Condom use in 2020 (% of acts protected)	Main/spousal	42	(34, 50)	A.3.5.2
	Casual	69	(64, 77)	A.3.5.2
	One-off sex work	86	(78, 94)	A.3.5.2
	Repeat sex work	74	(65, 84)	A.3.5.2
	Anal vs. vaginal sex	73	(55, 90)	A.3.5.2
Partnership duration (years)	Main/spousal	16.6	(15.0, 18.5)	A.3.15
	Casual	0.78	(0.48, 1.12)	A.3.15
	One-off sex work	0.08	—	A.3.15
	Repeat sex work	0.43	(0.21, 0.78)	A.3.15

Low: lowest activity; Med: medium activity; LR/HR: lower/higher risk; FSW: female sex workers; Clients: of FSW; p12m: past 12 months.

^a http://links.lww.com/QAD/D726

^bGUD prevalence declines universally after 2010 as described in A.3.5.3.

Antiretroviral therapy cascade

The modeled ART cascade includes states for undiagnosed HIV, diagnosis before ART, ART before viral suppression, viral suppression, and a generic ART failure/discontinuation state (Fig. 1c). HIV stage-specific infectiousness is then modelled as reduced/zero on ART before/after viral suppression, and modestly reduced during ART failure/discontinuation. We modeled rates of HIV diagnosis among people living with HIV as monotonically increasing over time. We defined a base rate for women with low/medium sexual activity, and fixed relative rates for men (RR < 1) and FSW (RR > 1), reflecting increased HIV testing access via antenatal care among women vs. men, and enhanced screening among FSW [22]. We modeled ART initiation, starting in 2003, similarly except the relative rate for ART initiation among FSW was RR < 1, reflecting barriers to uptake and engagement in care;[11] we defined additional relative rates by CD4⁺ cell count (0 ≤ RR ≤ 1) to reflect historical ART eligibility criteria (Figure A.5) [21]. We modeled viral suppression 4 months after ART initiation, on average, for all subpopulations [28]. We modeled treatment failure/discontinuation with a single monotonically decreasing rate for all subpopulations in the base-case, reflecting improving treatment success/retention over time [21]. Individuals with treatment failure/discontinuation could re-initiate ART at a fixed rate, reflecting re-engagement in care or detection of treatment failure and initiation of alternate regimens. We modeled rapid CD4⁺ recovery during the first 4 months of ART, followed by slower recovery while virally suppressed [29]. We modeled reduced HIV-attributable mortality among individuals on ART, in addition to mortality benefits of CD4⁺ recovery.

Calibration

We calibrated the model to available data from Eswatini on HIV prevalence, HIV incidence, and ART cascade of care, overall and stratified by subpopulation where possible (Tables A.6–A.10) [17–23]. We used an adapted version of Incremental Mixture Importance Sampling (IMIS) [30], yielding 1000 total model fits. Appendices A.4 and B.1 give full methodology and calibration results.

Scenarios and analysis

Objective 1: Influence of antiretroviral therapy cascade differences between subpopulations

For Objective 1, we defined the base-case scenario to reflect observed ART cascade scale-up in Eswatini, reaching 95-95-95 for the population overall by 2020 [17], and 88-98-(91) among FSW specifically [18].ⁱ

Next, we defined four counterfactual scenarios in which the overall population cascade reached 80-80-90 by 2020, and where FSW, clients, both, or neither were disproportionately left behind. We chose 80-80-90 to reflect potential attainment if investment in cascade scale-up had been slower, as in some other contexts [3]. In counterfactual scenarios, we altered cascade attainment among FSW, clients, and/or the remaining population (“all others”) by calibrating fixed subpopulation-specific relative rates of: diagnosis (0 ≤ RR_d ≤ 1), treatment initiation (0 ≤ RR_t ≤ 1), and treatment failure/discontinuation (0 ≤ RR_u ≤ 1). When FSW and/or clients were left behind, we calibrated their RRs such that these subpopulations attained approximately 60-40-80 by 2020, reflecting some of the lowest cascades recently observed among key populations [10]. When FSW and/or clients were not left behind, we fixed their RR = 1, reflecting maintenance of prioritized outreach/services from the base-case. In all counterfactual scenarios, we calibrated RRs for the remaining population such that the Swati population overall attained 80-80-90 in all four counterfactual scenarios, thus ensuring that a consistent proportion of the population overall attained viral suppression. Table 2 summarizes these scenarios, while Figure B.17 plots the modeled cascades over time.

Table 2.

Modelling scenarios for Objective 1 defined by 2020 calibration targets.

	ART cascade in 2020a			Re-scaled cascade ratesb
Scenarios	FSW	Clients	Overall	FSW	Clients	All Others
Base Case	88-98-(91)	(76-94-95)	95-95-95	—	—	—
Leave Behind: FSW	60-40-80	(68-93-92)	80-80-90	Yes	No	Yes
Leave Behind: Clients	(79-92-88)	60-40-80	80-80-90	No	Yes	Yes
Leave Behind: FSW & Clients	60-40-80	60-40-80	80-80-90	Yes	Yes	Yes
Leave Behind: Neither	(82-92-89)	(71-93-93)	80-80-90	No	No	Yes

^aCascade: % diagnosed among people living with HIV; % on ART among diagnosed; % virally suppressed among on ART; (*): median from model fits reported given no available data.

^bRates of: diagnosis; ART initiation; treatment failure; FSW: female sex worker; Clients: of FSW; All Others: all women and men not involved in sex work. Figure B.17 plots the modeled cascades over time.

We quantified ART prevention impacts via two outcomes: relative cumulative additional infections and additional incidence rate by 2020 in the counterfactual scenarios (k) vs. the base-case (0):

$$\begin{array}{l}\text{CAI},\text{AIR}(t)=\frac{{\Omega }_k\left(t\right)-{\Omega }_o\left(t\right)}{{\Omega }_o\left(t\right)},\\ \Omega \left(t\right)=\left\{\begin{array}{ll}{\int }_{t_0}^t\Lambda (\tau )d\tau \hfill & \text{CAI}:\text{cumulative additional infections}\hfill \\ \lambda \left(t\right)\hfill & \text{AIR}:\text{additional incidence rate}\hfill \end{array}\right.\end{array}$$ (1)

where $\Lambda \left(t\right)$ denotes absolute numbers of infections per year, λ(t) denotes incidence rate per susceptible per year, and t₀ = 2000. For each scenario, we report median (95% credible interval, CI) values across model fits, reflecting uncertainty.

Objective 2: Conditions that maximize the influence of antiretroviral therapy cascade differences

For Objective 2, we estimated via linear regression: the effects of lower ART cascade among FSW and clients on each outcome, plus potential effect modification by epidemic conditions. The hypothesized causal effects are illustrated as a directed acyclic graph in Figure B.20. Appendix B.2.2 explains how this approach aims to improve upon conventional sensitivity analyses for HIV transmission models.

For this regression, we generated 10 ,000 synthetic samples as follows. We explored a wider range of counterfactual scenarios vs. Objective 1 by randomly sampling relative rates (RR) for diagnosis and treatment initiation RR_d, RR_t ~ Beta with 95% CI (0.25–0.95), and treatment failure/discontinuation RR_u ~ Gamma with 95% CI (1.5–15) among: FSW, clients, and the remaining population (nine total values), yielding cascades in 2020 spanning approximately 60-60-90 through 85-90-95 (Figure B.19). For each of N_f = 1000 model fits, we generated N_k = 10 counterfactual scenarios per fit using Latin hypercube sampling of RRs[31], yielding N_fN_k = 10, 000 total counterfactual scenario samples.

For each sample, we defined each outcome as in Eq. (1). We further defined U_fki for subpopulations i as the proportions not virally suppressed among those living with HIV by 2020, reflecting a summary measure of ART cascade gaps. Using U_fki, we defined the main regression predictors as: D_fk = U_fk - U_fo > 0, reflecting differences in population overall viral nonsuppression in sample k vs. the base-case (denoted k = 0); and d_fki = U_fki - U_fk*, reflecting differences in subpopulation-i-specific viral nonsuppression in sample k vs. the population overall in sample k, that is, viral nonsuppression inequalities.

Next, we defined the following measures of epidemic conditions (C_fj), as hypothesized modifiers of the effect of unequal viral nonsuppression on our outcomes: FSW and client population sizes (% of population overall); average turnover rate among FSW and clients (reciprocal of duration selling/buying sex); and HIV incidence ratios in 2010 among FSW vs. women overall, and among clients vs. men overall. We used HIV incidence ratios to reflect summary measures of risk heterogeneity, rather than including all risk factors from the transmission model, which could lead to improper inference due to effect mediation.

Finally, we defined a general linear model for each outcome as:

$$\text{CAI},\text{AIR}={\beta }_0+{\beta }_{1}D+\sum _i{\beta }_i{d}_i+\sum _{ij}{\beta }_{ij}{d}_i{C}_j$$ (2)

such that each outcome was modeled as a liner sum of an intercept plus effects of: differential population-level nonsuppression in the counterfactual vs. the base-case scenario (D); differential nonsuppression among FSW and clients vs. the population overall within the counterfactual scenario (d_i); and effect modification of d_i by epidemic conditions (C_j). We fitted this model for each outcome using generalized estimating equations [32] to control for repeated use of model fits. We standardized all model variables (D, d_i, C_j) via $\stackrel{\circ }{x}=\left(x-\text{mean}\left(x\right)\right)/\text{SD}\left(x\right)$ to homogenize variable scales and reduce collinearity in interaction terms. Effect sizes can thus be interpreted as the expected change in outcome per standard deviation change in the variable.ⁱⁱ We verified the appropriateness of a linear link by examining residuals (Figures B.21).

Results

We first summarize modeled patterns of HIV transmission in the base-case, calibrated to reflect the Eswatini epidemic up to 2021. Our model suggests that transmission within repeat sex work partnerships was a dominant driver of early epidemic growth (Figures B.14–B.15). However, from approximately 1994 onward, the majority of yearly infections were transmitted within casual partnerships, including 50% (median) in 2020.ⁱⁱⁱ

Overall HIV prevalence in 2020 was median (95% CI): 23.8 (22.4–24.7) (Figure B.4a), and overall incidence was 6.6 (5.3–7.6) per 1000 person-years (Figure B.4c). The prevalence ratio between FSW and women overall was 1.78 (1.70–1.87), and between clients and men overall it was 1.92 (1.49–2.49) (Figure B.4b). Due to turnover and higher HIV incidence among FSW, achieving similar rates of diagnosis among FSW vs. other women (Figure B.6a) required approximately twice the rate of testing. Sex work contributed a growing proportion of infections over 2010–2020: from 17% to 32% (Figure B.14).

Objective 1: Influence of cascade differences between subpopulations

Figure B.16 illustrates ART cascade attainment over time in the base-case (95-95-95 overall by 2020) and each of the four counterfactual scenarios (80-80-90 overall by 2020), while Figure B.17 illustrates overall HIV incidence in each scenario. Figure 2 then illustrates cumulative additional infections and additional incidence rate in each counterfactual scenario vs. the base-case. If ART scale-up in Eswatini had been slower but relatively equal, we estimate there would have been 30,700 (22,200–38,800) additional infections by 2020, 8.8 (6.3–10.9)% more than the base-case. By contrast, if ART scale-up up had been slower and disproportionately left behind FSW and clients, we estimate there would have been 49,900 (39,200–62,700) additional infections by 2020, 14.3 (10.8–18.6)% more than the base-case and a 63 (31–128)% increase over the neither left behind scenario. Leaving behind either FSW or clients resulted in similar 10.9 (8.4–13.3)% or 12.4 (9.8–14.6)% additional infections vs. the base-case, respectively. Results were similar for overall incidence rate in 2020: 13 (11–16) vs. 19 (15–22) infections per 1000 person-years in the neither left behind vs. FSW and clients left behind scenarios, 106 (68–138)% vs. 193 (152–235)% higher than the base-case, respectively. Table B.1 summarizes Objective 1 results numerically. In all counterfactual scenarios, the majority of additional infections were transmitted via casual partnerships (Figure B.18a) and acquired among non-FSW women (Figure B.18c). Patterns of onward transmission were also similar across scenarios (Figure B.18b), though subpopulation contributions increased if they were left behind.

Fig. 2.

Modeled outcomes under counterfactual scenarios vs. the base case: cumulative additional infections and additional incidence rate.

Base case: 95-95-95 by 2020; counterfactual scenarios: 80-80-90 overall by 2020, with reduced cascade (60-40-80: left behind) among FSW, clients of FSW, both, or neither; whiskers, boxes, and midlines: 95% CI, 50% CI, median of model fits; see Eq. (1) for outcome definitions.

Objective 2: Conditions that maximize the influence of cascade differences

The regression analyses indicated that population-overall viral nonsuppression (D) and relative nonsuppression among FSW and clients (d_i) each had strong and positive effects on both cumulative additional infections and additional incidence rate (P <  10⁻⁵), corroborating the results of Objective 1 (Fig. 3). The effect of nonsuppression among FSW increased with FSW population size for both outcomes, and with client population size and HIV incidence ratio vs. men overall for incidence rate specifically. The effect of nonsuppression among clients increased with client population size and incidence ratio for both outcomes. Durations buying or selling sex did not appear to modify the impact of nonsuppression among either FSW or clients, and neither did HIV incidence ratio among FSW vs. women overall.

Fig. 3.

Estimated effects on relative additional infections of disproportionate viral nonsuppression (d) among FSW and clients vs. population overall, plus effect modification by epidemic conditions.

(a) Cumulative additional infections, (b) additional incidence rate by 2020 vs. base case; FSW: female sex workers; Clients: of FSW; IR: incidence ratio in 2020; d_i: difference in subpopulation-i-specific viral nonsuppression vs. population overall within counterfactual scenario; points and error bars: mean and 95% CI for each effect estimated via Eq. (2).

Discussion

The available data suggest that Eswatini not only achieved 95-95-95, but also minimized ART cascade inequalities across subpopulations while doing so [17,18]. Relative to this achievement, we estimated that slower but equitable ART scale-up, reaching only 80-80-90 by 2020, would have resulted in 6.3–10.9% more cumulative infections and 68–138% higher incidence in 2020. However, we also estimated that slower scale-up which disproportionately left behind FSW and their clients would have led to 31–128% more infections than slower scale-up alone. Perhaps unsurprisingly, the impact of leaving behind FSW and/or clients was largely determined by their population sizes and HIV incidence ratio among clients vs. men overall.

Eswatini surpassed 95-95-95 through numerous initiatives coordinated across sectors, including those led by the MaxART program [34,35]. Multiple stakeholders, including people with HIV, health care providers, traditional and religious leaders, community groups, and researchers were engaged via multiple channels, such as Technical Working Groups, Community Advisory Boards, and specific meetings for prioritized groups (men and adolescents) [34,35]. Drawing on this engagement and social science research to understand barriers to care, cascade services were comprehensively strengthened via investments in training, infrastructure, antistigma communication, demand creation, and monitoring [34,35].

Among FSW living with HIV in Eswatini, data suggest that 88% were diagnosed and 86% were on ART in 2021 (i.e., 98% ART coverage among those diagnosed) [18]. Data on viral suppression among FSW were lacking, and so assumed to be similar to other women in our model. Although lower than 95-95-95, this cascade among FSW is higher than in many other regions [10,13]. Considering that women enter and exit sex work (turnover) and likely experience highest risk of HIV acquisition during sex work, programs must achieve higher rates of HIV testing, ART initiation, and retention among FSW vs. other women to achieve similar cascades. For example, we inferred that rates of HIV testing in 2016 were 80–227% higher among FSW vs. other women to reproduce observed cascade data during model calibration.

In Eswatini, programs for key populations have included well tolerated access to tailored services via drop-in centers (locally known as TRUE), mobile outreach, venue-based, and one-on-one options [18]. Health and clinical services were also integrated with efforts to reduce structural vulnerabilities, including experiences of harassment, violence, and fear of seeking health care, through community empowerment, psycho-social and legal supports, and sensitization and training for police and health care workers [18]. These programs have been designed and refined with ongoing community leadership and engagement, allowing them to better meet the specific needs of key populations, for whom barriers to engagement in HIV care often intersect with drivers of HIV risk, including economic insecurity, mobility, stigma, discrimination, and criminalization [7,11–14,36]. Our data-informed modeling of cascade scale-up in Eswatini confirms that such an equity-focused approach to ART cascade scale-up can maximize prevention impacts, and accelerate overall reductions in HIV incidence. By contrast, it is alarming that recent funding disruptions to USAID and PEPFAR will disproportionately impact precisely these efforts, thereby threatening to erase decades of progress [15,37].

Our study highlights the importance of reaching both FSW and their clients, echoing recent modelling of South Africa and Cameroon [38,39], which found that gaps in HIV prevention and treatment for clients were among the largest contributors to onward transmission. Programmatically, such findings reiterate the need for improved data on both FSW and clients, including estimates of population size, sexual behavior, and ART cascade attainment, which can help prioritize settings and services to reduce inequalities. These estimates may be difficult to obtain because individuals are unlikely to report buying or selling sex in population-level surveys due to stigma and criminalization and because many clients are highly mobile (including transient seasonal/occupational migration) [36,40].

While numerous modeling studies have examined the potential prevention impacts of ART cascade scale-up (Appendix B.3) [5], our study is the first to explore the impact of inequalities in ART cascade across subpopulations with consistent population overall cascade across scenarios. Similar work by Marukutira et al. [41] illustrated the limited impact of achieving 95-95-95 for only citizens and not immigrants in Botswana, while Maheu-Giroux et al. [42]. illustrated the high cost-effectiveness of prioritizing key populations (including clients) for ART in Côte d’Ivoire. Thus, our findings are likely generalizable to other epidemic contexts. HIV prevalence ratios between key populations and the population overall are relatively low in Eswatini vs. elsewhere [43,44]; so, the impact of cascade inequalities among key populations in other contexts would likely be even greater than we found for Eswatini. Moreover, as HIV incidence declines in many settings, transmissions may become concentrated among key populations [45,46], further magnifying the impact of cascade inequalities. Indeed, we estimated a growing proportion of transmission via sex work partnerships in Eswatini going forward, as compared to casual or main/spousal partnerships. Similarly, the small number of growing epidemics globally also remain concentrated among vulnerable populations [3].

A primary strength of our analysis is the use of observed ART cascade scale-up to 95-95-95 in Eswatini as the base-case, with plausible cascade inequalities explored in counterfactual scenarios. As noted above, the available data suggest that Eswatini has minimized cascade inequalities which persist elsewhere [10]. Thus, our counterfactual scenarios directly estimate the consequences of failing to address these inequalities. Second, drawing on our conceptual framework for risk heterogeneity [5] and multiple sources of context-specific data [17–23], we captured several dimensions of risk heterogeneity, including heterosexual anal sex, four types of sexual partnerships, sub-stratification of FSW and clients into higher/lower risk strata, and subpopulation turnover (Table 1, Appendix A). Accurate modeling of risk heterogeneity has been shown to mediate model-estimated ART prevention impacts [47], and is especially important when considering differential ART scale-up across subpopulations. Finally, our analytic approach to Objective 2, in which epidemic conditions are conceptualized as potential effect modifiers represents a unique methodological contribution to the HIV modeling literature.

Our study also has limitations. First, we did not model PrEP. However, our analyses focus on the time period prior to widespread PrEP availability in Eswatini [18]. Second, we did not consider transmitted drug resistance (TDR). However, drug resistance is more likely to emerge in the context of barriers to viral suppression [48]; thus, lower cascade among those at higher risk would likely accelerate emergence of transmitted drug resistance, and thereby magnify our findings. Finally, our model structure did not include age, and we only considered heterosexual HIV transmission in Eswatini. Future work can explore adaptation of the model to consider PrEP, TDR, age stratification, and additional modes of HIV transmission. While the magnitude of our results may change with such adaptations, we do not expect that the qualitative interpretation would change. In fact, our findings would likely generalize to other transmission networks and determinants of risk heterogeneity, including other key populations and subpopulations such as highly mobile populations and young women [36,49].

In conclusion, the HIV response must remain rooted in context-specific understandings of inequalities in HIV risk and in access to HIV services, which often stem from common upstream factors. Thus, differences in ART cascade within and between subpopulations at higher risk of HIV must be monitored, characterized, and addressed to fully realize the anticipated benefits of ART at both the individual and population levels.

Acknowledgements

In memory of Zandile Mnisi.

The authors thank Kristy Yiu, Samantha Lo (Unity Health Toronto) for research coordination support; Amrita Rao, Carly Comins (Johns Hopkins University), Alex Whitlock, Korryn Bodner (Unity Health Toronto), and Leigh Johnson (University of Cape Town) for helpful discussions and feedback on model design.

The study was supported by the Natural Sciences and Engineering Research Council of Canada (CGS-D); the Ontario Ministry of Colleges and Universities (QEII-GSST); the Canadian Institutes of Health Research (FN-13455); the National Institute of Allergy and Infectious Diseases (R01AI170249).

Conceptualization: J.K., L.W., S.M.; formal analysis: J.K., H.M., L.W.; investigation: all authors; methodology: J.K., H.M., B.S., L.K., L.W., M.E., S.B., S.M.; project administration: J.K., H.M., L.W., R.K., M.E., S.B., S.M.; software: J.K.; supervision: R.K., M.E., S.B., S.M.; validation: J.K., S.M.; visualization: J.K.; original draft: J.K., S.M.; review & editing: all authors. All authors except Z.M. have read and approved the final manuscript.

The authors used only published aggregate data, except for individual-level data from two female sex worker surveys, which were accessed under approval from the Scientific and Ethics Committee of Eswatini Ministry of Health (MH/599B), and the Institutional Review Board of the Johns Hopkins Bloomberg School of Public Health (3508). All code and selected results are available at: github.com/mishra-lab/hiv-model-eswatini

Preprint: medRxiv: doi.org/10.1101/2024.02.16.24302584

Conflicts of interest

There are no conflicts of interest.