Defining strata of women at risk for HIV in sub-Saharan Africa: A pooled latent class analysis

Abstract

Purpose

Community-led programs have expanded sexual and reproductive health services for female sex workers (FSW) across sub-Saharan Africa. However, substantial heterogeneity in HIV risk and engagement with prevention services remains, underscoring the need for differentiated approaches to identify FSW who remain underserved by HIV programs.

Methods

We pooled 2013–2019 bio-behavioral survey data from nine countries in Sub-Sahara Africa. Latent class analysis defined underlying risk indicated by sexual behavior, client type, alcohol use, and violence. We profiled characteristics of classes and used mixture models to estimate adjusted odds ratios (aOR) and 95% confidence intervals (CI) for associations of class membership with prior HIV testing and biologically-confirmed HIV status.

Results

Among 5548 FSW (median age: 27, IQR 23–34), four classes were identified: Alcohol and Violence (16.8%), High Client Turnover (38.4%), Lower Client Volume (31.8%), and Limited Condoms and Lubricant Use (13.0%). Relative to other groups, FSW in Lower Client Volume were older and less likely to experience sex work interruptions or share earnings. Prior HIV testing was lower among the Limited Condoms and Lubricant Use class relative to others: Alcohol and Violence (aOR 4.99, 95%CI 3.44–7.24), High Client Turnover (aOR 4.09, 95%CI 3.19–5.24), and Lower Client Volume (aOR 3.58, 95%CI 2.80–4.59).

Conclusions

Pooled analyses demonstrated heterogeneous individual and structural influences on HIV prevention among FSW. Increased specificity in implementation of HIV services for FSW may increase efficiency to achieve more with less in the HIV response.

Introduction

Female sex workers (FSW) in countries across sub-Saharan Africa remain affected by high HIV incidence and prevalence threatening overall health and wellbeing. HIV prevalence estimates among FSW differ across countries and contexts, ranging from 7.5% in countries across West and Central Africa to more than 60% in countries across Southern Africa [1–3]. This heterogeneity in HIV burden is driven by a combination of factors, including differences in underlying HIV epidemic dynamics, prevalent HIV clades, socio-economic environments, and health systems capacity across sub-Saharan Africa [3].

HIV acquisition and transmission dynamics among FSW are shaped by interacting determinants across multiple social and structural levels [4, 5]. At a structural level, FSW face social and systems-level stigmas including criminalization of sex work [5], police harassment, and denial of safe workspaces in some contexts [6]. Shared social determinants of HIV acquisition among FSW across settings include economic vulnerability, stigma and discrimination, gender inequities, and violence [7, 8]. These upstream determinants are also known to result in adverse mental health outcomes including stress, depression, anxiety, trauma, which in turn precipitate substance use and inconsistent condom use with sexual partners [5, 9]. Additionally, these determinants also effect service delivery by preventing FSW from accessing critical and lifesaving health services [10], particularly in settings where sex work remains criminalized [11]. While these structural and social determinants are critical drivers of HIV vulnerability, they are often not directly measurable in routine data systems used for program planning. These upstream factors frequently manifest through observable behaviors that more proximally shape HIV acquisition risks and service engagement. This is consistent with previous frameworks such as the modified social ecological model which conceptualizes individual-level behaviors as downstream manifestations of broader structural and social risk environments [4]. As such, clustering of individual-level behaviors may provide a practical way to identify subgroups of FSW with distinct combinations of exposure risk, structural vulnerability, and unmet prevention or treatment needs, thereby informing differentiated HIV service delivery strategies.

HIV risks are also not consistent for all FSW and vary by factors such as age and economic circumstance, precipitating different support needs. Young FSW under the age of 25 have among the highest incidence of HIV relative to older counterparts [12]. Additionally, antiretroviral therapy (ART) coverage for young FSW remains limited across contexts, with people aged 18–19 least engaged in HIV-related services relative to other age groups [13]. Women who engage in sex work but who do not identify as sex workers are also less likely to access HIV programs compared to other women [14].[8] In the context of routine healthcare services, targeted assessments of HIV risk could inform programmatic priorities and additional opportunities for intervention [15]. For example, FSW in high-prevalence HIV settings have been shown to benefit from more frequent HIV screening through network-based and self-testing approaches, promoting more effective linkage to ART services [16]. Meanwhile, in contexts where sex workers face severe stigma or violence, support systems such as peer networks, legal protection, and mental health care remain essential health needs [17]. Further, for FSW in remote regions, there is heightened demand for mobile health services to overcome barriers posed by geographic isolation [18]. Similarly, FSW who report dependencies or other substance use could benefit from harm reduction strategies and specialized counseling support [19].

Community-led HIV programs have been effective in reaching FSW with comprehensive sexual and reproductive health (SRH) services and addressing social and structural determinants of HIV [6]. In the prevention space, community-led interventions have improved condom usage, promoted regular HIV testing and pre-exposure prophylaxis (PrEP) uptake [20]. Community-based ART programs for FSW living with HIV have also resulted in an increase in ART uptake, retention, and viral suppression when compared to facility-based ART [21, 22]. These community-led intervention models are delivered through strategies such as community mobilization, peer-mediated outreach and differentiated service delivery, which broadly have been shown to increase linkage to healthcare services for FSW [23]. However, amidst dramatic decreases in funding and resources for community-based HIV services and sustained or even increasing HIV infections [24, 25], there is a need to characterize the FSW who may benefit the most from more intensive HIV prevention and treatment programming.

Diverse HIV service delivery needs among FSW can be in part attributed to varying latent dimensions of risk, hard-to-observe dimensions inferred from patterns in behaviors that create highly variable barriers to health services [26]. Identifying population subgroups who share similar latent constructs can inform differentiated services that address the barriers that limit engaged in HIV prevention and treatment [20]. By modeling the probability that individuals belong to underlying, unobserved classes, latent class analysis (LCA) can capture the multidimensional and underlying latent factors that shape heterogeneity in risk profiles and service needs [27]. Guided by the modified social ecological model, we focus on key proximal individual-level domains through which structural vulnerability is expressed among FSW. First, sexual behavior, including frequency of sex work and client volume, directly determines exposure to HIV. Second, prevention practices, such as condom use and access to lubricant, reflect both individual agency and structural constraints on safer sex. Third, substance use, particularly alcohol consumption, has been linked to impaired risk perception and reduced condom negotiation. Finally, experiences of physical and sexual violence represent critical manifestations of structural vulnerability that directly increase HIV risk and limit access to services. Together, these domains represent measurable and comparable indicators across settings that capture how broader structural determinants translate into individual-level HIV risk. This makes it possible to move beyond one-size-fits-intervention strategies towards increasingly well-specified strategies addressing barriers faced by distinct subgroups of FSW, particularly in the context of decreasing resources[28].

The primary objective of this analysis is to identify underlying constructs of HIV-associated risk, and their associations with HIV-related outcomes, among FSW across countries in sub-Saharan Africa. Secondarily, we sought to determine social and structural factors associated with these risk strata, and assess the extent to which their composition differs across countries in sub-Saharan Africa.

Methods

Data collection and participants

This study used pooled data from nine cross-sectional bio-behavioral studies conducted between 2013 and 2019 in Burkina Faso, Cameroon, Côte d’Ivoire, Gambia, Guinea-Bissau, Lesotho, South Africa, Senegal, and Togo. Local study teams led primary data collection within each country. Eligibility criteria varied by country, but participants generally included cis-gender women who identified as sex workers aged 18 and above. Full eligibility criteria are reported in the Supplemental Table A. All recruitment was conducted via respondent-driven sampling (RDS). Recruitment chains were initiated at each site by seeds, or individuals selected in collaboration with local community-based organizations to reflect heterogeneity in demographic characteristics and geographic representation.

Trained study staff administered questionnaires in private locations with all participants in the language of the participant’s preference. Testing for HIV, including pre- and post-test counseling, was conducted following country-specific national guidelines. HIV testing was conducted prior to administering the socio-behavioral questionnaires. Study staff led post-test counseling and review of HIV test results with participants after completion of the questionnaire.

All participants provided either oral or written informed consent. Local ethical approval was obtained for each country. This analysis utilized de-identified data and was determined to be non-human subjects research by the Johns Hopkins Bloomberg School of Public Health Institutional Review Board (IRB0007442).

Measures

Measurement model inputs and outcomes

Nine total variables were included in the LCA measurement model to identify clustering of potential drivers of HIV risk, hereafter also referred to as risk strata. Consistent with the framework described in the Introduction, variables were selected a priori to represent four proximal individual-level domains: sexual behavior (frequency of sex work, number of new clients, and number of regular clients), prevention practices (condom use with new clients, condom use with regular clients, and lubricant use), substance use (alcohol use), and violence (physical and sexual violence). All were measured by self-report and are described in supplement appendix.

The primary outcome assessed through mixture modeling was biologically confirmed HIV status as living or not living with HIV based on test results. HIV testing followed a screening algorithm whereby an initial determination of HIV status was made (e.g., Determine HIV-1/2 test, Alere International Limited, Ballybrit, Ireland) followed by confirmatory testing (e.g., STATPAK HIV 1/2 test, Chembio Diagnostic). A secondary outcome of prior engagement in HIV testing was measured via self-report (yes/no).

Covariates

Broader social and structural factors, including sociodemographic characteristics and stigma items, were modeled separately as covariates in predictive models to evaluate their associations with latent class membership. All measures are described in detail in the supplemental appendix.

Statistical Analyses

We generated descriptive statistics to summarize the characteristics of the pooled sample. We evaluated item-level missingness for the latent class indicators and other measures; missingness was <5% for all items and we used full information maximum likelihood estimation to handle missingness [27]. To describe regional heterogeneity in the pooled sample, we summarized participant characteristics by region in the Supplementary Material. Chi-square tests were used to compare proportions for categorical variables, and t-tests were used to compare means for continuous variables, with statistical significance assessed at alpha = 0.05. All pooling and descriptive statistics were generated using RStudio (RStudio Team (2020). RStudio: Integrated Development for R. RStudio, PBC, Boston, MA) [29].

Latent Class Analysis

We conducted LCA, first specifying a measurement model that included the nine selected indicators of HIV risk. We considered up to five latent classes for FSW, and examined model diagnostics including Bayesian Information Criterion (BIC), Akaike Information Criterion (AIC), entropy, Vuong-Lo-Mendell-Rubin adjusted likelihood ratio test, bootstrap likelihood ratio test (LRT), and chi-squared p-values (see Supplemental Table B for detailed model fit statistics). A best-fitting model was selected based on model diagnostics (e.g., smallest values for AIC, BIC, entropy), overall class prevalences, and interpretability. To avoid violation of the conditional independence assumption among LCA indicators, we included women who reported at least one new and casual client, thus reducing the conditional dependence of condom use with new or casual clients on reported new and regular client volumes. Classes from the selected best-fitting model were descriptively named based on the estimated conditional probabilities for the nine included items. Individual-level characteristics, including known determinants of HIV risk, were descriptively profiled across these classes or risk strata. Sensitivity to country-level data was assessed in leave-one out analysis, whereby classes were iteratively re-estimated for n-1 countries to understand the extent to which the selected number of classes were driven by data from a single country.

Multinomial Logistic Regression

We used a maximum probability assignment rule whereby FSW were assigned to classes based on their highest LCA posterior probability of class membership. We then fit adjusted multinomial logistic regression models with assigned latent class membership as the outcome to evaluate predictors of class membership. Covariates were selected based on the modified social ecological model to capture social and structural factors potentially associated with latent class membership [4]. These predictors were not used to define the latent classes themselves but were included to characterize social and structural factors associated with class membership. Predictors included age, education, employment, childhood urban residence, number of biological children, pregnancies, unwanted pregnancies, financial dependents, shared earnings, sex work interruption, family stigma, healthcare stigma, and legal stigma. Beyond profiling individual characteristics across latent classes, these models were used to examine how social and structural factors differed across classes and to identify potential mechanisms of vulnerability relevant to resource allocation and program planning.

Mixture Modeling

Two distal outcomes were assessed: biologically confirmed HIV status and prior HIV testing. Biologically confirmed HIV status was considered the primary outcome, and prior HIV testing was considered the secondary outcome. To explore the association of group membership with these distal outcomes, we fit Bolck, Croon, and Hagenaars (BCH) weighted mixture models for each population [30]. The BCH method outperforms other three-step approaches for evaluating distal outcomes by applying weights to account for measurement error in latent class estimation and avoiding shifts in latent classes. We first used the BCH method to calculate the prevalence of each outcome by class. Covariates for inclusion as potential confounders in mixture models included age, education, and country of residence. We estimated adjusted odds ratios (aOR) and 95% confidence intervals (CI) to assess the association between class membership and each outcome. Mixture models were performed in Mplus Version 8 (Muthén & Muthén, Los Angeles, CA).

Results

Data comprised a total of 5,548 observations for FSW, with Cameroon being the country with the most participants (N=1717).

Sample characteristics

Demographic, behavioral, socio-structural, and clinical characteristics are presented in Table 1. FSW in the overall sample (n=5,548) were predominantly from West Africa (84.3%), with 40.1% aged 30 years or older and 84.1% having less than secondary education. Close to 30% of FSW reported living near an urban center during childhood (28.9%). A majority were unemployed (58.5%) and 34.7% reported binge drinking in the past 30 days. More than 80% reported consistent condom use with new or regular clients. HIV prevalence was 20.2% overall, and 78.7% had previously tested for HIV.

Table 1.

Characteristics among Female Sex Worker in 9 countries across Sub-Saharan Africa (n=5548)

Characteristics	Total sample (n=5548)
	No. (%)
Age
18–24	2002 (36.1)
25–29	1311 (23.6)
30+	2226 (40.1)
Missing	9 (0.2)
Number of Biological Children (Mean, SD†)	(1.7, 3.6)
Highest level of education
Less than secondary school	4665(84.1)
Secondary school or more	883(15.9)
Live near urban center during childhood
yes	1603 (28.9)
no	3936 (70.9%)
Missing	9 (0.2)
Employment
Employed	1370 (24.7)
Unemployed	3246 (58.5)
Other	367 (8.3)
Alcohol use(last 30 days)
No alcohol	1828 (32.9)
1–5 drinks in one day	1383 (24.9)
>5 drinks in one day	1923 (34.7)
Missing	414 (7.5)
Sex Frequency
Low	1805 (32.5)
Medium	1804 (32.5)
High	1804 (32.5)
Missing	135 (2.4)
Number of new clients
Low (< 4 partners)	1782 (32.1)
Medium (4 – 10 partners)	1781 (32.1)
High (>10 partners)	1781 (32.1)
Missing	204 (3.7)
Number of regular clients
Low (< 3 partners)	1807 (32.6)
Medium (3 – 8 partners)	1807 (32.6)
High (>8 partners)	1807 (32.6)
Missing	127 (2.3)
Condom Use with new clients
Inconsistent condom use	717 (12.9)
Consistent condom use	4831 (87.1)
Condom Use with regular clients
Inconsistent condom use	876 (15.8)
Consistent condom use	4672 (84.2)
Lubricant Use
No	2873 (51.8)
Yes	2674 (48.2)
Missing	1 (0.02)
Sexual Violence
No	3430 (61.8)
Yes	1663 (30.0)
Missing	455 (8.2)
Physical Violence
No	3426 (61.8)
Yes	1731 (31.2)
Missing	391 (7.05)
Sex work interruption
No	2103 (37.9)
Yes	3293 (59.4)
Missing	152 (2.7)
Number of pregnancies (Mean, SD) ‡	(2.2, 3.3)
Number of unwanted pregnancies (Mean, SD) §	(2.3, 2.2)
Financial Dependents (Mean, SD)	(3.6, 4.9)
Shared earnings
No	4281 (77.2)
Yes	1256 (22.6)
Missing	11 (0.2)
Stigma
Family stigma (Mean, SD)	(0.4, 0.8)
Health stigma (Mean, SD)	(0.2, 0.6)
Legal stigma (Mean, SD)¶	(0.7, 0.9)
HIV testing
No	997 (18.0)
Yes	4368 (78.7)
Missing	183 (3.3)
HIV Status
Positive	1120 (20.2)
Negative	4399 (79.3)
Missing	29 (0.5)

^†SD = standard deviation, >5% missing reported for continuous variable

^‡341 missing total

^§710 missing total

^¶320 missing total

Regional comparisons of participant characteristics are presented in the Supporting Information.

LCA: Measurement model

Supplementary Table B presents fit indices for 2–5 class models. A four-class model was preferred and selected as a final best-fitting model because of its fit indices (AIC: 89,275, BIC: 89,719; entropy: 0.648, p< 0.001) and interpretable differences in item-response probabilities between classes. Class estimations were generally similar in “leave one out analysis” (Supplemental Table C).

Table 2 presents the posterior class membership probabilities and item response probabilities, for the 4-class model. Class 1 (17% sample prevalence) was named the “Alcohol and Violence” class because it was characterized by a high frequency of sex work with a high volume of both new or regular clients, consistent condom use, lifetime lubricant use and having experienced both physical and sexual violence. This class reflects substantial structural vulnerability and heightened HIV risk related to violence and alcohol use in the presence of other sexual risk behaviors. Class 2 (34% sample prevalence) was named the “High Client Turnover ” class based on a medium frequency of sex work with a high volume of both new and regular clients, consistent condom use, and lifetime lubricant use. This class suggests highest HIV risk pattern primarily driven by high client volume and frequent client turnover. Class 3 (32% sample prevalence) was named “Lower Client Volume,” and was characterized by a low frequency of sex work with a low volume of both new and regular clients, consistent condom use, and generally no lifetime lubricant use. Relative to the other classes, this class reflects comparatively lower exposure-driven HIV risk and fewer overlapping behavioral and structural vulnerabilities. Lastly, Class 4 (13% sample prevalence) was named “Limited Condoms and Lubricant Use ” based on a low volume of new clients, a medium volume of regular clients, inconsistent condom use, and no lifetime lubricant use. This class reflects important HIV prevention gaps, with potential HIV risk driven less by client volume and more by limited HIV prevention uptake.

Table 2:

Class Prevalences and Conditional Probabilities for a 4-class Fitting Model

	Alcohol and Violence	High client Turnover	Lower Client Volume	Limited Condom and Lubricant Use
Class Prevalence	16.79%	38.41%	31.79%	13.01%
Frequency of sex work
Low (< 16 days)	25.12%	22.16%	51.89%	32.85%
Medium (16–24 days)	31.57%	40.03%	25.30%	34.95%
High (> 24 days)	43.31%	37.82%	22.81%	32.20%
Number of new clients, last 30 days
Low (< 4)	16.68%	6.79%	69.66%	46.05%
Medium (4–10)	32.65%	34.51%	27.85%	43.71%
High (> 10)	50.67%	58.70%	2.49%	10.25%
Number of regular clients, last 30 days
Low (< 3)	21.84%	24.02%	51.61%	31.59%
Medium (3–8)	27.33%	25.66%	38.48%	51.16%
High (> 8)	50.83%	50.32%	9.91%	17.25%
Condom use with new clients
Inconsistent	7.63%	1.72%	0.78%	82.51%
Consistent	92.37%	98.28%	99.22%	17.49%
Condom use with regular clients
Inconsistent	14.23%	2.72%	3.41%	86.61%
Consistent	85.77%	97.28%	96.59%	13.39%
Lubricant use (lifetime)
No	39.69%	42.96%	59.46%	74.74%
Yes	60.31%	57.04%	40.54%	25.26%
Binge drinking
Does not drink	17.56%	39.04%	45.89%	45.59%
<5 drinks	26.55%	26.14%	30.95%	23.87%
>= 5 drinks	55.89%	34.82%	23.16%	30.54%
Physical violence
No	18.40%	71.16%	79.84%	82.06%
Yes	81.60%	28.84%	20.16%	17.94%
Sexual violence
No	14.62%	83.27%	74.04%	73.54%
Yes	85.38%	16.73%	25.96%	26.46%

Predictors of the class membership

Adjusted multinomial logistic regression models evaluating the selected correlates of HIV risk and assigned class membership are presented in Table 3. These predictors suggested that the identified latent classes differed by social and structural factors. For all reported estimates of 95% CIs, the Lower Client Volume group is included as the reference.

Table 3:

Predictors of Class Membership using Maximum Probability Assignment Rule

Characteristic	aOR	95% CI	aOR	95% CI	aOR	95% CI
	Alcohol and Violence (ref: Lower Client Volume)		High Client Turnover (ref: Lower Client Volume)		Limited Condom and Lubricant Use (ref: Lower Client Volume)
Sex work interruption (ever)	1.22	1.04, 1.42	2.31	1.98, 2.70	1.30	1.07, 1.58
Share earnings	3.46	2.83, 4.23	1.76	1.43, 2.18	2.68	2.10, 3.44
No financial dependents	1.13	0.88, 1.45	1.24	0.97, 1.58	1.41	1.06, 1.88
More than middle school education	1.31	1.06, 1.64	1.00	0.80, 1.25	1.14	0.87, 1.49
Live near urban center during childhood	1.14	0.96, 1.36	1.05	0.89, 1.25	1.19	0.95, 1.47
Age
18–24	1.42	1.18, 1.70	1.31	1.10, 1.57	1.90	1.50, 2.40
25–29	1.28	1.04, 1.56	1.28	1.05, 1.55	1.71	1.32, 2.21
Family stigma	1.40	1.17, 1.68	0.78	0.65, 0.95	1.19	0.94, 1.50
Healthcare stigma	0.87	0.70, 1.08	0.59	0.46, 0.74	0.82	0.62, 1.10
Legal stigma	2.26	1.92, 2.66	2.07	1.77, 2.43	0.83	0.67, 1.04

FSW who experienced sex work interruptions were more likely to be in the Alcohol and Violence class (aOR = 1.22, 95% CI: 1.04–1.42), the Limited Condoms and Lubricant Use class (aOR = 1.30, 95% CI: 1.07–1.58), and the High Client Turnover class (aOR = 2.31, 95% CI: 1.98–2.70) compared to the Lower Client Volume group.

Those who shared their earnings had a significantly higher likelihood of being in the Alcohol and Violence class (aOR = 3.46, 95% CI: 2.83–4.23), Limited Condoms and Lubricant Use (aOR = 2.68, 95% CI: 2.10–3.44), and High Client Turnover class (aOR = 1.76, 95% CI: 1.43–2.18). Having no financial dependents was associated with higher odds of being in Limited Condoms and Lubricant Use class (aOR = 1.41, 95% CI: 1.06–1.88).

Younger FSW (aged 18–24) had a greater likelihood of being in the Alcohol and Violence class (aOR = 1.42, 95% CI: 1.18–1.70), Limited Condoms and Lubricant Use class (aOR = 1.90, 95% CI: 1.50–2.40), and High Client Turnover (aOR = 1.31, 95% CI: 1.10–1.57) classes.

FSW who experienced family stigma were more likely to be in the Alcohol and Violence class (aOR = 1.40, 95% CI: 1.17–1.68) but were less likely to be in the High Client Turnover class (aOR = 0.78, 95% CI: 0.65–0.95). Experiencing healthcare stigma was also associated with lower likelihood of being in the High Client Turnover class (aOR = 0.59, 95% CI: 0.46–0.74).

FSW experienced legal stigma were more likely to be in the Alcohol and Violence (aOR = 2.26, 95% CI: 1.92–2.66) and High Client Turnover (aOR = 2.07, 95% CI: 1.77–2.43) classes, compared to the Lower Client Volume group.

Profiles of observed demographic and behavioral characteristics within the 4-class model were generated using a maximum probability assignment rule and generally aligned with associations estimated through multivariable modeling (Supplemental Table D).

HIV-Related Outcomes

Overall, 4368/5548 (78.7%) FSW reported having tested for HIV previously, and 1,120/5548 (20.2%) were living with HIV based on biological testing. The prevalence of prior HIV testing was lowest (56.9%) in the Limited Condoms and Lubricant Use class compared to the other 3 classes. Compared to FSW in the Limited Condoms and Lubricant Use class, and as identified from the multivariable mixture model, FSW in the Alcohol and Violence class were more likely to have had a prior HIV test (aOR 3.01, 95% CI 1.90–4.80). Similarly, FSW in the High Client Turnover class and FSW in the Low Client Volume class were also more likely to have had a prior HIV test compared to the Limited Condoms and Lubricant Use class (High Client Turnover : aOR 2.62, 95 CI 1.94–3.55, Low Client Volume, aOR 2.75, 95 CI 2.07–3.64) (Table 4, see Supplemental Table F for additional aOR of covariates in the multivariable mixture model). HIV prevalence was highest (21.4%) in the High Client Turnover class compared to the other three classes but did not differ significantly by class membership.

Table 4.

Odds Ratios from a Multivariable Mixture Model Assessing the Association between LCA-identified Classes with HIV status and Prior HIV Testing among FSW

	Living with HIV			Prior HIV test
LCA-identified class	Pr. (%)	aOR†	95% CI	Pr. (%)	aOR	95% CI
1. Alcohol and Violence	20.7	1.04	0.76–1.43	87.8	3.01	1.90–4.80
2. High Client Turnover	21.4	1.06	0.70–1.62	86.0	2.62	1.94–3.55
3. Lower Client Volume	18.8	1.09	0.79–1.50	82.6	2.75	2.07–3.64
4. Limited Condom and Lubricant Use	20.4	Ref.	Ref.	56.9	Ref.	Ref.

^†Adjusted for age, prior education, and country of residence.

Discussion

In this pooled multi-country analysis of bio-behavioral data collected among FSW in sub-Saharan Africa, we identified four latent constructs of sexual risk characterized by overlapping socio-structural vulnerabilities. Classes were primarily defined by clustering around violence and alcohol, number of paid clients, and use of condoms and lubricant. Class membership was significantly associated with prior HIV testing experience, likely reflecting differences in access to health services including community-based HIV programs. While HIV prevalence estimates were consistently high, irrespective of class membership, these data also demonstrate varying individual, social, and structural vulnerabilities across classes. Findings from this analysis emphasize the importance of differentiating both proximal and distal risks within heterogeneous FSW populations to engage women who remain disconnected and marginalized from routine health services. Together, these data reinforce the need for intensified efforts to design and implement targeted strategies to optimize the effectiveness of HIV prevention with sex worker communities, particularly during a time of decreasing resources for global health.

Although FSW remain a priority population for HIV prevention, not all women who sell sex remain consistently at high risk of HIV. Importantly, HIV risks are not static, and evolve over time as a function of changing age, occupation, and sexual partnerships that naturally occur across the life course [31, 32]. We found distinct latent constructs that may speak to these shifts in risk within and across networks of FSW. While a majority of women in these analyses reported using condoms with new or regular clients, and had tested for HIV at least once, women in the Limited Condom and Lubricant Use class were the least likely to report prior HIV testing compared to other classes, indicating an important programmatic gap for women who may be at the highest risk of HIV acquisition and transmission. FSW in this class were also younger and reported less legal stigma and more shared earnings, suggesting newer entry into sex work and/or fewer interactions with the legal system. Younger women who sell sex experience a greater risk of HIV acquisition upon entering sex work given biological vulnerability, heightened experiences of physical and sexual violence, and a lack of autonomy/ability to negotiate safer sexual practices [12]. Thus, addressing these programmatic gaps will require youth-tailored HIV prevention approaches, including condom negotiation skill-building, empowerment and legal counseling [33], and strengthened peer support to build social cohesion [16, 20] and facilitate early connection to services.

Within the context of sex work, FSW who are more enmeshed within community-based HIV programs may have more opportunities to occupationally routinize effective HIV prevention strategies. For example, evidence from a community-empowerment–based combination prevention trial in Iringa, Tanzania, demonstrated that community-led programming, including peer mobilization, drop-in centers, provider sensitivity training, and linkage support, significantly reduced HIV incidence and improved care-continuum outcomes among FSW [34]. Within this pooled analysis, FSW who experienced fewer socio-structural vulnerabilities such as susceptibility to violence and alcohol use across classes demonstrated greater overall service engagement (HIV testing and condom use), underscoring the potential protective aspects of routinely engaging with HIV and other SRH services. Women in this class were also less likely to share earnings, indicating financial independence and possibly fewer exploitative relationships that can facilitate more successful condom negotiation [33]. In the context of evolving risk, community-based sex worker programs can support stable occupational environments, economic autonomy, and continued access to healthcare prevention services while providing FSW with the necessary scaffolding to optimize their sexual health and remain at a lower risk of HIV [35] [36]. Further aligning community preferences with service delivery strategies and strengthening local implementation capacity are also essential to ensure interventions remain responsive and sustainable in the context of constrained funding environments.

The identified latent classes can inform differentiated service delivery strategies tailored to distinct patterns of risk and unmet need among FSW. FSW in the Limited Condom and Lubricant Use class represent a critical gap in service engagement and may benefit from condom negotiation support, improved access to free condoms and lubricants, and integration with HIV testing and reproductive health services[6, 8]. FSW in the High Client Turnover class, reflecting elevated exposure risk, may benefit from community-based service delivery strategies that support more frequent HIV testing, expanded PrEP access, intensified outreach, and strengthened linkage to care to reduce transmission risk [14, 23].

FSW in the Alcohol and Violence class, despite relatively higher engagement with prevention practices, face substantial structural vulnerability and may benefit from integrated interventions addressing violence prevention, mental health and trauma counseling, and substance use support[16, 35]. FSW in the Lower Client Volume class, with comparatively lower exposure risk, may benefit from sustained prevention behaviors and continued engagement in care as risk environments evolve over time[32].

Given recent changes to foreign aid and global cuts to HIV funding worldwide [25, 28], the ability to identify FSW at the highest risk of HIV is increasingly critical for ensuring that limited resources are directed to those with the greatest unmet needs. Analyses described in this study represent a potentially scalable approach to describe risk heterogeneity among key populations and other historically marginalized communities using commonly collected data. Moreover, these approaches may be further leveraged to quantify and better address the needs of underserved populations more broadly, and inform large-scale policy decisions particularly amidst decreasing and shifting resources within the global HIV response. Evidence from economic modeling further demonstrates that targeted allocation of prevention resources to the most vulnerable subgroups yields greater reductions in HIV incidence than untargeted approaches, particularly during periods of budget cuts [37–39]. The findings are actionable and can guide program development for FSW at highest HIV risk, set performance targets, and assess coverage benchmarks [40, 41].

There are some limitations in this study. First, self-reported behavioral items may introduce potential recall and social desirability biases. Second, the cross-sectional design precludes any evaluation of causal relationships within these data, including between class membership and HIV-specific outcomes including prior testing and HIV seroprevalence. Additionally, assigning participants to their highest posterior probability of group membership may result in some misclassification across classes. Therefore, associations between social and structural covariates and class membership should be interpreted with some caution. Third, cultural variations and definition of sex work within countries that may have influenced indicator responses that were not captured in the combined analysis. Lastly, LCA inherently categorizes complex behaviors into distinct classes, potentially oversimplifying continuous risks.

Conclusions

Across sub-Saharan Africa, FSW experience heightened health and social inequities, resulting in critical need for tailored HIV services that are responsive to their distinct SRH needs. The intersecting dynamics around sexual risk along with higher-order factors such as substance use, violence, and socio-economic dependence result in diversity in lived experience than can influence the extent to which FSW need more intensified prevention and treatment services within and across different geographies. The methodological approach used in this analysis illustrates the practical value of applying scalable, data-efficient tools to differentiate risk within heterogeneous FSW populations. By differentiating risk profiles, these results offer an evidence-based approach to inform pragmatic risk stratification and direct limited resources where they will have the greatest impact.

Abbreviations

FSW

Female sex workers

SRH

Sexual and reproductive health

HIV

Human immunodeficiency virus

LCA

Latent class analysis

aOR

Adjusted odds ratio

CI

Confidence interval

IQR

Interquartile range

ART

Antiretroviral therapy

PrEP

Pre-exposure prophylaxis

RDS

Respondent-driven sampling

IRB

Institutional Review Board

BIC

Bayesian Information Criterion

AIC

Akaike Information Criterion

LRT

Likelihood ratio test

BCH

Bolck, Croon, and Hagenaars (weighted three-step method for distal outcomes)

Data statement

Due to the sensitive nature of the study population and data (including potentially identifiable information), the individual-level data cannot be made publicly available. Access to the underlying data may be granted to researchers upon approval.