The role of structural factors for preventing HIV risk practices among adolescents in South Africa: A three-wave analysis of caregiving, education, food security, and social protection

Background. Structural interventions are endorsed to enhance biomedical and behavioural HIV prevention programmes for adolescents. Aiming to inform future interventions, we evaluated longitudinal associations between six structural factors and five HIV risk practices in a cohort of adolescents in South Africa. Methods. We used three rounds of data between 2014–2018 on 1046 adolescents living with HIV and 483 age-matched community peers in South Africa’s Eastern Cape (Observations = 4402). We used multivariable random effects within-between logistic regression to estimate sex-specific associations between six time-varying structural factors − number of social grants, education enrolment, days with enough food, caregiver supervision, positive caregiving, and adolescent-caregiver communication − and five HIV risk practices − multiple sexual partners, transactional sex, age-disparate sex, condomless sex, and sex on substances. We calculated probability differences, contrasting predicted probabilities at average and maximum values of structural factors associated with multiple risk practices. Findings. The sample mean age was 15.29 (SD: 3.23) years and 58% were female. In females, compared to average, maximum positive caregiving scores were associated with lower probability of transactional sex (−1.06 percentage points [ppts], 95%CI=−1.60; −0.52ppts), and age-disparate sex (−0.73ppts; 95%CI=−1.26; −0.19ppts); maximum caregiver supervision scores were associated with lower probability of multiple sexual partners (−3.11ppts; 95%CI=−3.87; −2.35ppts) transactional sex (−1.07ppts, 95%CI=−1.42; −0.71ppts), age-disparate sex (−0.67ppts; 95%CI=−1.08; −0.25ppts), condomless sex (−3.96ppts; 95%CI=−5.65; −2.26ppts), and sex on substances (−0.93ppts; 95%CI=−1.50; −0.37ppts); and, seven days with enough food was associated with lower probability of multiple sexual partners (−1.18ppts, 95%CI=−2.06; −0.30ppts), and transactional sex (−0.91ppts; 95%CI=−1.41; −0.42ppts). Relative to non-enrolment, education enrolment was associated with lower probability of age-disparate sex (−3.18ppts; 95%CI=−5.35; −1.01ppts), and condomless sex (−11.32ppts; 95%CI=−19.15; −3.49ppts). In males, compared to average, maximum caregiver supervision scores were associated with lower probability of multiple sexual partners (−2.83ppts; 95%CI=−3.66; −2.00ppts), transactional sex (−0.90ppts; 95%CI=−1.20; −0.60ppts), age-disparate sex (−0.46ppts; 95%CI=−0.77; −0.15ppts), and sex on substances (−1.42ppts; 95%CI=−2.06; −0.78ppts). No other structural factors were associated with multiple risk practices. Interpretation. Structural interventions to improve food security and education enrolment among adolescent girls, and positive and supervisory caregiving among adolescent girls and boys are likely to translate into crucial reductions in HIV risk.


Introduction
In Southern and Eastern Africa, rates of new HIV infections among adolescents and young adults remain well above the Joint United Nations Programme on HIV/AIDS (UNAIDS) elimination targets, at around 220 per 100,000 in 2021 [1,2]. Within the region, South Africa concentrates 27% of new infections among eight covariates [37]. The Wald Test was used to compare pairs of within-and between-person coe cients for equality [38]. We then estimated a second model (Model 2) that only included separate within-and between-person effects where there was evidence that they were likely to differ signi cantly according to the Wald Test (p < 0.05). Throughout our analysis, structural factors were considered as either household-level (number of social grants, positive caregiving, caregiver supervision, and adolescent-caregiver communication), or individual-level (education enrolment and food security). We did not adjust for individuallevel factors when investigating household-level structural factors, as they were likely to lie on the causal pathway to study outcomes [39].
Third, for structural factors found to be signi cantly associated with lower odds of multiple HIV risk practices (either via between-person, within-person, or combined effects) we calculated adjusted probabilities of study outcomes xing structural factors to '0: No' and '1: Yes' for binary variables and to 'mean' and 'maximum' for continuous variables. In most cases, maximum scores lay within one standard deviation (SD) of the mean value. Adjusted probabilities were calculated overall, and xing HIV status to either "0. Not living with HIV" and "1. Living with HIV". Our prioritisation of structural factors associated with multiple outcomes is motivated by the UNDP 'accelerator' concept, which calls for greater focus on approaches to improve multiple outcomes simultaneously [40,41].
Fourth, as a robustness check for whether variation in structural factors temporally preceded HIV risk practices, we evaluated the association between prior (lagged) structural factors and outcomes measured at the subsequent waves. We used the same approach as for our main analysis, except that among males, we were unable to consider separate within-and between-person effects for education enrolment because of minimal within-person variation.

Results
The sample included 1563 adolescents, and the total number of observations included in the analysis was 4402, Figure S1. Respondents lost to follow-up data collection were older (p < 0.001) and lived in larger households (p = 0.04) in an urban location (p = 0.03), Table S3. Compared to adolescents living with HIV, community peers at baseline were on average six months older, less likely to be maternally or paternally orphaned, and lived in larger households (all p < 0.001), Table S4. Median time between rst and nal interview was 951 days. Missing values for all variables were < 10%, except for sex on substances, which was not measured at wave one, Table S5.
Summary of descriptive characteristics. Fifty-eight percent of respondents were female and 70% were living with HIV, Table 1. The average age of respondents was 15.29 (SD: 3.23), 26% lived in a rural area, 15% lived in informal housing, 41% were maternally orphaned, 34% were paternally orphaned, and the mean household size was 6.50 (SD: 4.06), Table 1. Compared to males, females were older (p < 0.001), lived in larger households (p = 0.022), and were more likely to live in informal housing (p = 0.003). They were less likely to be enrolled in education (p < 0.001), and on average, reported fewer days with enough food at home last week (p < 0.001). Between-and within-person variability in time-varying structural factors was higher among females, except for within-person variability in caregiver supervision, which was higher among males. The two most prevalent HIV risk practices were multiple sexual partners and condomless sex, followed by transactional sex, age-disparate sex, and sex on substances, Table 1. Females were more likely to report transactional sex (p < 0.001), age-disparate sex (p < 0.001), and condomless sex (p < 0.001). Compared to adolescents aged 11-19 years, prevalence of all ve practices were signi cantly higher among young adults aged 20-25 years (p < 0.001), Fig. 1.
Compared to their peers not living with HIV, females living with HIV were signi cantly less likely to report condomless sex (p < 0.001) and sex on substances (p = 0.03), and males living with HIV were less likely to report multiple sexual partners (p < 0.001), condomless sex (p < 0.001), and sex on substances (p < 0.001), Fig. 1. Correlations between study outcomes, and univariable associations between structural factors and outcomes, are summarised in Table S6 and Table   S7, respectively.
[ Figure 1 about here] Regression analyses. In models without explanatory variables, values of intra-class correlation (ICC) were > 0.5 for all outcomes − indicating a high correlation between latent HIV risk practices within the same participant across different waves − except for transactional sex in females (ICC = 0.44) and males (ICC = 0.20), age-disparate sex in males (ICC = 0.44), and condomless sex in males (ICC = 0.33), Table S8.
Multivariable associations between hypothesised structural factors and HIV risk practices. In model 1, among females, between-and within-person coe cients differed signi cantly for caregiver supervision in relation to multiple sexual partners, condomless sex, and sex on substances; education enrolment in relation to transactional sex and condomless sex; and number of days with enough food in relation to transactional sex, Table 2. Among males, between-and within-person coe cients only differed signi cantly for education enrolment in relation to condomless sex, Table 2.  Table S9.
Predicted probabilities of HIV risk practices at selected values of 'accelerator' structural factors. We summarise predicted probabilities for each of our HIV risk practices at selected values of statistically signi cant structural factors, and probability differences comparing adjusted probabilities among females in Fig. 2 and males in Fig. 3. [ Table 2 about here]

Multivariable associations between lagged hypothesised structural factors and HIV risk practices. Recombining between-and within-person
[ Figure 2 about here]

Discussion
This study found a high prevalence of ve HIV risk practices in a large cohort of adolescents in South Africa. Rates of condomless sex were much higher than the UNAIDS target of no more than 5% for priority groups [42,43], and young men reported high rates of multiple sexual partners [44]. Compared to their uninfected peers, females living with HIV were signi cantly less likely to engage in condomless sex, and males living with HIV were less likely to engage in both condomless sex and sex on substances. In gender-strati ed analyses, we found that among females, education enrolment, days with enough food at home, positive caregiving, and caregiver supervision were each associated lower probability of multiple HIV risk practices. Among males, caregiver supervision was associated with lower probability of multiple HIV risk practices, and in addition, positive parenting and days with enough food were associated with lower probability of single HIV risk practices. Education enrolment was associated with higher probability of transactional sex and condomless sex, but wide con dence intervals indicate a high level of uncertainty around these estimates. Finally, in lagged analyses we found that in both females and males, prior higher caregiver monitoring and number of days with enough food were associated with lower odds of subsequent HIV risk practices, suggesting that their protective in uence may be sustained.
Our analysis applies rigorous statistical methods to three waves of data to investigate relationships between six structural factors and ve HIV risk practices simultaneously, further unpacking differential associations for males and females. In females, associations between education enrolment and age-disparate sex and condomless sex, respectively, are consistent with previous analyses and theories supporting this factor's promotion of safer sexual networks and adolescent girl's negotiating power [45][46][47][48][49][50]. The observed association between days with enough food at home and transactional sex is supported by the 'sex for basic needs' paradigm of transactional sex [51], and the concurrent relationship between this factor and both multiple sexual partnership and transactional sex is supported by qualitative reports of the in uence of poverty on sexual practices among adolescent girls [52]. Associations between caregiver supervision and transactional sex and age-disparate sex are in line with theories suggesting that setting rules and monitoring peer-relationships can act as a 'protective shield', promoting the internalization of norms that foster healthy practices, mitigating sensation-seeking and impulsive decision-making, and deterring a liation with deviant peers [53][54][55][56]. Among males, the relationship between caregiver supervision and all ve study outcomes can be linked to masculine norms of independence and sexuality driving HIV risk in this population, and caregivers' gendered perceptions that independence should be encouraged among adolescent boys, while adolescent girls should remain restricted and protected [57,58].
Our ndings support continued emphasis on structural interventions for enhancing the effectiveness of core prevention programmes in HIV prevention [59].
Among females, the range of structural factors associated with HIV risk practices suggests a complex of motivations for engaging in HIV risk practices, and validates the need for comprehensive multi-component prevention approaches in contexts with su cient resources [19,60]. Comparing across structural factors, the strong association between education enrolment and age-disparate sex and condomless sex suggests that, in settings with limited resources, This study's use of three waves of data spanning an average of two and a half years for each participant enabled us to use advanced statistical models to unpack within-and between-person associations between structural factors and HIV risk practices. Nevertheless, there is still a risk of confounding from unmeasured time-varying factors for both between-and within-person associations, and estimated associations should not be interpreted causally. The study's exhaustive sampling strategy and small loss to follow-up minimises risk of selection bias among participants living with HIV. Although participants not living with HIV were recruited through invitation, the large number of respondents should also minimise selection bias in this group. Self-reported items may be subject to social desirability and recall bias, particularly those relating to sensitive topics [70], and participant subjectivity may also have been a source of measurement error. Because in our study almost all participants reported receipt of at least one social grant (~ 91%) we were unable to robustly evaluate the association between this structural factor and HIV risk practices. Similarly, because so few male participants left education during the study, observed associations between this risk factor and risk practices may not be robust, and should be investigated in future studies.
Building on this study, future mediation analysis could valuably inform the mechanisms via which structural factors act on HIV risk practices and provide stronger causal claim for associations identi ed in this study. Plausible pathways could include stronger a liation with a positive peer group, safer sexual networks, and greater negotiating strength. Little research has focused on the in uence of peers on sexual practices in South Africa or how peer a liation is shaped by structural factors [72,73]. Our nding that even when enrolled in education adolescent girls may still experience a high probability of condomless sex, highlights an urgent needed to identify other structural factors able to address this key sexual practice [74]. Further, evidence indicates that community disorder, caregiver stress, and young men's social inclusion could be candidates for structural 'accelerators' of HIV prevention [75].  Predicted probabilities of HIV risk practices and probability differences comparing two scenarios among females: (i) in the absence of education enrolment, or at the mean of number of days with enough food, positive caregiving, and caregiver supervision scores; and (ii) in the presence of education enrolment, or at the maximum of number of days with enough food, positive caregiving, and caregiver supervision. Lines connecting bars summarise the calculated difference between two predicted probabilities with 95% con dence intervals in brackets. We only calculated adjusted probabilities where there was evidence of signi cant associations between structural factors and lower odds of multiple HIV risk practices. Predicted probabilities were also calculated xing HIV status covariate to "0. Not living with HIV" and "1. Living with HIV". Values used to build Figure 2 are summarised in Table S11. a Sex on substances was only measured at wave two and wave three. b Predictions are based on between-person effects rather than combined effects. Abbreviations: ppts, percentage points.

Figure 3
Predicted probabilities of HIV risk practices and probability differences comparing two scenarios among males: (i) at the mean of caregiver supervision; and (ii) at the maximum of caregiver supervision. Lines connecting bars summarise the calculated difference between two predicted probabilities with 95% con dence intervals in brackets. We only calculated predicted probabilities where there was evidence of signi cant associations between structural factors and lower odds of multiple HIV risk practices. Adjusted probabilities were also calculated xing HIV status to "0. Not living with HIV" and "1. Living with HIV".
Values used to build Figure 3 are summarised in Table S11. a Sex on substances was only measured at wave two and wave three. Abbreviations: ppts, percentage points.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download. Supplementarymaterial.docx