The impact of the DREAMS partnership on HIV incidence among young women who sell sex in two Zimbabwean cities: results of a non-randomised study

Introduction Young women who sell sex (YWSS) in Zimbabwe remain at high risk of HIV infection. Effective HIV prevention strategies are needed. Through support to access a combination of evidence-based interventions, including oral pre-exposure prophylaxis (PrEP), the Determined, Resilient, Empowered, AIDS-free, Mentored and Safe (DREAMS) partnership aimed to reduce new HIV infections among adolescent girls and young women by 40% over 24 months. Methods Non-randomised ‘plausibility’ evaluation, powered to detect a 40% HIV incidence difference between DREAMS and non-DREAMS sites. Two large cities with DREAMS funding were included, and four smaller non-DREAMS towns for comparison. In all sites, YWSS were enrolled to a cohort through peer-referral. Women were followed up for 24 months. HIV seroconversion was the primary outcome, with secondary outcomes identified through a theory of change. Outcomes were compared between YWSS recruited in DREAMS cities and non-DREAMS towns, adjusting for individual-level confounders and HIV prevalence at enrolment. Results From April to July 2017, 2431 women were enrolled, 1859 of whom were HIV negative at enrolment; 1019 of these women (54.8%) were followed up from March to May 2019 and included in endline analysis. Access to clinical services increased, but access to socioeconomic interventions promoted by DREAMS was limited. A total of 79 YWSS HIV seroconverted, with HIV incidence among YWSS in DREAMS cities lower (3.1/100 person-years) than in non-DREAMS towns (5.3/100 person-years). In prespecified adjusted analysis, HIV incidence was lower in DREAMS cities but with weak statistical evidence (adjusted rate ratio (RR)=0.68; 95% CI 0.40 to 1.19; p=0.18). Women in DREAMS cities were more likely to report ever and ongoing PrEP use, consistent condom use, fewer sexual partners and less intimate partner violence. Conclusion It is plausible that DREAMS lowered HIV incidence among YWSS in two Zimbabwean cities, but our evaluation provides weak statistical evidence for impact and suggests any reduction in incidence was lower than the anticipated 40% decline. We identified changes to some important ‘pathways to impact’ variables, including condom use.


INTRODUCTION
Worldwide, especially in sub-Saharan Africa, young women are at high risk of HIV due to increased biological, economic and social vulnerability. In Zimbabwe, where the HIV burden is one of the highest in the world and HIV prevalence was 13.3% among adults aged 15-49 years in 2017, 1 adolescent girls and young women (AGYW) aged 15-24 years are at particularly high risk of HIV.
Recognising the vulnerability of AGYW and the social, economic and biological factors that shape women's risk, the DREAMS Partnership was developed to deliver a combined package of interventions targeted at AGYW, their partners, families and communities, to address the interacting factors that shape HIV risk among this particularly vulnerable population. One particular target population for the DREAMS Partnership was young women who sell sex (YWSS). In Zimbabwe, among this group, DREAMS included an offer of oral pre-exposure prophylaxis (PrEP), and condom promotion and provision.
We conducted an impact evaluation of DREAMS among YWSS in Zimbabwe.

AIM
The aim of this study was to estimate the impact of the DREAMS combined package of HIV prevention interventions on HIV incidence among YWSS aged 18-24 years. We also sought to evaluate the impact of DREAMS on a number of secondary outcomes.

STUDY DESIGN
This study was a non-randomised plausibility design to estimate the effect of DREAMS on HIV incidence and other secondary outcomes, highlighted in section 5.
We will compare HIV incidence among a cohort of YWSS recruited in two districts where DREAMS was being implemented and followed up over 2 years, to HIV incidence among a cohort of YWSS recruited in four districts where DREAMS was not being implemented and also followed up over 2 years. A similar approach will be used for comparison of secondary outcomes.

SAMPLING
In the two DREAMS sites, a network-based recruitment strategy (respondent-driven sampling (RDS)) was used to identify YWSS in the study communities, offer them HIV testing services and then inform and refer these YWSS to treatment and prevention services, including PrEP and the DREAMS package of HIV prevention interventions through the national programme for sex workers 'Sisters with a Voice' and then onward to the full range of DREAMS services. This process was also used to recruit these women into the evaluation cohort. More specific details of the recruitment process are provided in the protocol (e.g. seed selection, wave recruitment procedures, remuneration).
In two DREAMS sites, eligible YWSS were asked for written informed consent to be interviewed at enrolment into the study in 2017, and then followed-up at 12 and 24 months after the initial enrolment survey. At each time point, rapid HIV testing and counselling was offered to YWSS to ascertain the HIV status of the study participants. A detailed working definition of YWSS is provided in the protocol, but we note here that we included both young women who did and did not selfidentify as sex workers, anticipating that the outcome profile might differ among these two groups of young women.
The same network-based recruitment strategy was used to identify and recruit a cohort of YWSS in the four non-DREAMS districts. YWSS recruited in these four districts were also offered HIV counselling and testing services and were referred to the existing national HIV programme for sex BMJ Publishing Group Limited (BMJ) disclaims all liability and responsibility arising from any reliance Supplemental material placed on this supplemental material which has been supplied by the author(s) BMJ Global Health doi: 10.1136/bmjgh-2020-003892 :e003892.
workers, run by the Centre for Sexual Health and HIV/AIDS Research (CeSHHAR). Through this programme they could access HIV prevention services, including condoms, STI treatment and health education, but did not have access to PrEP or other initiatives offered under the DREAMS Partnership. These YWSS were recruited during the same time frame as women in the two DREAMS districts, and were interviewed at enrolment and followed up at 24-months post-enrolment. Although planned, no follow up was done at 12 months among this group due to budgetary constraints.
The difference, therefore, between DREAMS and non-DREAMS sites was that, in DREAMS sites, YWSS had access to comprehensive SRH and HIV services through the Sisters programme PLUS access to PrEP and other DREAMS services, such as social protection interventions. In the non-DREAMS sites, YWSS had access to comprehensive SRH and HIV services through the 'Sisters' programme but were not referred for PrEP / other DREAMS services.

PRIMARY AND SECONDARY OUTCOMES
The primary outcome is incident HIV infection over the 24-month study period, defined as:

Number of new HIV infections among YWSS who tested HIV-negative at enrolment / Total person-years of follow-up accumulated during the 24-month study period.
Measurement of the primary outcome was restricted to women testing HIV negative at enrolment and followed-up 24-months post-enrolment. HIV status at each round was ascertained through the rapid HIV tests delivered during the counselling and testing process with results returned to participants. No confirmatory testing procedures were conducted for the purpose of the impact evaluation.
Person-years of follow-up was defined as the total follow-up time between the enrolment survey in 2017 and the follow-up survey 24 months after enrolment (2019). For women who seroconverted during the study, the time of seroconversion was set as the mid-point between the enrolment survey and follow-up at 24 months.
The secondary objective of the impact evaluation is to explore whether the DREAMS package of interventions had an impact on the on the secondary outcomes listed in Table 1 and in line with the hypothesised causal pathway through which the DREAMS package of interventions would reduce HIV incidence ( Figure 1). analyses reflecting this design. In this study the primary aim is to estimate the impact of the DREAMS interventions, by making fair comparisons across the six clusters, and so we will not weight the data to account for recruitment using RDS -on the basis that the approach to sampling was the same/standardised in each cluster. However, we plan a sensitivity analysis in which weighting will be applied; this approach adds an additional layer of complexity, but aims to provide estimates of rates/prevalences for the total study population.
Adjustment for confounding strategy: In the absence of randomisation, and also because the number of study clusters is low, it is appropriate to adjust the analysis for potential confounding variables. If we find evidence for a difference in HIV incidence between the two study arms, caution is needed when considering whether this difference is attributable/partly attributable to DREAMS. We investigated a priori what confounding factors to adjust for in order to obtain the fairest comparisons among study clusters. To do this, we used HIV prevalence at baseline as a "proxy" for the background level of HIV incidence.
We found that HIV prevalence was different between the two study arms at baseline. We then identified a set of variables that we a priori thought might be associated with HIV prevalence as explored elsewhere, 2 including age at enrolment, highest level of education attained, marital status, self-identification as a sex worker, STI symptoms and number of sexual partners in the past month. We included each variable in a univariable logistic regression model to confirm that they were associated with HIV prevalence and then modelled HIV prevalence against DREAMS arm adjusting for all six variables. Adjusting for these variables attenuated the difference in HIV prevalence between the two arms substantially (see Table 2). We will therefore adjust our primary analysis for these six variables, measured at baseline among the cohort, so as to make fairer comparisons between the two DREAMs and four non-DREAMs study clusters, and a fairer attribution of any difference we see to DREAMS intervention. The number of sero-conversions over the two-year study period may be small relative to the number of parameters if all six variables are included in a model. We will, therefore, need to consider the "rule-of-thumb" that the number of sero-conversions should be approximately ten-times higher than the number of parameters included in our model. If the number of sero-conversions is smaller than the number of parameters if all six variables are included, we will prioritise adjusting for the variables that were the strongest confounders of the association between prevalent HIV and DREAMS, including age and educational attainment if there is evidence that the model becomes unstable with the addition of more parameters.
We will not be able to formally adjust our analysis for cluster level factors that may differ between the DREAMS and non-DREAMS districts, for example background HIV prevalence, because we have "only" included six clusters in the study. This is a limitation of our study described in further detail below At baseline, we also analysed how HIV prevalence increased with age among the YWSS recruited in DREAMS and non-DREAMS districts (see Figure 2). We fitted a simple linear regression line to HIV prevalence by age. We noted that, while there was a difference in the HIV prevalence at each age, as described above, the rate of increase of HIV prevalence with age was almost identical at approximately an average of 6% per single year of age increase in both groups. Among young women, HIV prevalence increasing with age can be cautiously interpreted as reflecting the underlying HIV incidence.  The combination of our data suggesting an i) HIV prevalence that differed between the arms at baseline, but that could be largely adjusted away through the addition of six key covariates measured at baseline, and which we will adjust our primary outcome analyses for, and ii) our finding that at baseline HIV prevalence increasing by age was near identical between the groups, provide strength to the argument we will make post-analysis, that an adjusted difference in measured HIV incidence between the arms can be plausibly interpreted as reflecting the effect of the DREAMs intervention. The further the adjusted rate ratio is from 0.8, the more credible it is that DREAMS had an effect on HIV incidence. Our findings will be interpreted in the context of DREAMS uptake. We will explore uptake of DREAMS interventions across the two study sites (as described in section iv); where there is evidence of differential uptake across the two groups, this would add credibility to any finding that the DREAMS intervention had an effect on HIV incidence.
Reporting: Analyses will be reported in line with the Transparent Reporting of Evaluations with Nonrandomized Designs (TREND) statement.

STAGES OF ANALYSIS i. COHORT RECRUITMENT AND RETENTION
We will first describe recruitment into the cohort study, and present a flow chart showing the number of women recruited through the RDS in each district and by study arm (DREAMS and non-DREAMS), and retention at 24-month follow-up.
We will generate RDS recruitment trees by site, colour coding women by whether they tested HIV negative, HIV positive at enrolment, or tested HIV negative at enrolment and seroconverted at 24month follow up. We have already done the detailed RDS diagnostics which found no evidence of Using data collected on efforts to contact the women at the 24-month follow up, we will describe the reported reasons why women were lost to follow, for example they reportedly migrated and/or married, or were contacted but refused to participate.

ii. PARTICIPANT CHARACTERISTICS AT ENROLMENT
We will describe, by site and study arm, key demographic and behavioural characteristics of the women recruited to the cohort, including those that were identified to be associated with HIV prevalence at enrolment. We will repeat this analysis for YWSS who tested HIV negative at enrolment. iii.

UPTAKE OF SERVICES
The services delivered by DREAMS implementing partners may also be available to the women in non-DREAMS sites, particularly HIV testing service but also educational subsidies, and vocational skills training. Using 24-month follow-up data, we will describe the uptake of other services that may be available in and accessible to women in non-DREAMS sites, by arm and by site.

iv. PARICIPANT CHARACTERISTICS AT 24-MONTH FOLLOW UP
We will describe the proportion of women retained in the study at the 24-month follow-up, and describe follow-up by enrolment characteristics, including: age, marital status, highest level of education attainment and whether women self-identified as a female sex worker.
We will repeat the descriptive analysis presented for enrolment for YWSS testing HIV negative at enrolment and followed up at 24 months.

v. HIV INCIDENCE RATE BY DISTRICT AND STUDY ARM
We will describe the number of new infections observed over the 24-month study period, and describe person-years of follow-up by district and study arm (Table 10). Subsequently, we will estimate the HIV incidence rate among the YWSS by district and by study arm.

vi. UNADJUSTED ANALYSIS
We will use Poisson regression to compare the HIV incidence rates across the two study arms.
We will fit three models, namely:

vii. ADJUSTED ANALYSIS
We will use Poisson regression models to compare the HIV incidence rates across the two study arms, adjusting for covariates. The models we will fit will be: 2) An age-adjusted Poisson regression model 3) A fully adjusted model including age and individual level potential confounders measured at baseline and found to be associated with HIV prevalence. We will conduct exploratory analysis, stratifying the analysis by age and by self-identification as a sex worker, to get a sense of difference in the effect size among these subgroups, and will report the findings of these analyses. We however recognise that these analyses will likely be underpowered and should be interpreted cautiously.

SENSITIVITY ANALYSIS
We will perform 2 sets of sensitivity analyses: 1) Our primary analysis strategy does not weight the data. We will exclude seed participants and weight data by the inverse of women's reported YWSS-network size and normalise these weights by site.
2) Our primary analysis excludes data collected from women followed up at 12-months postenrolment in the DREAMS districts. In our second sensitivity analysis, we will add data from the 12-month survey in the DREAMS districts to obtain information on women not followed up at 24 months. In these DREAMS districts, if someone seroconverted by the time of the 12-month follow-up, we will place the seroconversion date at the mid-point of enrolment and 12 months, and if someone tested HIV-negative at 12 months but HIV-positive at 24 months, we will place the seroconversion at the mid-point 12 months and 24 months. We will then use this information to compare HIV incidence between the two arms.

STRENGTHS AND LIMITATIONS
Our primary analysis is relatively simple, providing descriptions of HIV incidence in each site and unadjusted and adjusted statistical comparisons between the DREAMS and non-DREAMS study arms.
Our adjustment strategy includes a step-wise adjustment process, with few a priori variables adjusted for if they were associated with HIV prevalence at enrolment. We therefore consider the analysis to be transparent, and to build an evidence base for whether it is plausible that DREAMS had an impact on HIV incidence.
A limitation of our analysis is that we may not have collected data on important covariates that are associated with HIV and important risk factors for the outcome (HIV incidence), and that differ among study clusters / by arm. As such, if we find evidence for an effect of DREAMS on HIV incidence after adjustment, we cannot be entirely sure that this effect is attributable to DREAMS.
We are unable to conduct cluster level analysis or adjust for cluster level covariates, despite the fact that the intervention was allocated at cluster level. This means we will be unable to adjust for important potential confounders at cluster level; for example, the background HIV prevalence in the DREAMS clusters may differ from that in the non-DREAMS clusters owing to both the lack of randomisation and because of chance variability. Our interpretation will thus need to be cautious, commenting on any differences in cluster level factors that are observed at baseline.
Our intention to treat approach will compare women eligible to receive DREAMS interventions in districts where DREAMS was operating with women in districts where it was not. If DREAMS delivery was weak, or did not reach the specific target populations of women who are the focus for our impact evaluation (YWSS), then we may conclude that DREAMS did not have an impact but this may reflect limited delivery rather than the maximum potential effect. We will describe the delivery and uptake of DREAMS interventions in order to support our interpretation.

APPENDIX 5: RDS DIAGNOSTICS AND RESULTS
RDS diagnostics were based on enrolment (2017) data. Combined convergence and bottleneck plots (reported elsewhere), 2 suggested that key characteristics and outcomes, including age, whether women self-identified as FSW and HIV prevalence, stabilised with increasing sample sizes in five of the six sites. There was also little evidence of disconnected networks of YWSS in each site. In this study, we report the proportion of women who said that they were recruited into the study by strangers, to have an understanding of reciprocity, and assessed recruitment homophily with respect to HIV status, to understand if HIV positive recruiters preferentially recruited HIV positive peers from amongst their personal networks.
We generated RDS recruitment trees by site, colour coding women by whether they tested HIV negative at enrolment, HIV positive at enrolment, or tested HIV negative at enrolment and seroconverted at 24-month follow up. Detailed RDS diagnostics for these data are presented elsewhere. 2 Non-DREAMS Site F (N=120) n/N (%) Apart from the woman who did recruit you to the study, has anyone else approached you to give you a coupon?

APPENDIX 6: SENSITIVITY ANALYSES
We conducted sensitivity analyses in which we (i) included data collected from women followed up at 12-months post-enrolment in the DREAMS sites, and (ii) RDS-II weighted our data, where the 24-momth data was weighted using RDS weights that were generated using enrolment. For (i), we only present the primary outcome results which are very similar to that of the primary analysis results, and we do not present the results of uptake of DREAMS and secondary outcomes results because they are also very similar to that of primary analysis.