HIV incidence estimation among female sex workers in South Africa: a multiple methods analysis of cross-sectional survey data
[ad_1]
Study design and participants
We conducted an analysis of the data from the national survey by use of multiple methods.
Briefly, a computerised simple random sampling procedure (SAS PROC SURVEYSELECT) selected 12 districts from the 22 eligible districts in South Africa that had active sex worker programmes (among 54 districts), stratified by province to ensure at least one district in each of South Africa’s nine provinces was represented. Both provincial and district sample sizes were proportional to estimates of female sex-worker population size.
A sample of mapped hotspots was drawn per district, ensuring representation across hotspot types (eg, brothel, tavern, or street). Seeds, defined as initial venue recruits, were initially identified by peer educators at sex work venues and, once enrolled by the sex work programmes, recruited subsequent female sex workers. Surveys were completed at the programme site or at their stations of sex work. A chain referral method was used for recruitment to enrol female sex workers, where every participant who was enrolled into the study was asked to distribute three coupons at random to fellow female sex workers. For ethical reasons, potential victims of human trafficking were excluded, and the relevant sex work programme was notified, enabling the provision of legal and psychosocial support.
Voluntary written informed consent was undertaken in a language of the participants choosing, from among the 11 official languages in South Africa, before data collection. This study and the questionnaire were approved by the University of the Witwatersrand Human Research Ethics Committee (reference number 180809).
Procedures
A cross-sectional interviewer-administered survey was completed. The data captured included demographic characteristics, self-reported previous sexual history (eg, age at first sex and first commercial sex), client sexual violence (ie, during the past year), known HIV status, and previous treatment history. Whole blood samples were collected from all participants and couriered within 24 h, at a stable temperature, to a centralised district-level state laboratory dispatch, at which the samples were processed and refrigerated. From these National Health Laboratory Services laboratories, the samples were then couriered to the National Institute for Communicable Diseases incidence laboratory in Johannesburg, South Africa.
Parallel testing by two rapid HIV point-of-care tests was performed on each participant (from among ABON HIV 1/2/O Tri-Line HIV Rapid Test Device [ABON Biopharm, Hangzhou, China; 99·6% sensitivity and 100% specificity]; First Response HIV 1-2.O Card Test [Premier Medical Corporation, Gujarat, India; 99·2% sensitivity and 100% specificity]; and Toyo Anti-HIV 1/2 Test [Turk Lab, Izmir, Turkey; 99·2% sensitivity and 100% specificity]). Discordant results were resolved by a laboratory ELISA conducted by the National Health Laboratory Services.
Among participants who were HIV positive, viral load was quantified by use of 1 mL plasma on the COBAS AmpliPrep/COBAS TaqMan HIV-1 Test version 2.0 instrument (Roche Diagnostics, Mannheim, Germany) for automated extraction, amplification, and detection. The dynamic linear range of the assay is 20–107 copies per mL with dual-target detection of the gag and long terminal repeat regions to ensure broad genotype inclusivity.
Statistical analysis
All analyses were performed in R (version 3.6.3), with a combination of standard packages, including glm, and custom-developed code. Listwise deletion of missing values was performed, as, on review of the testing process, laboratory results could reasonably be assumed to be missing completely at random (≤5% missing). Data were generally analysed as is (ie, self-weighting), and without stratification due to the small sample size. For prevalence and incidence estimation, bootstrap percentile 95% CIs were calculated, with 50 000 bootstrap samples drawn to approximate the survey design (ie, sampling chains per district until district sample sizes were reached, then sampling people within chains).
Figure 1Methods of analysis used to estimate HIV incidence from the cross-sectional survey data
Most estimation methods require the prevalence of HIV as a function of age as an input, and method 3 requires the prevalence of recent infection by age. Prevalence by age was analysed by use of binomial regression, with each model form guided by statistical goodness-of-fit Akaike information criterion values and likelihood ratio tests.
Method 1 analysed long-term mean incidence. A rough estimate of mean historical HIV incidence was produced by considering a model world, defined to be one in which HIV incidence began abruptly at a chosen age and was constant for a predetermined period, while maintaining an excess mortality among individuals with HIV infection.
Method 2 used self-report testing history to analyse incidence. Female sex workers who self-reported their last HIV test as negative provided a virtual cohort of female sex workers, followed up from the negative test result to the survey, when new infections could be counted. To select women into the cohort, we considered a range of upper bounds for the time between the negative test and the survey date.
this method required ascertaining recent infection among individuals who were HIV positive, by any chosen algorithm with sufficiently well estimated values for mean duration of recent infection (MDRI) and false recent rate (FRR). MDRI is the mean time after per-protocol detectable infection for which participants show the markers of recent infection, within some chosen time (T; typically 1–2 years) after infection (after which, recent infection would be considered spurious or false), ideally at least six months.
FRR is the context-specific proportion of participants who are HIV positive who, despite being detectably infected for more than time T, nevertheless show the recent infection markers, ideally close to zero.
In this method, HIV incidence was then estimated from the prevalence of HIV and recent HIV, and the MDRI and FRR.
The LAg Avidity EIA was used as the primary infection-staging test and produced an ODn.
As per convention,
ODn was dichotomised into high (ie, non-recent infection) and low (ie, recent infection) categories by the cutoff of 1·5.
However, serological assays, such as variants on the LAg Avidity EIA, are highly prone to false recent results among individuals who are virally supressed,
and a large FRR leads to imprecise and thus uninformative incidence estimates.
Since many respondents who were HIV positive in this study were virally suppressed (1143 [64·4%] of 1774 participants had viral load ≤1000 copies per mL), mainly due to treatment (1453 [86·8%] of 1673 participants self-reported antiretroviral therapy use in an analysis of data from the same survey, which focused on the HIV cascade
), we adopted the widely used mitigation of including a viral load criterion in the definition of recent infection: cases with a viral load below a set threshold (ie, 1000 copies per mL) were defined as non-recent, independent of serological result.
This mitigation reduced the FRR and, although often neglected, the applicable MDRI.
,
The results are, therefore, not directly generalisable to settings in which a substantial portion of the population initiated antiretroviral therapy and achieved viral suppression within a few months of becoming infected.
,
MDRI (and FRR) estimates were, therefore, adapted for this analysis to account for estimated testing and treatment initiation rates inferred from the intensive follow-up in sex worker programmes, by use of a method that has not previously been proposed (appendix pp 8–10).
The incidence estimates were obtained by three structurally distinct analyses: pooling all data into a single risk group, fitting prevalence of both HIV and recent infection as a function of age and inferring age-specific incidence, and estimating incidence by client sexual violence during the past year and in two groups of districts on the basis of their observed HIV prevalence. There were insufficient data to produce high statistical power for analyses by age or subgroups, which are therefore primarily illustrative.
,
,
and made formally rigorous by Mahiane and colleagues,
the age and time dependence of HIV prevalence, read in conjunction with mortality estimates, places strong constraints on what incidence could have occurred in a given population. Particularly, it has been shown that age-specific and calendar-time-specific HIV incidence is related to HIV prevalence, the rate of change of HIV prevalence, and excess mortality, which are all also age-specific and time-specific.
The rate of change of prevalence should be estimated accounting for the evolution of prevalence both over age and time, that is, by adding how prevalence varies with age to how it varies with the passing of time. In this analysis, we estimated HIV prevalence and its dependence on age from our data. Since data were derived from a single cross-sectional survey, providing no information on the dependence of prevalence on the passing of time, sensitivity analyses were used to consider a plausible range of values for this term. Mindful of the uncertainty around inputs required for this incidence estimation approach, we chose to include only 20–30-year-old female sex workers for this method.
[ad_2]
Source link
