Study data and sampling
The data used in the analysis were obtained from a nationally representative population-based household survey that was conducted in 2017 using a multi-stage stratified random cluster sampling design described in detail elsewhere . A total of 1 000 small area layers (SALs) were used as the primary sampling unit, drawn from the master sample through stratified, disproportionate sampling. The selection of SALs was stratified by province, locality type (urban areas, rural informal and formal areas), and race group. A total of 15 visiting points (VPs) were randomly selected from each of 1 000 SALs, targeting 15 000 VPs. Of these, 12 435 (82.9%) VPs were approached. Among these VPs, 11 776 (94.7%) were valid VPs. A household response rate of 82.2% was achieved from the valid VPs (Simbayi et al, 2019). All consenting members of the selected household formed the ultimate sampling unit.
The survey collected data using a household questionnaire and three age-appropriate questionnaires were administered to consenting individuals. For those younger than 18 years of age, consent was given by parents/guardians and assent by the participant. The interview instruments solicited information among others on socio-demographic characteristics, HIV-related knowledge, attitudes, and behaviours, including questions on HIV testing. The questionnaires were fieldworker administered and electronically captured using CSPro software on Mercer tablets. Fieldworkers also collected dried blood specimen samples from participants using a finger prick.
Fieldworkers also collected dried blood specimen samples from participants using a finger prick. Samples were sent to a centralised laboratory for HIV antibodies testing using an algorithm with three different enzyme immunoassays (EIAs). All samples testing HIV positive during the first two EIAs (Roche Elecys HIV Ag/Ab assay, Roche Diagnostics, Mannheim, Germany and Genescreen Ultra HIV Ag/Ab assay, Bio-Rad Laboratories, California, USA) were subjected to a nucleic acid amplification test (COBAS AmpliPrep/Cobas Taqman HIV-1 Qualitative Test, v2.0, Roche Molecular Systems, New Jersey, USA) for the final interpretation of test results. Testing for exposure to antiretroviral drugs (ARVs) in HIV-positive specimens was performed using High-Performance Liquid Chromatography (HPLC) coupled with Tandem Mass Spectrometry.
The survey protocol was approved by the Human Sciences Research Council (HSRC) Research Ethics Committee (REC: 4/18/11/15), and both the Division of Global HIV and TB (DGHT) and the Center for Global Health (CHG) of the Centers for Disease Control and Prevention (CDC). Ethical clearance was also obtained from the University of KwaZulu-Natal’s Biomedical Research Ethics Committee (BE 646/18). Verbal or written informed consent was sought before undertaking both the behavioural data and blood specimen collection.
The primary outcome variable, the first 90 of the UNAIDS 90–90–90 targets  was defined as people who have been diagnosed HIV positive in the central laboratory and knew their status or were exposed to antiretrovirals, dichotomized as diagnosed and aware of HIV status=1 and diagnosed and not aware of HIV status=0.
Explanatory variables were socio-demographic and HIV related behavioural characteristics. Socio-demographic characteristics included age group (15–19, 20–24, 25–49, 50 years and older), race groups (African and other race groups), marital status (married, never married), level educational qualification (no education/primary, secondary, matric, tertiary), employment (yes, no), and locality type (urban areas, rural informal/tribal areas, rural formal/farms). HIV related behaviour characteristics included condom use last sex act (yes, no), correct HIV knowledge and myth rejection (yes, no), and self-perceived risk of HIV infection (yes, no).
Descriptive statistics were used to summarize the sample characteristics. The Pearson’s chi-square tests were used to test for differences between the explanatory variable and the first 90 target by gender. A series of hierarchical multiple logistic regression models structured by sex (males and females) were fitted, and the estimates of the contributions of each independent variable were computed against the dependent variable in successive models. The best-fitting models with variables that reliably predict the first 90 target were determined by assessing changes in R-squared (R2) values and best predictors by adjusted odds ratios (aOR) with 95% confidence intervals (CIs) and p≤0.05. The ‘svy’ command was used to introduce weights that take into account the complex design of the survey. All data analyses were conducted using STATA version 15.0 (STATACORP, College Station, TX) software.