Investigating the effect of different weighting methods on HIV prevalence estimation in the SABSSM V survey of 2017 in South Africa

Abstract

Introduction: In the health systems, prevalence surveys play a key role in identifying key priority areas. They provide valuable information which is important for monitoring and evaluation of health programs and interventions. The South African Behavioural, Sero-surveillance and Media communication (SABSSM) survey is one of the biggest health surveys in South Africa. However, the SABSSM continues to experience a high number of participants refusing to take HIV tests (non-response). This high non-response can introduce bias in HIV prevalence estimates. Sampling weights are introduced to address bias taking into account non-response. This study was aimed to investigate the effect of sampling weights and explore the use of weights derived from predictors of non- response as an alternative method to reduce bias. Methods: A cross-sectional population-based household survey data analysis was conducted. Stratified multistage cluster random sampling was used. Two sets of HIV prevalence estimates derived from questionnaire sampling weights and HIV data sampling weights were compared. Factors associated with non- response were investigated and used to determine a set of weights. The third set of weights were used to calculate new HIV prevalence estimates. Results: A change in HIV prevalence estimates when using different weights were observed. Using HIV sampling data resulted in a lower HIV prevalence estimate, slightly high SEs and wide 95%CI compared to using questionnaire sampling weights. Factors associated with non-response include mental depression and being sexually active. Third weights resulted in slightly lower HIV prevalence estimates compared to estimates obtained from HIV sampling weights with a lower SE and narrower 95%CI. Conclusion: Using different sampling weights changed prevalence estimates. The study also demonstrate that non-response is not entirely at random.