P28.01
Background: Self-weighted sampling designs are not always feasible in sample surveys. Sample weights are calculated and applied to have unbiased estimates. However, it is important to understand the extent of difference between weighted and un-weighted estimates. We examined these differences among men who have sex with men (MSM) participants using data from an integrated bio-behavioral survey conducted in 2009 in four districts of Andhra Pradesh, India.
Methods: Two-stage time location cluster sampling approach was used to recruit 1,608 (around 400 from each district) men aged ≥18 years who had sex with another male in one month prior to survey. Sample size was calculated based on 95% confidence level, 90% power, and design effect=1.7. Consented participants provided behavioural data in structured questionnaires and blood and urine specimens. Specimens were tested for HIV and syphilis. Multivariate analysis was performed to determine correlates of CCU and HIV prevalence adjusted for socio-demographic variables.
Results: Weighted and un-weighted estimates of consistent condom use (CCU) with regular male partners, paying male partners, and casual male partners, HIV prevalence, and syphilis prevalence varied across the four districts. The absolute difference between the weighted and un-weighted estimates ranged from 0.7% to 13.2% for CCU and from 0.2% to 5.3% for HIV and syphilis. Z-scores show a significant difference between weighted and un-weighted estimates in many cases. Weighted and un-weighted multivariate analysis shows opposite trends in the adjusted odds ratio (AOR) in few cases and relatively large differences in the significance level (p-value).
Conclusions: Weighted analysis for a probability-based sample survey provides unbiased population estimates whereas un-weighted estimates are representative of the sample. If possible, a self-weighted sampling design should be adopted, allowing for unbiased estimation for the sample and population.
