Assessing Consistency of Respondent-driven Sampling Estimators by Using Repeated Surveys among People Who Inject Drugs (PWID) in New Jersey

Respondent-driven sampling (RDS) is widely used to sample populations with higher risk of HIV infection for whom no sampling frames exist. However, few studies have been done to assess the consistency of RDS estimators in real world settings. We conducted an assessment study on the consistency of RDS estimators using data from the National HIV Behavioral Surveillance – People Who Inject Drugs surveys in Newark, New Jersey from 2005 to 2018. Population parameter estimates based on RDS-I, RDS-II, Gile’s SS, and HCG were compared longitudinally and cross-sectionally. Population homophily statistics and differential recruitment statistics were estimated and compared. Convergence plots were used for RDS diagnosis. Sensitivity analyses were conducted on population size estimates and seeds biases. By comparing time-insensitive population parameters and population homophily statistics estimated by four RDS estimators, the study found that RDS-II and Gile’s SS could provide longitudinally and cross-sectionally consistent estimates and population homophily statistics on gender and sexual orientation. Cross-sectional comparison of time-sensitive population parameter estimates also supported the consistency of RDS-II and Gile’s SS. However, RDS-I and HCG did not perform well in those comparisons. In conclusion, RDS estimators may not address all inconsistencies, but RDS-II and Gile’s SS are recommended to weight RDS samples given enough consistency was observed in them.