Assessing the performance of respondent-driven sampling under (reasonably) realistic conditions.

Respondent-driven sampling (RDS) is a technique for studying hidden or hard-to-reach human populations such as drug injectors, sex workers, and men who have sex with men (three groups that are often at highest risk for HIV). The sample is collected through a form of snowball sampling where current sample members recruit future sample members. Despite being introduced quite recently, RDS has already been used in more than 120 studies around the world. However, little is known about how RDS preforms in situations where it is likely to be used. All analytic results are asymptotic and may not hold for sample sizes that are typically collected. Further, recent results suggest that RDS estimates can have extremely high variance when the social network connecting the hidden population has certain features, features that may be common in real networks. In talk we will review the background behind RDS and then explore its properties in (reasonably) realistic conditions by simulating sampling from real social networks: the high risk heterosexuals from the Project 90 study and the middle and high school students from the National Longitudinal Study of Adolescent Health. The substantive implications of these results will be described.