Respondent-driven sampling (RDS) is a network-based sampling strategy useful for studying hard-to-reach populations, such as low-wage workers. Respondent-driven sampling designs prompt respondents to recruit other members of the population of interest in their social network to the survey. RDS methods can collect large samples from hard-to-reach populations by leveraging social ties within these communities to facilitate recruitment. However, these designs are prone to being affected by many sources of bias, including seed bias–the effect of starting the recruitment chains with a biased convenience sample.

Previous work utilizing RDS to sample low-wage workers has suffered from issues of seed bias, making inference difficult. To address this problem, we propose a new design that collects seeds in a probability sample, and study this design's resilience to network homophily, or the tendency for similar people to cluster within social networks. The structure of this design is novel in its focus on estimation within multiple sub-populations of interest (for example, low-wage industries), and in its formulation of complex constraints imposed on recruitment to limit bias. We study and model the population networks and recruitment sampling, propose a modified estimator, and, via simulation, analyze the validity of inference.

Results indicate that inference in this design is feasible, and that modifications to a popular RDS estimator to account for the sampling constraints improve the accuracy of estimation. While the accuracy of the estimator is promising, further improvements to this estimator and the network generation algorithm are likely necessary to properly assess the validity of inference. These improvements include incorporating the sub-population structure of the sampling more fully into the estimator and implementing non-uniform homophily effects estimation and correction within the estimator.