Adaptive Respondent Driven Sampling of Social Networks: A Simulation-Based Study Using Machine Learning

Dr. Douglas Heckathorn introduced respondent-driven sampling (RDS) in 1997 to survey the social and behavioral attributes of hidden and hard-to-reach networked populations. In conducting RDS, certain norms have emerged, e.g., giving exactly three coupons to each respondent (Goel and Salganik 2010). This norm is problematic, as it often leads to a very clustered network sample, where it would be preferable to have long chains that reach mixing equilibrium deep in the network and beyond the social characteristics of the seeds (Heckathorn 1997). It would be preferable for the researcher to improve the rate at which researchers hand out coupons in relation to the number of interviews the researcher gets, maximizing the efficiency of the RDS. The primary aim of this project is to create a variant of the RDS method that engages machine learning techniques during the sampling process to dynamically adjust the number of coupons given to each respondent, with the aim of producing long referral chains. The secondary aim is to create a variant of the RDS method that will favor efficiency, allowing researchers to maximize the data they collect for the fixed number of coupons they distribute. Here I propose that, in practice, characteristics of the participants be used to build and continuously update predictive models of participants’ propensity to generate future successful referrals. Forecasts of these predictive models can then be used to adjust the number of coupons allocated to each participant to drive the production of long referral chains. Experiments, in the form of simulations, will systematically evaluate the impact of parameter settings on the relative advantage that the adaptive RDSs provide in terms of performance measures compared to classical RDS. In this paper, I report that using an adaptive RDS is successful at producing long referral chains and systematically describe the contextual factors that drive its relative advantage over the classical respondent driven sampling paradigm