|
Abstract:
|
Respondent-driven sampling (RDS) is a common method for sampling from and estimating networks of hidden populations. RDS starts from an initial sample of subjects who are provided with coupons to distribute among their contacts in the population of interest. Coupon recipients can redeem their coupons with study coordinators for a small reward at which time their information is collected and they are given coupons with which to recruit additional participants. This process continues until budget, sample size, or other stopping criteria are met. Standard RDS uses a fixed number of coupons, incentives, and the same prompt when recruiting new subjects. Thus, standard RDS fails to utilize accumulating information about network structure and dynamics which may decrease efficiency, incur excess cost, and negatively affect policy decisions based on collected data. We propose to use reinforcement learning (RL) to adaptively tailor the number, type, and/or participant prompts used in RDS. We show that optimizing coupon allocation under partially observable Markov decision process model improves efficiency relative to classic RDS.
|
