Respondent-driven Sampling in the Real World

In many countries, HIV/AIDS infections are concentrated in 3 high-risk groups: injection drug users, sex workers, and men who have sex with men. Accurate information about disease prevalence and risk behaviors in these groups is therefore critical for designing and evaluating HIV prevention programs. Respondent-driven sampling (RDS) is a relatively new method that allows researchers to collect such information about “hidden” or “hard-to-reach” groups, and the work of McCreesh and colleagues provides an important contribution to our understanding of RDS. An RDS sample is collected through a peer-to-peer recruitment process, akin to “snowball sampling” and “link-tracing sampling.” Once these data are collected, they might not be directly representative of the target population, and thus RDS also provides researchers a set of statistical procedures to adjust the observed data, with the hope that these adjustments will result in estimates that are more reflective of the target population. Given the importance of the public health problem and the limitations of available alternatives, respondent-driven sampling has been rapidly adopted by the international public health community. Despite its widespread use, however, little is currently known about the actual (as opposed to the theoretical) performance of RDS, and some recent research suggests cause for concern. Given this, many researchers may be left to wonder: How well does RDS actually work in the real world? Our limited ability to address this question is not for lack of effort; in the past, researchers have taken several approaches to the problem, and the paper of McCreesh et al is a new and important contribution to this stream of research. RDS as a method of data collection—what I will call “RDS sampling”—and RDS as a method of data analysis—what I will call “RDS inference”—were both introduced by Heckathorn in 1997. Although RDS sampling has remained largely unchanged, RDS inference has been an area of active research resulting in the RDS-I estimator, the RDS-II estimator, the RDS-MR estimator, the RDS-SS estimator, estimators currently in development, and several approaches to variance estimation. Previous efforts to assess the performance of RDS generally fall into 3 categories: (1) analytic results, (2) simulation studies, and (3) studies using data from hidden populations. Each of these approaches has strengths and weaknesses, but it can be helpful to consider them in terms of the trade-off between precision and relevance. Some approaches—analytic results and simulation studies—allow for precise, definitive conclusions (eg, under these 5 conditions, this estimator has these 3 properties), but these conclusions may be irrelevant to actual RDS studies because they could depend on assumptions that bear little relationship to what actually happens in practice. On the other hand, studies involving data from hidden populations, while certainly relevant, rarely yield precise, unequivocal results because the underlying truth is not known. The work of McCreesh et al, building on innovative work by Wejnert and Heckathorn, attempts