Online Program

Some diagnostics for respondent-driven sampling: Illustrations and results from 12 studies in the Dominican Republic

*Krista Jennifer Gile, University of Massachusetts 
Matthew Salganik, Princeton University 
Lisa Johnston, Tulane University/UCSF 

Keywords: respondent-driven sampling, diagnostics, snowball, link-tracing, sampling

Respondent-Driven Sampling is now a widely-used method for collecting data on hard-to-reach human populations that are socially connected. While this method is powerful in collecting large diverse samples in many settings, valid inference is difficult and relies on many (sometimes dubious) assumptions. Much recent literature has highlighted sensitivities of RDS estimates to violations of these assumptions, however there are few available tools for evaluating those violations in data. This paper introduces several such approaches, and illustrates their application in the context of 12 parallel RDS studies. Many of these diagnostics take advantage of specific features of RDS studies that are not typically utilized: information about the time sequences of responses, the second contact with respondents who return to collect recruitment incentives, and the multiple initial samples used to begin the respondent-driven process. We use some data commonly available in RDS studies, as well as novel types of data collected in the interest of allowing for these diagnostics. We use these diagnostics to assess RDS assumptions in data collected among drug users, female sex workers, and men who have sex with men in four cities in the Dominican Republic (total n = 3, 862). Unlike a single study with a single population, the 3 × 4 design allows us to identify repeated patterns. We conclude with a summary of our assessment of the validity of assumptions in the 12 sites considered, and with recommendations for data collection and analysis, both during and after the data collection.