RESUMEN
BACKGROUND: Respondent-driven sampling (RDS) is a successful data collection method used in hard-to-reach populations, like those experiencing or at high risk of drug dependence. Since its introduction in 1997, identifying appropriate methods for estimating population means and sampling variances has been challenging and numerous approaches have been developed for making inferences about these quantities. To guide researchers and practitioners in deciding which approach to use, this article reviews the literature on these methodological developments. METHODS: A systematic review using four electronic databases was conducted in order to summarize the progress of RDS inference over the last 20 years and to provide insight to researchers on using the appropriate estimators in analyzing RDS data. Two independent reviewers selected the relevant abstracts and articles; thirty-two studies were included. The content of the studies was further categorized into developing and evaluating RDS mean and variance estimators. RESULTS: The population mean estimator RDSIEGO and the sampling variance estimators associated with tree boot strapping were identified as promising methods as the most robust population mean and variance estimate, respectively; as these estimators rely on a fewer assumptions. CONCLUSIONS: RDS holds substantial promise as a sampling method for understanding populations at high risk. The varied approaches to inference with RDS data each rely on different assumptions, but some require fewer assumptions than others and provide more robust and accurate inferences, when their corresponding assumptions are met.