Spatial Autocorrelation and Qualitative Sampling: The Case of Snowball Type Sampling Designs

Hidden populations are defined as subsets of a larger population that are hard to target with traditional (e.g., random) sampling methods. For qualitative research, difficulties of achieving a good sample could include the time of day surveys are conducted, the safety of interviewers in areas with high crime rates, or the unwillingness of members in a hidden population to interact with researchers. Various chain-driven methods, such as snowball sampling (SS) and respondent-driven sampling (RDS), have been developed as techniques to reach hidden populations. Such methodologies have been implemented in previous research for investigations into the networks of people associated with illicit drug use and other risky behavior. To date, some of these studies have considered the contribution to variance inflation attributed to the effects of social network (SN) autocorrelation but not to spatial autocorrelation. This article implements a probabilistic simulation based on two RDS network data sets: one from Rio de Janeiro and another from the Colorado Springs metropolitan region. The network configurations are studied with respect to their associated geographic landscapes and a set of selected census variables. The results of the simulations demonstrate a lack of bias on the mean of the demographic variables and impacts on sample-to-sample variability attributed to both SN autocorrelation and spatial autocorrelation in the presence of other sources of excess variance. Findings reported in this article offer insights into designing future studies using network-based sampling strategies.