Online and Network Sampling Methods for Survey Research: A Total Survey Error Framework

Abstract

Survey research faces an existential crisis. Response rates for major surveys have declined steadily over several decades, and the costs associated with maintaining high-quality probability sampling panels have become prohibitive for all but the most well-resourced organizations. Populations of increasing scientific and policy relevance, such as individuals experiencing homelessness and those who use drugs, are systematically excluded from address-based sampling frames. These challenges necessitate the development of non-probability methods that support valid population inference. This dissertation extends the Total Survey Error (TSE) framework to evaluate two alternatives: population estimates derived from algorithmically targeted online samples and survey data collected through network sampling and aggregated relational data. Application of the TSE framework to these methods demonstrates that each introduces distinct error sources that require careful attention to design and statistical adjustment. Social media recruitment via advertising platforms can rapidly yield large, low-cost samples; however, algorithmic optimization for engagement introduces selection bias, leading to samples that overrepresent certain demographic groups, such as college graduates. Propensity score adjustment combined with calibration using a probability sampling reference survey can substantially correct these biases, but its effectiveness depends on the relationship between selection mechanisms and the outcome of interest. Adjustment performs well for time-invariant measures, such as chronic health conditions, but less effectively for time-variant measures, such as health behaviors that may correlate with social media use and information exposure. Survey mode, whether interviewer-administered or self-administered, shapes measurement error in population-specific ways. Interviewer presence improves response rates but can influence response content, and the direction of this influence depends on the social expectations associated with the population. Among individuals experiencing homelessness, social desirability bias appears to reverse: the socially expected role of demonstrating need may create pressure to report worse health when an interviewer is present. These findings support a hybrid approach in which interviewers administer questions where completeness is paramount, while respondents complete sensitive questions independently. Network-based data collection methods, particularly those using aggregated relational data, offer a cost-effective approach to characterizing the social networks of hidden populations. When combined with social media recruitment, these methods can produce diverse samples at a competitive cost, provided that researchers attend to design decisions and post-adjustment strategies. The resulting network data reveal that participants maintain broad acquaintance ties to other people who use drugs but report far fewer trusted contacts, a distinction with direct implications for how harm reduction resources might be disseminated. However, the overrepresentation of highly active users remains a limitation. As traditional probability sampling becomes increasingly difficult to sustain, online and network sampling methods are likely to continue to grow in adoption. The central question is no longer whether researchers will use these methods, but how to use them effectively and responsibly. This dissertation provides practical tools and conceptual clarity to support this effort, while recognizing that translating the TSE vocabulary into practice requires ongoing methodological development across survey methodology, demography, public health, and the social sciences.