Combining Probability and Nonprobability Web Surveys for Studying Voter Turnout

Web surveys conducted via volunteer panels have become a popular means of collecting survey data for studying politics and political decisions. As no interviewers are involved, these panels offer a distinctively cheaper and faster means of data collection compared to more traditional modes. However, because volunteer online panels do not utilize probability sampling, concerns prevail that such panels are inferior to their probability-based counterparts in terms of meeting the high quality standards expected of scientific research. Indeed several comparative studies have shown that samples drawn from volunteer panels are less representative than samples drawn using probability-based methods. Given that the popularity of volunteer samples is unlikely to abate, it is prudent to consider ways of integrating them with probability samples in order to overcome their weaknesses. We contribute to this idea by implementing a Bayesian approach that integrates both sampling types. We demonstrate the method using nationally-representative probability-based samples collected in Germany alongside eight concurrently collected non-probability samples from different survey vendors. To evaluate the method, the error properties of voter turnout estimates derived from the integrated samples are compared with estimates derived from probability-only samples. We show that the integrated approach yields estimates with smaller mean-squared error (MSE) compared to the standard (non-integrated) approach. We also explore the cost implications of this approach to determine whether combining both data sources makes sense from a cost perspective.