Mode Effects in American Trends Panel: Bayesian Analysis of a Cross-classified Item-person Mixed Model
Jeff Gill
Washington University
Stas Kolenikov
Abt SRBI
Kyley McGeeney
Pew Research Center
Nancy Mathiowetz
Independent Consultant
American Trends Panel is a probability panel with RDD recruitment developed by Pew Research Center and Abt SRBI. Over the life of the panel, surveys have been conducted primarily via web mode, with mail mode for those who do not have access to the Internet or do not provide an email address. We analyze the results of the July 2014 wave (Wave 5) that included a comprehensive, large-scale mode-of-interview experiment that randomly assigned web respondents to telephone and web modes, with approximately 1,500 respondents in each mode. To quantify the contributions to the mode effects of the different question characteristics, we build a cross-classified location-scale random effects model with effects of person and question characteristics to identify the properties of survey questions that make them susceptible to mode effects, as well as the demographic groups that tend to exhibit mode effects. The model was estimated by Markov chain Monte Carlo computational Bayesian methods using a combination of R and JAGS packages.