Using Double Machine Learning to Understand Nonresponse in the Recruitment of a Mixed-Mode Online Panel

This article introduces the use of double machine learning (DML) in survey statistics to understand nonresponse in a high-dimensional setting. It explains how traditional machine learning methods are good for prediction but can produce biased estimates of causal relationships, while DML can provide approximately unbiased estimators by effectively handling numerous potential confounding variables. The authors apply DML to analyze nonresponse in the GESIS panel's welcome survey, identifying significant socio-demographic factors, mode choice, and interviewer ratings that influence dropout rates, and discuss the implications for preventing nonresponse in panel surveys. The paper also highlights the limitations of DML, such as the reliance on the unconfoundedness assumption and the inability to handle missing data or measurement error, while suggesting potential future applications in various social science fields.