Multivariate assessment of interviewer-related errors in a cross-national economic survey

Interviewers have long been identified as a source of error in face-to-face surveys. However, previous studies have typically focused on a single source of interviewer-related error and single-country cross-sectional surveys. We extend this literature by investigating the influence of interviewers from multiple dimensions in the Oesterreichische Nationalbank (OeNB) Euro Survey, a cross-national survey conducted annually in ten Central, Eastern, and Southeastern European countries. Using data from ten rounds (i.e., 100 country-years), we first analyze the extent of interviewer variance in financial literacy measures and how these effects compare to other questionnaire items. Building on the previous literature, we also evaluate the stability of these estimates over time and across countries. Second, we apply several data quality indicators on various dimensions of interviewer-related error and investigate country-years with particularly exceptional patterns. Finally, we use a multivariate tree-based outlier detection method (isolation forest) that flags country-years and interviewers with outlying values and combine it with methods from the interpretable machine learning literature to identify the respective exceptional feature values.