Identification of partial falsifications in survey data

摘要

Survey data allow constructing indicators, which differ for real and falsified interviews. It could be shown in previous research that applying cluster analysis to a set of indicators helps to identify potential falsifications at the interviewer level. The current work analyzes to what extent a differentiation remains feasible when interviewers falsify only a part of their interviews. An experimental dataset containing both real and falsified data for each respondent allows to construct bootstrap samples with the required properties, i.e., a predefined share of falsified interviews for those interviewers doing (partial) falsifications. The bootstrap approach allows measuring how robust the method works when the share of falsified interviews per interviewer decreases while taking into account also other relevant factors such as the total number of interviews per interviewer, the share of falsifiers, and the number of interviewers. The presented results demonstrate that the method loses power with decreasing share of falsifications, but remains a valuable tool for ensuring high data quality in surveys.