Robustness of clustering methods for identification of potential falsifications in survey data
Falsifications of survey data might result in specific statistical properties of the generated data differing from those of the surveyed population. Clustering methods have been proposed to identify potential falsifications based on such indicators. As any statistical procedure, the classification might entail errors, i.e. misclassification of ... honest interviewers as potential falsifiers and failing to identify all falsifications as such. Typically, the robustness of a statistical classification procedure is studied using a large number of problem instances with known allocation to the groups. However, given the sensitivity of falsifications in survey data, the access to datasets comprising correctly identified falsifications is very limited. Consequently, a bootstrap based approach is introduced and applied to assess the clustering method. This approach also allows modifying settings such as number of interviews per interviewer or share of falsifications in the dataset and to study the impact of these settings on the quality of the assignments. Results based on a small real dataset with identified falsifications are reported.