Identifying Suspicious Respondents in an Online Survey
By Diane Maillot-Tchofo, Data Analyst, Tom Devynck*, Fabienne Le Sager, Director of the Statistics Department, and Louis Marec - Médiamétrie
*Médiamétrie and Toulouse School of Economics
15th edition of INSEE’s Statistical Methodology Days (JMS 2025), November 25–27, 2025
Abstract
The most familiar observational error within surveys is associated with respondents’ inability or unwillingness to provide the correct answer[1]. In this context, the will to correctly estimate digital devices ownership (e.g. TV, smartphone) drove Médiamétrie to develop an ambivalent method to detect neglectful respondents in a Computer-Aided Web Interview (CAWI). We drew from the literature (Laura Gamble, 2023 [2] and Anvita Mahajan, 2023 [3]) works and derived an ambivalent method combining both approaches. Our first approach makes use of the questionnaire’s completion times. The second approach is a two-step clustering algorithm focused on the ownership of digital equipment. AK-Means algorithm was applied on the respondent’s household socio-demographic characteristics.
Then, machine learning models were applied to each cluster to contain the models’ shortcomings.
Our final list of sloppy respondents was obtained by combining the results of the two approaches.
[1] Groves, Robert M., Survey Errors and Survey Costs, John Wiley & Sons, 2005.
[2] Gamble, Laura. Evaluation of Key Performance Indicators for Suspected Interviewer Falsification
from Paradata and Interview Data, Ottawa 2023.
[3] Mahajan, Anvita Anjay. Hybrid Model Using LOF and iForest Algorithms for Detection of Insider
Threats, Authorea Preprints, 2023.