Notice: the WebSM website has not been updated since the beginning of 2018.

Web Survey Bibliography

Title Evaluating a New Proposal for Detecting Data Falsification in Surveys
Year 2016
Access date 06.03.2016
Full text pdf (459 KB)
Abstract

Concern about data falsification is as old as the profession of public opinion polling. However, the extent of data falsification is difficult to quantify and not well documented. As a result, the impact of falsification on statistical estimates is essentially unknown. Nonetheless, there is an established approach to address the problem of data falsification which includes prevention, for example by training interviewers and providing close supervision, and detection, such as through careful evaluation of patterns in the technical data, also referred to as paradata, and the substantive data.

In a recent paper, Kuriakose and Robbins (2015) propose a new approach to detecting falsification. The measure is an extension of the traditional method of looking for duplicates within datasets. What is new about their approach is the assertion that the presence of respondents that match another respondent on more than 85% of questions, what we refer to as a high match, indicates likely falsification. They apply this threshold to a range of publicly available international survey datasets and conclude that one-in-five international survey datasets likely contain falsified data.

The claim that there is widespread falsification in international surveys is clearly concerning. However, an extensive investigation conducted by Pew Research Center and summarized in this report finds the claim is not well supported. The results demonstrate that natural, benign survey features can explain high match rates. Specifically, the threshold that Kuriakose and Robbins propose is extremely sensitive to the number of questions, number of response options, number of respondents, and homogeneity within the population. Because of this sensitivity to multiple parameters, under real-world conditions it is possible for respondents to match on any percentage of questions even when the survey data is valid and uncorrupted. In other words, our analysis indicates the proposed threshold is prone to generating false positives – suggesting falsification when, in fact, there is none. Perhaps the most compelling evidence that casts doubt on the claim of widespread falsification is in the way the approach implicates some high-quality U.S. surveys. The threshold generates false positives in data with no suspected falsification but that has similar characteristics to the international surveys called into question.

This paper proceeds as follows. First, we briefly review the problem of data falsification in surveys and how it is typically addressed. Second, we summarize Kuriakose and Robbins’ argument for their proposed threshold for identifying falsified data and discuss our concerns about their evidence. Third, we outline the research steps we followed to evaluate the proposed threshold and then review in detail the results of our analysis. Finally, we conclude with a discussion of the findings and other ways the field is working to improve quality control methods.

Year of publication2016
Bibliographic typeReports, seminars
Print