Identification of Falsifications in Survey Data

在项目第一阶段的基础上，我们从三个方面扩展我们的分析。首先，考虑到了部分造假的情况，即面试官只伪造了一些面试和/或部分面试。分析了现有的方法在多大程度上仍然适用于这种情况。其次，结合心理学中的动机理论，拓展了对面试官动机的研究。在此理论背景下，提出了关于部分造假的类型和程度的假设。我们将使用一个实验来测试这些假设，并得出更多关注部分证伪的指标。第三，在存在部分证伪的情况下，确定性聚类法的使用似乎不够充分。取而代之的是，考虑了允许将一个面试者的数据描述为两个分布的混合的概率聚类方法。为了克服这些方法的数值问题，采用了优化启发式方法。此外，建模方法扩展到两个以上的组，允许根据动机或经验等因素进行不同类型的证伪。所有结果都被用来改进问卷和调查方法，以通过增强面试者的积极性和增加事后发现的机会来阻止造假。

Building on the first phase of the project, we extend our analysis in three aspects. First, the case of partial falsifications is taken into account, i.e. that interviewers falsify only some interviews and/or only parts of interviews. It is analyzed to what extent the existent methods are still useful in this setting. Second, the study of interviewers’ motivation is extended taking into account also theories of motivation from psychology. Based on this theoretical background, hypothesis about the type and extent of partial falsifications are derived. We will use an experiment to test these hypotheses and to derive further indicators focusing on partial falsifications. Third, in the presence of partial falsifications, the use of deterministic clustering methods does not appear adequate. Instead, probabilistic clustering methods are considered which allow describing the data for one interviewer as a mixture of two distributions. To overcome numerical problems of these methods, optimization heuristics are used. Furthermore, the modelling approach is extended to more than two clusters, allowing for different types of falsifications depending on factors such as motivation or experience. All results are used to improve questionnaires and survey methodology in order to deter falsifications by enhancing interviewers’ motivation and increasing the chance of ex post detection.