Statistical modeling using mouse movements to model measurement error and improve data quality in web surveys

该项目的总体目标是通过检测工具中的问题和参与者个人的问题，改进网络调查数据的收集。在第一个资助期内，我们已根据一个案例研究，证明了鼠标跟踪paradigm在这方面的有用性，并开发了先进的分析技术。在本延续中，我们的目标是进一步发展和扩展这些方法，并评估其在该领域的鲁棒性。我们将采取双管齐下的跨学科方法，建立在我们以前的工作和经验。工作包的第一股涉及的多维时间序列，使互动paradigm分析的统计方法。在这里，我们将扩展我们的分析方法，不仅仅是鼠标移动，以处理所有可观察到的用户交互（包括但不限于键盘输入，触摸屏数据和其他设备信息），应用深度学习来发现尚未发现的功能，这些功能超出了我们手动定义的指标，并调查用户组之间的异质性。作为第二个独立但紧密相连的调查研究链，我们将在该领域应用和评估我们的旧方法和新开发的方法，研究它们如何推广到更大，更多样化的样本，并指导方法开发以满足调查从业者的要求。此外，我们还将评估我们的方法不仅能检测出棘手的问题，还能检测出粗心和欺诈的受访者，从而解决调查提供商面临的两个相关问题。最后，我们将通过开发和测试同意程序，为调查从业人员提供收集这类数据的最佳做法。与此同时，我们将创建数据收集和处理基础设施，以支持我们的工作，我们将向研究人员提供开源软件，并添加文档，教程和学习材料。在这两个方面，我们将与维护调查小组的项目合作伙伴密切合作，并与调查提供商建立合作关系，以确保我们的计划可行。我们将共同创建一个新的数据集，这将有助于我们和其他人未来对paradox的调查。

The overarching goal of this project is the improvement of web survey data collection by detecting issues in the instrument and on part of individual participants. In the first funding period, we have demonstrated the usefulness of mouse-tracking paradata for this purpose based on a case study, and developed sophisticated techniques for their analysis. In this continuation, we aim to further develop and extend these methods, and evaluate their robustness in the field. We will take a two-pronged interdisciplinary approach, building on our prior work and experience. The first strand of work packages concerns the statistical methodology for analyzing the multidimensional time-series that make up interaction paradata. Here, we will extend our analysis methods beyond mouse movements alone to handle all observable user interactions (including, but not limited to, keyboard input, touchscreen data, and further device information), apply deep learning to uncover yet-undiscovered features that go beyond our manually defined indicators, and investigate heterogeneity between user groups. As a second, independent but closely linked, survey research strand, we will apply and evaluate our old and newly developed methods in the field, investigate how they generalize to larger and more diverse samples, and guide the methods development to meet the requirements of survey practitioners. In addition, we will assess our methods’ potential to detect not only hard questions, but inattentive and fraudulent respondents, addressing two pertinent issues facing survey providers. Lastly, we will supply survey practitioners with best practices for collecting this class of data, through development and testing of consent procedures. In parallel, we will create data collection and processing infrastructure to underpin our efforts, and that we will make available to researchers as open-source software with added documentation, tutorials and learning materials. Throughout both strands, we will work closely with project partners who maintain survey panels, and have established collaborations with survey providers to ensure that our plans are feasible. Together, we will create a new dataset that will aid our and others’ future investigations of paradata.