Interactive visualization to facilitate monitoring longitudinal survey data and paradata
Jul 26, 13:45
In the age of paradata, survey researchers can be overwhelmed by the amount and variety of information available to inform decisions during data collection. This is especially true in the case of longitudinal surveys, which may contain survey data and paradata from multiple past waves. These large quantities of data can present challenges to project staff responsible for analyzing and extracting actionable information from the data. Furthermore, when a survey management strategy contains elements of adaptive, responsive, or tailored designs, the survey team must rely on monitoring critical-to-quality indicators during data collection to minimize total error across multiple data sources while efficiently utilizing project resources. These indicators, based on survey data and paradata, inform decisions which may involve switching to an alternative protocol, halting the pursuit of some or all cases, initiating additional contact attempts, or fielding other interventions.
To mitigate challenges presented by decision-making in a data-rich context, visualization can serve as a valuable tool to express data as readily interpretable media. Furthermore, interactive visualizations which allow the end-user to easily and rapidly produce ad hoc reports may provide additional utility to project staff. In this presentation, we describe the process of designing a suite of visualization tools called the Adaptive Total Design (ATD) Dashboard. This system is designed for monitoring and visualizing data from multiple sources to track experimental, multimode, and longitudinal survey designs in near-real time. We use as our case study in this presentation the National Longitudinal Study of Adolescent to Adult Health (Add Health). By employing an extensible web application framework for R (namely, Shiny), the ATD platform standardizes the approach to, and production of, easily interpretable data visualizations and reports. Users are presented with an array of display options and mechanisms for categorizing, subsetting, and aggregating data. For longitudinal studies, graphs that overlay projections, display survey outcomes from prior rounds, and incorporate model-derived estimates are available. Given that data inputs may be derived from disparate systems and may exist at multiple units of analysis (e.g., sample member, interviewer, day, PSU, etc.), we have constructed a data taxonomy embedded into display and selection mechanisms to allow only logical instantiations. Critical-to-quality indicators are prominently displayed while extraneous information is minimized, using best practices of visual design.