A Scalable Approach to Measuring the Impact of Nonignorable Nonresponse with An Application to EMA Data
Modern measurement and collection methods generate a large amount of data that become increasingly instrumental for public health and medicine, social sciences and business. Missing data are ubiquitous in these new types of data, and there is often a strong need to adjust inference for nonignorable data incompleteness. However, unlike in traditional studies, nonignorable missingness in these intensive data poses significant new analytic challenges and calls for more general, flexible and robust methods that are applicable in these studies to quantify and improve the reliability, validity and usability of the collected data. To meet these challenges, we develop principled and parsimonious statistical index measures that are scalable to these new types of data to quantify the reliability and validity of empirical findings, while alternative methods can become computationally prohibitive. We illustrate the method in a dataset collected using Ecological Momentary Assessment (EMA) methods where missing data arise because of study participants'nonresponses to those random prompts on their handheld electronic devices, and discuss the implications of the findings for developing more powerful translational interventions.