Online Program

Return to main conference page
Saturday, May 19
Data Science
Data Science in Health
Sat, May 19, 1:15 PM - 2:45 PM
Grand Ballroom G

A Proposed Framework to Assess the Sensitivity of Network-Based Estimands to Non-Ignorable Non-Response, for Networks Ascertained With Non-Ignorable Sampling (304633)

*Kenneth J Wilkins, National Institutes of Health, National Institute of Diabetes & Digestive & Kidney Diseases 

Keywords: informatively missing data, latent space models, missing not at random, network analysis, network data, non-ignorable, sensitivity analysis.

Organizational data scientists must account for network dependencies more often than not. As an example, research funders seek input from a broad (networked) community of stakeholders through public workshops or solicit comments under an official Request for Information (RFI). Such surveys yield a range of responses from which funders must extract opinions, either to judge funding priorities in future research support, or to weigh decisions concerning proposed changes in policy. As one instance of this latter case, RFI NOT-OD-17-015, "Strategies for NIH Data Management, Sharing, and Citation" presumably garnered a vigorous response from stakeholders – those who either felt strongly for increased requirements for data sharing, or were incentivized to maintain the status quo (e.g., extending exclusive access to data they helped collect). In addition to this suspected association between the probability of response and content of such responses, there is also the unavoidable fact that stakeholders often influence others to share their views in a manner that's likely disseminated along social and professional networks. Thus, the collection of responses obtained ought neither to be deemed to reflect a random sample of stakeholders, nor would collected responses be reasonably assumed to yield unbiased estimates of stakeholder opinion. We explore the intersection of non-ignorable sampling with non-ignorable non-response, and how adaptations of recently proposed network methods may provide a framework for judging the extent to which conclusions about elicited stakeholder input are sensitive to unverifiable assumptions about these inherent issues. We adapt the framework for valid inference from non-ignorable network sampling mechanisms (Lunagómez & Airoldi, 2014) in conjunction with variational Bayes fitting of latent space joint models for network views (Gollini & Murphy, 2014); we illustrate it using plasmode data of surveys on a trans-NIH program's funding priorities.