Analytical Approaches to Targeted Sub-sampling Designs with Longitudinal Continuous Responses
Pre-existing data from cohort studies, administrative databases, and electronic medical records are increasingly being used to address novel study hypotheses. While the resources may contain longitudinal outcome and covariate data, often an important exposure or confounder variable must be ascertained retrospectively, and the associated costs compel judicious sampling strategies. We propose sampling strategies where exposure information is only obtained on a subsample of subjects who are chosen on the basis of the already observed longitudinal outcomes. We will describe these designs in detail.  We will also discuss several analytical approaches including an ascertainment corrected likelihood (ACL) and multiple imputation  extensions.  We examine scenarios under which imputation approaches improve estimation efficiency over maximum ACL analyses and importantly the extent to which the targeted sampling designs improve estimation efficiency when multiple imputation is the default analytical approach.

Bio: Jonathan Schildcrout is an associate professor in the Department of Biostatistics at the Vanderbilt University School of Medicine.  His methodological research interests involve longitudinal data with specific emphasis on unplanned and planned biased selection mechanisms at the level of subjects and at the level of observation times within subjects.  He has worked on a variety of collaborative research projects including the health effects of air pollution on children with asthma, the evaluation of a pre-emptive genotyping program for personalizing cardiac medication treatments, the examination of the mechanisms by which high body mass impacts kidney function after thoracic surgery, and the role of social and environmental factors on outcomes and readmission risk in patients with acute coronary syndrome.