Outcome Dependent Sampling Designs for Longitudinal Studies Utilizing Existing Cohort Studies: Quantifying Possible Biases Due to Study Dropout

• Activity Number: 529 - Contributed Poster Presentations: WNAR
• Type: Contributed
• Date/Time: Wednesday, July 31, 2019 : 10:30 AM to 12:20 PM
• Sponsor: WNAR
• Author(s): Melissa Wilson* and Samantha MaWhinney and Jose Castillo-Mancilla and Kristine Erlandson
• Companies: Colorado School of Public Health - Denver|Anschutz, Dept. Biostatistics and Informatics and University of Colorado Anschutz Medical Campus and University of Colorado - Denver|Anschutz, Dept. of Medicine, Division of Infectious Diseas and University of Colorado - Denver|Anschutz, Dept. of Medicine, Division of Infectious Diseas
• Keywords: Outcome-dependent sampling; Missing data

Abstract

Prospective clinical trials, case-control, cohort studies and electronic medical records (EMRs) gather clinical data that is often paired with biospecimens. An increasing amount of statistical and epidemiological research is aimed at utilizing existing databases to address new research questions and/or inform study designs beyond the original aims. This outcome dependent sampling approach, informed by existing data (outcome, exposure or related covariates), will have the greatest benefit when the outcome is rare, conservation of sample is desired or assay cost precludes a large sample size. ODS design methods have been motivated by longitudinal studies where loss to follow-up and death are rare, or ignored. However, longitudinal studies (clinical cohorts in particular) are often plagued by dropout which, if ignored in the analysis, may mask associations. If it is plausible that data is missing not at random, non-standard statistical methods and sensitivity analyses should be considered to avoid underestimating the study effect. Here, we investigate the effects of missing data on ODS analyses through data simulated to represent drop-out patterns common to longitudinal cohort data.