'Loss-to-Follow-Up' in Observational Clinical Cohorts: Longitudinal Effect Estimation in the Presence of Informative Monitoring and Non-Monotone Missingess

• Activity Number: 324
• Type: Topic Contributed
• Date/Time: Tuesday, July 31, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Maya Petersen*+
• Companies: University of California at Berkeley School of Public Health
• Keywords: causal inference ; missing data ; loss to follow up ; censoring ; clinical cohorts

Abstract

In observational clinical cohorts subjects are typically seen at irregular and informative intervals. While "loss-to-follow-up" is common in the sense that many patients fail to return to clinic and have unknown outcomes, no clear right censoring time is measured. Common practice when estimating the causal effects of longitudinal treatments with such data is to artificially define a right censoring time based on some minimum duration during which no clinic visits were recorded. Data between observation times are imputed, and data are truncated at the assigned right censoring time, with future measurements discarded. This approach can result in bias (due to reliance on an imputed outcome and failure to address the underlying informative monitoring process) and loss of efficiency (because data observed after a subject's assigned censoring time are not used). We describe alternative approaches to estimation of longitudinal causal effects using data subject to informative non-monotone missingness, based on defining and estimating causal parameters indexed by an intervention on the underlying monitoring process. Results are illustrated using HIV clinical cohort data and simulations