Department of Biostatistics Seminar/Workshop Series
When Death was Not the Outcome of Interest: Communicating a match between longitudinal data analysis methods and research aims when follow-up is truncated by death
Laura Lee Johnson, PhD
Statistician, Office of Clinical and Regulatory Affairs, National Center for Complementary and Alternative Medicine (NCCAM)
NIH US Department of Health and Human Services
Wednesday, May 21, 1:30-2:30pm, MRBIII Conference Room 1220
Intended Audience: Persons interested in applied statistics, statistical theory, epidemiology, health services research, clinical trials methodology, statistical computing, statistical graphics, R users or potential users
Distinguishing data missing due to death and other forms of nonresponse has been mentioned in many contexts. Without care, standard statistical analysis methods may yield answers that are not in line with study aims. Every model provides an answer, but is it to the question patients, clinicians, and researchers want answered? Unconditional models, such as random effects models, may implicitly impute data beyond time of death. Fully conditional models are effective for describing individual trajectories in terms of aging or dying. Partly conditional models reflect average response in survivors at a given time point rather than individual trajectories. Joint models of survival and response describe the evolving health status of the entire cohort. None of the models are inherently wrong, but each one may be the wrong way to accommodate deaths consistent with research aims and clinically important questions. Those reading research articles and analysis plans need to be vigilant of the subtle impact analysis differences may have on study results and their relevance in answering patient questions.