Please see the following information on our Symposium at the 2011 Annual Meeting of the Gerontological Society of America in Boston (November 18-22, 2011).
The symposium introduces novel analytic methods to account for competing risk of death. Many longitudinal and time-to-event studies of non-fatal health outcomes in the geriatric population face challenges of non-ignorable missing data and informative censoring resulting from competing risk of death. Standard methods such as Cox models and random effects models (REM) may lead to biased inference because of their unlikely assumption of independence between death and the outcome of interest. The first talk gives an overview of statistical approaches for modeling trajectories of longitudinal data truncated by death, and illustrates how standard techniques such as REM and generalized estimating equations (GEE) may be applied or extended to address different research aims in this setting. The second talk introduces a competing risk model for survival analysis. Instead of treating death as a censored event, the model estimates the cumulative incidence of the event conditional on event-free survival. The third talk discusses joint modeling of both event and event-free death through a mixed likelihood approach, the mixing parameter being the proportion of subjects who experienced the event (vs. died before event onset). This model aims to assess how interventions modify the mixture in terms of both the mixing parameter and the survival time distributions of the event and death. Real data examples are used to illustrate the methods, with the goal of improving our understanding of competing risk and its implications in the prediction and prevention of poor health outcomes among older adults.
Brenda F. Kurland, Fred Hutchinson Cancer Research Center; Laura L. Johnson; Brian L. Egleston; Paula H. Diehr
Diverse analysis approaches for longitudinal data truncated by death are illustrated using cognitive functioning data (MMSE) from the Cardiovascular Health Study. Unconditional models, such as random effects models, average longitudinal responses over the survival distribution and may implicitly impute data beyond the time of death. Fully conditional models stratify by time of death and describe individual trajectories in terms of either aging (age, or years from baseline) or dying (years from death). Modern causal models (principal stratification) describe group differences at one timepoint for a cohort that will survive past a later timepoint. Partly conditional models reflect the average response in survivors at a given timepoint, rather than individual trajectories. Joint models of survival and longitudinal response describe the evolving health status of the entire cohort. Researchers should consider which method of accommodating deaths is consistent with study aims, and analyze data accordingly.
Presentation Slides (PDF)
Sarah D. Berry, MD MPH1,2; Long Ngo, PhD3; Elizabeth J. Samelson, PhD1,2; Douglas P. Kiel, MD MPH1,2
1IFAR, Hebrew SeniorLife; 2Dept of Medicine, Division of Gerontology, BIDMC; 3Dept of Medicine, Division of General Internal Medicine, BIDMC.
Clinical studies of elderly persons are often faced with the difficult problem of how to account for participants who die without experiencing the study outcome of interest. Traditional approaches to describe risk of disease include Kaplan-Meier survival analysis and Cox proportional hazards regression. In these methods individuals who die without experiencing the study outcome are censored, yet censored individuals are assumed to still be at risk of disease. Consequently, these methods overestimate risk. We illustrate this concept by comparing a competing risk approach with a Kaplan-Meier approach to estimate risk of second hip fracture in the Framingham Osteoporosis Study. In this example, Kaplan-Meier survival analysis overestimated the five-year risk of second hip fracture by 37% and the ten-year risk by 75% compared with competing risk estimates. We conclude a competing risk approach should be used to determine risk of disease in studies of elderly persons with long follow-up.
Presentation Slides (PDF)
William Checkley MD PhD, Roy G Brower MD, Alvaro Muñoz PhD for the NIH ARDS Network Investigators Johns Hopkins University
Time-to-event data with two or more types of endpoints are common in epidemiological settings. Instead of treating the times for one of the endpoints as censored observations for the other, we treat competing events as distinct outcomes in a mixture of survival distributions. The model aims to determine if and how the mixture was modified in response to an intervention. Using data from patients with acute lung injury, we compared the effects of liberal vs. conservative fluid-management strategies (LFM vs. CFM) and low vs. traditional tidal volumes strategies (LTV vs. TTV) on the frequency and survival times of two competing events: unassisted breathing (UAB) followed by hospital discharge alive and in-hospital death. We found that LTV lowered the frequency of death (p=0.005), but did not affect time to UAB or death (p>0.4).CFM shortened time to UAB (p<0.001), but did not affect frequency of death or time to death (p>0.2).
Presentation Slides (PDF)