Frailty Dynamics from Midlife to Late Life: Methodological and Substantive Insights from Simulation and Cohort Studies

Abstract

Frailty is a common age-related condition characterized by loss of physiologic reserve and increased vulnerability. Since it was first proposed in the geriatrics literature, frailty has been conceptualized as a dynamic characteristic; existing evidence supports that frailty is progressive and that frailty trajectories are heterogeneous across individuals. Assessment of long-term frailty trajectories is necessary to understand the natural course of frailty, and repeated frailty assessments may improve risk stratification over single-time measures. However, assessment of long-term frailty outcomes is challenging due to data requirements and methodological challenges involving missing data and accounting for mortality during follow-up. Further, little is known about the early stages of frailty that may emerge starting in midlife and how frailty evolves into late life. This dissertation aims to address both of these challenges by applying rigorous statistical methods and utilizing the rich data available in the Nurses’ Health Study (NHS) cohort to improve the evidence on long-term frailty trajectories beginning in midlife. In Chapter 1, we focus on methodological challenges common in epidemiological aging research. Longitudinal studies in older adults are often subject to missing data as well as truncation due to mortality; however, it is not known how commonly used methods for addressing missing data may be impacted by mortality. In a simulation study, we assessed single and multiple imputation methods for addressing missing data under two forms of dropout: missing at random (MAR) dropout intended to mimic study attrition and missing not at random (MNAR) dropout intended to mimic mortality. We found that multiple imputation using linear mixed effects models substantially reduced bias under both dropout mechanisms; however, each mechanism resulted in distinct patterns of residual bias. Single imputation methods may provide low-computational burden alternatives in settings where specific underlying assumptions are met; we discuss these assumptions and research settings in which they may hold. In Chapter 2, we characterized frailty trajectories from 1992 – 2016 in the full NHS cohort, building a 29-item cumulative deficits frailty index (FI) to measure frailty. We used two-stage multiple imputation to address missing FI items and full FI values and inverse probability weighting to account for attrition. We modeled trajectories partly conditional on death to acknowledge mortality and estimate trajectories in the evolving cohort of survivors. We found that frailty increases were linear and were larger in those over 60 than under 60 at baseline. Trajectories over 24 years differed by degree of frailty at baseline, and those who were prefrail had the most rapid frailty increases. However, trajectories did not differ according to the deficits present at baseline, suggesting that frailty increases are not driven by specific health deficits but instead reflect a more global decline in health status. In Chapter 3, we assessed long-term frailty trajectories and their association with mortality in breast cancer survivors in the NHS, finding that trajectories increased linearly as in the general population. At all times, current FI value was strongly predictive of mortality, while historic measures of frailty did not predict mortality. These findings support routine frailty screenings throughout the full cancer survivorship journey. Taken together, this dissertation provides insight into methodological challenges encountered in longitudinal frailty studies and provides examples of how such challenges can be addressed, while improving our understanding of the natural course of frailty over decades in both the general population and an important clinical population.