Evidence is accumulating that cognitive abilities are inversely associated with mortality, and this association seems to be present irrespective of study population, cohort, and the cognitive tests used. Cognitive abilities are not traditionally considered in epidemiology, however, and their additional value to the prediction of mortality is not clear. On the other hand, cognitive decline is viewed as one of the multifaceted aspects of the aging process, but few studies have related characteristics in the changing process of cognition to mortality. The examination of these problems largely has been impeded by methodological difficulties. This project will examine several novel methods to investigate the relations of cognition (in particular, episodic memory and verbal ability) and mortality using longitudinal data in a national sample of U.S. adults aged 70 years and older. These methods will be employed in a progressive order and will include latent curve models, mixed-effects change point models, joint growth and survival models, traditional survival analysis, and recursive partitioning techniques (survival trees and random forests).
The project will examine cognitive change in the oldest population cohort and relate it to mortality. The increasing proportion of this group in the population is impacting many social and economic issues. The substantive results from this project may be especially informative for policy makers in considering hospice care plans. Methodologically, this project will compare alternative approaches to analyzing longitudinal and survival data and will employ exploratory data mining in survival analysis. The methodological finding may have broad applications for many fields in the behavioral and health sciences. As a Doctoral Dissertation Research Improvement award, support is provided to enable a promising student to establish a strong, independent research career.
This project examined the interrelations between cognition and mortality in a nationally representative sample of the U.S. population aged 70 years and older (N=5,963), using publicly available data from the Asset and Health Dynamics of the Oldest Old Study. Over seven waves of longitudinal data collection between 1993 and 2006, 73% of the study participants passed away. The first half of the project focuses on modeling longitudinal trajectories of cognitive performance in domains of episodic memory and verbal abilities. Three different longitudinal models were fit to the data: (1) latent curve model, which does not use death information at all; (2) change point model, which hypothesizes a sudden drop in cognitive abilities within a few years prior to death, the so-called "terminal decline"; and (3) joint growth and survival model, which adjusts longitudinal trajectories for individual distance to death. Through the comparison of results from different models, we found that the shape of the trajectories could be biased if death attrition was neglected. The rate of cognitive decline was estimated to be faster after adjusting for the attrition. This implies that, when studying an aged population, plotting available data only can be an overly optimistic depiction of the true process of aging. The second half of the project focuses on the prediction of survival using both the traditional method--Cox proportional harzards model, and the exploratory data mining technique--survival trees and random forests. Potential predictors considered are demographic characteristics (including age, gender and education), health and disease variables, functional variables, and cognitive performance (initial level and change rate). Results from the two approaches consistently demonstrated that cognition was a useful predictor of mortality above and beyond demographics, comorbidities, risk behaviors and functional status. The prognostic value of cognition can be potentially useful in informing clinical and policy decision making. This project illustrated an application of data mining techniques in psychological research. Big data is one of the most important trends in today's world including the scientific community, and data mining techniques are receiving increasing attention. These new methods are rarely used in social science research despite that more and more studies are collecting large amount of data. This project may serve as an introduction and example for the use of data mining in big data.