Enhancing Plausibility and Interpretability of Cause-specific Models of Mortality

New York University, New York NY

Investigators

Abstract

? DESCRIPTION (provided by applicant): We seek R03 funding to establish proof of concept and to generalize a computationally- intensive method to enhance the plausibility and interpretability of estimates from continuous- time competing-risk models of cause-specific mortality. The significance and innovation of our brute force method is that it can be used in situations not handled well by even state-of-the-art life table methods. First and most importantly, we will generalize our brute force method to models in which cause-specific risks are correlated, thus relaxing strong and implausible assumptions of independence maintained in nearly all demographic methods to date, including heavily-used cause-deleted and cause-reduce life table methods. Progress in developing and applying such models has been impeded by the intractability of life table derivations when independence does not hold. Thus, a significant contribution of this R03 is that it provides researchers with the means to formulate, apply, and interpret estimates under assumptions that are far more plausible. Second, our method lets researchers more easily understand what is implied by estimates from competing-risk regressions, which are otherwise difficult to interpret. Consider a hypothetical treatment that, net of controls, is estimated to reduce mortality risks by 50% for one cause and by 5% for a second cause. It is in fact not obvious from a simple inspection of estimated coefficients what is implied in that answers will depend, at least in part, on levels. That is, if one cause is common but the other rare, then even modest reductions in risks for a common cause can imply a substantial increase in life expectancy whereas a sizeable improvement for a rare condition may increase life expectancy only negligibly. The core idea of this R03 is to tackle such problems by using raw computing power to transform what is otherwise an inherently difficult high-dimensional problem possessing no general closed-form solution into one involving the minima of easily computed one-dimensional quantities. These methods will thus provide researchers with a highly general set of tools by which to advance knowledge in aging, health, mortality, and other areas in which multiple-decrement demographic phenomena arise naturally.