Bayesian Statistical Learning for Robust and Generalizable Causal Inferences in Alzheimer Disease and Related Disorders Research

SUMMARY To develop multi-faceted interventions for Alzheimer’s Disease and Related Disorders (ADRD) prevention, it is key to quantify joint effects of environmental exposures throughout the life-span as well as the mechanisms through which the exposures operate, which involve dynamic disease processes. These causal investigations in observational studies face methodological challenges. This application, in response to PAR-22-093, will accomplish the following goals: (1) develop robust Bayesian machine learning methods for causal mediation analyses in life-course observational studies of ADRD; (2) apply these approaches in the analysis of two 30+ year cohort studies of Native Americans (Strong Heart Study) and of the Greater Boston area (Normative Aging Study) to determine whether and to what extent the onset and severity of hypertension and cardiovascular disease mediate the harmful effects of air pollution and heavy metals on ADRD; (3) develop and disseminate computationally efficient and user-friendly software for widespread application of the methods. The proposed work will address methodologic gaps in the causal investigation of health effects. First, no mediation analyses approaches are available that simultaneously allow for multiple exposures and multiple time-to-event and longitudinal mediators. Investigators can currently only build models that do not reflect real- life conditions, considering a single exposure, or a single mediator producing segmented results, limited in informing prevention. Second, most of observational studies in ADRD research are plagued by selection bias. Participants may die before an ADRD diagnosis (attrition due to death) or may drop-out due to cognitive impairment. Third, multicollinearity, skewness of exposures, complex exposure-response relationships and time dependent confounding challenge valid estimation and inference. Fourth, no statistical approaches are available to evaluate the generalizability of findings on determinants of and mechanisms leading to ADRD. We propose to fill these gaps by developing and applying Bayesian machine learning approaches for quantifying the total, direct and indirect effects of environmental and health factors on ADRD outcomes under the counterfactual framework. We will develop and apply the new methods to estimate complex exposure- response relationships of pollutants with time-to-event or longitudinal outcome potentially mediated by a single time-to-event or longitudinal mediator (Aim 1), and mechanisms through multiple longitudinal and time-to-event mediators (Aim 2). Furthermore, we will develop Bayesian data fusion algorithms to evaluate unmeasured confounding bias and to generalize evidence from the study sample to the target population (Aim 3). In the SHS and NAS we will investigate the role of hypertension and CVD trajectory, onset, and severity in mid and late life as mediators of the neurotoxic effects of AP and metals. We will develop user-friendly and efficient R packages that implement the proposed new methods. Our work has great potential to have broad impact on life-course epidemiology research and prevention for ADRD and other chronic health outcomes.