Data-Driven Discovery of Heterogeneous Treatment Effects of Statin Use on Dementia Risk

PROJECT SUMMARY (ABSTRACT) Alzheimer’s Disease and Related Dementias (ADRD) currently affects more than 4 million Americans and over 50 million individuals worldwide. The identification of prevention strategies for dementia is critical, particularly due to the lack of effective treatments. In parallel, there is growing consensus that lipid metabolism is a major contributor to ADRD and may be an important strategy for risk reduction and prevention. Antihyperlipidemic agents (i.e., statins) are widely used, yet evidence on the relationship between antihyperlipidemic agents (i.e., statins) and ADRD has been largely inconclusive. One possible explanation for the mixed findings is heterogeneity in study populations and their characteristics. For example, the effectiveness of statins is evidenced to vary by age, ApoE4 status, and pre-existing disease status.34-36 Accordingly, there is a growing need to identify the factors (i.e., effect modifiers) which influence heterogeneities in the effect of statins on dementia. The objective of this study is to triangulate evidence on the identification and estimation of heterogeneous treatment effects by using three causal machine learning methods, specifically the honest causal forest/policy tree, doubly robust adaptive LASSO, and Bayesian Adaptive Regression Trees (BART), to identify novel effect modifiers and optimal subgroups for the effect of statins on dementia. While traditional parametric regression approaches are designed to test a priori hypotheses regarding effect modification, such approaches are not suitable for yielding novel hypotheses. The causal machine learning methods described in this proposal fill this gap; not only do such approaches help identify novel effect modifiers, but they can also facilitate the subsequent identification of optimal treatment rules across those modifiers. In this study, I propose to use a cohort of 307,719 individuals from the UK Biobank data who were at least 55 when they were initially recruited from 2006 to 2010. The analytical sample will be large, allowing me to rigorously investigate heterogeneous treatment effects across different subgroups. Specifically, in Aim 1, I propose to estimate the real-world average treatment effect (ATE) of statins on ADRD across the entire sample. I will then, in Aim 2, apply three causal machine learning algorithms to identify novel effect modifiers and corresponding optimal subgroups for the effect of statin use on ADRD risk. Finally, in Aim 3, I will quantify the reduction in ADRD cases that would result from implementing each of the optimal treatment rules generated under Aim 2 and compare them to the reduction in ADRD cases observed under Aim 1. This F31 proposal application will support my dissertation research, as well as my interest in gaining training in causal machine learning, as well as substantive training in dementia, cognitive aging, and its psychometric methods. Under the guidance of my mentorship team, I look forward to advancing dementia prevention research while also pursuing my goal of becoming an independent investigator in research methods on cognitive aging.