Abstract
Antihyperlipidemic agents, including statins, currently rank as the second most frequently prescribed drugs in
the U.S. While there is adequate evidence supporting the benefits of statins for the prevention of CVD events,
evidence to support the beneficial or harmful effects of statins on cognition and the risk of Alzheimer’s disease
and related dementias (ADRD) is inconclusive. Findings from two large randomized controlled trials of statins
found no effect on cognitive outcomes over the short-term (4 to 5 years). Results from observational studies
have been mixed. Given the widespread use of statins, this knowledge gap represents a huge opportunity for
illuminating viable prevention strategies for ADRD. The most plausible explanation for these inconsistent results
is that the effectiveness of statin treatment varies across patient characteristics such as age, sex, and the
presence of chronic conditions. Importantly, most trials do not have adequate statistical power to examine these
heterogeneous treatment effects (HTEs). Further, trials typically include participants who are healthier than the
general population of statin-users and less likely to experience side effects, and often exclude participants who
do not tolerate statins. Large, administrative observational data sources can provide sufficient sample sizes to
address these limitations, but traditional analytical techniques are insufficient in the presence of strong
confounding. In this study, we propose to leverage a widespread clinical prescription guideline and established
statistical methods in economics, specifically regression discontinuity (RD) designs, to address confounding and
approximate a randomized trial. Our data (N=175,234) will come from the Health Improvement Network (THIN)
database which includes general practices in the UK covering about 5% of the total population. In 2008, the
National Institute for Health and Care Excellence in the UK passed a guideline which recommends statin use
when a patient’s 10-year CVD risk score exceeds 20%. Since treatment is given or withheld according to this
guideline, we assume that patients who are “near” the 20% cutoff will be similar except for the treatment received,
thus creating the ideal setup to estimate the causal effect of statins. We will also apply an honest causal forest,
a state-of-the-art causal inference and supervised learning method, to flexibly identify novel patient subgroups
who could most benefit from statin use. The proposed research will (Aim 1) estimate the effect of statins on
ADRD risk, using an RD design, accounting for adherence. We will (Aim 2) then estimate the effect of statins on
ADRD risk across a priori (hypothesis-driven HTEs) and newly identified (data-driven HTEs) patient subgroups,
using RD and machine-learning methods. Finally, we will (Aim 3) quantify the reduction in ADRD cases that
could be achieved if specific subgroups of the population are treated with statins. Using contemporary methods,
this innovative study will provide more valid and public health relevant estimates of the effects of statins on ADRD
risk. By also examining HTEs and identifying groups that may particularly benefit or be harmed from statins, this
study will allow us to move towards targeted “precision medicine” approach for the prevention of ADRD.