Antihypertensive and statin therapy represent imminently scalable interventions for preventing Alzheimer’s
disease and related dementias (ADRD). Evidence from randomized trials for the effect of these medications on
preventing ADRD provided mixed results and has often been limited by short follow-up times, low statistical
power, and a lack of representativeness for the broader population and routine care. Using large-scale
electronic health record (EHR) data from three integrated health systems in the US as well as nation-wide data
from the UK and Denmark, this project will employ an innovative and well-validated method for causal effect
estimation to fill this evidence gap. Specifically, we will exploit thresholds in blood pressure, LDL, and
cardiovascular disease risk in clinical guidelines. Patients close to these thresholds are similar to each other
except that crossing a given threshold results in a sudden jump in the probability of receiving the medication.
Importantly, the validity of this approach can be fully verified in the data, including by comparing patients’
characteristics just above versus just below each threshold. Our datasets have follow-up times between 17 and
35 years, include detailed pharmacy records, and have minimal loss to follow-up due to complete linkage of all
US datasets to Medicare claims data and the single-provider nature of the UK and Danish health system.
Generated as part of an NIH New Innovator Award to the PI, our extensive preliminary analyses demonstrate
that this innovative approach in EHR data is both feasible and valid and that all key variables are reliably
recorded in our EHR datasets. Our approach in EHR data has also successfully replicated findings from clinical
trials on the effect of antihypertensive and statin therapy on cardiovascular disease events, giving us a high
degree of confidence that we can extend our approach to ADRD. Bringing together a team with deep expertise
in ADRD, causal inference, and EHR data, we will establish the causal effect of antihypertensive (Aim 1) and
statin (Aim 2) therapy on the incidence of ADRD. Aim 3 will determine whether antihypertensive and statin
prescriptions have adverse behavioral effects, which is critical knowledge for interpreting findings (regardless
of their direction) in Aims 1 and 2. All analyses will ascertain the causal effects both of receiving a prescription
and of long-term adherence to these medications, use novel machine-learning approaches to determine how
effects vary across detailed patient subgroups, and be compared to more common analytical approaches that
make stronger assumptions than our threshold-based method. Although validation studies have generally
found the reliability of ADRD diagnoses in our EHR datasets to be high, we will also conduct extensive
probabilistic bias analyses for potential under- and overdiagnosis of ADRD. This project will inform clinical and
public health guidelines as well as directions for the development of new ADRD preventive interventions. In
demonstrating the value of our approach, this project could also encourage efforts in EHR data that exploit
threshold-based decision making in other areas of medicine.
