2E-02 - Comparison of Approaches for Regression Discontinuity Analysis of the Effect of a Treatment Policy for Statin Prescribing on Atherosclerotic Cardiovascular Disease Incidence

Background: Regression discontinuity analysis (RDA) is a useful method for estimating the effect of a treatment policy when treatment is determined by a continuous measure with a decision threshold. Nevertheless, effect estimates from this approach have large variance because only individuals within a specified bandwidth of the decision threshold are used for analysis. Studies of effect estimates in experimental and observational contexts suggest that G-computation reduces variance, which may be useful for RDA.

Objectives: We aimed to compare effect estimates from a conventional approach and G-computation using RDA to assess the effect of a statin prescribing treatment policy on 5-year atherosclerotic cardiovascular disease (ASCVD) incidence.

Methods: We used electronic health record data to identify adults who completed at least one in-person or telehealth encounter at one of 12 community health centers in a safety-net health system between July 2013 and December 2018 with follow-up through December 2023, and had a 10-year ASCVD risk score between 8.0% and 12%. The decision threshold of interest was 10%, which corresponds to a guideline-based treatment policy of assignment to statin therapy. We assessed initiation of stain therapy within 180 days of the index encounter. We estimated 5-year ASCVD risk difference (RD) and 95% confidence limits (CL) using a generalized linear model with binomial distribution, identity link, binary discontinuity variable, and restricted cubic spline of the running variable in both approaches. For G-computation, we also standardized estimates (by age, sex, race/ethnicity, body mass index, current smoking, total cholesterol level and hypertension) and estimated bootstrapped 95% CL using 1000 replicates.

Results: Our study population comprised 2,992 patients, 1,253 (42%) of whom had a 10-year ASCVD risk score ≥10%. The median age was 57 years (interquartile range: 52 – 62), 55% were male, 40% were non-Hispanic Black, and 70% were diagnosed with hypertension. Statins were prescribed to 10.3% of patients above the decision threshold and 7.65% below the threshold. The estimated RD for 5-year ASCVD incidence was -0.37% (95% CL: -6.0%, 5.2%) from conventional RDA and -0.82% (95% CL: -1.0%, -0.64%) from G-computation.

Conclusions: Our results suggest that G-computation substantially reduces variance in effect estimates from RDA and presents a double-robust option, which may inform analytic choices in future studies. Nevertheless, we caution against interpreting the effect estimates in our study as valid causal effects of the statin prescribing policy because we observed only a modest increase in treatment prescribing at the decision threshold in our population.