1G-02 - Improving confounding control by statin utilization prior to follow-up for 1-year mortality in US older adults – proof of concept

Background: Pharmacoepidemiologic studies often control for baseline prevalence of prescription (Rx) drug classes (yes vs. no). We have previously shown that conditioning on the initiation of certain drug classes (e.g., statins) rather than prevalent use can reduce confounding by health-related behaviors, social determinants, and frailty.

Objectives: Estimate statin use trajectories (Trajs) in US older adults for confounding control for 1-year mortality

Methods: We used group-based Traj models to identify statin use patterns <= 9 months prior to an index date (T0: the first outpatient visit in 2018) as a proxy for frailty in 3080881 US Medicare beneficiaries aged 65+ using their 2017-2019 claims data. Our models had months before T0 as independent variable and the proportion of days with pills available in each month as outcome. We chose the model with the largest Bayesian Information Criterion to define the number of Traj groups. We estimated propensity scores (PSs) for each Traj vs. the largest statin Traj (reference) using 44 variables including demographics, comorbidities, heathcare use, and frailty predictors. We weighted each Traj group to resemble the distribution of features of the reference group by Standardized Mortality Ratio (SMR) weights PS/(1-PS). We estimated crude 1-year mortality risks for each Traj group and adjusted 1-year risk difference (aRD%) for the Traj groups vs the reference. Next, we assessed the impact of including statin Traj groups as a polytomous variable vs. a binary use variable in logistic regressions predicting 1-year mortality. Sensitivity analyses used a 12-month statin Traj period.

Results: Our selected model identified 7 Traj groups. Their sample size proportion, 1-year crude and aRD for mortality are: early stoppers (n=2%, 5.8%, 0.88 [0.83, 0.94]), late stoppers (n=2%, 5.4%, 0.71 [0.66, 0.76]), nonusers (n=47%, 4.3%, 0.89 [0.84, 0.94]), high adherers (n=7%, 4.2%, -0.12 [-0.17, -0.07]), late initiators (n=3%, 4.1%, -0.03 [-0.07, 0.02]), early initiators (n=2%, 3.7%, -0.19 [-0.24, -0.14]), and complete adherers (n=37%, 3.7%, reference). When controlling for all other confounders, the predictive value of statin Traj variable and binary statin use variable were similar (C-statistics 0.852). Sensitivity analyses confirmed consistency in findings and identified lower mortality in complete adherers with 90-d compared to 30-d Rx.

Conclusions: Traj models can capture meaningful variations in statin use patterns prior to a T0. Although statin Traj does not add to the C-statistic, our analysis suggests heterogeneity in mortality risk across the 7 trajs independent of “usual” covariates. The Traj of Rx, diagnoses, and procedures prior to a T0 may contain additional information for confounding control that warrants further exploration.