Matched Design with Sensitivity Analysis for Observational Survival Data in Cardiovascular Patient Management using EMR Data

Time-to-event is a ubiquitous outcome measure in clinical diagnosis and assessment of therapeutic effects in many disease areas including stroke (time-to-stroke), respiratory (time-to-first medication for worsening asthma) and sleep diseases (time-to-insomnia-related mortality). The hazard rate is commonly seen in survival analysis as it has the convenient interpretation of instantaneous risk. The hazard ratio (HR) is routinely used as an effect measure when comparing between two treatment groups, largely due to the popular Cox proportional hazards (PH) model. However, the HR is vulnerable to selection bias and not collapsible, which make it a questionable marginal causal effect measure. Restricted mean survival time (RMST) is an alternative measure, defined as the area under the survival curve up to a fixed time point. RMST difference is a more adequate causal effect measure than the HR because (i) it is a collapsible measure, thus avoids discrepancy between marginal and conditional effects; (ii) it does not depend on the PH assumption; (iii) it is essentially a mean difference with simpler interpretation. RMST has become a popular metric of treatment effects in randomized trials recently. However, the development of RMST methodology for observational survival data is lacking. The goal of this proposal is to develop a comprehensive matching-based RMST difference estimation strategy to infer causal effects in observational survival data, and apply such tools to evaluate causal effects of direct oral anticoagulants (DOAC) vs. warfarin on the risk of cardiovascular events in a secondary data analysis. We plan to develop propensity score matching-based RMST estimation methodology and corresponding sensitivity analysis, which do not rely on strong outcome modeling assumptions. The matching method will use an optimal algorithm to create matched sets to mimic a block randomized design and an asymptotically valid post-matching inferential procedure will be developed by accounting for the correlation introduced in matching. Built upon the matched data, the sensitivity analysis will address how much association an unmeasured confounder would need to have with both the exposure and the outcome, to explain away the observed effect. In the secondary data analysis, we will apply our methods to examine the hypothesis that using DOAC has lower risk of composite cardiovascular events including stroke, venous thromboembolism, myocardial infarction, and death, using electronic medical record (EMR) data. We will also explore subgroup causal effects related to gender and race to examine potential health disparity issues. Our proposed work will not only result in novel and valid research methodology for estimating causal effects in observational survival data, but also advance the understanding of how different anticoagulant drugs would impact patient outcomes using a large secondary database. Our general-purpose methodology will be widely applicable to study survival data in heart, lung, blood and sleep disease treatment, and disparity research. This will also enable clinical researchers to rigorously identify causal evidence using increasingly available real-world data.