Keywords: random survival forest, ischemic cardiomyopathy, causal inference, potential outcomes, survival analysis
Estimation of treatment effect in observational survival data is complicated due to the challenges of confounding and selection bias. We define individual treatment effect (ITE) and average treatment effect (ATE) directly through comparison of two survival functions using counterfactual treatment assignments. Treatment effect is viewed as a dynamic causal procedure. Our requirement is complete overlap, which permits only individuals who are eligible for both treatments to be included in such comparisons. Unfortunately, not all treatments have a clearly defined evidence based eligibility criteria. Therefore, we propose random forest methods to estimate an individual's eligibility for each treatment. Then using a novel counterfactual random survival forest, we model an individual's survival function directly to obtain the ITE. We provide a practical illustration on a real dataset of ischemic cardiomyopathy under four surgical treatment options.