Direct estimation of the mean outcome on treatment when treatment assignment and discontinuation compete
提出一种不依赖治疗模型的新估计量，用于估计目标治疗时长策略下的平均结果，适用于离散或连续治疗时长数据，避免因粗化连续时间数据导致的偏差，并通过数值研究和真实数据验证其精度优于逆概率加权估计量。
Several authors have investigated the challenges of statistical analyses and inference in the presence of early treatment termination, including a loss of efficiency in randomized controlled trials and a connection to dynamic regimes in observational studies. Popular estimation strategies for causal estimands in dynamic regimes lend themselves to studies where treatment is assigned at a finite number of points and the extension to continuous treatment assignment is nontrivial. We re-examine this from a different perspective and propose a new estimator for the mean outcome of a target treatment length policy that does not involve a treatment model. Because this strategy avoids modelling the treatment assignment mechanism, the estimator works for both discrete and continuous treatment length data and eschews bias and imprecision that arise as a result of coarsening continuous time data into intervals. We show how the competition of treatment length assignment and terminating event lead to a competing risks problem. We exemplify the direct estimator through numerical studies and the analysis of two real datasets. When all modelling assumptions for both the direct and inverse weighted estimators are correct, our simulation studies suggest that the direct estimator is more precise.