國立政治大學統計學系
學 術 演 講
學 術 演 講
主講人：余日彰助理教授(台北大學統計學系)
題 目：From Separable Effects under Competing Risks to Flexible Semiparametric
Modeling
時 間：民國115年4月13日 (星期一) 下午2：30
地 點：國立政治大學逸仙樓050101教室
摘 要：
The separable effects framework has recently provided a useful causal perspective for distinguishing the direct effect of an exposure on a primary event from the indirect effect operating through an intervening event. Our research develops this framework along a broader and more flexible direction. In earlier work, we extended separable effects from the classical setting of competing risks to the more general setting of semicompeting risks, where a nonterminal intermediate event and a terminal primary event may both be observed. Within this setting, we established identification results for the causal quantities of interest and developed both nonparametric and semiparametric estimation methods. The nonparametric approach offers robustness under weaker modeling assumptions, whereas the semiparametric approach facilitates confounder adjustment through covariates and can achieve greater statistical efficiency. Building on this foundation, we further revisit the semiparametric methodology. Our initial approach was developed under the Cox proportional hazards model, but this model class may be too restrictive in applications where the proportional hazards assumption is questionable. To address this limitation, we extend the semiparametric framework to a broader class of generalized transformation models, thereby allowing substantially greater modeling flexibility while preserving a tractable structure for confounder adjustment and statistical inference. In the current stage, we focus on the competing risks setting in order to develop the methodology and study its theoretical properties. Nevertheless, the proposed framework is sufficiently general that it can be carried back naturally to the semicompeting risks setting. Overall, this line of work traces a coherent progression: from separable effects in competing risks, to their extension in semicompeting risks, and then to a more flexible semiparametric framework beyond the proportional hazards model.
Keywords: causal mediation analysis; semi-competing risks; semiparametric inference