[資料科學演講] 干擾機制對資料科學帶來的挑戰與契機：以抽煙及肺癌存活的因果關係為例

Complexity of the data may provide challenges for making valid causal inference about a scientific hypothesis (smoking vs. lung cancer mortality), particularly in the presence of unknown confounding (underlying lung function). Mendelian randomization (MR) addresses the issue of unknown confounding by using genetic information as an instrumental variable (IV) to estimate the causal effect of an exposure of interest on an outcome. Despite the popularity of IV analyses in fully observable outcomes, methodology is limited for time-to-event survival outcomes with censoring, a common data structure in biomedical sciences. We propose an IV analysis method in the survival context, estimating causal effects on a transformed survival time and the survival probabilities using semiparametric transformation models. We construct unbiased estimating equations to circumvent the difficulty in deriving joint likelihood of the exposure and the outcome, due to the unknown covariance by confounding. Asymptotic properties of the proposed estimators are established. We apply our methods to conduct an MR study for lung cancer survival, which suggests a harmful prognostic effect by smoking intensity (p=0.0067) that would have been missed by the conventional methods.