Skip to Main Content
Principal Investigator
Youjin Lee
University of Pennsylvania
Position Title
Postdoctoral Fellow
About this CDAS Project
PLCO (Learn more about this study)
Project ID
Initial CDAS Request Approval
Dec 16, 2019
Instrumental variable methods in survival outcome
The instrumental variable (IV) methods have brought an inspiring opportunity to rule out unmeasured confounding in causal inference. Compared to discrete or continuous outcomes, however, there is a dearth of IV methods in censored survival outcome where we cannot directly enjoy seamless, two-stage regression-based estimation that is dominantly developed in linear model settings. In this project we propose a nonparametric estimator for a local average treatment effect on survival probabilities or hazards when censoring is conditionally independent on the observed covariates. We provide an efficient influence function-based estimator and suggest a simple estimation procedure when IV is either binary or continuous. The proposed estimator was shown to enjoy double-robustness and efficiency. We also demonstrate how Cox proportional hazards model and additive hazards model can be absorbed in our nonparametric estimation combined with machine learning techniques.

The goal of our project is as following:
1. Using an intervention assigned as an instrument and an intervention received as a treatment variable, apply our nonparametric method to evaluate the causal effect of the intervention on the survival probabilities including time to deaths to colorectal cancer.
2. Compare the above results with/without instrumental variable approaches.
3. Compare our proposed method to other model-based approaches.


Nandita Mitra (University of Pennsylvania) and Edward Kennedy (Carnegie Mellon University)

Related Publications