Online Program Home
  My Program

All Times EDT

Abstract Details

Activity Number: 320 - Electronic Health Records, Causal Inference and Miscellaneous
Type: Contributed
Date/Time: Wednesday, August 11, 2021 : 3:30 PM to 5:20 PM
Sponsor: Section on Statistics in Epidemiology
Abstract #318534
Title: Efficient and Robust Semi-Supervised Learning: Estimating ATE with Partially Annotated Treatment and Response
Author(s): Jue Hou* and Tianxi Cai and Rajarshi Mukherjee
Companies: Harvard T.H. Chan School of Public Health and Harvard T.H. Chan School of Public Health and Harvard T.H. Chan School of Public Health
Keywords: Semi-supervised learning; Semi-parametric efficiency; Multiple robustness; Treatment effect; Nonparametric estimation; High-dimensional regression

A notable challenge of emulating Electronic Health Records (EHR) for treatment comparison is the noise in clinical variables, including the treatment and the response. The proportion of annotation is often limited by the labor cost. We develop a Semi-supervised Learning (SSL) framework with missing label rate potentially approaching one through the investigation into the first order efficient influence function. We apply the framework to estimating the average treatment effect (ATE) when the treatment and the response are partially annotated and showcase that our SSL estimator is 1) semi-parametrically efficient under low-dimensional smooth nonparametric model, 2) multiply robust under high-dimensional regression models. Simulation studies have justified the validity of our SSL method and its superiority over supervised and unsupervised benchmarks. We apply our method to the comparison of average 1-year progression free survival probability between 5-FU based chemotherapies and targeted therapies as the first-line therapies for advanced colorectal cancer patients using the MGH EHR.

Authors who are presenting talks have a * after their name.

Back to the full JSM 2021 program