Skip to Content

College of Arts & Sciences
Department of Statistics

Butch Tsiatis Palmetto Lecture 1

Thursday, March 24, 2016 - 2:45pm

Statistics Department Palmetto Lecture 1

Where: LeConte College 210A

Speaker: Anastasios (Butch) Tsiatis

Affiliation: North Carolina State University, Department of Statistics

Title:  2016 Palmetto Lecture 1: Optimal Treatment Regimes for Survival Endpoints from a Classification Perspective

Abstract: A treatment regime is a decision rule which takes an individual's baseline information and dictates the treatment to be given to that patient among the available options. The value of a treatment regime is the expected outcome (function of survival time) for a population of patients if they were treated consistent with that regime. Assuming large outcomes are good, the optimal treatment regime is the one with the largest value.

The most common method for estimating an optimal treatment regime using data from an observational study of clinical trial is by using regression methods where one develops a regression model for the mean outcome as a function of treatment and covariates. However, regression methods may lead to poor treatment regimes if the regression model is misspecified. We instead consider the "value search method" where we derive a "robust" estimator for the value of any treatment regime and then find the regime among a class of restricted feasible regimes that maximizes this estimator.

When the primary outcome of interest is a function of the survival time, we show how to derive an estimator for the value of any regime with censored survival data using doubly-robust augmented inverse probability weighted complete-case estimators. Unfortunately, deriving an estimator for the optimal regime within a restricted class using the "value search method" may be computationally prohibitive. We show how to recast this problem as a weighted classification problem which then allows us to use off-the-shelf software to derive the optimal restricted regime.

These methods are illustrated by using data from the ASCERT study to decide whether patients with two or three vessel coronary artery  disease should get CABG or PCI based on their baseline characteristics.

This work is joint with Xiaofei Bai, Wenbin Lu and Rui Song.