Machine learning to control confounding in personalized medicine

Denis Talbot (ULaval)

Wednesday 5th November 14:00-15:00
Maths 311B

Abstract

Causal inference methods for personalized medicine aim to estimate the optimal adaptive treatment strategy, that is, they take as an input the characteristics of the patient and produce as an output the treatment option that optimizes a health outcome of interest. When estimating causal effects with non-experimental data, a key challenge is to adequately control for confounding covariates that may simultaneously affect treatment and outcome. The use of machine learning is becoming increasingly popular to control confounding data-adaptively. However, relatively little work has been done on this topic when the goal is specifically to estimate an optimal adaptive treatment strategy. I will introduce a nonparametric efficient estimator of the parameters of a causal model for optimal adaptive treatment strategies. I will show that this estimator is doubly robust, that it can make use of machine learning to control confounding while retaining ease of interpretation of the target parameter, and remains root-n consistent under machine learning estimation of the nuisance function. A simple variance estimator based on the efficient influence function is available. Both theoretical and simulation results will be presented. This work is motivated by the treatment of breast cancer with hormonal therapy. We aim to discover improved rules for deciding whom to recommend hormonal therapy to, using data from a breast cancer registry.

Add to your calendar

Download event information as iCalendar file (only this event)