Background

Individual Treatment Effects (ITE) estimation methods have become increasingly popular in recent years. However, most of the time, individual effects are better presented as Conditional Average Treatment Effects (CATE). Representation balancing techniques have gained considerable momentum in causal inference from observational data but are still limited to continuous (and binary) outcomes. In numerous pathologies, the outcome of interest is a (possibly censored) survival time. Our paper proposes theoretical guarantees for a representation balancing framework applied to counterfactual inference in a survival setting using a neural network capable of predicting the factual and counterfactual survival functions (and then the CATE), in the presence of censorship, at the individual level.

Problem

The problem we faced was that representation balancing techniques were limited to continuous and binary outcomes, while many pathologies require survival time outcomes. We needed to develop a framework that could predict factual and counterfactual survival functions using a neural network while accounting for censorship at an individual level.

Solution

We proposed a representation balancing framework applied to counterfactual inference in a survival setting using a neural network capable of predicting the factual and counterfactual survival functions (and then the CATE), in the presence of censorship, at an individual level. We also conducted extensive experiments on synthetic and semisynthetic datasets that showed our proposed extensions outperformed baseline methods.

Impact

Our proposed framework provides theoretical guarantees for representation balancing techniques applied to counterfactual inference in a survival setting.
We developed a neural network capable of predicting factual and counterfactual survival functions while accounting for censorship at an individual level.
Our experiments on synthetic and semisynthetic datasets showed that our proposed extensions outperformed baseline methods.
Our framework has the potential to improve survival outcomes in numerous pathologies by allowing for more accurate individual treatment effects estimation.

Team

Ayoub Abraich - Lead Researcher
- Developed the theoretical guarantees for our proposed framework.
- Oversaw the development and implementation of the neural network.
Agathe Guilloux - Researcher
- Conducted experiments on synthetic and semisynthetic datasets to test our proposed extensions.
- Analyzed results and provided insights for further improvements.
Blaise Hanczar - Researcher
- Contributed to the development of the theoretical guarantees for our proposed framework.
- Oversaw data preprocessing and cleaning for experiments.

Neural Network for Survival CATE Estimation