A Loss-Function for Causal Machine-Learning

TL;DR

This paper introduces a universal loss function for causal machine learning that enables direct training of models like neural networks to predict true causal effects, overcoming previous limitations due to lack of point-wise true values.

Contribution

It proposes a novel mean-square-error based loss function for causal inference, allowing direct gradient descent training without meta-learner strategies.

Findings

The loss function is applicable to various models including deep neural networks.

Gradient descent can be performed directly on the proposed loss function.

The method provides a standard for evaluating causal prediction models.

Abstract

Causal machine-learning is about predicting the net-effect (true-lift) of treatments. Given the data of a treatment group and a control group, it is similar to a standard supervised-learning problem. Unfortunately, there is no similarly well-defined loss function due to the lack of point-wise true values in the data. Many advances in modern machine-learning are not directly applicable due to the absence of such loss function. We propose a novel method to define a loss function in this context, which is equal to mean-square-error (MSE) in a standard regression problem. Our loss function is universally applicable, thus providing a general standard to evaluate the quality of any model/strategy that predicts the true-lift. We demonstrate that despite its novel definition, one can still perform gradient descent directly on this loss function to find the best fit. This leads to a new way to train any parameter-based model, such as deep neural networks, to solve causal machine-learning problems without going through the meta-learner strategy.