Core Concepts
The core message of this paper is to propose a novel causal inference approach called Causal-StoNet that can effectively handle high-dimensional covariates and unknown data generation processes by leveraging deep learning techniques, including sparse deep learning theory and stochastic neural networks.
Abstract
The paper addresses the challenges of causal inference with high-dimensional complex data, where the data dimension can be extremely high and the underlying data generation process can be unknown and highly nonlinear. Existing methods for causal inference often rely on the assumptions of low-dimensional covariates or linear/approximately linear data generation processes, which may not hold in practice.
To tackle these challenges, the authors propose a novel approach called Causal-StoNet, which is based on deep learning techniques, including sparse deep learning theory and stochastic neural networks. The key features of Causal-StoNet are:
A natural forward-modeling framework: Causal-StoNet is formulated as a composition of multiple simple linear and logistic regressions, providing a natural forward-modeling framework for complex data generation processes.
Universal approximation ability: The authors prove that the stochastic neural network (StoNet) used in Causal-StoNet possesses a valid approximation to a deep neural network, thereby enabling it to approximate the outcome and propensity score functions.
Consistent sparse learning: By imposing an appropriate sparse penalty/prior on the structure of the StoNet, Causal-StoNet can identify relevant variables to the outcome and propensity score along with the training, even under high-dimensional settings.
The authors also show that Causal-StoNet can be easily extended to handle various causal inference scenarios, such as missing covariates, multi-level or continuous treatments, and mediation analysis.
Extensive numerical studies demonstrate that Causal-StoNet outperforms existing methods in estimating the average treatment effect (ATE) and conditional average treatment effect (CATE), as well as in covariate selection accuracy.
Stats
The paper does not provide specific numerical values or statistics. The focus is on the methodological development of the Causal-StoNet approach.
Quotes
The paper does not contain any striking quotes that support the key logics.