Published on:

17 August 2020

Primary Category:

Methodology

Paper Authors:

Sergio Garrido,

Stanislav S. Borysov,

Jeppe Rich,

Francisco C. Pereira

•

Proposes neural networks to estimate causal effects without functional form assumptions

•

Uses neural autoregressive density estimators to model nonlinear variable relationships

•

Recovers accurate linear and nonlinear causal effects when causal graph is known

•

Provides reasonable approximations with limited data support if causal graph is accurate

•

Causal graph topology is critical; unobserved confounding decreases quality

Estimating Causal Effects with Neural Networks

This paper proposes using neural networks to estimate causal effects, avoiding assumptions about functional relationships between variables. Neural autoregressive density estimators model complex nonlinear interactions. Simulations show the method recovers linear and nonlinear effects when the causal graph is known. It provides reasonable approximations with limited data support if the graph is accurate. Causal topology assumptions are critical; unobserved confounding decreases quality regardless of estimator flexibility.

Neural network approximates likelihood of discrete time series

Detecting causal connections in time series models

Deep Learning: A Clear Explanation of Causal Neural Networks

Inferring causal relationships from time series data

Using causal graphs to improve neural network treatment effect estimation

Estimating network vector autoregression under uncertainty

No comments yet, be the first to start the conversation...

Sign up to comment on this paper