Causal Inference for Observational Time-Series with Encoder-Decoder Networks

This paper proposes a method for estimating the causal effect of a discrete intervention in observational time-series using encoder-decoder networks. Encoder-decoder networks, which are special class of recurrent neural networks (RNNs) suitable for handling variable-length sequential data, are used to predict a counterfactual time-series of treated unit outcomes using only the pre-intervention outcomes of control units as inputs. Unlike the synthetic control method, the proposed method does not rely on pretreatment covariates, allows for nonconvex combinations of control units, and can handle multiple treated units. Encoder-decoder networks outperform the synthetic control method in simulated and empirical data applications.