Learning Decomposed Representation for Counterfactual Inference

Abstract

One fundamental problem in the learning treatment effect from observational data is confounder identification and balancing. Most of the previous methods realized confounder balancing by treating all observed variables as confounders, ignoring the identification of confounders and non-confounders. In general, not all the observed variables are confounders which are the common causes of both the treatment and the outcome, some variables only contribute to the treatment and some contribute to the outcome. Balancing those non-confounders would generate additional bias for treatment effect estimation. By modeling the different relations among variables, treatment and outcome, we propose a synergistic learning framework to 1) identify and balance confounders by learning decomposed representation of confounders and non-confounders, and simultaneously 2) estimate the treatment effect in observational studies via counterfactual inference. Our empirical results demonstrate that the proposed method can precisely identify and balance confounders, while the estimation of the treatment effect performs better than the state-of-the-art methods on both synthetic and real-world datasets.Preprint. Under review.

Disentangled Counterfactual Recurrent Networks for Treatment Effect Inference over Time

“…Subsequently, this was extended to enforce the preservation of local similarity [Yao et al, 2018], to learn overlapping representations [Zhang et al, 2020] and to incorporate importance weighting [Hassanpour and Greiner, 2019]. Instead of learning one representation of all inputs, Hassanpour and Greiner [2020] and Wu et al [2020] identify disentangled representations that separate input covariates by the effects they have on treatment assignment and outcome. However, none of these methods are fit for use in dynamic settings, as they model neither time-dependent confounding nor other dynamic relationships between variables.…”

“…and J, L ∈ {I, C, O}. Wu et al [2020] improve over Hassanpour and Greiner [2020]'s method by augmenting their loss function to enforce that the factors do not overlap, an approach that we adapt to the dynamic setting. Note that forcing such disentanglement is not more restrictive than assuming complete entanglement -should there be no underlying disentangled factors in the data-generating process, then DCRN will simply not learn these factors and will represent the entire covariate space through the confounding factor, and hence able to handle any violations against the assumed existence of disentangled factors.…”

“…Thus, while the above loss components will suffice to achieve some disentanglement, we adapt Wu et al [2020]'s orthogonal regularizer to enforce better identification of the factors, and add to our loss function,…”

Auto IV: Counterfactual Prediction via Automatic Instrumental Variable Decomposition

“…Recently, causal representation learning [18], [19], [22], [33], [40], [42] has attracted lots of attention. Among these works, Yao et al [42] propose to reduce prediction bias by filtering out the nearly IVs.…”

“…Inspired by the recent works [18], [40], [42] on causal disentangled representation learning, we argue that although invalid IV candidates do not satisfy the conditions of the valid IVs strictly, one might decompose and utilize a part of their information to generate IV representations. Therefore, in this paper, we propose a novel Automatic Instrumental Variable decomposition (AutoIV) algorithm to automatically generate representations serving the role of IVs for counterfactual prediction with fewer constraints for the IV candidates.…”