Sequential Deconfounding for Causal Inference with Unobserved Confounders

Tobias Hatt,Stefan Feuerriegel

DOI: https://doi.org/10.48550/arXiv.2104.09323

2022-02-28

Abstract:Using observational data to estimate the effect of a treatment is a powerful tool for decision-making when randomized experiments are infeasible or costly. However, observational data often yields biased estimates of treatment effects, since treatment assignment can be confounded by unobserved variables. A remedy is offered by deconfounding methods that adjust for such unobserved confounders. In this paper, we develop the Sequential Deconfounder, a method that enables estimating individualized treatment effects over time in presence of unobserved confounders. This is the first deconfounding method that can be used in a general sequential setting (i.e., with one or more treatments assigned at each timestep). The Sequential Deconfounder uses a novel Gaussian process latent variable model to infer substitutes for the unobserved confounders, which are then used in conjunction with an outcome model to estimate treatment effects over time. We prove that using our method yields unbiased estimates of individualized treatment responses over time. Using simulated and real medical data, we demonstrate the efficacy of our method in deconfounding the estimation of treatment responses over time.

Methodology,Machine Learning

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to estimate the change of individualized treatment effects over time from observational data in the presence of unobserved confounding factors. Specifically, the paper focuses on how to adjust the influence of unobserved confounding factors through the proposed method - Sequential Deconfounder, so as to achieve an unbiased estimate of treatment effects when only a single treatment can be applied at each time point.

The paper mentions that in fields such as medicine, due to the high cost or difficulty in implementing randomized experiments, observational data are usually used to estimate treatment effects. However, observational data often lead to biases in treatment effect estimates because treatment assignment may be interfered with by unobserved variables. These unobserved confounding factors make it difficult for standard methods to provide effective treatment effect estimates. Therefore, the paper proposes a novel method - Sequential Deconfounder, which uses the time - dependence of treatment assignment to infer surrogate variables of unobserved confounding factors and uses these surrogate variables in combination with the outcome model to estimate treatment effects that change over time.

The main contributions of the paper include:
1. Developing a theoretical framework for dealing with the influence of unobserved confounding factors on treatment effect estimates when only a single treatment can be applied at each time point.
2. Proposing the Sequential Deconfounder method, which can infer surrogate variables of unobserved confounding factors and provide unbiased estimates of individualized treatment effects that change over time based on this.
3. Based on a new Gaussian Process Latent Variable Model (GPLVM), achieving a specific instantiation of the Sequential Deconfounder. This model can capture the sequence - dependence between treatment assignments, thus effectively dealing with unobserved confounding factors.

Through experiments on simulated data and real - world medical data, the paper verifies the effectiveness of the proposed method and proves that in the presence of unobserved confounding factors, the Sequential Deconfounder can significantly improve the accuracy of treatment effect estimates.