Encode-Decoder-based GAN for Estimating Counterfactual Outcomes under Sequential Selection Bias and Combinatorial Explosion

Yoshiyuki Norimatsu, Masaaki Imaizumi

Keywords: Causal Inference, GAN, Time-varying treatment types and dosages

Abstract:

Estimating counterfactual outcomes of time-varying treatment types and associated dosages is important for addressing medical problems.
This task becomes more challenging when both the treatment type and dosage assignment are biased due to the presence of time-varying confounders,
as compared to estimating outcomes for treatment types alone.
Specifically, the setup yields the following two obstacles:
first, treatment types and dosages are selected sequentially, causing observed outcomes to be biased at each time step,
leading to $2 \times \tau$ biases for a $\tau$-step-ahead prediction (sequential selection bias);
second, the number of treatment trajectories increases exponentially with $\tau$ (combinatorial explosion).
In this paper, we introduce Encoder-Decoder Time-SCIGAN (EDTS),
which combines a longitudinal encoder-decoder transformer with a Generative Adversarial Network (GAN) for estimating counterfactuals.
The encoder-decoder architecture predicts outcomes for one-step- and multi-step-ahead predictions separately,
while the GAN generates counterfactual outcomes that cannot be distinguished from observed outcomes by the discriminators to handle sequential selection bias.
To address combinatorial explosion, we propose a novel discrimination method,
Sequential Counterfactual Discrimination (SCD) for EDTS discriminators.
Our evaluation of synthetic and semi-synthetic datasets demonstrate that EDTS outperforms the current baselines.
To the best of our knowledge, this is the first study to propose an architecture
for estimating counterfactual outcomes of both time-varying treatment types and dosages.
Implementation is available at \url{https://github.com/ynorimat/EDTS}.

Submission Number: 28