[Paper Review] Estimating the Effects of Continuous-valued Interventions using Generative Adversarial Networks

Ioana Bica, James Jordon+1 more|arXiv (Cornell University)|Feb 27, 2020

Advanced Causal Inference Techniques38 references49 citations

TL;DR

SCIGAN introduces a GAN-based framework to estimate dose–response curves for continuous interventions, using a hierarchical discriminator and multi-task generator to learn counterfactual outcomes and infer personalized treatment effects.

ABSTRACT

While much attention has been given to the problem of estimating the effect of discrete interventions from observational data, relatively little work has been done in the setting of continuous-valued interventions, such as treatments associated with a dosage parameter. In this paper, we tackle this problem by building on a modification of the generative adversarial networks (GANs) framework. Our model, SCIGAN, is flexible and capable of simultaneously estimating counterfactual outcomes for several different continuous interventions. The key idea is to use a significantly modified GAN model to learn to generate counterfactual outcomes, which can then be used to learn an inference model, using standard supervised methods, capable of estimating these counterfactuals for a new sample. To address the challenges presented by shifting to continuous interventions, we propose a novel architecture for our discriminator - we build a hierarchical discriminator that leverages the structure of the continuous intervention setting. Moreover, we provide theoretical results to support our use of the GAN framework and of the hierarchical discriminator. In the experiments section, we introduce a new semi-synthetic data simulation for use in the continuous intervention setting and demonstrate improvements over the existing benchmark models.

Motivation & Objective

• Motivate the estimation of personalized effects for continuous (dosage) interventions from observational data.
• Develop a GAN-based approach to learn counterfactual outcome distributions across multiple treatments and dosages.
• Propose a hierarchical discriminator and a multi-task generator to handle continuous interventions.
• Provide theoretical justification for the GAN framework and hierarchical discriminator.
• Create a semi-synthetic data simulation to benchmark continuous-intervention methods.

Proposed method

• Define a counterfactual generator G that outputs a dose–response curve over all treatment-dosage pairs given x, observed factual treatment t_f and outcome y_f, plus random noise z.
• Train a hierarchical discriminator D_H composed of a treatment discriminator D_W and dosage discriminators D_w to identify factual treatment and dosage from generated curves, acting on finite sets of dose points.
• Regularize generator with a supervised loss that anchors its output at the observed factual dose (L_S).
• Use a multi-task generator with per-treatment heads to allow heterogeneous response curves across treatments.
• Provide theoretical justification (global GAN optimum implies matching marginal distributions of counterfactuals) and prove results under a hierarchical-discriminator setup.
• Inference: train an inference network I using generated counterfactuals to estimate outcomes for new samples across all treatments and dosages.

Experimental results

Research questions

• RQ1 Can a GAN-based framework learn the distribution of counterfactual outcomes for continuous treatments (dosages)?
• RQ2 Does a hierarchical discriminator improve stability and accuracy for learning dose–response curves across multiple treatments?
• RQ3 Can the learned counterfactuals be used to build accurate inference models for new samples with continuous interventions?
• RQ4 How does SCIGAN compare to existing methods (GPS, DRNets) in semi-synthetic settings for continuous dosages?

Key findings

• SCIGAN outperformed GPS and DRNets across semi-synthetic datasets (TCGA, News, MIMIC) on multiple metrics.
• A supervised loss, multitask generator, hierarchical discriminator, and invariant/equivariant discriminator components each contributed to performance gains, with the full model achieving the best results.
• SCIGAN maintains performance when increasing the number of discrete dosages and outperforms GANITE in discrete settings, illustrating applicability to both discrete and continuous interventions.
• SCIGAN shows robustness to treatment and dosage bias across varying bias levels.
• The method achieves statistically significant improvements over benchmarks on all datasets tested.