Abstract
Perhaps the most pressing concern of a patient diagnosed with cancer is her life expectancy under various treatment options. For a binary-treatment case, this translates into estimating the difference between the outcomes (e.g., survival time) of the two available treatment options – i.e., her Individual Treatment Effect (ITE). This is especially challenging to estimate from observational data, as that data has selection bias: the treatment assigned to a patient depends on that patient's attributes. In this work, we borrow ideas from domain adaptation to address the distributional shift between the source (outcome of the administered treatment, appearing in the observed training data) and target (outcome of the alternative treatment) that exists due to selection bias. We propose a context-aware importance sampling re-weighing scheme, built on top of a representation learning module, for estimating ITEs. Empirical results on two publicly available benchmarks demonstrate that the proposed method significantly outperforms state-of-the-art.
Introduction: The Challenge of Selection Bias
In precision medicine, the ultimate goal is to determine the Individual Treatment Effect (ITE): the difference in outcomes if a specific patient received Treatment A versus Treatment B. However, researchers face the "fundamental problem of causal inference"—we only ever observe the outcome of the treatment actually administered (the factual). The outcome of the alternative treatment (the counterfactual) remains hidden.
When using historical "observational" data to train models, we encounter selection bias. In clinical practice, treatments are rarely assigned randomly; for instance, doctors might prescribe surgery to younger, healthier patients while opting for medication for older ones. This creates a significant distributional shift: the population receiving Treatment A looks fundamentally different from the population receiving Treatment B.
Standard machine learning models struggle to generalize across this shift, often leading to poor predictions for counterfactual scenarios where the model must estimate outcomes for a patient type it rarely saw receiving that specific treatment.
Methods: CFR-ISW
The authors propose CFR-ISW (CounterFactual Regression with Importance Sampling Weights), a framework that integrates deep representation learning with a novel weighting strategy to mitigate selection bias.
1. Representation Learning
The model employs a neural network to map high-dimensional patient features ($x$) into a latent representation space ($\Phi(x)$). The objective is to find a transformation where the distributions of the treated and control groups overlap more closely, facilitating more reliable comparisons.
2. Context-Aware Weighting
The primary contribution of this work is the "context-aware" importance sampling scheme. While traditional methods like Inverse Propensity Weighting (IPW) use global weights based on the overall probability of treatment assignment, CFR-ISW utilizes weights tailored to the individual's specific "context" within the latent representation space. This ensures the model prioritizes training on samples that most closely resemble the counterfactual cases it needs to predict.
3. Bayesian Weight Estimation and Optimization
To avoid the computational complexity of estimating high-dimensional density functions, the authors leverage Bayes' theorem to redefine their weights. They train a secondary "propensity network" to estimate the probability of treatment given the latent representation. The training process then alternates between updating the representation/outcome networks and refining the propensity network weights.
Experimental Evaluation
The authors evaluated CFR-ISW on two standard benchmarks where ground-truth counterfactuals are available via simulation:
- IHDP (Infant Health and Development Program): A semi-synthetic dataset (747 instances, 25 covariates) simulating the effect of home visits on cognitive scores. Artificial selection bias is introduced to mimic real-world conditions.
- ACIC’18 (Atlantic Causal Inference Conference): A large-scale suite of 24 synthetic datasets (1,000 to 50,000 instances, 177 features), providing a rigorous test of scalability and robustness.
Key Results
CFR-ISW demonstrated state-of-the-art performance across multiple metrics, including PEHE (Precision in Estimation of Heterogeneous Effect) and ENoRMSE (Effect-Normalized Root Mean Squared Error).
- Performance on IHDP: CFR-ISW significantly outperformed standard CounterFactual Regression (CFR) and Re-weighted CFR (RCFR). The authors also noted that using BART (Bayesian Additive Regression Trees) as a surrogate model for hyperparameter validation was more effective than traditional nearest-neighbor approaches.
- Performance on ACIC’18: The results were particularly striking on this complex dataset. On the aggregate "ALL" suite, CFR-ISW achieved an ENoRMSE of 1.03, vastly outperforming the standard CFR model (5.43) and BART (9.35).
- Robustness: The method maintained superior performance across varying dataset sizes, from 1,000 to 50,000 instances.
Conclusion and Future Work
CFR-ISW offers a powerful approach to causal inference by combining the strengths of representation learning with a sophisticated, context-aware weighting mechanism. While currently limited to binary treatment scenarios and reliant on the "Strong Ignorability" assumption (no unobserved confounders), it sets a new benchmark for ITE estimation. Future research aims to extend this importance sampling logic to multi-categorical and continuous treatment variables.
Summary updated