IGC-Net for conditional average potential outcome estimation over time

1. Problem Statement

The paper addresses the challenge of estimating Conditional Average Potential Outcomes (CAPOs) from observational data over time. This is critical for personalized medicine, where clinicians need to predict patient outcomes under hypothetical future treatment sequences (e.g., "What will happen to Patient X if they receive Treatment A, then B, then C?").

Core Challenges:

• Time-Varying Confounding: In longitudinal settings, future covariates and outcomes influence future treatment decisions. These "future confounders" are unobserved during inference (runtime confounding), making standard regression approaches biased.
• Limitations of Existing Methods:
  • Methods without proper adjustment (e.g., CRN, CT, TE-CDE): These rely on balanced representations but fail to target the correct causal estimand, leading to infinite-sample bias (errors that do not vanish even with infinite data).
  • Inverse Propensity Weighting (IPW) methods (e.g., RMSNs): These adjust for confounding but rely on products of inverse propensity scores. In multi-step predictions, this leads to division by values close to zero, causing high variance and instability.
  • Distribution-based G-computation (e.g., G-Net, G-transformer): These use G-computation but require estimating the entire joint distribution of all future time-varying confounders (all higher-order moments). This is computationally inefficient and suffers from the curse of dimensionality, often requiring Monte Carlo (MC) sampling which introduces noise.

2. Methodology: Iterative G-Computation Network (IGC-Net)

The authors propose IGC-Net, a novel neural end-to-end model that performs regression-based iterative G-computation.

Theoretical Foundation

IGC-Net reformulates the G-computation formula (Equation 3) as a sequence of nested conditional expectations. Instead of integrating over high-dimensional distributions, it defines pseudo-outcomes recursively:

1. Base Case: The final pseudo-outcome is the observed outcome $G^{\bar{a}}_{t+\tau} = Y_{t+\tau}$.
2. Recursive Step: For earlier time steps $\delta < \tau$, the pseudo-outcome is the expected value of the next pseudo-outcome given the history and treatment:
   $$G^{\bar{a}}_{t+\delta} = E[G^{\bar{a}}_{t+\delta+1} \mid \bar{H}_{t+\delta}, A_{t+\delta}]$$
   The goal is to learn the function $g^{\bar{a}}_t(\bar{h}_t)$ that estimates the CAPO by iteratively regressing these pseudo-outcomes.

Architecture

IGC-Net consists of two main components (see Figure 1 in the paper):

1. Neural Backbone ($z_\phi$): A multi-input Transformer (or LSTM) that encodes the full observed history $\bar{H}_t$ (outcomes, covariates, past treatments) into hidden states. It uses separate sub-transformers for different input types (outcomes, covariates, treatments) with cross-attention mechanisms to share information.
2. G-Computation Heads ($\{g^\phi_\delta\}$): A set of feed-forward networks that take the hidden states and current treatments as input to predict the pseudo-outcomes for each step in the future.

Training Algorithm (Iterative Scheme)

Since ground-truth pseudo-outcomes for intermediate steps ($\delta < \tau$) are unobserved, IGC-Net employs an iterative training loop consisting of two alternating steps:

• Step A (Generation): The network generates predictions for the missing pseudo-outcomes $\tilde{G}^{\bar{a}}_{t+\delta}$ using the current model parameters. These are detached from the computational graph to act as fixed targets for the next step.
• Step B (Learning): The network performs a regression to minimize the squared error between the predicted pseudo-outcomes and the generated targets (or observed outcomes for the final step).
• End-to-End Optimization: This process is unrolled and optimized jointly. As training progresses, the predictions for earlier steps become more accurate because they are trained on increasingly accurate targets generated by the later steps (which are trained on ground truth).

3. Key Contributions

1. First Neural End-to-End G-Computation: IGC-Net is the first neural model to perform fully regression-based iterative G-computation for CAPOs in a time-varying setting without Monte Carlo sampling.
2. Variance Reduction: By avoiding Inverse Propensity Weighting (IPW), IGC-Net eliminates the instability caused by dividing by small propensity scores (Proposition 3).
3. Dimensionality Efficiency: Unlike G-Net, IGC-Net does not estimate the full joint distribution of future confounders. It only estimates the first moment (mean) of a lower-dimensional random variable, avoiding the curse of dimensionality.
4. Theoretical Guarantees: The paper provides proofs (Propositions 1 & 2) demonstrating that the iterative regression recovers the true CAPO and properly adjusts for time-varying confounding under standard identifiability assumptions (Consistency, Positivity, Sequential Ignorability).

4. Experimental Results

The authors evaluated IGC-Net against state-of-the-art baselines (CRN, CT, TE-CDE, RMSNs, G-Net, G-transformer) on three types of datasets:

• Synthetic Data (Tumor Growth Model):

  • Result: IGC-Net consistently outperformed all baselines across varying levels of confounding strength ($\gamma$).
  • Improvement: Achieved up to 17.4% relative improvement in RMSE over the best baseline.
  • Stability: Showed significantly lower variance compared to IPW-based methods (RMSNs) and distribution-based methods (G-Net).

• Semi-Synthetic Data (MIMIC-III Extract):

  • Setup: High-dimensional covariates ($d_x=25$) and varying prediction horizons ($\tau=2$ to $6$).
  • Result: IGC-Net remained robust as the prediction horizon increased, whereas baselines like G-Net and RMSNs degraded significantly due to the curse of dimensionality and overlap violations.
  • Improvement: Up to 26.7% relative improvement in RMSE.

• Real-World Data (MIMIC-III ICU):

  • Task: Factual outcome prediction (sanity check).
  • Result: IGC-Net achieved state-of-the-art performance, matching or beating CT and CRN, demonstrating its effectiveness even when causal adjustments are not strictly required.

• Ablation Studies:

  • Replacing the Transformer backbone with an LSTM (IGC-LSTM) still outperformed baselines, proving the efficacy of the iterative G-computation algorithm itself.
  • A "biased transformer" (without iterative generation) performed poorly, highlighting the necessity of the proper adjustment mechanism.

5. Significance and Impact

• Personalized Medicine: IGC-Net provides a reliable tool for simulating "what-if" scenarios in clinical settings, enabling doctors to optimize treatment sequences for individual patients based on electronic health records (EHRs).
• Methodological Advancement: It bridges the gap between classical causal inference (G-computation) and modern deep learning, offering a solution that is both theoretically sound (unbiased) and practically efficient (low variance, no MC sampling).
• Robustness: The method is particularly valuable in scenarios with high-dimensional data and long prediction horizons, where existing neural causal methods typically fail due to bias or variance explosion.

In conclusion, IGC-Net represents a significant step forward in causal machine learning for time-series data, offering a robust, scalable, and accurate framework for estimating the effects of dynamic treatment regimes.