Robust and Data-Adaptive Integration of Nonconcurrent Data in Platform Trials via Gaussian Processes

Technical Summary: Robust and Data-Adaptive Integration of Nonconcurrent Data in Platform Trials via Gaussian Processes

Problem Statement
Platform trials offer an efficient design for evaluating multiple treatments simultaneously under a single master protocol. However, statistical inference in these trials faces a critical tension between efficiency and validity. Traditional robust methods restrict analysis to "concurrently randomized" participants—those enrolled during periods when all treatments being compared are available—to avoid bias from temporal trends (e.g., shifts in patient populations or standard of care). While this ensures validity, it discards "nonconcurrent" data (typically from control arms enrolled before a new treatment arm was introduced), resulting in a loss of statistical power and precision.

Existing approaches to incorporate nonconcurrent data often suffer from limitations: they may lack formally defined causal estimands, rely on parametric models that are either too simplistic or opaque, or depend on strong assumptions like conditional exchangeability. Furthermore, many methods do not provide transparent mechanisms for how individual observations contribute to the final estimate, making it difficult to diagnose whether borrowing information is beneficial or harmful in a specific application.

Methodology
The authors propose a flexible, interpretable Gaussian Process (GP) framework designed to incorporate nonconcurrent data while controlling for bias. The methodology is grounded in causal inference and machine learning, targeting the "entirely concurrently eligible" (ECE) estimand defined by Qian et al. (2025).

1. Framework and Estimand:
   The target estimand is the contrast between expected potential outcomes for two treatments within the ECE population (individuals eligible for both treatments during periods when both are available). The framework utilizes enrollment calendar time as a critical covariate, exploiting the "temporal smoothness" inherent in platform trials where patient distributions evolve gradually over time.

2. Gaussian Process Models:

   • Single-Task GP: Fits a separate GP for each treatment arm, modeling the outcome trajectory as a latent function of enrollment time plus noise. This captures temporal smoothness via a kernel function (e.g., squared exponential or Matérn).
   • Multi-Task GP: Decomposes outcomes into a shared baseline function (common across arms) and a treatment effect function. This structure exploits the fact that while baseline risks may shift over time, treatment effects often remain relatively stable. The model uses a composite kernel to capture both within-arm variance and between-arm covariance.

3. Estimation and Inference:
   The framework fits the GP using all eligible participants (concurrent and nonconcurrent) but averages the treatment effect estimate only over the concurrently eligible population. This anchors the estimand to the ECE definition while allowing nonconcurrent data to adaptively inform the GP fit. Inference is conducted via Bayesian bootstrap to account for both posterior uncertainty and sampling variability.

4. Frequentist Interpretation:
   The authors establish a connection between the GP posterior mean and Kernel Ridge Regression (KRR). This yields a transparent frequentist interpretation where the estimator is a weighted linear combination of observed outcomes. The weights depend on the kernel hyperparameters and the proximity of enrollment times, explicitly quantifying the contribution of nonconcurrent data.

5. Extensions:
   The framework is extended to handle discrete outcomes (binary or count) via generalized linear models with exponential family likelihoods and to adjust for baseline covariates through additive or product kernels.

Key Theoretical Guarantees
The paper establishes two primary theoretical results regarding the properties of the proposed estimators:

• Variance Reduction: Incorporating nonconcurrent controls is proven to reduce the posterior variance of both the outcome mean (single-task) and the treatment effect function (multi-task). This holds because conditioning a Gaussian process on additional variables never increases conditional variance.
• Bias Control: The bias introduced by nonconcurrent data is bounded by a quantity that does not increase when nonconcurrent controls are added. While the bias itself is not guaranteed to decrease, the bound is non-increasing, ensuring that the efficiency gains do not come at the cost of uncontrolled bias inflation, provided the working model is reasonable.

Results
The authors validate the method through simulation studies and a real-data application:

• Simulations: In a simulated platform trial with two enrollment windows, the proposed GP methods (both single-task and multi-task) demonstrated superior performance compared to concurrent-only benchmarks (ANOVA, ANHECOVA) and other nonconcurrent methods (Bayesian Time Machine, LR-EW).
  • The multi-task GP incorporating all data achieved a smaller standard error (0.395) than any concurrent-only method (e.g., 0.420 for ANHECOVA) while maintaining near-nominal coverage (94.4%).
  • Competing methods showed significant drawbacks: the Bayesian Time Machine exhibited high sensitivity to prior specification with overly wide credible intervals, while the linear model with enrollment window adjustment (LR-EW) suffered from bias when temporal trends were nonlinear.
  • Even under misspecified conditions (abrupt temporal shifts), the GP methods maintained controlled bias and outperformed competitors, though coverage decreased slightly.
• Real Data Application: The method was applied to a hypothetical platform trial constructed from the SURMOUNT-1 obesity trial. The GP methods incorporating nonconcurrent controls reduced the within-imputation variance by 12–16% compared to concurrent-only analyses, with stable point estimates, suggesting successful efficiency gains without introducing meaningful bias.

Significance and Claims
The paper claims to offer a principled solution to the efficiency-validity trade-off in platform trials. By leveraging the temporal smoothness of enrollment, the proposed GP framework allows for the robust integration of nonconcurrent data without discarding valuable information. Key contributions include:

• Transparency: The connection to kernel ridge regression provides explicit weights for each observation, allowing researchers to diagnose the influence of nonconcurrent data.
• Theoretical Rigor: The establishment of variance reduction and non-increasing bias bounds provides a formal guarantee of the method's robustness.
• Flexibility: The framework accommodates various outcome types and covariate adjustments, extending beyond the limitations of previous parametric or nonparametric approaches.
• Implementation: The authors provide an implementation in the R package RobinCID, facilitating practical application.

The authors conclude that while the method improves estimation efficiency, it remains essential to examine temporal trends, as the degree of borrowing and potential bias depends on the validity of the smoothness assumption. The paper does not claim to eliminate bias entirely but rather to control it within a non-increasing bound while significantly reducing variance.