Statistical inference for heterogeneous treatment effect with right-censored data from synthesizing randomized clinical trials and real-world data

Guangcai Mao, Shu Yang, Xiaofei Wang

TL;DR

This paper introduces a novel statistical framework to estimate Heterogeneous Treatment Effects (HTE)—how medical treatments affect different patients differently—by combining high-quality Randomized Clinical Trial (RCT) data with large-scale, but potentially biased, Real-World Data (RWD). By using a flexible "omnibus bias function" and a penalized sieve estimation method, the authors provide a way to "borrow" strength from messy real-world evidence to make precision medicine more accurate and statistically efficient.

What Problem Does This Paper Address?

In the era of precision medicine, doctors want to know not just if a drug works on average, but if it works for this specific patient based on their age, genetics, or tumor size. This is the study of Heterogeneous Treatment Effects (HTE).

The "gold standard" for this is the Randomized Clinical Trial (RCT). However, RCTs are often small, expensive, and have strict entry criteria that don't reflect the general population. Real-World Data (RWD), such as electronic health records or insurance registries, offers massive sample sizes and diversity but comes with a major catch: bias. RWD is plagued by unmeasured confounders (factors that influence both treatment and outcome that aren't recorded), different censoring patterns (patients dropping out of study), and outcome heterogeneity.

Existing methods often assume RWD is unbiased or only address one type of bias at a time. This paper tackles the challenge of survival data (time-to-event) where outcomes are "right-censored" (we don't know when an event happens for everyone), and it accounts for multiple sources of RWD bias simultaneously.

Methods

The authors focus on a metric called Conditional Restricted Mean Survival Time (CRMST), which represents the average life expectancy of a patient up to a specific time horizon (e.g., 3 years), given their specific characteristics.

The key innovations include:

  1. The Omnibus Bias Function ($\lambda(X)$): Instead of trying to model every source of bias individually, the authors define a single, flexible function that captures the total discrepancy between the RCT and the RWD.
  2. Penalized Sieve Estimation: To estimate the HTE and the bias function without assuming they follow a simple linear shape, the authors use "sieves"—a method that approximates complex functions using a combination of simpler basis functions (like B-splines). They add "roughness penalties" to prevent the model from overfitting to noise.
  3. Double Robustness: The framework is designed so that the estimates remain consistent even if some parts of the statistical model (like the censoring mechanism) are slightly misspecified, provided other parts are correct.
  4. Theoretical Grounding: The authors use Reproducing Kernel Hilbert Space (RKHS) theory to prove that their estimator converges to the truth at an optimal rate and follows a normal distribution, allowing for the construction of reliable confidence intervals.

Datasets

The methodology was evaluated using both simulations and a high-stakes medical application:

  • Simulated Data: The authors ran extensive tests with varying numbers of covariates ($p=2$ and $p=4$) and different bias scenarios to prove their method outperforms RCT-only or RWD-only approaches.
  • CALGB 140503 (RCT): A Phase III trial of 697 patients with early-stage non-small-cell lung cancer (NSCLC) comparing two surgical techniques: lobectomy (removing a whole lung lobe) vs. limited resection (removing a smaller section).
  • National Cancer Database (NCDB) (RWD): A massive registry of 17,995 patients used as the supplementary real-world evidence.

Key Insights and Ideas

  • Information Borrowing: The core "aha" moment is that even biased data contains a "signal" about the structure of treatment effects. By explicitly modeling the bias, you can extract that signal to reduce the uncertainty (variance) of the RCT estimate without letting the RWD's biases pull your answer in the wrong direction.
  • Model-Free Flexibility: By using nonparametric sieve methods, the model can discover complex, non-linear interactions between patient traits and treatment outcomes that a standard regression might miss.
  • Efficiency Gains: The math proves (and the data confirms) that the integrative estimator is always at least as efficient as using the RCT data alone.

Key Results

In the lung cancer case study, the integrative approach provided much more granular insights than the RCT data alone:

  • Increased Precision: The proposed method reduced the standard error of treatment effect estimates by up to 52% in certain patient subgroups compared to using only trial data.
  • Broader Significance: While the original RCT showed limited resection was "non-inferior" on average, the new method identified specific regions (younger patients with larger tumors) where lobectomy provided a statistically significant survival benefit.
  • Narrower Confidence Intervals: The "blue regions" in their visualizations—representing statistically significant findings—were much more extensive in the integrative model, giving clinicians more confidence in tailoring surgical choices.

Limitations and Future Work

  • Boundary Effects: Like many nonparametric methods, the estimates can be less accurate at the "edges" of the data (e.g., for extremely old or extremely young patients).
  • The Curse of Dimensionality: Sieve methods work well for a handful of covariates, but if a patient has hundreds of relevant traits, the model's performance may degrade.
  • Static Treatments: The current model assumes a one-time treatment. The authors suggest that future work could extend this to "marginal structural models" to handle treatments that change over time.

Summary updated

Abstract

The heterogeneous treatment effect plays a crucial role in precision medicine. There is evidence that real-world data, even subject to biases, can be employed as supplementary evidence for randomized clinical trials to improve the statistical efficiency of the heterogeneous treatment effect estimation. In this paper, for survival data with right censoring, we consider estimating the heterogeneous treatment effect, defined as the difference of the treatment-specific conditional restricted mean survival times given covariates, by synthesizing evidence from randomized clinical trials and the real-world data with possible biases. We define an omnibus bias function to characterize the effect of biases caused by unmeasured confounders, censoring, and outcome heterogeneity, and further, identify it by combining the trial and real-world data. We propose a penalized sieve method to estimate the heterogeneous treatment effect and the bias function. We further study the theoretical properties of the proposed integrative estimators based on the theory of reproducing kernel Hilbert space and empirical process. The proposed methodology outperforms the approach solely based on the trial data through simulation studies and an integrative analysis of the data from a randomized trial and a real-world registry on early-stage non-small-cell lung cancer.
BibTeX
@inproceedings{mao2025statistical,
  author = {Guangcai Mao and Shu Yang and Xiaofei Wang},
  title = {{Statistical inference for heterogeneous treatment effect with right-censored data from synthesizing randomized clinical trials and real-world data}},
  booktitle = {Biometrics},
  year = {2025},
  doi = {10.1093/biomtc/ujaf131},
}

Loading related content...

Figure