Data fusion methods for the heterogeneity of treatment effect and confounding function

Shu Yang, Donglin Zeng, Xiaofei Wang, Wang, XiaofeiarXiv 2020 · DOI · arXiv

Abstract

The heterogeneity of treatment effect (HTE) lies at the heart of precision medicine. Randomized controlled trials are gold-standard for treatment effect estimation but are typically underpowered for heterogeneous effects. In contrast, large observational studies have high predictive power but are often confounded due to the lack of randomization of treatment. We show that an observational study, even subject to hidden confounding, may be used to empower trials in estimating the HTE using the notion of confounding function. The confounding function summarizes the impact of unmeasured confounders on the difference between the observed treatment effect and the causal treatment effect, given the observed covariates, which is unidentifiable based only on the observational study. Coupling the trial and observational study, we show that the HTE and confounding function are identifiable. We then derive the semiparametric efficient scores and the integrative estimators of the HTE and confounding function. We clarify the conditions under which the integrative estimator of the HTE is strictly more efficient than the trial estimator. Finally, we illustrate the integrative estimators via simulation and an application.

Introduction

The heterogeneity of treatment effect (HTE) is a cornerstone of precision medicine, aiming to tailor medical interventions to individual patient characteristics. While randomized controlled trials (RCTs) are the gold standard for estimating treatment effects due to their internal validity, they are frequently limited by small sample sizes, making them underpowered for detecting heterogeneous effects across subgroups. Conversely, large-scale observational studies (OS), such as those derived from electronic health records, offer vast datasets with high predictive power. However, these studies are often plagued by hidden confounding—unmeasured factors that influence both treatment assignment and outcomes—leading to biased estimates. This paper proposes a novel data fusion framework to integrate these two disparate data sources, leveraging the strengths of each to identify and estimate HTE with high precision.

The Data Fusion Framework

The central challenge in combining RCT and OS data is the presence of unmeasured confounding in the latter. To address this, the authors introduce the concept of a "confounding function." This function, denoted as $\lambda(X)$, mathematically represents the bias introduced by unmeasured confounders in the observational study. It quantifies the difference between the observed treatment effect and the true causal treatment effect for a given set of observed covariates $X$.

Conceptual framework for data fusion, illustrating the sampling process from a superpopulation into randomized trial and observational study cohorts.

As shown in Figure 1, both the trial sample and the observational sample are viewed as being drawn from a common "superpopulation." While the trial employs randomization ($A_i \sim$ Randomization), the observational study's treatment assignment is governed by an unknown mechanism ($A_i \sim$ Unknown) that may depend on unmeasured variables. By "coupling" these two samples, the RCT serves as an unbiased anchor that allows researchers to identify the confounding function within the OS. Once identified, this function can be used to "de-bias" the large-scale observational data, effectively unlocking its statistical power for HTE estimation.

Statistical Methodology

The researchers derive semiparametric efficient scores for both the HTE and the confounding function. This approach ensures that the resulting "integrative estimator" is mathematically optimal, achieving the lowest possible variance among a broad class of estimators. A key feature of this method is its "rate-double robustness." The estimator remains consistent even if the nuisance models (such as the propensity score or the outcome mean model) are misspecified, provided their convergence rates are sufficiently fast.

Furthermore, the paper introduces an over-identification test. This goodness-of-fit diagnostic allows researchers to verify whether the structural assumptions made about the HTE and confounding functions are supported by the combined data. This is crucial for ensuring that the integration process does not introduce new biases due to model misspecification.

Empirical Validation

The proposed method was rigorously tested using both simulated data and a real-world clinical application. In simulations involving a trial of 1,000 participants and an observational study of 5,000, the integrative estimator demonstrated significantly lower bias and variance compared to estimators using trial data alone or standard meta-analysis techniques that ignore hidden confounding.

The practical utility of the framework was demonstrated in a study of adjuvant chemotherapy for non-small cell lung cancer. The researchers fused data from the CALGB 9633 trial ($n=319$) with a large observational sample from the National Cancer Database ($n=15,166$). While the original trial was underpowered to show a statistically significant benefit for the overall population, the integrative analysis revealed a significant treatment benefit for a specific subgroup: patients with tumor sizes between 3.6 cm and 6.8 cm. This finding highlights the power of data fusion to uncover clinically relevant insights that would remain hidden in smaller trial datasets.

Discussion and Insights

A critical insight from this work is the "efficiency gain" condition. The authors clarify that the integrative estimator is strictly more efficient than the trial-only estimator when certain predictors of the treatment effect are not present in the confounding function. This provides a theoretical roadmap for when data fusion is most beneficial.

The method also relies on the assumption of "transportability" — the idea that the fundamental mechanism of how a treatment works is consistent across the trial and the general population, even if the distributions of patient characteristics differ. By moving beyond simple sensitivity analysis and using trial data to actively estimate and remove bias, this framework represents a significant step forward in the robust application of big data to clinical decision-making.

Limitations and Future Directions

Despite its strengths, the method has limitations. It relies on parametric assumptions for the HTE and confounding functions, which may be restrictive in complex real-world scenarios. Additionally, in cases where there is no overlap in covariate distributions between the trial and the observational study, the results depend on model extrapolation. Future research aims to extend this framework to non-parametric models, allowing for even greater flexibility in capturing complex treatment effect patterns without predefined functional forms.

Summary updated

Abstract

The heterogeneity of treatment effect (HTE) lies at the heart of precision medicine. Randomized controlled trials are gold-standard for treatment effect estimation but are typically underpowered for heterogeneous effects. In contrast, large observational studies have high predictive power but are often confounded due to the lack of randomization of treatment. We show that an observational study, even subject to hidden confounding, may be used to empower trials in estimating the HTE using the notion of confounding function. The confounding function summarizes the impact of unmeasured confounders on the difference between the observed treatment effect and the causal treatment effect, given the observed covariates, which is unidentifiable based only on the observational study. Coupling the trial and observational study, we show that the HTE and confounding function are identifiable. We then derive the semiparametric efficient scores and the integrative estimators of the HTE and confounding function. We clarify the conditions under which the integrative estimator of the HTE is strictly more efficient than the trial estimator. Finally, we illustrate the integrative estimators via simulation and an application.
BibTeX
@misc{yang2020data,
  author = {Shu Yang and Donglin Zeng and Xiaofei Wang and Wang, Xiaofei},
  title = {{Data fusion methods for the heterogeneity of treatment effect and confounding function}},
  year = {2020},
  eprint = {2007.12922},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2007.12922},
}
Figure