Exploration of Heterogeneous Treatment Effects via Concave Fusion

Shujie Ma, Jian Huang, Zhiwei Zhang, Mingming LiuThe International Journal of Biostatistics 2016

Abstract

Understanding treatment heterogeneity is essential to the development of precision medicine, which seeks to tailor medical treatments to subgroups of patients with similar characteristics. One of the challenges to achieve this goal is that we usually do not have a priori knowledge of the grouping information of patients with respect to treatment. To address this problem, we consider a heterogeneous regression model by assuming that the coefficient for treatment variables are subject-dependent and belong to different subgroups with unknown grouping information. We develop a concave fusion penalized method for automatically estimating the grouping structure and the subgroup-specific treatment effects, and derive an alternating direction method of multipliers algorithm for its implementation. We also study the theoretical properties of the proposed method and show that under suitable conditions there exists a local minimizer that equals the oracle least squares estimator with a priori knowledge of the true grouping information with high probability. This provides theoretical support for making statistical inference about the subgroup-specific treatment effects based on the proposed method. We evaluate the performance of the proposed method by simulation studies and illustrate its application by analyzing the data from the AIDS Clinical Trials Group Study.

Introduction

In the quest for precision medicine, the primary objective is to tailor medical interventions to individual patients. However, treatment effects are frequently heterogeneous (HTE)—a therapy that is highly effective for one patient may be inert or even harmful for another due to underlying genetic, environmental, or lifestyle factors.

Traditional approaches to HTE often rely on "subgroup analysis," where researchers examine pre-defined cohorts (e.g., age or gender). The fundamental challenge arises when the characteristics defining these subgroups are unknown beforehand, particularly when they are driven by unmeasured variables or complex interactions. This paper addresses the "unknown grouping" problem: identifying latent subgroups and estimating their specific treatment effects simultaneously without prior knowledge of group membership.

Methodology: The Concave Fusion Approach

The authors propose a heterogeneous regression model where each patient $i$ is assigned an individual treatment effect coefficient, $\beta_i$. To make this model estimable in a high-dimensional context, they introduce a concave fusion penalized method.

The Penalty and Clustering

The core innovation lies in applying a penalty to the pairwise differences between all patients' treatment coefficients ($|\beta_i - \beta_j|$). This encourages coefficients that are numerically close to collapse into the exact same value, effectively clustering patients into subgroups automatically. Unlike the standard Lasso penalty, which can over-shrink coefficients toward zero and introduce bias, the authors utilize concave penalties like the Minimax Concave Penalty (MCP) or the Smoothly Clipped Absolute Deviation (SCAD). These are "nearly unbiased" for large differences, allowing for sharper separation between distinct subgroups.

Visualizing Group Discovery: The Fusiongram

To visualize this process, the authors introduce the "fusiongram." Similar to a dendrogram in hierarchical clustering, the fusiongram illustrates how patient coefficients merge as the regularization strength ($\lambda$) increases.

Fusiongrams for MCP, SCAD, and LASSO penalties showing the path of coefficient fusion as the regularization parameter $\lambda$ increases.

As shown in the figure, concave penalties like MCP and SCAD provide a clearer path for subgroup identification compared to LASSO. The optimal number of subgroups ($K$) is selected along this path using data-driven criteria such as the Bayesian Information Criterion (BIC).

Implementation and Theoretical Guarantees

Solving the optimization problem created by $n(n-1)/2$ pairwise penalties is computationally demanding. The authors derive an efficient algorithm using the Alternating Direction Method of Multipliers (ADMM) to handle this complexity.

Crucially, the paper proves that the estimator possesses the "oracle property." This means that, under suitable conditions, the method performs as well as a model where the true subgroup memberships were known from the start. This theoretical foundation allows for robust statistical inference regarding subgroup-specific effects.

Empirical Results

Simulation Studies

In controlled simulations with two latent subgroups ($n=400$), the method correctly identified the number of groups over 92% of the time, achieving a high Rand Index (0.826) for clustering accuracy. It consistently outperformed standard logistic-normal mixture models, which often fail when the underlying group membership logic is mis-specified.

Application to HIV Treatment (ACTG 175)

The method was applied to the AIDS Clinical Trials Group Study 175, which compared zidovudine (AZT) with other therapies. While a standard global model suggested that the drug didanosine was not significantly different from AZT, the concave fusion method revealed two distinct subgroups: one where didanosine had no significant effect, and another where it was highly effective (treatment effect of 1.687, $p < 0.001$). This explains why the average effect appeared insignificant in traditional analysis.

Conclusion and Future Directions

This framework provides a powerful tool for discovering latent heterogeneity in treatment responses. By basing clusters on the treatment effect itself rather than baseline covariates, it can capture heterogeneity driven by unmeasured confounders.

Future work aims to address the computational scaling of the $O(n^2)$ pairwise penalties for very large datasets ($n > 10,000$) and to extend the framework to high-dimensional covariates ($p, q > n$) and non-linear models, such as Cox proportional hazards for survival data.

Summary updated

Abstract

Understanding treatment heterogeneity is essential to the development of precision medicine, which seeks to tailor medical treatments to subgroups of patients with similar characteristics. One of the challenges of achieving this goal is that we usually do not have a priori knowledge of the grouping information of patients with respect to treatment effect. To address this problem, we consider a heterogeneous regression model which allows the coefficients for treatment variables to be subject-dependent with unknown grouping information. We develop a concave fusion penalized method for estimating the grouping structure and the subgroup-specific treatment effects, and derive an alternating direction method of multipliers algorithm for its implementation. We also study the theoretical properties of the proposed method and show that under suitable conditions there exists a local minimizer that equals the oracle least squares estimator based on a priori knowledge of the true grouping information with high probability. This provides theoretical support for making statistical inference about the subgroup-specific treatment effects using the proposed method. The proposed method is illustrated in simulation studies and illustrated with real data from an AIDS Clinical Trials Group Study.
BibTeX
@article{ma2016exploration,
  author = {Shujie Ma and Jian Huang and Zhiwei Zhang and Mingming Liu},
  title = {{Exploration of Heterogeneous Treatment Effects via Concave Fusion}},
  journal = {The International Journal of Biostatistics},
  year = {2016},
}
Figure