Correction to: Graphing and reporting heterogeneous treatment effects through reference classes

James A Watson, Chris HolmesTrials 2020 · DOI · OpenAlex

Abstract

This document serves as a formal correction to the methodology presented in "Graphing and reporting heterogeneous treatment effects through reference classes," originally published in Trials in 2020. The amendments rectify fundamental mathematical errors in the Individual Treatment Effect (ITE) estimation formula and a recurring scaling typo. These corrections are essential for ensuring that treatment and control outcomes are appropriately weighted when visualizing heterogeneous patient responses.

Introduction

The original research introduced a framework for analyzing Heterogeneous Treatment Effects (HTE) using "reference classes"—subgroups of patients with similar characteristics. This approach is vital for precision medicine, as it allows clinicians to visualize how interventions vary in efficacy across diverse patient populations. However, the mathematical foundation for calculating the ITE within these classes contained errors that could lead to biased results, particularly in datasets with imbalanced treatment and control arms.

Revised Methodology

The correction implements two critical changes to the mathematical framework to ensure statistical rigor.

1. Decoupled Normalization Constants

In the original publication, the formula for estimating the Individual Treatment Effect at a given point $q_i$ utilized a single normalization constant ($C$) for both the treatment and control groups. This approach assumes a balanced distribution of weights and sample sizes between the two arms within a reference class—an assumption that rarely holds in practice. If a reference class contains significantly more treated patients than control patients, a global normalization factor would heavily distort the perceived treatment effect.

To resolve this, the authors have decoupled the normalization. The corrected formula for $ITE(q_i)$ is:

$$ITE(q_i) = C_1 \sum_k w_k y_k \mathbf{1}(t_k = T_1) - C_0 \sum_k w_k y_k \mathbf{1}(t_k = T_0)$$

Where the constants $C_1$ and $C_0$ are defined independently for the treatment group ($T_1$) and control group ($T_0$) as the inverse of the sum of weights for each respective group:

$$C_j = \left(\sum_k w_k \mathbf{1}(t_k = T_j)\right)^{-1} \text{ for } j \in \{0, 1\}$$

This adjustment ensures that the estimator calculates a true weighted average for each arm before determining the difference, preventing imbalances in group size or weight distribution from skewing the results.

2. Parameter Scaling Correction

A recurring typo was identified in Equation (2) and subsequent sections regarding the parameter $\gamma$. The original text incorrectly represented the term as $\gamma/N$. The corrected term is $\gamma N$.

This change is significant for the model's sensitivity to sample size ($N$). In many statistical contexts, parameters representing smoothing or regularization must scale upward with $N$ to maintain consistent behavior as more data is added. The shift from an inverse relationship to a direct relationship fundamentally alters the model's scaling logic.

Conclusion

These mathematical corrections are necessary for the accurate application of the reference class method in clinical research. By implementing group-specific normalization and correcting the scaling of the $\gamma$ parameter, researchers can produce more reliable visualizations of treatment effect heterogeneity. This ensures that the reported ITE reflects a rigorous comparison of subgroup averages, providing a more accurate foundation for clinical decision-making.

Summary updated

Abstract

An amendment to this paper has been published and can be accessed via the original article.
BibTeX
@inproceedings{watson2020correction,
  author = {James A Watson and Chris Holmes},
  title = {{Correction to: Graphing and reporting heterogeneous treatment effects through reference classes}},
  booktitle = {Trials},
  year = {2020},
  doi = {10.1186/s13063-020-04782-5},
}

Loading related content...

Figure