Median Optimal Treatment Regimes

Liu Leqi, Edward H. KennedyarXiv 2021 · OpenReview

Abstract

Optimal treatment regimes (OTRs) aim to personalize medical decisions to maximize patient outcomes. While standard OTRs focus on maximizing the population mean, this can lead to policies sensitive to outliers and influenced by unrelated subgroups. We propose a median optimal treatment regime that maximizes the Average Conditional Median Effect (ACME). This framework provides robustness to extreme outcomes and ensures treatment decisions for one group are not affected by distribution shifts in another. We derive a nonparametric efficiency bound and a doubly robust estimator for the ACME, demonstrating its performance through simulations and an application to HIV clinical trial data.

Introduction

In personalized medicine, an Optimal Treatment Regime (OTR) is a decision rule that assigns treatments based on individual patient characteristics (covariates). Traditionally, these regimes are designed to maximize the mean outcome across a population. However, mean-based approaches suffer from two significant drawbacks:

  1. Sensitivity to Outliers: Means are easily skewed by a small fraction of patients with extreme reactions, leading to decisions that may not benefit the "typical" patient.
  2. Across-Group Unfairness: Using marginal medians or quantiles can lead to "spooky action at a distance," where a treatment decision for one demographic group changes solely because of outcome shifts in a completely unrelated group.

This paper introduces a framework centered on the conditional median, ensuring that treatment decisions are both statistically stable and locally determined by the relevant patient subgroup.

The ACME Framework and Methodology

The core of this approach is the Average Conditional Median Effect (ACME). Unlike the mean, which averages all expected outcomes, ACME averages the medians of outcomes across different covariate groups.

To implement this, the authors developed three technical innovations:

  • Doubly Robust-Style Estimators: They derived the Efficient Influence Function (EIF) for the ACME to create an estimator that remains accurate even when nuisance models (like propensity scores) contain some error.
  • Pseudo-Outcome Regression: To learn the optimal policy, the authors estimate the Conditional Median Treatment Effect (CMTE). This involves generating "pseudo-outcomes" for patients and regressing them against covariates to identify the decision boundary where treatment utility exceeds the control.
  • Efficiency Bounds: The paper establishes a rigorous semi-parametric efficiency bound, proving their estimator achieves optimal statistical efficiency under weak conditions.

Key Insights: Robustness and Fairness

A primary motivation for this work is "Within-Group Robustness." By focusing on the median, the policy ignores "miracle cures" or "catastrophic failures" that affect only a tiny fraction of a specific patient type.

Figure 1: Comparison of outcome densities for treatment (A=1) and control (A=0) in a specific covariate group, illustrating how outliers can skew mean-based decisions.

As shown in Figure 1, for a specific group $X=G_2$, the treatment ($A=1$) results in a bimodal distribution. While a small subgroup experiences a very high outcome (near $Y=22$), the majority of patients have outcomes centered around $Y=-5$. A mean-based policy might be pulled toward $A=1$ by the outlier peak, whereas a median-based policy correctly identifies that the typical patient fares better under the control ($A=0$, median near 0).

Furthermore, the conditional median approach addresses "Across-Group Fairness." It prevents distribution shifts in one part of the population from flipping treatment decisions for a different, unrelated part of the population—a common failure mode of marginal median methods.

Results and Applications

The authors validated their framework using both synthetic simulations and real-world clinical data.

Simulation Results

In data-generating processes with skewed lognormal distributions, the proposed doubly robust (DR) estimator significantly outperformed standard "plug-in" estimators. The DR estimator's error decreased rapidly as sample size increased, demonstrating high efficiency even when nuisance models were slightly misspecified.

HIV Clinical Trial (ACTG 175)

The algorithm was applied to data from the ACTG 175 trial, which compared HIV monotherapy to combination therapies. The outcome, CD4 T cell count, was highly skewed. The analysis revealed that:

  • The mean-based policy was heavily influenced by a small subgroup with exceptionally high responses.
  • The ACME-based policy provided a more stable, conservative recommendation that maximized the outcome for the typical patient.
  • The ACME policy achieved the highest estimated value (342.40) compared to mean-based or "treat-all" strategies.

Conclusion and Future Directions

While the median OTR offers significant advantages in robustness and fairness, it presents unique challenges. Unlike mean-based models, the median estimator is not "strictly" doubly robust and generally requires consistent estimation of the conditional median. Future work may explore "V-specific" regimes—restricting policies to a subset of available covariates—and seek to reduce the computational complexity associated with estimating conditional densities at the median.

Summary updated

Abstract

Optimal treatment regimes are personalized policies for making a treatment decision based on subject characteristics, with the policy chosen to maximize some value. It is common to aim to maximize the mean outcome in the population, via a regime assigning treatment only to those whose mean outcome is higher under treatment versus control. However, the mean can be an unstable measure of centrality, resulting in imprecise statistical procedures, as well as unrobust decisions that can be overly influenced by a small fraction of subjects. In this work, we propose a new median optimal treatment regime that instead treats individuals whose conditional median is higher under treatment. This ensures that optimal decisions for individuals from the same group are not overly influenced either by (i) a small fraction of the group (unlike the mean criterion), or (ii) unrelated subjects from different groups (unlike marginal median/quantile criteria). We introduce a new measure of value, the Average Conditional Median Effect (ACME), which summarizes across-group median treatment outcomes of a policy, and which the median optimal treatment regime maximizes. After developing key motivating examples that distinguish median optimal treatment regimes from mean and marginal median optimal treatment regimes, we give a nonparametric efficiency bound for estimating the ACME of a policy, and propose a new doubly robust-style estimator that achieves the efficiency bound under weak conditions. To construct the median optimal treatment regime, we introduce a new doubly robust-style estimator for the conditional median treatment effect. Finite-sample properties are explored via numerical simulations and the proposed algorithm is illustrated using data from a randomized clinical trial in patients with HIV.
BibTeX
@misc{leqi2021median,
  author = {Liu Leqi and Edward H. Kennedy},
  title = {{Median Optimal Treatment Regimes}},
  year = {2021},
  eprint = {2103.01802},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2103.01802},
}
Figure