📊 statistics

Doubly Robust Targeted Estimation of Subgroup Average Treatment Effects for Time-to-event Outcomes with Competing Risks

This paper proposes a novel, doubly robust Targeted Maximum Likelihood Estimation (TMLE) framework for estimating subgroup average treatment effects in time-to-event data with competing risks, featuring newly derived variable importance measures to optimize precision medicine strategies.

Original authors: Runjia Li, Victor B. Talisa, Chung-Chou H. Chang

Published 2026-08-13

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Runjia Li, Victor B. Talisa, Chung-Chou H. Chang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor trying to decide which medicine to give a patient. In the old days, doctors often treated everyone the same, like handing out the same size shoe to every foot. But people are different; a shoe that fits a giant might crush a child's foot. This is the heart of "precision medicine": finding the right treatment for the right person.

However, life in the hospital is messy. Patients don't just face one danger; they face a race against time where multiple bad things can happen at once. If a patient is in the Intensive Care Unit (ICU) with a severe infection, they might die from the infection, or they might get so sick they have to be discharged, or they might recover. These are called "competing risks" because they compete to be the event you are watching. It's like watching a race where runners can trip, drop out, or win, and you want to know exactly how a specific training method (like a new drug) changes the odds of winning versus tripping.

The big challenge is figuring out who benefits from the treatment. Does a drug help young people but hurt older ones? Does it help people who had surgery but not those who didn't? To answer this, scientists need a super-accurate way to calculate the "Average Treatment Effect" (the average benefit) for specific groups of people, even when the data is messy and some patients leave the study early. If the math is wrong, doctors might give the wrong medicine to the wrong person.

The New "Double-Check" Tool for Life-or-Death Decisions

In this paper, researchers Runjia Li, Victor B. Talisa, and Chung-Chou H. Chang have built a brand-new mathematical tool to solve this exact problem. They call it a "Doubly Robust Targeted Estimation" framework. Think of it as a high-tech, self-correcting GPS for medical decisions.

Usually, when scientists try to predict how a drug works, they have to build a model of the patient's health (the "outcome model") and a model of who gets the drug (the "treatment model"). The problem is, if you get even one of these models wrong, your prediction can be totally off. It's like trying to navigate a city with a map that has the wrong streets; you'll get lost.

This new tool is special because it is "doubly robust." Imagine you are trying to guess the winner of a race. You have two different ways to make your guess: you can look at the runners' past speeds, or you can look at the weather conditions. This new method says, "I don't care if your weather guess is wrong, as long as your speed guess is right. Or, if your speed guess is wrong, as long as your weather guess is right, I will still get the correct answer." Only if both guesses are wrong does the method fail. This makes it incredibly reliable, even when scientists aren't 100% sure about every detail of the patient's history.

The researchers tested this tool using computer simulations, creating thousands of fake ICU patients with severe infections (sepsis) to see how the tool performed. They found that when they messed up the models on purpose (to simulate real-world mistakes), their new tool still gave the right answer, while older methods got confused and gave wrong results. It was like a navigator that kept you on the right path even when the map was torn.

The "Subgroup" Detective Work

The paper doesn't just look at the average effect for everyone; it digs deeper to find out how the treatment works for specific "subgroups." They used a method called Targeted Maximum Likelihood Estimation (TMLE), which is a fancy way of saying they tweaked their estimates over and over again until they were perfectly tuned to the specific question they were asking.

To make this work, they had to invent a new way to measure "Variable Importance." Imagine you are trying to figure out why some cars get better gas mileage than others. You might guess it's the engine size, the weight, or the driver's style. The researchers created two different "scorecards" to figure this out:

The Predictive Scorecard: This measures which patient characteristics (like age or surgery history) cause the biggest differences in how well the drug works. It answers: "Who is the drug good for, and who is it bad for?"
The Prognostic Scorecard: This measures which characteristics are most important for predicting the patient's outcome, regardless of the drug. It answers: "What details do we absolutely need to know to get an accurate calculation?"

Testing it on Real Sepsis Data

To see if their tool worked in the real world, the team applied it to data from 3,245 ICU patients with sepsis. They wanted to know if giving steroids (a type of anti-inflammatory drug) helped or hurt patients. They looked at different groups based on age, whether they had surgery, and their level of consciousness (measured by a score called the Glasgow Coma Scale, or GCS).

The results were eye-opening. The tool found that for patients who had not had surgery and had a relatively high level of consciousness (GCS of 7 or higher), steroids actually seemed to increase the risk of death within 7 days. Specifically, for patients 65 or older in this group, the risk went up by about 0.118 (meaning an 11.8% higher chance of death compared to no treatment). For younger patients in the same group, the risk went up even more, by about 0.173.

However, for patients who were older (65+) and had a low level of consciousness (GCS below 7), the story changed. In these patients, the tool suggested that steroids might actually be helpful, lowering the risk of death, though the researchers noted there was still some uncertainty in this specific finding.

The study also used their new "scorecards" to figure out which variables mattered most. They discovered that the patient's level of consciousness (GCS) was the single most important factor in deciding who would benefit from steroids, explaining more than 60% of the differences in treatment effects. Age was the next most important factor. Interestingly, for patients with low consciousness, the level of "serum lactate" (a chemical in the blood) was crucial for getting an accurate estimate of the treatment's effect.

Why This Matters

This paper doesn't claim to have solved the mystery of sepsis forever. Instead, it provides a powerful, flexible, and safer way to ask the question: "Does this treatment work for this specific type of patient?" By using a method that can handle messy data and wrong guesses, and by giving doctors a clear way to see which patient details matter most, this tool helps move medicine closer to the goal of truly personalized care. It suggests that one size does not fit all, and with the right math, we can finally start matching the right treatment to the right person.

Technical Summary: Doubly Robust Targeted Estimation of Subgroup Average Treatment Effects for Time-to-event Outcomes with Competing Risks

Problem Statement
Precision medicine increasingly demands methods capable of addressing patient heterogeneity, particularly in critical care settings like the ICU where outcomes are time-to-event with competing risks (e.g., death vs. discharge). While the Fine and Gray (1999) subdistribution hazards model is widely used for competing risks, it suffers from non-collapsibility issues when inferring causality. Furthermore, existing Targeted Maximum Likelihood Estimation (TMLE) approaches for survival data often rely on cause-specific hazards, where covariate effects do not directly translate to the Cumulative Incidence Function (CIF), or they focus on average treatment effects (ATE) rather than subgroup-specific effects. There is a need for a robust framework that estimates subgroup ATEs defined by differences in potential CIFs, accommodates flexible machine learning, and provides valid inference under model misspecification.

Methodology
The authors propose a novel TMLE framework designed to estimate subgroup ATEs in the presence of competing risks. The methodology proceeds through the following steps:

Causal Estimand Definition: The target parameter is the subgroup ATE, defined as the difference in potential CIFs for the primary event at a specific time $t_0$ between treatment arms ( $A=1$ vs. $A=0$ ) within a prespecified subgroup defined by predictive variables $V$ . The framework assumes standard causal identification conditions: consistency, no unmeasured confounding, positivity, and coarsening at random censoring.
Initial Estimation: The procedure begins with initial estimates of nuisance parameters: the subdistribution hazard function for the main event ( $\lambda_1$ ), the censoring survival function ( $G$ ), and the treatment mechanism ( $\pi$ ). These can be estimated using various methods, including penalized subdistribution hazards models (e.g., fastcmprsk), Cox models, logistic regression, or flexible Super Learner ensembles.
Targeting Step (Iterative Updates): To achieve asymptotic efficiency and double robustness, the initial estimates are updated via a fluctuation model. The authors derive the Efficient Influence Function (EIF) for the target parameter. A "clever covariate" is constructed based on the inverse probability of censoring weighting (IPCW) and the treatment mechanism. The subdistribution hazard is updated iteratively using a log-linear fluctuation model ( $\lambda_{1,\epsilon} = \lambda_1 \exp\{\epsilon h\}$ ) until the empirical mean of the EIF score function converges to zero.
Inference and Variable Importance:
- Inference: The asymptotic normality of the estimator is established, allowing for the construction of Wald-type confidence intervals based on the empirical variance of the EIF.
- Variable Importance Measures (VIMs): Two distinct VIMs are developed:
  - Predictive Importance ( $VIM_1$ ): Quantifies a variable's contribution to the variance of individual-level treatment effects (heterogeneity).
  - Prognostic Importance ( $VIM_2$ ): Quantifies the absolute change in the estimated subgroup ATE when a prognostic variable is omitted, assessing its impact on estimation accuracy.

Key Contributions

Derivation of the EIF: The primary theoretical contribution is the derivation of the efficient influence function for the subgroup ATE defined by differences in potential CIFs under competing risks.
Doubly Robust Subgroup Estimation: The proposed estimator is doubly robust; it yields consistent estimates if either the outcome model (CIF) is correctly specified or both the treatment and censoring models are correctly specified.
Substitution Estimator Properties: As a substitution estimator, the method guarantees that the estimated ATE remains within the natural bounds of probability differences $(-1, 1)$ , a critical feature for interpretable clinical metrics.
Flexible Framework: The approach accommodates diverse regression and machine learning algorithms for nuisance parameter estimation, making it adaptable to high-dimensional data and complex relationships.
Dual Variable Importance: The introduction of complementary VIMs allows researchers to distinguish between variables driving treatment effect heterogeneity and those critical for accurate prognostic estimation.

Results

Simulation Studies: Extensive simulations across various scenarios (correctly specified models, misspecified outcome models, misspecified treatment/censoring models, and varying sample sizes) demonstrate that the proposed TMLE estimator maintains low bias and Root Mean Squared Error (RMSE) even under model misspecification, confirming its double robustness. In contrast, S-learner and T-learner meta-algorithms exhibited significant bias and increased RMSE when models were misspecified. The TMLE method consistently achieved approximately 95% coverage for confidence intervals across all scenarios.
Sepsis Data Application: The method was applied to ICU sepsis data to evaluate the effect of steroids on 10-day mortality (with discharge as a competing event). The analysis identified treatment effect heterogeneity based on age, surgical history, and Glasgow Coma Scale (GCS).
- Steroids were associated with a significantly increased risk of mortality (positive ATE) in patients who did not undergo surgery with a GCS $\ge$ 7, and in surgical patients with GCS < 7.
- Conversely, for patients aged $\ge$ 65 with GCS < 7, the ATE was negative (suggesting a protective effect), though the confidence intervals included zero.
- Variable importance analysis revealed GCS as the dominant predictor of treatment effect heterogeneity, while serum lactate and SOFA scores were identified as critical prognostic variables for specific subgroups.

Significance
The paper claims that this work provides a powerful, theoretically grounded tool for optimizing personalized treatment strategies in the era of precision medicine. By directly targeting the CIF and ensuring double robustness, the framework offers a reliable alternative to existing methods that may yield biased results under model misspecification. The development of specific VIMs further aids clinicians and researchers in prioritizing variables for data collection and intervention strategies. The authors position this framework as a step toward more accurate, subgroup-specific treatment decisions in complex time-to-event settings with competing risks.

The New "Double-Check" Tool for Life-or-Death Decisions

The "Subgroup" Detective Work

Testing it on Real Sepsis Data

Why This Matters

More like this