📊 statistics

Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE

This paper proposes and validates two Targeted Maximum Likelihood Estimators (TMLE-PR and A-TMLE) that leverage data from non-subgroup participants within the same randomized controlled trial to significantly improve the precision of subgroup-specific treatment effect estimates without introducing external bias, as demonstrated by successful application to the LEADER cardiovascular trial.

Original authors: Sky Qiu, Nerissa Nance, Rachael Phillips, Jens Tarp, Maya Petersen, Mark van der Laan

Published 2026-05-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sky Qiu, Nerissa Nance, Rachael Phillips, Jens Tarp, Maya Petersen, Mark van der Laan

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 detective trying to solve a mystery: Does a new medicine work for a specific, small group of people?

In the world of medical trials (Randomized Controlled Trials), researchers test drugs on thousands of people. Usually, they have enough data to say, "This drug works for the average person." But sometimes, they want to know if it works specifically for a small subgroup, like people of a certain race or age.

The Problem: The "Small Group" Dilemma
The problem is that these small groups are often too tiny to draw a confident conclusion. It's like trying to guess the average height of a specific type of rare tree by measuring only three saplings. The data is too "noisy," and the results might just be a fluke.

Traditionally, to get more data, scientists might look outside the trial for real-world records (like hospital databases). But this is risky. Real-world data is messy, unorganized, and might have hidden biases that ruin the trial's strict scientific rules.

The Solution: "Borrowing" from the Neighbors
This paper proposes a clever new way to solve the problem without leaving the safety of the original trial. Instead of looking outside, the researchers suggest borrowing information from the "neighbors"—the thousands of other participants in the same trial who don't belong to the small group.

Think of it like this: You are trying to estimate the average temperature in a tiny, cold corner of a large, heated house. You don't have enough thermometers in that corner. But you do have hundreds of thermometers in the rest of the house. The new methods allow you to use the data from the rest of the house to make a smarter guess about the corner, while carefully correcting for the fact that the corner is naturally colder.

The Two New Tools
The authors introduce two mathematical "tools" (estimators) to do this borrowing safely:

TMLE-PR (The "Big Picture" Learner):
Imagine you are trying to predict how a specific type of plant grows. You have very few examples of this plant, but you have thousands of examples of other plants growing in the same garden.
- How it works: This tool first learns the general rules of gardening by studying all the plants in the garden (the whole trial). Then, it applies those general rules to predict how the specific plant should grow. Finally, it makes a tiny adjustment based only on the actual few plants it has.
- The Benefit: It uses the massive amount of data from the whole trial to build a strong foundation, making the guess for the small group much more precise.
A-TMLE (The "Smart Corrector"):
This tool is more sophisticated. It realizes that just using the "Big Picture" might be wrong if the small group is fundamentally different from the rest.
- How it works: It starts with the "Big Picture" guess (like the first tool). Then, it calculates a "Bias Correction." It asks: "How much does the small group actually differ from the rest of the house?" It measures this difference and subtracts it out.
- The Analogy: Imagine you are estimating the speed of a race car on a specific track. You look at the average speed of all cars on all tracks (the Big Picture). But you know this specific track has a sharp turn. A-TMLE calculates exactly how much that sharp turn slows the car down and adjusts the estimate accordingly. It gets the best of both worlds: the power of the big data and the accuracy of the specific correction.

The Test Drive: The LEADER Trial
To prove these tools work, the authors tested them on real data from a famous heart disease trial called LEADER.

The Scenario: The trial had 9,340 people, but only about 900 were Asian and 700 were Black. The original trial wasn't big enough to say for sure if the drug worked specifically for these groups.
The Result: Using their new "borrowing" tools, the researchers found clear evidence that the drug reduced heart risks for both Asian and Black participants.
- For the Asian group, the drug lowered the risk of major heart events by about 1.5 percentage points.
- For the Black group, it lowered the risk by about 2.0 percentage points.
- Crucially, the standard methods (looking only at the small groups) were too shaky to find these results, but the new methods found them clearly.

Why This Matters
The paper argues that we don't need to risk the integrity of a trial by mixing in messy outside data. By treating the "non-subgroup" participants as helpful neighbors rather than strangers, we can make our conclusions about small groups much sharper and more reliable.

In Summary:

Old Way: "We have too few people in this group to know if the drug works."
New Way: "Let's use the data from everyone else in the trial to help us figure it out, but we'll do the math carefully to make sure we aren't fooled by the differences between the groups."
Outcome: We get stronger, more precise answers for the people who need them most, without breaking the rules of the experiment.

Technical Summary: Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE

Problem Statement
Subgroup analyses within Randomized Controlled Trials (RCTs) are frequently underpowered due to limited sample sizes within specific subpopulations. While Real-World Data (RWD) offers a potential avenue to augment trial subpopulations, integrating RWD with RCT data introduces substantial statistical challenges, including distributional shifts, structural differences in data generation, and confounding mechanisms that can bias treatment effect estimates. Consequently, there is a critical need for methods that improve the precision of subgroup-specific treatment effect estimates using only the data collected within the same trial, thereby preserving the internal validity and protection against bias afforded by randomization without relying on external sources.

Methodology
The authors propose two Targeted Maximum Likelihood Estimators (TMLEs) designed to borrow information from participants outside the subgroup of interest (the "non-subgroup" population) while targeting a prespecified subgroup causal estimand. Both estimators operate within the semi-parametric TMLE framework, allowing for the integration of flexible machine learning algorithms (e.g., Super Learner, Highly Adaptive Lasso) to avoid restrictive parametric assumptions.

TMLE with Pooled Regression (TMLE-PR):
- Mechanism: This estimator fits an initial outcome regression ( $Q_0$ ) on the entire pooled trial sample (including both subgroup and non-subgroup participants), treating subgroup membership ( $S$ ) as a predictor. This initial regression is then evaluated specifically on the subgroup sample ( $S=1$ ).
- Targeting: A standard TMLE targeting step is applied only to the subgroup observations to update the initial estimate and solve the efficient influence curve equation.
- Rationale: By fitting the initial regression on the full sample, TMLE-PR leverages the larger sample size to obtain a more precise estimate of the outcome regression function, which is a key component driving the variance of the estimator.
Adaptive TMLE (A-TMLE) for Subgroup Analysis:
- Mechanism: Inspired by data integration methods used for RCT-RWD hybrid studies, this approach decomposes the target subgroup estimand ( $\Psi$ $Ψ$ ) into two components:
  - A pooled estimand ( $\tilde{\Psi}$ ): The treatment effect estimated over the pooled population but averaged over the subgroup's covariate distribution. This component allows for maximum efficiency by utilizing the full trial data.
  - A bias estimand ( $\Psi^\#$ ): The difference between the pooled estimand and the true subgroup-specific target. This captures the bias induced by borrowing information from the non-subgroup population.
- Estimation: The method estimates projection parameters for both components using working models (e.g., via Highly Adaptive Lasso). Crucially, it employs a "bias correction" step where the estimated bias is subtracted from the pooled estimate.
- Regularization: The A-TMLE framework regularizes the targeting step, which is particularly beneficial in finite samples and when covariate overlap is limited. The authors propose specific strategies for estimating nuisance functions and selecting working models, including the use of Group Lasso to screen HAL subspaces and ensure the treatment variable is included in the working model to prevent degenerate bias estimates.

Key Contributions

Novel Estimators: The introduction of TMLE-PR and a tailored A-TMLE specifically for subgroup analysis within a single RCT, avoiding the need for external RWD.
Theoretical Framework: Derivation of the canonical gradients and exact remainder terms for both estimators, demonstrating that under randomization, the remainder terms are favorable (often vanishing or of higher order) when the treatment mechanism is known.
Bias-Variance Trade-off Management: The A-TMLE explicitly models the bias introduced by information borrowing. This allows the method to achieve efficiency gains even when the subgroup-specific treatment effect differs from the pooled effect, provided the bias structure can be approximated by the working model.
Practical Implementation: Detailed algorithmic steps (pseudo-code) and recommendations for nuisance function estimation (Super Learner, HAL) and model selection (Group Lasso screening) to ensure computational feasibility and statistical robustness.

Results

Simulation Studies:
- In scenarios where subgroup-specific effects differed substantially from pooled effects (stress-testing the bias), both TMLE-PR and A-TMLE demonstrated improved efficiency (lower Mean Squared Error and Standard Error) compared to unadjusted and covariate-adjusted subgroup-only estimators (AIPW, standard TMLE).
- TMLE-PR achieved the lowest MSE but exhibited conservative confidence interval coverage.
- A-TMLE maintained approximately nominal coverage while offering significant efficiency gains, particularly in settings where baseline covariates were strongly predictive of subgroup membership. In these cases, A-TMLE reduced MSE by more than half compared to unadjusted estimators.
- The simulations confirmed that A-TMLE performs a non-trivial bias correction, distinguishing it from naive pooling approaches.
Case Study (LEADER Trial):
- Applied to the LEADER cardiovascular outcome trial (N=9,340) to estimate the risk reduction of Major Adverse Cardiac Events (MACE) for Asian and Black subgroups (each <10% of the population).
- Findings: A-TMLE provided stable, precise estimates of absolute risk reduction (e.g., ~1.5–1.6 percentage points for Asians; ~2.0–2.1 percentage points for Blacks at 365, 540, and 730 days) with 95% confidence intervals excluding the null.
- Comparison: In contrast, the subgroup-only AIPW estimator yielded similar point estimates for Black participants but with much larger variance, and estimates close to zero with wide intervals for Asian participants, failing to exclude the null.

Significance and Claims
The paper claims that these methods represent a practical approach to enhancing the precision of subgroup-specific evidence generation using only internal trial data. By treating non-subgroup participants as an auxiliary source and explicitly modeling the resulting bias, the proposed estimators advance regulatory priorities (e.g., FDA) for equitable labeling and post-market evaluation. The authors emphasize that their approach improves efficiency without introducing external sources of bias, thereby facilitating rigorous inference in subgroups that were originally underpowered. The work suggests that data integration concepts can be fruitfully repurposed within a single RCT to address the common challenge of small subgroup sample sizes.

Technical Summary: Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE

More like this