📊 statistics

Joint Estimation of Marginal and Heterogeneous Treatment Effects

This paper proposes a joint modeling framework that extends nonparanormal adjusted marginal inference to simultaneously estimate marginal treatment effects and rank prognostic and predictive covariates, thereby preserving marginal interpretability while improving statistical efficiency across various outcome types.

Original authors: Leticia Wuethrich, Torsten Hothorn

Published 2026-05-25

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Leticia Wuethrich, Torsten Hothorn

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 figure out if a new medicine works better than a sugar pill. In a perfect world, you'd give the medicine to one group of people and the pill to another, and the only difference between the groups would be the treatment. But in real life, people are different. Some are older, some are sicker, some have different lifestyles. These differences are called covariates.

This paper introduces a new mathematical "kitchen" (a statistical framework) that helps researchers cook up a clearer picture of whether a treatment works, while also understanding who it works best for, without messing up the main recipe.

Here is the breakdown of their method using simple analogies:

1. The Problem: The "Non-Collapsible" Trap

Usually, when researchers want to be more precise, they adjust for these differences (like age or weight) in their math.

The Linear World (The Easy Way): If you are measuring something simple, like height, adding "age" to your math is like adding a pinch of salt. It makes the measurement sharper, but the main result (the average height difference) stays the same.
The Non-Linear World (The Tricky Way): Many medical outcomes (like survival time or pain scores) are like a rubber sheet. If you stretch the sheet (add covariates), the shape changes. In standard math, adjusting for these factors often accidentally changes the definition of the result you are looking for. You might end up measuring "how the drug works for a specific type of person" (conditional) instead of "how the drug works for the whole population" (marginal). This is called non-collapsibility. It's like trying to measure the average temperature of a room, but every time you open a window to let in fresh air (adjust for a variable), the definition of "room temperature" changes.

2. The Solution: The "Joint Model" (The Big Tent)

The authors, Wüthrich and Hothorn, built a new framework called NAMI-HTE. Think of this as a Giant Transparent Tent that covers two things at once:

The outcome (did the patient get better?).
The patient's starting characteristics (age, weight, baseline health).

Instead of looking at the patient's starting stats and then separately looking at the outcome, this method looks at them as a single, connected unit. It uses a clever mathematical trick called a Gaussian Copula.

The Analogy: Imagine you have a bunch of different shaped balloons (different types of data: continuous numbers, yes/no answers, time until an event). Usually, you can't easily tie them together. This method inflates all of them into a standard, round shape (a "latent normal scale") so they can be tied together easily. Once they are tied, the researchers can look at the whole bundle.

3. The Magic Trick: Keeping the "Marginal" View

The coolest part of this tent is that even though they are looking at the complex bundle of data inside, they can still step outside and see the Marginal Treatment Effect.

The Metaphor: Imagine you are looking at a 3D sculpture through a foggy window. Usually, if you try to clean the glass (adjust for covariates), the fog moves and distorts the view of the sculpture's overall shape. This new method cleans the glass without moving the fog or changing the shape of the sculpture. You get a clearer view of the whole group's average improvement, even while you are accounting for individual differences.

4. Sorting the Ingredients: Prognostic vs. Predictive

The paper also helps sort out two types of "ingredients" (covariates) that affect the result:

Prognostic Covariates: These are like weather forecasts. If a patient starts with a high fever, they are likely to stay sick longer, regardless of which medicine they take. The medicine doesn't change this; the patient's starting state just predicts the outcome.
Predictive Covariates: These are like special keys. They tell you if a specific medicine works better for a specific person. For example, a medicine might only work for people with a specific gene.

The authors' method puts these two types of ingredients on the same measuring stick. It allows researchers to say, "This factor is a strong weather forecast (prognostic)," and "This factor is a strong key (predictive)," and compare them directly.

5. What They Found (The Results)

Efficiency: By using this "Giant Tent" method, they found that the estimates of whether the drug works are sharper and more precise (smaller error bars) than just ignoring the patient's background.
The Source of Improvement: Most of this improvement comes from prognostic factors (the "weather forecasts"). Knowing a patient's starting health helps predict the outcome better, which makes the drug's effect look clearer.
Predictive Factors: Finding "keys" (predictive factors) is much harder. The method can do it, but it requires a lot of data. In their tests, even with a strong "key," it was hard to prove it existed unless the sample size was huge.
Real World Test: They tested this on a real study about acupuncture for headaches.
- It confirmed the original study's finding: Acupuncture helps.
- It showed that the "baseline headache score" was the biggest predictor of how a patient would do (prognostic).
- It hinted that people with worse headaches at the start might get slightly more benefit from acupuncture (predictive), but the evidence for this was weak and depended on how they modeled the data.

Summary

This paper offers a new statistical "kitchen" that lets researchers:

Adjust for patient differences to get a clearer, more precise answer about whether a treatment works.
Keep the answer simple (it still applies to the whole population, not just a subgroup).
Simultaneously identify which patient traits predict the outcome generally (prognostic) and which traits determine if the treatment works specifically for them (predictive).

It's like having a map that shows you the best route for the whole fleet of ships (the marginal effect) while also telling you which specific ships need extra fuel based on their cargo (the heterogeneous effects), all without getting lost in the fog.

Technical Summary: Joint Estimation of Marginal and Heterogeneous Treatment Effects

Problem Statement

Randomized clinical trials (RCTs) typically aim to estimate the marginal treatment effect (the average effect in the trial population). While covariate adjustment is standard for improving precision, it introduces a methodological challenge in nonlinear models due to noncollapsibility. In such models (e.g., logistic regression, Cox proportional hazards, or models using Cohen's $d$ ), adjusting for covariates changes the estimand from a marginal effect to a conditional effect, even in the absence of confounding.

Conversely, identifying prognostic covariates (associated with the outcome regardless of treatment) and predictive covariates (which modify the treatment effect) is crucial for understanding heterogeneity and guiding personalized medicine. Existing methods generally fall into two camps:

Marginal-focused methods (e.g., semiparametric augmentation, standardization, IPTW, TMLE) preserve marginal interpretability but do not explicitly parameterize or rank prognostic and predictive effects.
Heterogeneity-focused methods (e.g., causal forests, model-based forests) estimate conditional average treatment effects (CATEs) but often lack formal type I error control, struggle to distinguish prognostic from predictive effects in small samples, and do not inherently preserve a marginal estimand.

The paper addresses the need for a unified framework that estimates a marginal treatment effect while simultaneously adjusting for covariates to improve efficiency and explicitly quantifying and ranking prognostic and predictive effects.

Methodology: Nonparanormal Adjusted Marginal Inference with Heterogeneous Treatment Effects (NAMI-HTE)

The authors propose extending the Nonparanormal Adjusted Marginal Inference (NAMI) framework (Dandl and Hothorn, 2026) to accommodate heterogeneous treatment effects. The core innovation is embedding the marginal treatment effect directly into a joint model for the outcome and baseline covariates using a Gaussian copula (nonparanormal) construction.

1. Theoretical Framework

Transformation Models: The method utilizes univariate transformation models where the outcome distribution is mapped to a latent scale via a monotone transformation function $h(\cdot)$ . The treatment effect $\tau$ is defined as a shift on this latent scale.
Joint Modeling: The framework models the joint distribution of the outcome $Y$ and covariates $X$ on a latent standard normal scale. The joint cumulative distribution function (CDF) is defined as:
$P(X \le x, Y \le y \mid W = w) = \Phi_{\Sigma(w)}(h(x, y \mid W = w))$
where $\Phi_{\Sigma(w)}$ is a multivariate normal CDF with a treatment-dependent correlation matrix $\Sigma(w)$ .
Preserving Marginality: By constraining the diagonal of the correlation matrix to unity ( $\text{diag}(\Sigma(w)) \equiv 1$ ), the model ensures that the marginal distribution of the outcome remains unchanged by the inclusion of covariates. This guarantees that the treatment effect $\tau$ retains its marginal interpretation (collapsibility) even after adjustment.
Parameterization of Effects:
- Prognostic Effects: Captured by the correlation between covariates and the outcome in the control arm (elements of the last row of the inverse Cholesky factor $\Lambda(0)$ ).
- Predictive Effects: Captured by the difference in correlations between the treatment and control arms (elements of the treatment-dependent component $\Gamma(\gamma)$ ).
- Standardization: Because covariates are mapped to a standard normal scale, prognostic and predictive strengths are quantified on a common scale, allowing for direct comparison and ranking.

2. Applicability

The framework is flexible and applies to:

Continuous outcomes: With $\tau$ interpreted as Cohen's $d$ (standardized mean difference).
Binary and Ordinal outcomes: With $\tau$ interpreted as a log-odds ratio.
Time-to-event outcomes: With $\tau$ interpreted as a log-hazard ratio.
Censoring: It accommodates right-, left-, and interval-censoring, as well as missing data, within a single likelihood-based estimation procedure.

3. Estimation

Parameters are estimated via maximum likelihood. The method allows for formal hypothesis testing of prognostic and predictive effects with multiplicity adjustment, providing a confirmatory approach to heterogeneity analysis.

Key Results

Theoretical Findings (Continuous Outcomes)

For continuous outcomes measured by Cohen's $d$ , the authors derive closed-form expressions for the asymptotic variance of the marginal treatment effect under covariate adjustment.

Efficiency: The asymptotic variance of the adjusted estimator is never worse and often better than the unadjusted estimator.
Drivers of Efficiency: Efficiency gains are primarily driven by prognostic effects. Realistic predictive effects contribute only modest additional improvement.
Worst-Case Scenario: The only scenario where no efficiency gain is achieved is when prognostic and predictive effects have equal magnitude but opposite signs in the two treatment arms.

Simulation Studies

Simulations were conducted across continuous, binary, and survival outcomes with varying strengths of prognostic and predictive effects.

Bias: The marginal treatment effect $\hat{\tau}$ was estimated without bias under both unadjusted and adjusted models, even in the presence of predictive effects.
Precision: The adjusted model (NAMI-HTE) consistently yielded smaller standard errors and higher statistical power compared to unadjusted marginal inference.
Predictive Power: While the method correctly estimates predictive parameters, the power to detect predictive effects was found to be limited (e.g., <10% for moderate effects, <40% for strong effects) in realistic sample sizes, consistent with the general difficulty of detecting interactions in RCTs.
Type I Error: Empirical sizes were generally close to the nominal level, though slight deviations were observed in small samples for continuous and survival outcomes, attributed to Wald statistic approximations.

Application: Acupuncture Trial

The method was applied to a randomized trial of acupuncture for chronic headaches ( $N=401$ ).

Reproduction of Results: The method reproduced the original trial findings (a significant benefit of acupuncture over usual care) while improving precision through covariate adjustment.
Prognostic Ranking: Baseline headache score was consistently identified as the strongest prognostic factor across all models.
Predictive Ranking: Baseline headache score was also the top candidate for a predictive effect, suggesting patients with more severe baseline pain may benefit more. However, statistical evidence for this interaction was weak and sensitive to distributional assumptions, highlighting the difficulty of confirming heterogeneity in this dataset.
Robustness: Results remained consistent across models assuming normality, flexible transformations, and ordinal outcomes.

Significance and Claims

The paper claims to bridge the gap between marginal inference and heterogeneity analysis by offering a unified framework that:

Preserves Marginal Interpretability: Unlike standard nonlinear regression adjustments, the proposed method ensures the treatment effect remains a marginal estimand by construction.
Explicit Quantification: It explicitly separates and quantifies prognostic and predictive effects on a standardized scale, enabling direct ranking of covariates.
Confirmatory Inference: It provides a framework for formal hypothesis testing of treatment-covariate interactions with controlled type I error, contrasting with exploratory machine learning approaches.
Generality: It unifies the analysis of continuous, binary, ordinal, and time-to-event outcomes, along with various censoring mechanisms, within a single modeling structure.

The authors modestly acknowledge limitations, specifically the low power to detect predictive effects in typical trial sample sizes and the reliance on a latent Gaussian correlation structure (implying monotonicity on original scales). They suggest future work should investigate robustness under model misspecification.