Optimal estimation of generalized causal effects in cluster-randomized trials with multiple outcomes

Technical Summary: Optimal Estimation of Generalized Causal Effects in Cluster-Randomized Trials with Multiple Outcomes

Problem Statement
Cluster-randomized trials (CRTs) are increasingly utilized to evaluate group-level interventions, often collecting multiple outcomes to capture complementary dimensions of benefit and risk. While investigators frequently seek a single global summary of treatment effects, existing literature largely focuses on single-outcome estimands or relies on model-based procedures that may suffer from unclear causal interpretations or limited robustness. Furthermore, current methods for multiple outcomes in CRTs often fail to explicitly account for heterogeneous outcome types, potentially informative cluster sizes, and the integration of baseline covariates without compromising target estimands. There is a lack of well-defined potential outcome estimands for generalized causal effects that accommodate these complexities while offering optimal efficiency and robustness.

Methodology
The authors develop a unified potential outcomes framework for generalized treatment effects in CRTs with multiple outcomes, addressing both non-prioritized and prioritized settings.

1. Causal Estimands:
   The paper defines two complementary estimands based on flexible, possibly nonlinear pairwise contrast functions, $w(\cdot, \cdot)$, which map multivariate comparisons to a single numerical value:

   • Cluster-Pair Generalized Causal Effect (cp-GCE): $\lambda_{C,a}$, which gives equal weight to each pair of clusters regardless of their sizes.
   • Individual-Pair Generalized Causal Effect (ip-GCE): $\lambda_{I,a}$, which assigns equal weight to each pair of individuals, thereby accounting for informative cluster sizes.
     These estimands generalize conventional average treatment effects and probabilistic indices. They are constructed via cross-cluster comparisons to avoid the fundamental problem of causal inference regarding within-cluster contrasts under nonlinear functions.

2. Nonparametric Estimation:
   Initial estimators are derived as weighted clustered two-sample U-statistics. These estimators are consistent and asymptotically normal, with variance estimators proven to be consistent under cross-fitting.

3. Efficient Influence Functions (EIFs) and Model-Robust Estimation:
   The authors derive the EIFs for both cp-GCE and ip-GCE. These EIFs motivate model-robust estimators that incorporate baseline covariates through nuisance functions (outcome means). By leveraging the Neyman orthogonality of the EIFs, these estimators remain consistent and asymptotically normal even if the working nuisance model is misspecified. If the nuisance model is correctly specified, these estimators attain the semiparametric efficiency lower bound.

4. Debiased Machine Learning (DML):
   To handle complex nuisance functions, the paper proposes DML estimators using cross-fitting. The nuisance functions are estimated using flexible, data-adaptive machine learning methods (e.g., SuperLearner). The authors establish that these DML estimators achieve semiparametric efficiency under mild regularity conditions (specifically requiring the nuisance estimator to converge at an $m^{1/4}$ rate). Crucially, the paper provides an analytical variance estimator for these DML U-statistics under a cluster-splitting scheme, proving its consistency without relying on the stronger Donsker conditions often required for bootstrap methods.

5. Computational Efficiency:
   To address the computational burden of U-statistics (which scale with $m^2$), the authors propose an estimation procedure based on incomplete U-statistics using subsamples. They demonstrate that these subsample-based estimators retain the same asymptotic properties (consistency and normality) as full-sample estimators, with negligible asymptotic efficiency loss.

Key Results

• Theoretical Properties: The proposed estimators (nonparametric, model-robust, and DML) are shown to be consistent and asymptotically normal. The DML estimators are proven to attain the semiparametric efficiency bounds under mild regularity conditions.
• Variance Estimation: The analytical variance estimators derived for the DML approach are proven consistent under cross-fitting, offering a computationally efficient alternative to nonparametric bootstrapping.
• Simulation Studies: Extensive simulations with $m=30$ and $m=60$ clusters demonstrate that:
  • The proposed estimators exhibit little to no bias.
  • The DML estimator is the most efficient, outperforming nonparametric and existing model-based approaches (e.g., Smith et al., 2025) that ignore informative cluster sizes or covariates.
  • Variance estimators are consistent, though finite-sample adjustments (degrees-of-freedom correction) improve coverage probabilities in smaller samples.
• Data Application: The methods are applied to a CRT evaluating a cognitive behavioral therapy intervention for chronic pain. In both prioritized and non-prioritized outcome settings, the DML estimator provided the most precise estimates (smallest standard errors and narrowest confidence intervals) for win ratios, consistently suggesting the intervention was more effective than usual care. The model-robust estimator sometimes yielded non-significant results, highlighting the potential for model misspecification in finite samples, whereas the DML approach maintained robustness and precision.

Significance and Claims
The paper claims to provide a unified framework for causal inference in CRTs with multiple outcomes that explicitly handles informative cluster sizes and heterogeneous outcome types. Its primary contributions are:

1. Unified Framework: Defining generalized causal effect estimands (cp-GCE and ip-GCE) that unify various summary measures (net benefit, win ratio, win odds) and accommodate both prioritized and non-prioritized outcomes.
2. Optimal Estimation: Developing estimators that integrate debiased machine learning with U-statistics to achieve semiparametric efficiency while remaining robust to model misspecification.
3. Technical Advancement: Providing a rigorous treatment of semiparametric efficiency theory for CRTs, including the derivation of EIFs and the construction of consistent analytical variance estimators for cross-fitted DML U-statistics, overcoming limitations of previous literature that relied on stronger conditions or bootstrapping.
4. Practical Utility: Demonstrating that leveraging baseline covariates through flexible machine learning significantly improves estimation precision in complex, multilevel trial structures, as evidenced by the data application and simulation studies.

The authors position this work as a step toward more robust and efficient causal inference in modern CRTs, moving beyond single-outcome, model-dependent analyses toward flexible, data-adaptive summaries of treatment effects.