Optimal estimation of generalized causal effects in cluster-randomized trials with multiple outcomes
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Optimal estimation of generalized causal effects in cluster-randomized trials with multiple outcomes".
Jane: Here's the gist of the paper. Optimal estimation of generalized causal effects in cluster-randomized trials with multiple outcomes develops a unified potential outcomes framework for generalized treatment effects with multiple outcomes in CRTs, accommodating both non-prioritized and prioritized outcome settings The proposed cluster-pair and individual-pair causal estimands are defined through flexible pairwise contrast functions and explicitly account for potentially informative cluster sizes The resulting estimators are consistent and asymptotically normal, attain the semiparametric efficiency bounds under mild regularity conditions, and have analytically tractable variance estimators that are proven to be consistent under cross-fitting Framework for Generalized Causal Effects The paper introduces a unified potential outcomes framework for generalized causal effects with multiple outcomes in CRTs, explicitly acknowledging the possibility of informative cluster sizes Within this framework, two complementary estimands are introduced: the cluster-pair generalized causal effect and the individual-pair generalized causal effect These estimands are defined through a general, possibly nonlinear contrast function and naturally accommodate multiple outcomes of mixed types The cp-GCE estimand gives equal weight to each cluster-pair regardless of their cluster sizes, and addresses the expected change in outcome associated with treatment for the population of clusters In contrast, the ip-GCE estimand assigns equal weight to each individual-pair and accounts for the expected change in the outcome associated with treatment across the population of individuals, regardless of their cluster membership Estimation Methodology The paper establishes nonparametric identification and estimation results for these generalized causal effect estimands under cluster randomization The resulting estimators take the form of the two-sample Ustatistics for clustered data (Lee and Dehling, 2005), with the weighting structure reflecting the natural variation in cluster sizes For inference, efficient influence functions (EIFs) are derived to motivate model-robust U-statistics based on semiparametric outcome models and debiased machine learning (DML) U-statistics The DML approach attains the semiparametric efficiency lower bounds under mild regularity conditions Model Robustness and Efficiency The EIFs in Theorem 2 suggest approaches to leverage covariates to develop potentially semiparametrically efficient estimators, which are robust to misspecification of the working nuisance model The model-robust estimator is consistent and asymptotically normal regardless of whether the working model correctly specifies the true nuisance function Furthermore, cross-fitted analytical variance estimators for DML U-statistics are proposed, which is a marked improvement over previous developments These subsample-based estimators based on incomplete U-statistics are shown to have the same asymptotic properties as their complete sample counterparts Finite Sample Performance and Application Simulation studies demonstrate that the DML estimator is the most efficient, confirming theoretical efficiency results under cluster randomization The analysis of a primary care-based cognitive behavioral therapy intervention for long-term opioid users with chronic pain illustrates the practical utility of the proposed methods The DML estimator improves the efficiency of the nonparametric estimator by significant percentages across all three outcome cases, demonstrating superior precision The work develops a unified potential outcomes framework for causal inference with multiple outcomes in CRTs, anchoring inference on explicitly defined generalized causal effect estimands rather than model-dependent or outcome-dependent summaries The distinction between cluster-pair and individual-pair estimands clarifies how scientific questions, units of inference, and potentially informative cluster sizes jointly shape causal interpretation in CRTs Future research should extend this framework to covariate-adaptive cluster randomization and cluster rerandomization The paper concludes by showing that the interdisciplinary care model is more effective than usual care across all outcome cases considered References Balzer, L.
Tom: B., Zheng, W., van der Laan, M.
Jane: J., and Petersen, M.
Tom: L.
Title and authors: Tom: Let's go back to the start of this paper, "Optimal estimation of generalized causal effects in cluster-randomized trials with multiple outcomes." The authors are Xinyuan Chen and Fan Li.
Jane: They come from different departments, which is interesting because one is in Mathematics and Statistics and the other in Biostatistics. That suggests they’re bridging the gap between pure theory and real-world medical data.
Lu: They are tackling a problem where cluster randomized trials, or CRTs, are becoming so common that we need better tools for analyzing them when you have multiple outcomes to measure simultaneously.
Meng: So, what's the specific problem they’re solving beyond just having multiple results? Is it something deeper about how we calculate the effect?
Tom: It’s not just having multiple results; it’s that existing methods largely focus on single-outcome estimands or rely heavily on model-based procedures where the causal interpretation is often unclear.
Jane: That means they are moving away from those older methods and aiming for something more direct about what the treatment actually does to a group of people.
Lu: They propose a unified potential outcomes framework for generalized treatment effects that can handle both non-prioritized and prioritized outcome settings in these trials.
The paper's summary: Tom: So, what’s the actual substance of this unified framework? It seems to be built around defining two key estimands: the cluster-pair generalized causal effect and the individual-pair generalized causal effect.
Jane: Those are the core concepts. The cluster-pair effect gives equal weight to every pair of clusters regardless of their size, addressing the expected change in outcome for a population of clusters.
Lu: In contrast, the individual-pair generalized causal effect assigns equal weight to every individual pair and accounts for the expected change across all individuals, no matter if they are in a cluster or not.
Meng: So it’s like they’re offering two different ways to look at the effect, one focused on the group structure and one focused on the individual structure.
Tom: Right, and this allows them to use flexible pairwise contrast functions that naturally accommodate multiple outcomes of mixed types within that framework.
Jane: That flexibility is what makes it powerful; it lets you define your comparison in a way that suits your specific scientific question without being locked into one outcome type or another.
The paper's improvements: Tom: Beyond just defining the estimands, the authors highlight several specific improvements to how they actually perform the estimation. They establish nonparametric identification via weighted clustered Ustatistics for this framework.
Lu: This means they are using these standard two-sample Ustatistics for clustered data, but crucially, the weighting structure is specifically designed to reflect how variable those cluster sizes naturally are.
Meng: So it’s not just a default weighting scheme; it’s tailored to the known variation in cluster sizes, which is necessary because unequal cluster sizes are a known issue.
Jane: And then for inference, they derive efficient influence functions that motivate model-robust Ustatistics based on semiparametric outcome models and debiased machine learning Ustatistics.
Tom: Those debiased machine learning statistics are motivated by leveraging covariates to develop potentially semiparametrically efficient estimators, which is robust to misspecification of the working nuisance model.
Lu: The paper shows that these model-robust estimators are consistent and asymptotically normal regardless of whether the working model correctly specifies the true nuisance function.
Conclusion: Tom: So, to wrap up this discussion on "Optimal estimation of generalized causal effects in cluster-randomized trials with multiple outcomes," we’re seeing a paper that gives us a clear way to define our questions.
Jane: The main implication is that by anchoring inference on these specific generalized causal effect estimands rather than outcome-dependent summaries, we get a much more principled approach.
Lu: It clarifies how the distinction between cluster-pair and individual-pair estimands dictates what scientific question or unit of inference we are actually asking about in these trials.
Meng: From an engineering side, it means having a clear definition for the analysis helps us design things that are actually robust when dealing with messy, real-world data.
Lalam: And ultimately, they show this framework can be applied to complex clinical settings like opioid pain studies to provide measurable gains in efficiency for treatment effect estimation.
Xinyuan Chen, *, Fan Li, –
Department of Mathematics and Statistics, Mississippi State University · Department of Biostatistics, Yale School of Public Health
stat.ME, math.ST, stat.TH
Submitted: 2026-01-19
Updated: 2026-11-03
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 80/100
The gist: Optimal estimation of generalized causal effects in cluster-randomized trials with multiple outcomes develops a unified potential outcomes framework for generalized treatment effects with multiple…
Key concepts
- Cluster-Pair Generalized Causal Effect (cp-GCE)
- This estimand weights each pair of clusters equally, regardless of their size. It measures the expected change in outcome associated with treatment across the entire population of clusters, focusing on cluster-level effects.
- Individual-Pair Generalized Causal Effect (ip-GCE)
- This estimand weights each individual pair equally. It accounts for the expected change in outcome associated with treatment across all individuals, irrespective of which cluster they belong to.
- Model-Robust Estimators via EIFs
- Efficient Influence Functions (EIFs) are used to derive estimators that use covariates. These estimators are robust, meaning they remain consistent even if the assumed model for the data's underlying functions is slightly incorrect.
- Debiased Machine Learning (DML) U-statistics
- DML U-statistics provide highly efficient estimates by combining machine learning techniques with statistical methods. They achieve semiparametric efficiency bounds, providing precise estimates for causal effects even when using flexible nuisance models.
Terminology
Summary
Optimal estimation of generalized causal effects in cluster-randomized trials with multiple outcomes develops a unified potential outcomes framework for generalized treatment effects with multiple outcomes in CRTs, accommodating both non-prioritized and prioritized outcome settings The proposed cluster-pair and individual-pair causal estimands are defined through flexible pairwise contrast functions and explicitly account for potentially informative cluster sizes The resulting estimators are consistent and asymptotically normal, attain the semiparametric efficiency bounds under mild regularity conditions, and have analytically tractable variance estimators that are proven to be consistent under cross-fitting
Framework for Generalized Causal Effects
The paper introduces a unified potential outcomes framework for generalized causal effects with multiple outcomes in CRTs, explicitly acknowledging the possibility of informative cluster sizes Within this framework, two complementary estimands are introduced: the cluster-pair generalized causal effect and the individual-pair generalized causal effect These estimands are defined through a general, possibly nonlinear contrast function and naturally accommodate multiple outcomes of mixed types The cp-GCE estimand gives equal weight to each cluster-pair regardless of their cluster sizes, and addresses the expected change in outcome associated with treatment for the population of clusters In contrast, the ip-GCE estimand assigns equal weight to each individual-pair and accounts for the expected change in the outcome associated with treatment across the population of individuals, regardless of their cluster membership
Estimation Methodology
The paper establishes nonparametric identification and estimation results for these generalized causal effect estimands under cluster randomization The resulting estimators take the form of the two-sample Ustatistics for clustered data (Lee and Dehling, 2005), with the weighting structure reflecting the natural variation in cluster sizes For inference, efficient influence functions (EIFs) are derived to motivate model-robust U-statistics based on semiparametric outcome models and debiased machine learning (DML) U-statistics The DML approach attains the semiparametric efficiency lower bounds under mild regularity conditions
Model Robustness and Efficiency
The EIFs in Theorem 2 suggest approaches to leverage covariates to develop potentially semiparametrically efficient estimators, which are robust to misspecification of the working nuisance model The model-robust estimator is consistent and asymptotically normal regardless of whether the working model correctly specifies the true nuisance function Furthermore, cross-fitted analytical variance estimators for DML U-statistics are proposed, which is a marked improvement over previous developments These subsample-based estimators based on incomplete U-statistics are shown to have the same asymptotic properties as their complete sample counterparts
Finite Sample Performance and Application
Simulation studies demonstrate that the DML estimator is the most efficient, confirming theoretical efficiency results under cluster randomization The analysis of a primary care-based cognitive behavioral therapy intervention for long-term opioid users with chronic pain illustrates the practical utility of the proposed methods The DML estimator improves the efficiency of the nonparametric estimator by significant percentages across all three outcome cases, demonstrating superior precision
Conclusion and Future Directions
The work develops a unified potential outcomes framework for causal inference with multiple outcomes in CRTs, anchoring inference on explicitly defined generalized causal effect estimands rather than model-dependent or outcome-dependent summaries The distinction between cluster-pair and individual-pair estimands clarifies how scientific questions, units of inference, and potentially informative cluster sizes jointly shape causal interpretation in CRTs Future research should extend this framework to covariate-adaptive cluster randomization and cluster rerandomization The paper concludes by showing that the interdisciplinary care model is more effective than usual care across all outcome cases considered
References
Balzer, L. B., Zheng, W., van der Laan, M. J., and Petersen, M. L. (2019), “A new approach to hierarchical data analysis: targeted maximum likelihood estimation for the causal effect of a cluster-level exposure,” Statistical Methods in Medical Research, 28, 1761–1780
Bebu, I. and Lachin, J. p. 5 Under the
joint approach, one directly defines a rule w on Y ×Y , e. , with a prioritized user-specified
order w(u, v) = 1(u ≻ v) (Bebu and Lachin, 2016) or the non-prioritized Pareto partial
order w(u, v) = 1(u ⪰P v).Where the paper says it
Blom, G. (1976), “Some properties of incomplete U-statistics,” Biometrika, 573–580
Brunner, E., Vandemeulebroecke, M., and M¨utze, T. (2021), “Win odds: An adaptation of the win ratio to include ties,” Statistics in Medicine, 40, 3367–3384
Buyse, M. (2010), “Generalized pairwise comparisons of prioritized outcomes in the twosample problem,” Statistics in Medicine, 29, 3245–3257
Chen, X. and Li, F. (2026), “Principal stratification with U-statistics under principal ignorability,” Journal of the Royal Statistical Society: Series B (Statistical Methodology), qkag044
Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., and Robins, J. (2018), “Double/debiased machine learning for treatment and structural parameters,” The Econometrics Journal, 21, C1–C68
DeBar, L., Mayhew, M., Benes, L., Bonifay, A., Deyo, R. A., Elder, C. R., Keefe, F. J., Leo, M. C., McMullen, C., Owen-Smith, A. p. 1 The first type concerns
non-prioritized outcomes, in which components may be measured on different scales, but
none is assigned intrinsic clinical precedence. In this setting, summary measures that map
heterogeneous outcomes onto a common probability scale are appealing, for example, by
using the global treatment effect summary defined in Smith et al. The first type concerns
non-prioritized outcomes, in which components may be measured on different scales, but
none is assigned intrinsic clinical precedence. In this setting, summary measures that map
heterogeneous outcomes onto a common probability scale are appealing, for example, by
using the global treatment effect summary defined in Smith et al.Where the paper says it
Where the paper says it
Escanciano, J. C. and Terschuur, J. R. (2023), “Machine learning inference on inequality of opportunity,” arXiv preprint arXiv. p. 0 Simulations
and an application to a CRT for chronic pain management illustrate the practical
utility of the proposed methods. Keywords: causal inference, covariate adjustment, debiased machine learning, efficient influence function, probabilistic index model, U-statistics
1 Introduction
Cluster randomized trials (CRTs) are a fundamental design for evaluating interventions
that are implemented at the group level, such as clinics, schools, or communities, and
1
arXiv:2601.Where the paper says it
Fang, X., Cao, Z., and Li, F. (2025), “Sample size determination for win statistics in cluster-randomized trials,” arXiv preprint arXiv:2510. p. 0 Simulations
and an application to a CRT for chronic pain management illustrate the practical
utility of the proposed methods. Keywords: causal inference, covariate adjustment, debiased machine learning, efficient influence function, probabilistic index model, U-statistics
1 Introduction
Cluster randomized trials (CRTs) are a fundamental design for evaluating interventions
that are implemented at the group level, such as clinics, schools, or communities, and
1
arXiv:2601.Where the paper says it
Kahan, B. C., Li, F., Copas, A. J., and Harhay, M. O. (2023), “Estimands in clusterrandomized trials: choosing analyses that answer the right question,” International Journal of Epidemiology, 52, 107–118
Lee, M.-L. T. and Dehling, H. p. 3 Second, we establish nonparametric identification and estimation results for these generalized causal effect estimands
under cluster randomization. The resulting estimators take the form of the two-sample Ustatistics for clustered data (Lee and Dehling, 2005), with the weighting structure reflecting
the natural variation in cluster sizes.Where the paper says it
Li, F., Tong, J., Fang, X., Cheng, C., Kahan, B. p. 2 Despite these recent developments, several fundamental gaps remain in causal inference
with multiple outcomes in CRTs. First, there is a lack of well-defined potential outcome
estimands for generalized causal effects that explicitly account for heterogeneous outcome
types and potentially informative cluster sizes, a defining feature that complicates both
interpretation and estimation (Kahan et al. For example, Su and Ding (2021) established the asymptotic properties of model-assisted analysis-of-covariance estimators for
weighted average treatment effects in CRTs. Wang et al.Where the paper says it
Where the paper says it
MacKinnon, J. G. and White, H. (1985), “Some heteroskedasticity-consistent covariance matrix estimators with improved finite sample properties,” Journal of Econometrics, 29, 305–325
Mao, L. (2017), “On causal estimation using U-statistics,” Biometrika, 105, 215–220
Ostelo, R. W., Deyo, R. A., Stratford, P., Waddell, G., Croft, P., Von Korff, M., Bouter, L. M., and De Vet, H. C. (2008), “Interpreting change scores for pain and functional status in low back pain: towards international consensus regarding minimal important change,” Spine, 33, 90–94
Pocock, S. J., Ariti, C. A., Collier, T. J., and Wang, D. p. 1 For example, Su and Ding (2021) established the asymptotic properties of model-assisted analysis-of-covariance estimators for
weighted average treatment effects in CRTs. Wang et al.Where the paper says it
Reed II, D. E., Stump, T. E., Monahan, P. O., and Kroenke, K. (2024), “Comparable minimally important differences and responsiveness of brief pain inventory and PEG pain scales across 6 trials,” The Journal of Pain, 25, 142–152
Rubin, D. B. (1974), “Estimating causal effects of treatments in randomized and nonrandomized studies,” Journal of Educational Psychology, 66, 688–701
Schochet, P. Z., Pashley, N. E., Miratrix, L. W., and Kautz, T. (2022), “Design-based ratio estimators and central limit theorems for clustered, blocked RCTs,” Journal of the American Statistical Association, 540, 2135–2146
Smith, E. D., Jairath, V., and Zou, G. (2025), “Rank-based estimators of global treatment effects for cluster randomized trials with multiple endpoints on different scales,” Statistical Methods in Medical Research, 34, 1267–1289
Su, F. and Ding, P. p. 1 , 2017), an emerging literature has focused on estimating average treatment effect estimands defined under
the potential outcomes framework and on developing causal inference procedures that are
agnostic to the choice of working outcome models. For example, Su and Ding (2021) established the asymptotic properties of model-assisted analysis-of-covariance estimators for
weighted average treatment effects in CRTs.Where the paper says it
Thas, O., Neve, J. D., Clement, L., and Ottoy, J.-P. (2012), “Probabilistic index models,” Journal of the Royal Statistical Society Series B: Statistical Methodology, 74, 623–671
Tsiatis, A. A. (2006), Semiparametric Theory and Missing Data, Springer New York, NY
Turner, E. L., Prague, M., Gallis, J. A., Li, F., and Murray, D. M. (2017), “Review of recent methodological developments in group-randomized trials: part 2—analysis,” American Journal of Public Health, 107, 1078–1086
van der Laan, M. J., Polley, E. C., and Hubbard, A. E. (2007), “Super learner,” Statistical Applications in Genetics and Molecular Biology, 6
van der Vaart, A. (1998), Asymptotic Statistics, Cambridge, UK: Cambridge University Press
Wang, B. and Li, F. (2024), “Asymptotic inference with flexible covariate adjustment under rerandomization and stratified rerandomization,” arXiv preprint arXiv:2406. p. 0 Simulations
and an application to a CRT for chronic pain management illustrate the practical
utility of the proposed methods. Keywords: causal inference, covariate adjustment, debiased machine learning, efficient influence function, probabilistic index model, U-statistics
1 Introduction
Cluster randomized trials (CRTs) are a fundamental design for evaluating interventions
that are implemented at the group level, such as clinics, schools, or communities, and
1
arXiv:2601.Where the paper says it
Wang, B., Park, C., Small, D. S., and Li, F. (2024), “Model-robust and efficient covariate adjustment for cluster-randomized experiments,” Journal of the American Statistical Association, 1–13
Wolfe, D. A. and Hogg, R. V.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Unified Potential Outcomes Framework for Generalized Causal Effects: The AI system can define generalized causal effect estimands through a unified potential outcomes framework accommodating both non-prioritized and prioritized outcome settings, including
cluster-pair and individual-pair causal estimands.
-
Nonparametric Estimation via Weighted Clustered U-statistics: The system can utilize nonparametric estimation by constructing estimators that
take the form of the two-sample Ustatistics for clustered data (Lee and Dehling, 2005), with the weighting structure reflecting the natural variation in cluster sizes.
-
Model-Robust Estimators via Efficient Influence Functions: The AI can construct covariate-adjusted estimators using efficient influence functions to
attain the semiparametric efficiency bounds under mild regularity conditions,
which are motivated by the derived EIFs. -
Debiased Machine Learning Estimators with Cross-Fitting: The system can employ debiased machine learning U-statistics, which leverage flexible nuisance function estimates and
attain the semiparametric efficiency lower bounds under mild regularity conditions
through cross-fitting schemes such as sample splitting for U-statistics. -
Computationally Efficient Estimation via Subsampling: The AI can implement subsample-based estimators motivated by incomplete U-statistics to
alleviate the computational resource constraint,
allowing for sequential estimation on smaller subsets of data while maintaining asymptotic properties. -
Prioritized Outcome Analysis: The system can perform win ratio analyses in prioritized settings by using contrast functions like "w(u, v) = 1(u > v)" and deriving the corresponding cluster-pair and individual-pair win ratio estimands, as shown in Remark 1.
-
Non-Prioritized Outcome Analysis: The system can evaluate non-prioritized outcomes by applying joint comparison contrast functions like the Pareto partial order w(u, v) = 1(u <= P v), leading to summary measures such as
net benefit
andwin ratio.
Abstract
Cluster-randomized trials (CRTs) are used to evaluate group-level interventions and can collect multiple outcomes capturing complementary dimensions of benefit and risk. Investigators typically seek a single global summary of treatment effect, yet existing methods largely focus on single-outcome estimands or rely on model-based procedures with unclear causal interpretation or limited robustness. We develop a unified potential outcomes framework for generalized treatment effects with multiple outcomes in CRTs, accommodating both non-prioritized and prioritized outcome settings. The proposed cluster-pair and individual-pair causal estimands are defined through pairwise contrast functions and explicitly account for informative cluster size. We establish nonparametric estimation via weighted clustered U-statistics and derive efficient influence functions to construct covariate-adjusted estimators that integrate debiased machine learning with U-statistics. The resulting estimators are consistent and asymptotically normal, attain semiparametric efficiency bounds under mild regularity conditions, and admit analytically tractable variance estimators that are consistent under cross-fitting. Simulations and an application to a CRT for chronic pain management illustrate the practical utility of our proposed methods.
Sources
- Sample size determination for win statistics in cluster-randomized trials
- Asymptotic inference with flexible covariate adjustment under rerandomization and stratified rerandomization
Related papers
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes
- Covariance-Calibrated Prediction-Powered Inference