A Bayesian approach to the survivor average causal effect in cluster-randomized crossover trials
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: "A Bayesian approach to the survivor average causal effect in cluster-randomized crossover trials".
Jane: Here's the gist of the paper. In cluster-randomized crossover (CRXO) trials, researchers often want to assess the effect of treatments on secondary non-terminal outcomes but frequently in these studies, there are patients who do not survive to have these measurements fully recorded <ref:2505.21447#pg2>.
Tom: This paper provides a causal inference framework and treatment effect estimation methods for addressing truncation by death in the setting of CRXO trials <ref:2505.21447#pg4>.
Jane: Target Estimand Definition The authors target the survivor average causal effect (SACE), a well-defined subgroup treatment effect obtained via principal stratification <ref:2505.21447#pg4>.
Tom: The SACE is defined as a causal contrast, often expressed in the literature as a difference in means, among the sub-population of “always-survivors” – individuals that would survive regardless of assigned treatment such that their non-mortality outcomes under either treatment are potentially observable <ref:2505.21447#pg4>.
Title and authors: Tom: So this paper is called "A Bayesian approach to the survivor average causal effect in cluster-randomized crossover trials," written by Isenberg, Harhay, Forbes, Young, Li, Mitra.
Jane: That title tells us right away that they're looking at how to measure an effect in a specific type of trial called a cluster-randomized crossover trial.
Lu: The authors are from a bunch of different places including the University of Pennsylvania and Monash University; that suggests they bring together different statistical traditions.
Meng: It sounds like they're trying to solve a problem where you have this complex randomization scheme, the CRXO design, but you want to know something specific about treatment effects that are hard to get out cleanly.
Lalam: They're focusing on something called the survivor average causal effect, or SACE; it’s a specific way they define an effect among people who survive regardless of what treatment they got.
Tom: Exactly, so it’s not about the average person in the trial, but this carefully selected subgroup—the always-survivors—and figuring out how the treatment affects them.
Jane: It's a refinement because when you look at standard trials, like individual-randomized trials, you need smaller sample sizes to get precise estimates than what we can achieve with this CRXO design.
The paper's summary: Tom: This paper explains that to tackle the problem of people dying before we measure their outcome, they are using a Bayesian approach where they add latent membership information about which group an individual belongs to.
Jane: They augment the posterior distribution with this principal stratum membership so that the conditional posteriors become easier to handle when they use Markov chain Monte Carlo sampling.
Lu: The core idea is that you can define this SACE as a causal contrast, which means it’s expressed as a difference in means or a ratio of means among those always-survivors.
Meng: They are using the concept of principal stratification to isolate this subgroup, which is the only group where you don't have to worry about whether their death status blocks your measurement of the outcome.
Lalam: Essentially, they use data augmentation with Polya-Gamma random variables to represent how much each individual contributes to each principal stratum they belong to.
Tom: They then run a simulation study on a two-period CRXO design and find that Model one which includes cluster random effects in the principal strata modeling and both cluster and cluster period random effects in the outcome modeling, shows low bias <ref:2505.21447#pg1>.
Jane: The implication here is that they show you how to estimate this SACE robustly even when dealing with truncation by death, which is a big hurdle in these types of clinical trials.
The paper's improvements: Tom: They suggest several improvements to their method, including evaluating performance across different model specifications like Model two or Model three <ref:2505.21447#pg1>.
Jane: One key improvement they highlight is that while it’s important to model the full hierarchy of random effects in the outcome modeling, you can simplify the principal strata modeling to keep estimation stable in small samples.
Lu: They found that dropping cluster-period random effects in the principal strata modeling doesn't hurt coverage of these SACE estimators as much as dropping them in the outcome modeling does.
Meng: From a deployment standpoint, it’s good news if you can simplify the model structure without losing statistical power or coverage for something critical like this effect.
Lalam: They also mention that they need to assume survival monotonicity, which is a structural assumption—that being assigned the active treatment doesn't make survival worse compared to the control treatment.
Tom: That assumption, survival monotonicity, is untestable and could be challenged in comparative effectiveness studies, which is a real limitation they flag.
Conclusion: Jane: So to wrap up, this paper proposes a Bayesian framework with data augmentation to estimate the SACE in cluster-randomized crossover trials by defining it as a causal contrast among always-survivors.
Tom: They showed that using Model one—the one with both cluster random effects in both modeling components—works really well in simulation, showing low bias and near nominal coverage for the SACE estimands across different scenarios <ref:2505.21447#pg1>.
Lu: The key contribution is showing how to use this method to estimate the survivor average causal effect in a setting where you have truncation by death, which is a big problem when looking at non-terminal outcomes.
Meng: For those of us shipping models, the practical point is that they have a framework that can handle this complexity without needing an impossibly complex model structure to run on real data.
Lalam: And for the PEPTIC trial they applied this, getting estimates for the difference in means in log days of zero point sixty-eight and a ratio of means in days of one point sixty-three among always-survivors.
Jane: Those numbers suggest a statistically significant increase in hospital stay for patients who got proton pump inhibitors compared to histamine two receptor blockers among the always-survivors.
Tom: That means the expected one day length among those on H2RBs is increased by about zero point sixty-three days, or one point five twelve hours for those on PPIs.
Lu: They also showed that Model one performs best overall compared to simpler models in terms of bias and root mean square error across the simulations they ran with one thousand simulated data sets <ref:2505.21447#pg1>.
Meng: It sounds like a solid way to go if you're trying to estimate these effects in real clinical settings where you can't afford massive computational overhead.
Lalam: In short, the paper "A Bayesian approach to the survivor average causal effect in cluster-randomized crossover trials" gives us a way to estimate this specific effect using a robust Bayesian framework.
Dane Isenberg, Michael O. Harhay, Andrew B. Forbes, Paul J. Young, Fan Li, Nandita Mitra
Department of Biostatistics, Epidemiology and Informatics, University of Pennsylvania · Division of Quantitative Research Methodology, School of Public Health and Preventive Medicine, Monash University · Australian and New Zealand Intensive Care Research Centre Department of Epidemiology and Preventive Medicine, School of Public Health and Preventive Medicine, Monash University · Medical Research Institute of New Zealand Australian National Institute for Medical Research
stat.ME
Submitted: 2025-05-27
Updated: 2026-11-08
Comments: Revisions for clarity of assumptions and methods. Expansion of appendix, including addition of assumptions for stepped-wedge trials. New authorship
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 88/100
The gist: In cluster-randomized crossover (CRXO) trials, researchers often want to assess the effect of treatments on secondary non-terminal outcomes but frequently in these studies…
Key concepts
- Survivor Average Causal Effect (SACE)
- This is a specific treatment effect calculated only among 'always-survivors'—patients who would survive regardless of which treatment they received. It represents the causal effect for this highly specific subgroup, allowing researchers to estimate outcomes where death prevents full measurement.
- Principal Stratum Membership
- This refers to classifying individuals into distinct groups based on their potential survival status under different treatments. The Bayesian approach uses latent variables to model this membership, helping the analysis account for the fact that not all patients can be fully observed due to mortality.
- Bayesian Augmentation
- The method improves standard Bayesian modeling by adding latent principal stratum membership information directly into the posterior distribution. This makes complex conditional posteriors more manageable for sampling algorithms like MCMC, allowing for more accurate estimation of the SACE.
Terminology
Summary
In cluster-randomized crossover (CRXO) trials, researchers often want to assess the effect of treatments on secondary non-terminal outcomes but frequently in these studies, there are patients who do not survive to have these measurements fully recorded <ref:2505. p. 021447#pg2>. This paper provides a causal inference framework and treatment effect estimation methods for addressing truncation by death in the setting of CRXO trials <ref:2505. p. 021447#pg4>.
Target Estimand Definition
The authors target the survivor average causal effect (SACE), a well-defined subgroup treatment effect obtained via principal stratification <ref:2505.21447#pg4>. The SACE is defined as a causal contrast, often expressed in the literature as a difference in means, among the sub-population of “always-survivors” – individuals that would survive regardless of assigned treatment such that their non-mortality outcomes under either treatment are potentially observable <ref:2505. p. 121447#pg4>. This subgroup is the only principal stratum comprised of individuals whose death status under either treatment would not preclude their complete measurement of the non-terminal outcome <ref:2505. p. 321447#pg4>.
Causal Assumptions and Identification
The identification of the SACE requires a set of structural and modeling assumptions to point identify the SACE so that it is estimable using data from a CRXO trial <ref:2505.21447#pg4>. These assumptions include six specific points, such as Assumption 1 stating that individuals’ potential outcomes are only governed by the treatment assigned to their cluster-period <ref:2505. p. 421447#pg4>, and Assumption 6 which states that each individual’s potential survival status under the active treatment is no worse than their potential survival status under the control treatment <ref:2505. p. 421447#pg4>.
Bayesian Estimation Methodology
The authors propose a Bayesian approach where they augment the posterior with latent principal stratum membership so that conditional posteriors are more tractable for MCMC sampling <ref:2505.21447#pg4>. They complete the observed-data likelihood by including each individual’s principal stratum membership, allowing them to use the EM-algorithm or define an augmented posterior distribution which is itself more tractable and suited to simpler Markov chain Monte Carlo (MCMC) methods <ref:2505. p. 521447#pg4>. The conditional posteriors for the P´olya-Gamma variables used for augmentation are themselves P´olya-Gamma, where they are related to the base distribution PG(b, 0) through an exponential tilting <ref:2505. p. 721447#pg4>.
Simulation Results and Application
In simulation studies, the authors evaluate the small-sample frequentist performance of their Bayesian approach to SACE estimation for a two-period CRXO design <ref:2505.21447#pg4>. Model 1, which includes cluster random effects in PS modeling and cluster and cluster-period random effects in outcome modeling, generally shows low bias and achieves close to nominal coverage for both the difference and ratio of means SACE estimands across all the scenarios considered <ref:2505. p. 921447#pg4>. The authors apply their methods to the PEPTIC trial where they estimate the SACE comparing the effect of PPIs to H2RBs on hospital LOS for mechanically ventilated critical care patients <ref:2505. p. 221447#pg4>.
Discussion and Limitations
The discussion notes that while it is imperative to account for the full random effects hierarchy in the nonmortality outcome modeling, they can reduce such complexity in the principal strata modeling to ensure stable estimation in small samples without measurably sacrificing performance <ref:2505.21447#pg4>. Limitations include assuming survival monotonicity, which is untestable and could be challenged in comparative effectiveness studies <ref:2505. p. 321447#pg4>, and the risk of model mis-specification due to length of stay being notoriously difficult to model parametrically <ref:2505. p. 621447#pg4>. The results suggest that dropping cluster-period random effects in the principal strata modeling is less consequential for coverage of the SACE estimators than dropping them in outcome modeling <ref:2505. p. 921447#pg4>.
Algorithmic Framework
The algorithmic framework for drawing from the posterior distribution using the Gibbs sampler involves ten steps, including initialization of necessary components of theta ext, initialization of augmentation variable g, and subsequent updates for non-mortal regression and SACE parameters <ref:2505.21447#pg4>. The updates involve drawing from conditional posterior distributions conditioned on current draws (i. p. 9e. within the iteration or from the last iteration) of all requisite parameters <ref:2505. p. 1 p. 121447#pg4>.
Data Augmentation Detail
The data augmentation approach uses P´olya-Gamma random variables to represent the contributions of individuals to each principal stratum <ref:2505.21447#pg4>. The conditional posterior of ωijk,zuijk,11,. p. 8.. is given by a formula involving an integral identity in Equation (16) <ref:2505. p. 821447#pg4>. This allows for drawing from conditional posterior distributions involving parameters of zijk using the P´olya-Gamma distribution <ref:2505. p. 821447#pg4>.
Inference on PEPTIC Trial
For the PEPTIC trial, the authors obtain estimates and 95% HPD credible intervals for the SACE in difference in means in log-days of 0.068 (0.016, 0.118) and ratio of means in days of 1. p. 3 p. 3063 (1.008, 1.118) <ref:2505. p. 0 p. 021447#pg4>. These credible intervals suggest a statistically significant increase in hospitalization LOS for those assigned to PPIs versus H2RBs among always-survivors <ref:2505. p. 1 p. 121447#pg4>. The naive estimates obtained from the LMM among observed survivors are similar in direction and magnitude to the SACE estimates (see Table 3) <ref:2505. p. 921447#pg4>.
Supplemental Results Summary
The simulation study results show that Model 1 and/or Model 2 may also give rise to some undercoverage <ref:2505.21447#pg4>. Furthermore, when the composition of our strata is (0. p. 325, 0.25, 0.5), the performance of the models appear to be more impacted when we reduce the number of never-survivors and protected-patients with composition (0. p. 418, 10%, 72%) <ref:2505. p. 021447#pg4>. The authors demonstrate that Model 1 performs best overall relative to simpler comparator models in terms of bias, RMSE, and coverage in aggregate <ref:2505. p. 921447#pg4>. The results are from 1,000 simulated data sets with Bayesian inference on 10,000 MCMC iterations and 2,500 burn-ins <ref:2505. p. 9 p. 921447#pg4>. The trace plots for key parameters show sufficient mixing <ref:2505. p. 9 p. 921447#pg4>. The results are from 1,000 simulated data sets with Bayesian inference on 10,000 MCMC iterations and 2,500 burn-ins <ref:2505.21447#pg4>. The trace plots for key parameters show sufficient mixing <ref:2505.21447#pg4>.
Parameter Specification
The parameter vector theta is given by a long list of components including regression coefficients, random effects vectors, and variance terms <ref:2505.21447#pg4>. The specification for the posterior distribution includes normal and inverse gamma prior specifications on its components, which enable draws from closed-form conditional posteriors for all parameters per Gibbs sampling <ref:2505. p. 821447#pg4>. The posterior distributions are conjugate to the prior distributions as written out in the Supplement A <ref:2505. p. 8 p. 821447#pg4>. The posterior distributions are conjugate to the prior distributions as written out in the Supplement A <ref:2505.21447#pg4>.
Prior Specifications
For all random effects, error terms, and their respective variances are mean zero multivariate normal distributed with diagonal variances per the design of the hierarchical model <ref:2505.21447#pg4>. All variance terms are ascribed diffuse inverse gamma priors <ref:2505. p. 821447#pg4>. The prior specifications on theta and the data augmentation steps are conjugate to the prior distributions as written out in Supplement A <ref:2505. p. 8 p. 821447#pg4>. The prior specifications on theta and the data augmentation steps are conjugate to the prior distributions as written out in Supplement A <ref:2505.21447#pg4>.
MCMC Algorithmic Steps
The algorithmic framework for drawing from the posterior distribution using the Gibbs sampler involves ten steps, including initialization of necessary components of theta ext, initialization of augmentation variable g, and subsequent updates for non-mortal regression and SACE parameters <ref:2505.21447#pg4>. The updates refer to draws from conditional posterior distributions conditioned on current draws (i. p. 9e. within the iteration or from the last iteration) of all requisite parameters <ref:2505.21447#pg4>. The MCMC algorithm is written out below <ref:2505. p. 521447#pg4>.
Posterior Update for SACE
The update for SACE in difference in means (log-time scale), LDiff, and ratio of means (time scale), ROM, which are contrasts of averages of functions among always-survivors is performed <ref:2505.21447#pg4>. The posterior mean estimate for LDiff is approximated by E(µLDiffdata) = Z α(1,1),1 + β T (1,1),1 Z XijkdF(XijkGijk = (1, 1)) dPpost ≈ αˆ(1,1), 1 + βˆ T (1, 2) X¯ (2) <ref:2505. p. 921447#pg4>. The posterior mean estimate for ROM is given by µˆROM = Z exp α(1,1),1 exp β T (1, 1)X + δ T (1, 1)κ + β T (1, 2)Xijko exp β T (2)X + δ T (2)κ <ref:2505. p. 921447#pg4>. The posterior mean estimate for LDiff is obtained by E(µLDiffdata) = Z α(1,1),1 + β T (1,1),1 Z XijkdF(XijkGijk = (1, 1)) dPpost ≈ αˆ(1, 2) X¯ (2) <ref:2505. p. 921447#pg4>. The posterior mean estimate for ROM is given by µROM = exp α(1,1),1 E n exp β T (1, 1)X + δ T (1, 1)κ + β T (2)Xijko E n exp β T (2)X + δ T (2)κ <ref:2505. p. 921447#pg4>. The posterior mean estimate for LDiff is obtained by E(µLDiffdata) = Z α(1,1),1 + β T (1, 1), 1 X¯ (1, 1) <ref:2505. p. 9 p. 9 p. 921447#pg4>. The posterior mean estimate for ROM is given by µˆROM = Z exp α(1, 2) X + δ T (2)κ + β T (2),1Xexp β T (2)X + δ T (2)κ <ref:2505. p. 921447#pg4>. The posterior mean estimate for LDiff is obtained by E(µLDiffdata) = Z α(1,1),1 + β T (1, 1), 1 X¯ (1, 1) <ref:2505.21447#pg4>. The posterior mean estimate for ROM is given by µˆROM = Z exp α(2)X + δ T (2)κ + β T (2),1Xexp β T (2)X + δ T (2)κ <ref:2505. p. 9 p. 921447#pg4>. The posterior mean estimate for LDiff is obtained by E(µLDiffdata) = Z α(1,1),1 + β T (1, 1), 1 X¯ (1, 1) <ref:2505.21447#pg4>. The posterior mean estimate for ROM is given by µˆROM = Z exp α(2)X + δ T (2)κ + β T (2),1Xexp β T (2)X + δ T (2)κ <ref:2505.21447#pg4>.
PEPTIC Trial Estimates
For the PEPTIC trial, the authors obtain estimates and 95% HPD credible intervals for the SACE in difference in means in log-days of 0.068 (0.016, 0.118) and ratio of means in days of 1.063 (1.008, 1.118) <ref:2505.21447#pg4>. These credible intervals suggest a statistically significant increase in hospitalization LOS for those assigned to PPIs versus H2RBs among always-survivors <ref:2505.21447#pg4>. The expected LOS of one day among individuals assigned to H2RBs is estimated to be increased by about 0. p. 1063 days or 1.512 hours for those assigned to PPIs <ref:2505. p. 121447#pg4>.
Discussion on Robustness
The authors demonstrate that Model 1 performs best overall relative to simpler comparator models in terms of bias, RMSE, and coverage in aggregate <ref:2505.21447#pg4>. They show that including cluster-period random effects in the outcome model was necessary to maintain coverage across all datagenerating processes <ref:2505. p. 921447#pg4>. Model 1 performs just as well if not better than the “correctly specified” fuller model, Model A, which explicitly includes those cluster-period random effects <ref:2505. p. 721447#pg4>. The authors caution that the SACE effects described in this paper can only be interpreted from an intention-to-treat perspective <ref:2505. p. 9 p. 921447#pg4>. The proximity of these estimates might be explained by the fact that the classified always-survivors are comprised of similar individuals as the observed survivors <ref:2505. p. 321447#pg4>. The results show that Model 1 and/or Model 2 may also give rise to some undercoverage <ref:2505. p. 421447#pg4>. The authors suggest that a remedy for this undercoverage might be to employ a different family of priors, such as half-Cauchy priors <ref:2505.21447#pg4>. The development of a sensitivity analysis approach to the survival monotonicity assumption in CRXOs is a fruitful area for future research <ref:2505. p. 421447#pg4>. The authors caution that the SACE effects described in this paper can only be interpreted from an intention-to-treat perspective <ref:2505.21447#pg4>.
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
-
Improved causal inference framework for truncation by death in cluster randomized crossover trials allows for estimating the survivor average causal effect (SACE) using a Bayesian paradigm that augments posterior distributions with latent principal stratum membership to make conditional posteriors tractable for Markov chain Monte Carlo sampling.
-
The improved system can estimate the SACE by defining it as a causal contrast, often expressed as
a difference in means summary measure and setting h(x, y) = x - y generates the difference in means
orthe ratio of means summary measure (i.e., the relative risk for a binary outcome)
within the subgroup of always-survivors. -
The system can perform simulation studies to evaluate the finite sample operating characteristics, demonstrating that Model 1, which includes cluster random effects in principal strata modeling and both cluster and cluster-period random effects in outcome modeling, generally shows minimal bias and achieves near nominal coverage for SACE estimands across various scenarios.
-
The improved system can be used to compare different model specifications by evaluating the empirical bias and root mean square error for SACE estimation across reduced models, such as Model 2 (removes cluster random effects from PS modeling only) or Model 3 (removes cluster-period random effects from outcome modeling only).
-
The system can estimate the SACE for specific clinical scenarios, such as the PEPTIC trial comparing proton pump inhibitors versus histamine-2 receptor blockers on length of hospital stay, providing posterior mean estimates and 95 percent highest posterior density credible intervals for both difference in means on log time scale and ratio of means on time scale.
-
The system can incorporate survival monotonicity as a structural assumption, which "simplifies modeling assumptions by removing the stratum of harmed patients, thereby identifying individuals observed to have died under the active treatment or have survived under the control treatment with a principal stratum (never-survivors and always-survivors respectively)."
Related papers
- Learning Robust Treatment Rules for Censored Data
- 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