A Bayesian Framework for Latent Compliance Modeling in Cluster Randomized Trials with One-Sided Noncompliance
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 Framework for Latent Compliance Modeling in Cluster Randomized Trials with One-Sided Noncompliance".
Paper summary: Tom: So we’re looking at this paper titled "A Bayesian Framework for Latent Compliance Modeling in Cluster Randomized Trials with One-Sided Noncompliance." Jane, what's the core idea they are pushing here?
Jane: Well, it tackles a problem where you have a cluster randomized controlled trial, or PCRCT, and you have both the healthcare providers and the patients potentially deviating from what they were supposed to do. The authors say that standard complier average causal effect estimands only look at how the individual patient complies, ignoring how different clusters actually implemented their intervention.
Tom: Right, so it’s about that heterogeneity in implementation across clusters, which is usually a big blind spot when we look at those standard CACE estimands. What exactly do they claim this framework does?
Jane: They propose using a latent mixture model to summarize that heterogeneity in cluster-level implementation based on baseline characteristics and observed metrics. Then, they link those latent implementation types directly to the individual compliance and the final outcomes through a joint model. This lets you estimate effects not just for the whole population, but also within different implementation types.
Lu: I like this because it moves beyond just looking at one type of noncompliance, like binary noncompliance. They are modeling partial compliance at the cluster level alongside binary compliance at the individual level, which is a step up from previous work that used a single measure for both > Page three of THIS PAPER — "We propose a Bayesian framework for analyzing PCRCTs with one-sided binary noncompliance at the individual level and one-sided partial compliance at the cluster level <ref:2509.15280#pg3,We propose a Bayesian framework for analyzing PCRCTs with one-sided binary>. The method uses a latent mixture model to summarize heterogeneity in cluster-level implementation based on pre-treatment characteristics and observed implementation measures, and links these latent implementation strata to individual compliance and outcomes through a joint model, there"
Meng: From my side, what I need to know is how this translates from theory to something we can actually deploy. If you have these latent strata, how does that affect the computational cost? Are we talking about something that runs smoothly on real traffic or is it just for theoretical exercises?
Paper summary: Lalam: Based on what I see in the structure of this paper, the framework is built to handle complexity by using a Dirichlet prior to regularize the estimation and reflect uncertainty in how many implementation strata exist > Page three of THIS PAPER — "The method uses a latent mixture model to summarize heterogeneity in cluster-level implementation based on pre-treatment characteristics and observed implementation measures, and links these latent implementation strata to individual compliance and outcomes through a joint model, there" <ref:2509.15280#pg3,The method uses a latent mixture model to summarize heterogeneity in cluster>
Tom: It sounds like they’re setting up a very detailed statistical structure. Jane, if we boil it down for someone just listening, what’s the big payoff here? Why does characterizing these different implementation patterns matter for the real world?
Jane: The payoff is being able to estimate intervention effects specifically among providers with similar implementation patterns and also among individuals who receive the intervention exactly as intended. It allows you to see how compliance behavior and outcome dependence are governed by that shared environment of how things get implemented > Page one of THIS PAPER — "In pragmatic cluster randomized controlled trials (PCRCTs), healthcare providers are randomized while both providers and patients may deviate from the assigned intervention <ref:2509.15280#pg1,In pragmatic cluster randomized controlled trials (PCRCTs), healthcare providers are randomized while>. In many PCRCTs, cluster-level implementation is measured using multiple continuous metrics, while individual compliance is recorded as a binary indicator"
Lu: It’s about moving beyond the naive view where you assume everyone in a cluster acts the same way. This framework treats the way an intervention gets put into practice—that implementation strategy—as a source of variation that directly influences what happens to patients > Page two of THIS PAPER — "Agitation and aggressive behaviors are common among cognitively impaired individuals residing in nursing homes (NHs) (Zuidema et al <ref:2509.15280#pg1>., two thousand seven) <ref:2509.15280#pg2>. These behaviors can decrease quality of life and accelerate cognitive decline (Garc´ıa-Alberca et al., two thousand nineteen; Shin et al <ref:2509.15280#pg2>., two thousand five), while also serving as a major source of stress for NH staff (Karlsen et al <ref:2509.15280#pg2,2005), while also serving as a major source of stress for NH>. two thousand twenty-three) and other residents (Rosen et al <ref:2509.15280#pg2,2023) and other residents (Rosen et al>. two thousand eight) <ref:2509.15280#pg2>. Antipsychotic and mood-stabilizing medications are commonly used to manage aggressive behaviors; however, these medications are associated with adverse events, including falls and increased mortality (Huybrechts et al. two thousand twelve) <ref:2509.15280#pg2>."
Meng: So you’re saying if we can identify that a specific cluster type—say, one with low targeting proportions—has a different outcome profile than another type, we can tailor the intervention strategy for those specific environments? That sounds like something I could actually use to deploy more efficiently.
Paper summary: Lalam: The authors show this by applying it to METRIcAL data and finding two distinct latent implementation strata based on facility size and targeting proportions > Appendix three Prior distributions for METRIcAL application — "Latent implementation strata are modeled as Si ∼ Categorical(π1, π2), (π1, π2) ∼ Diriclet(one one) Conditional on Si = k, the facility-level implementation measure (logit clinical proportion) and baseline facility characteristic follow Ci Zi Si = k ∼ MVN(µS k, Σ), k = one two"
Tom: Okay. So what’s the conclusion then? What is the ultimate value proposition of this whole "A Bayesian Framework for Latent Compliance Modeling in Cluster Randomized Trials with One-Sided Noncompliance"?
Jane: The framework itself is principled because it jointly models implementation heterogeneity and individual compliance > Page one of THIS PAPER — "We propose a Bayesian framework for PCRCTs with onesided binary noncompliance at the individual level and one-sided partial compliance at the cluster level <ref:2509.15280#pg1,We propose a Bayesian framework for PCRCTs with onesided binary noncompliance at>. The method uses a latent mixture model to summarize heterogeneity in cluster-level implementation based on pre-treatment characteristics and observed implementation measures, and links these latent implementation strata to individual compliance and outcomes through a joint model, there"
Tom: So it’s essentially giving us a more nuanced way to look at why an intervention might work differently depending on the context of its delivery. It moves us past just asking if the treatment worked overall.
Jane: Exactly. It enables estimation of effects among providers with similar implementation patterns and among individuals who receive the intervention as intended, which is something standard intent-to-treat analyses often miss > Page one of THIS PAPER — "The framework enables estimation of finite- and super-population intent-to-treat (ITT) and CACE estimands, both marginally and within latent implementation types" <ref:2509.15280#pg1>
Tom: That’s it. It gives us better causal estimates that are more specific to the context. We’ll be keeping an eye on how they handle those partial compliance assumptions in future work > Page three of THIS PAPER — "In many pragmatic trials, however, compliance may be partial rather than binary <ref:2509.15280#pg3,In many pragmatic trials, however, compliance may be partial rather than binary>. For example, individuals may receive different doses of an intervention (Jin and Rubin, two thousand eight), and clusters may vary in the intensity or scope of implementation" <ref:2509.15280#pg2>
Conclusion: Tom: So we've seen how they built this whole Bayesian framework for looking at cluster randomized trials when you have one-sided noncompliance, and now we need to talk about what it actually means for us as a way of thinking about these studies.
Jane: Exactly. The title is pretty long, right? "A Bayesian Framework for Latent Compliance Modeling in Cluster Randomized Trials with One-Sided Noncompliance." It sounds like a lot of math piled on top of a clinical problem, and we need to make sure we understand the core idea without getting lost in the notation.
Lu: The paper is essentially proposing a way to handle the mess when you measure implementation at the group level but compliance at the individual level, especially when people don't always follow through perfectly. They use this latent mixture model to sort those clusters into different types based on how they actually implemented things.
Meng: So, instead of just saying "this intervention worked for everyone," we can now say "this intervention worked differently depending on *how* the facility set it up." That’s a big difference when you’re trying to deploy something at scale.
Tom: Yeah, the implication is that we stop treating all clusters as one giant unit and start seeing the variation in delivery itself as a key driver of outcomes. It moves us from just looking at patient behavior to looking at system behavior.
Jane: It also lets researchers estimate effects for groups of providers who are similar in their implementation style, which is something standard intent-to-treat analyses completely miss. So we can get more specific causal effect estimates for those subgroups.
Lalam: I'm pretty confident in this direction because they show how the model handles the uncertainty and links that cluster heterogeneity directly to the individual outcomes, giving us super-population estimates that are much richer than what we usually get.
Tom: It gives us a principled way to analyze these kinds of pragmatic trials where both the setup and the execution are variable. We're seeing how this framework helps isolate what really drives change in patients when there's so much noise in how the treatment is actually delivered.
Jane: That leads perfectly into how they applied this to METRIcAL data, which showed that smaller facilities had different results than larger ones, giving us concrete examples of what these latent strata look like in practice.
Department of Biostatics, Brown University School of Public Health · Department of Health Services, Policy & Practice, Brown University School of Public Health
stat.AP, stat.ME
Submitted: 2025-09-18
Updated: 2026-11-17
License: http://creativecommons.org/licenses/by-sa/4.0/
Importance score: 83/100
The gist: A Bayesian framework for latent compliance modeling in cluster randomized trials with one-sided noncompliance proposes a method to characterize how intervention effects vary across sources of…
Key concepts
- Latent Implementation Strata
- This concept refers to unobserved groups of clusters defined by shared characteristics and observed implementation measures. The model uses a latent mixture model to summarize this heterogeneity, grouping clusters based on how they actually put the intervention into practice, rather than just their assigned treatment group.
- Joint Model
- A joint model is a statistical approach that simultaneously models multiple related variables—in this case, cluster-level implementation heterogeneity and individual compliance/outcomes. By linking these strata together, the framework can account for how the environment (implementation) influences both what individuals comply with and what their final outcomes are.
- CACEk
- CACEk stands for the complier average causal effect within a specific latent implementation stratum k. This measures the intervention's effect on individuals who complied, specifically within that particular implementation environment. It helps characterize how the treatment works differently depending on the cluster's underlying operational style.
- Pragmatic Cluster Randomized Trials (PCRCTs)
- These are clinical trials where healthcare providers are randomized, but both providers and patients might deviate from the assigned intervention. The framework is designed to handle this complexity by focusing on how implementation differences across clusters affect individual compliance and outcomes.
Terminology
Summary
A Bayesian framework for latent compliance modeling in cluster randomized trials with one-sided noncompliance proposes a method to characterize how intervention effects vary across sources of compliance by using a latent mixture model to summarize heterogeneity in cluster-level implementation and linking these strata to individual compliance and outcomes through a joint model
Pruning redundant reasoning tokens
In pragmatic cluster randomized trials (PCRCTs), healthcare providers are randomized while both providers and patients may deviate from the assigned intervention In many PCRCTs, cluster-level implementation is measured using multiple continuous metrics, while individual compliance is recorded as a binary indicator Standard complier average causal effect (CACE) estimands focus on individual-level compliance and do not account for heterogeneity in implementation across clusters When intervention uptake is shaped by both provider- and patient-level processes, it is of scientific interest to characterize how effects vary across these sources of compliance The method uses a latent mixture model to summarize heterogeneity in cluster-level implementation based on baseline characteristics and observed implementation measures, and links these latent implementation types to individual compliance and outcomes through a joint model Because compliance is only observed in treated clusters, the model imputes unobserved compliance behavior for clusters and individuals assigned to control The framework enables estimation of finite- and super-population intent-to-treat (ITT) and CACE estimands, both marginally and within latent implementation types
How the training signal is built
The model defines a joint distribution for the complete data, (Y(1, D), Y(0, 0), D, C, Z, X, S), with the observed data arising through randomization and the assumptions in Section 3.1 The variable Si ∈ (1. p. 0 The analysis illustrates how accounting for implementation heterogeneity and individual
compliance can provide insights beyond standard ITT analyses. Key words: Causal inference; Principal stratification; Complier average causal effect; Cluster randomized trials;
Noncompliance; Bayesian methods; Latent variable models; Interference. , K} represents qualitatively distinct implementation contexts, reflecting differences
in organizational capacity, staff engagement, leadership support, and other unmeasured features that influence implementation measures and individual behavior. Latent implementation strata follow a categorical
distribution,
Si ∼ Categorical(π1, . Latent implementation strata follow a categorical
distribution,
Si ∼ Categorical(π1, . , πK),
with a Dirichlet prior placed on (π1, . , πK),
with a Dirichlet prior placed on (π1, . , πK) to regularize estimation and reflect a-priori uncertainty in the
prevalence of each implementation stratum. Potential outcomes are denoted by Yij (w, r) and represent the outcome for individual j in cluster i under
cluster assignment w and cluster receipt vector r ∈ {0, 1}
ni
. This notation allows for partial interference
within clusters, as a patient’s outcome may be influenced by the outcomes of other patients who comply
with the intervention.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the gains come from
Causal estimands summarize the effects of an intervention in a population of interest Principal stratification estimands characterize effects within subpopulations defined by post-assignment variables such as treatment receipt We define the complier average causal effect within latent implementation stratum k as CACEk = E[Yij (1, 1, Si) − Yij (0, 0, Si) Dij = 1, Si = k], k = 1..., K A marginal complier average causal effect is obtained by averaging over the distribution of Si: CACE = E[Yij (1, 1, Si) − Yij (0, 0, Si) Dij = 1]
Application to METRIcAL data
We apply the proposed method to METRIcAL to estimate super-population ITT and CACE of the personalized music intervention on change in CMAI from baseline to 4-month follow-up and antipsychotic use at 4-month follow-up Two latent implementation strata were identified The first consisted of lower-occupancy nursing homes with higher proportions of residents targeted for the intervention, while the second included higher-occupancy facilities with lower targeting proportions At the population level, point estimates suggest that residents in intervention facilities experienced smaller reductions in CMAI scores relative to standard care, but also exhibited lower rates of antipsychotic use at four-month follow-up
Discussion
The latent implementation variable summarizes the shared environment that governs the compliance behavior and outcome dependence This can be viewed as a latent analogue of the exposure mapping, in which interference operates through unobserved implementation mechanisms rather than a pre-specified function of peers’ treatment Through simulated case studies, we examined the operating characteristics of the proposed method under both correctly specified and misspecified data-generating mechanisms When the model was correctly specified, estimates were relatively accurate with approximately nominal posterior coverage The procedure also demonstrated robustness to targeted misspecification of key model components, with posterior credible intervals retaining near-nominal coverage, although bias increased under misspecification of the compliance and outcome models
Posterior predictive checks
We assess model fit using posterior predictive checks (Gelman et al., 1996; Rubin, 1984) based on predictive discrepancy measures proposed by Barnard et al. p. 6 4 Bayesian model
Model-based estimation is a principled way to impose structure on the joint distribution of observed and
unobserved variables and to incorporate covariates flexibly (Imbens and Rubin, 2016). Bayesian approaches
also enable the assessment of model fit using posterior predictive checks to ensure sensible model specifications
(Mattei et al. 4 Bayesian model
Model-based estimation is a principled way to impose structure on the joint distribution of observed and
unobserved variables and to incorporate covariates flexibly (Imbens and Rubin, 2016). Bayesian approaches
also enable the assessment of model fit using posterior predictive checks to ensure sensible model specifications
(Mattei et al.Where the paper says it
Where the paper says it
Conclusion
In conclusion, the proposed Bayesian framework is a principled approach for analyzing pragmatic CRCTs in which intervention implementation is assessed at the cluster level using multiple measures, and individual-level compliance is assessed with a binary indicator By jointly modeling implementation heterogeneity and individual compliance, the method enables estimation of intervention effects among providers with similar implementation patterns and among individuals who receive the intervention as intended Applied to METRIcAL, the approach yields insights that are not accessible through standard intent-to-treat analyses An important extension for future work would be to incorporate models allowing for partial or continuous individual-level compliance, and more complex interference structures
References
James H. Albert and Siddhartha Chib. Bayesian analysis of binary and polychotomous response data. p. 0 The analysis illustrates how accounting for implementation heterogeneity and individual
compliance can provide insights beyond standard ITT analyses. Key words: Causal inference; Principal stratification; Complier average causal effect; Cluster randomized trials;
Noncompliance; Bayesian methods; Latent variable models; Interference. The trial evaluated the effects of a
personalized music intervention on agitated behaviors and medication use among NH residents with moderate to severe dementia. As in many pragmatic trials involving individuals with dementia, a substantial
proportion of participants assigned to the intervention did not comply with the intervention protocol (Allore
et al. Under monotonicity, exclusion restrictions, and individual-level randomization, instrumental variable (IV) methods can
be used to estimate the CACE (Angrist et al. Bayesian approaches have demonstrated favorable operating characteristics relative to classical IV estimators and provide a framework for assessing sensitivity to
identifying assumptions (Imbens and Rubin, 1997; Hirano et al. Principal stratification estimands
characterize effects within subpopulations defined by post-assignment variables such as treatment receipt. We define the complier average causal effect within latent implementation stratum k as
CACEk = E[Yij (1, 1, Si) − Yij (0, 0, Si) | Dij = 1, Si = k] , k = 1, . Naive alternatives such as as-treated
and per-protocol analyses, which condition on observed receipt, can lead to biased causal estimates because
they break the original randomization (Ellenberg, 1996; Matilde Sanchez and Chen, 2006; Goldberg and
Belitskaya-Levy, 2014). With binary compliance, an estimand that preserves the randomization structure is the complier average
causal effect (CACE), which describes the effect of an intervention among units that would comply with their
assigned treatment (Imbens and Rubin, 1997; Frangakis, 2002). However, in pragmatic cluster randomized trials where healthcare providers
are the units of randomization, noncompliance may arise at both the cluster and individual levels. The principal stratification framework provides a flexible approach for defining causal estimands that
account for post-treatment compliance (Frangakis and Rubin, 2002). Gutman
1 Introduction
Agitation and aggressive behaviors are common among cognitively impaired individuals residing in nursing
homes (NHs) (Zuidema et al. These behaviors can decrease quality of life and accelerate cognitive
decline (Garc´ıa-Alberca et al. Bayesian Latent Compliance Model for Cluster RCTs with One-Sided Noncompliance 7
on the mixture specification for Si
, and it is interpreted as a model-based summary of effect heterogeneity
across latent implementation strata. A marginal complier average causal effect is obtained by averaging over the distribution of Si
:
CACE = E[Yij (1, 1, Si) − Yij (0, 0, Si) | Dij = 1] . In the first stage of METRIcAL, no statistically significant effects of the personalized music intervention
on agitated behaviors were observed (McCreedy et al. One potential explanation for the lack of
statistical significance is limited adherence to the assigned intervention. Across all case studies, posterior inference is conducted using the Bayesian model described
in Section 4. The prior distributions for all model parameters are similar across simulation scenarios and
are specified in Appendix A2. In the first stage of METRIcAL, no statistically significant effects of the personalized music intervention
on agitated behaviors were observed (McCreedy et al. One potential explanation for the lack of
statistical significance is limited adherence to the assigned intervention. Across all case studies, posterior inference is conducted using the Bayesian model described
in Section 4. The prior distributions for all model parameters are similar across simulation scenarios and
are specified in Appendix A2. While ITT estimands preserve the benefits of randomization, they may not adequately
characterize the effects of intervention receipt or provide insight into how treatment effects vary across
settings with different compliance patterns (Hern´an and Robins, 2017). Naive alternatives such as as-treated
and per-protocol analyses, which condition on observed receipt, can lead to biased causal estimates because
they break the original randomization (Ellenberg, 1996; Matilde Sanchez and Chen, 2006; Goldberg and
Belitskaya-Levy, 2014). Pragmatic randomized controlled trials (PRCTs) are designed to estimate intervention effects under realworld conditions, where participants and providers may exhibit varying degrees of compliance (Patsopoulos,
2011). Analyses of PRCTs typically rely on the intention-to-treat (ITT) principle, whereby units are analyzed
according to their assigned treatment regardless of compliance (Sheiner and Rubin, 1995; Gupta, 2011;
Jo et al. While the 95% credible intervals
for the estimated effects include 0, the analysis illustrates how accounting for cluster-level implementation
heterogeneity and individual compliance can yield insights that are not apparent from ITT analyses alone. 2 Motivating Example: METRIcAL
METRIcAL is a PCRCT that examined the effects of a personalized music intervention on agitated behaviors
and antipsychotic use among nursing home (NH) residents with dementia. While ITT estimands preserve the benefits of randomization, they may not adequately
characterize the effects of intervention receipt or provide insight into how treatment effects vary across
settings with different compliance patterns (Hern´an and Robins, 2017). Naive alternatives such as as-treated
and per-protocol analyses, which condition on observed receipt, can lead to biased causal estimates because
they break the original randomization (Ellenberg, 1996; Matilde Sanchez and Chen, 2006; Goldberg and
Belitskaya-Levy, 2014). Because compliance is only observed in treated
clusters, the model imputes unobserved compliance behavior for clusters and individuals assigned to control. The framework enables estimation of finite- and super-population intent-to-treat (ITT) and CACE estimands,
both marginally and within latent implementation types.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Appendix 3 Prior distributions for METRIcAL application
Latent implementation strata are modeled as Si ∼ Categorical(π1, π2), (π1, π2) ∼ Diriclet(1, 1) Conditional on Si = k, the facility-level implementation measure (logit clinical proportion) and baseline facility characteristic follow Ci Zi Si = k ∼ MVN(µS k, Σ), k = 1, 2, with priors µS k ∼ MVN0 0, 1000 0 / 100 I2, Σ ∼ Inv-Wishart1/10 cubed 0 / I/1.5,5 Individual-level compliance is modeled using a multilevel probit regression: Dij Si = k, Xij ∼ Bernoulliprobit−1 µD k + X⊤ijα + ϕD i, with priors µD k ∼ N (0, 3 2), α ∼ MVN(0, 10 I), ϕD i ∼ N (0, τ squared D), τD ∼ Uniform(0, 4) Potential outcomes for change in CMAI are modeled as Yij (1, Dij, Si) ∼ N µY k + X⊤ijβ0 + Dij X⊤ijβ1 + δ1 k + ϕY i, σ2, Yij (0, 0, Si) ∼ N µY k + X⊤ijβ0 + Dij δ0 k + ϕY i, σ2, if Si = k, with priors µY k ∼ N (0, 10 2), δ1 k ∼ N (0, 10 2), δ0 k ∼ N (0, 4 2), β0 ∼ MVN(0, 10 squared I), β1 ∼ MVN(0, 1. p. 8 , K,
Yij1, Dij , Si
| Si = k, Dij , Xij ∼ N µ
Y
k + X⊤
ijβ0,k + Dij δ1,k + X⊤
ijβ1,k
+ ϕ
Y
i
, σ2
k
, (4. 11)
Yij (0, 0, Si) | Si = k, Dij , Xij ∼ N µ
Y
k + X⊤
ijβ0,k + Dij δ0,k + ϕ
Y
i
, σ2
k
.Where the paper says it
Appendix A6 Super-population Estimand Derivation as a Function of Model Parameters
ITTsp k = E[Yij (1, Dij, Si) − Yij (0, 0, Si) Si = k] = 1/n XijpijkXijβ1 k + δk CACEsp k = E[Yij (1, Dij, Si) − Yij (0, 0, Si) Dij = 1, Si = k] = 1/n P i∈Ikni Xijβ1 kpijk + δk These expressions show how the super-population estimands are derived as functions of the model parameters under alternative choices of the super-population covariate distribution
Table A1 Super-population estimands expressed as functions of the model parameters under alternative choices of the super-population covariate distribution
ITTsp k = 1/n Xijpijk (Xijβ1,k + δk) CACEsp k = P kπk P ij pijk P ijP Xijβ1,kpijk + δk
Table A4 Gelman-Rubin statistic for the 3 chains of posterior samples for each estimand presented in the data analysis in Section 6
The Gelman-Rubin statistic for each estimand is close to 1, and indicates adequate model fit
Fig. A1 Overlaid line plots of the 4 MCMC chains of the samples used for posterior inference in Section 6. p. 8 Estimation procedures 1a and 1b target ITT effects, while procedures 2a and 2b target CACEs. All procedures rely on posterior draws from the model described in Section 4 using Markov chain Monte Carlo
(MCMC) algroithms.Where the paper says it
Fig. A1 illustrates the posterior predictive distribution of the clinical proportion and average daily census under the fitted mixture model alongside the observed facility-level data The posterior predictive draws indicate separation between the two latent implementation strata, with stratum 1 facilities characterized by smaller average daily census and higher clinical proportions relative to stratum 2 facilities This pattern suggests systematic differences in how facilities operationalized the personalized music intervention, with smaller facilities more likely to target a larger fraction of residents’ behaviors
Table A3 Mean and standard deviation of facility characteristics and facility compliance metrics used in each model pair
Outcome ITT ITT1 ITT2 ITT1 − ITT2 CACE CACE1 CACE2 CACE1 − CACE2 CMAI 1.019 1.015 1.007 1.003 1.018
CMAI Antipsychotics
Outcome ITT ITTsp: -2.9 percentage points; 95% CrI: [-6.1, 0.4] CACEsp: -7.8 percentage points; 95% CrI: [-16.9, 0.7]
Table A5 Posterior distribution of super-population estimands for antipsychotic use at 4-month follow-up
The posterior mean of the ITT estimand suggests lower antipsychotic use among residents in nursing homes assigned to the personalized music intervention relative to standard care (ITTsp: -2.9 percentage points; 95% CrI: [-6.1, 0.4]) The corresponding complier average causal effect indicates a larger estimated reduction among residents who complied with the intervention (CACEsp: -7. p. 1 Naive alternatives such as as-treated
and per-protocol analyses, which condition on observed receipt, can lead to biased causal estimates because
they break the original randomization (Ellenberg, 1996; Matilde Sanchez and Chen, 2006; Goldberg and
Belitskaya-Levy, 2014). With binary compliance, an estimand that preserves the randomization structure is the complier average
causal effect (CACE), which describes the effect of an intervention among units that would comply with their
assigned treatment (Imbens and Rubin, 1997; Frangakis, 2002).Where the paper says it
Table A2 Mean and standard deviation of patient level characteristics derived from the MDS
Age, y, mean (SD) 80.5 (12.1) 80.2 (11.8) 80.8 (12.3) CMAI, mean (SD) 50. p. 3 , 2019) (Appendix Table A2). Study outcome measures included the Cohen–Mansfield Agitation Inventory (CMAI) and antipsychotic
use, which were collected at baseline and four months after study initiation for each participant. , 2019) (Appendix Table A2). Study outcome measures included the Cohen–Mansfield Agitation Inventory (CMAI) and antipsychotic
use, which were collected at baseline and four months after study initiation for each participant. McCreedy and R. Gutman
(ADL) (Graf, 2008), measures of cognitive status such as the Cognitive Function Scale (CFS) (Thomas
et al. McCreedy and R. Gutman
(ADL) (Graf, 2008), measures of cognitive status such as the Cognitive Function Scale (CFS) (Thomas
et al. McCreedy and R. Gutman
(ADL) (Graf, 2008), measures of cognitive status such as the Cognitive Function Scale (CFS) (Thomas
et al.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Age, y, mean (SD) 80.5 (12.1) 80.2 (11.8) 80.8 (12.3) CMAI, mean (SD) 50.4 (19.4) 51.0 (20.7) 49.9 (18.1) ADL, mean (SD) 17.0 (6.1) 17.17 (5.9)
CFS, mean (SD) 2.7 (0.9) 2.8 (0.9)
Outcome ITT ITT1 ITT2 ITT1 − ITT2 CACE CACE1 CACE2 CACE1 − CACE2 CMAI 1.019
Antipsychotics
Outcome ITTsp: -2.9 percentage points; 95% CrI: [-6.1, 0.
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 AI can perform causal inference in pragmatic cluster randomized trials by using a Bayesian framework that jointly models latent implementation heterogeneity and individual compliance as a latent variable Si. This allows for estimation of CACEk, which characterizes effects
within latent implementation strata.
-
The system can estimate finite-sample ITT and CACE estimands, both marginally and within latent implementation types. This moves beyond standard ITT analyses to characterize
how effects vary across these sources of compliance.
-
The AI can impute unobserved compliance behavior for clusters and individuals assigned to control based on the latent implementation strata model. This addresses the challenge that
compliance is only observed in treated clusters
by using a joint model linking implementation types to individual outcomes. -
It can estimate super-population causal estimands, which are defined as functions of the model parameters and assumed distributions of covariates and latent variables, integrating over uncertainty in the underlying data-generating process. This provides
super-population causal estimands
that account for variability beyond the realized sample. -
The system can assess model fit using posterior predictive checks based on discrepancy measures such as SI study k, NO study k, and SINO study k. This helps ensure that
the fitted model can reproduce key features of the observed data relevant to the ITT and CACE estimands.
Abstract
In pragmatic cluster randomized controlled trials (PCRCTs), healthcare providers are randomized, but intervention delivery and patient receipt may vary after assignment. In many PCRCTs, cluster-level implementation is measured using continuous implementation metrics, while individual compliance is recorded as a binary indicator. Standard complier average causal effect (CACE) estimands focus on individual-level compliance and do not account for heterogeneity in implementation across clusters. When intervention uptake is shaped by both cluster- and individual-level processes, it is of scientific interest to characterize how effects vary across these sources of compliance. We propose a Bayesian framework for PCRCTs with one-sided binary noncompliance at the individual level and one-sided partial compliance at the cluster level. The method uses a latent mixture model to summarize heterogeneity in cluster-level implementation based on baseline characteristics and treatment-arm implementation measures, and links these latent implementation strata to individual compliance and outcomes through a joint model. Because individual compliance and treatment-defined implementation measures are observed only under treatment assignment, the model imputes these quantities for control clusters and individuals assigned to control. The framework enables estimation of finite- and super-population intent-to-treat (ITT) and CACE estimands, both marginally and within latent implementation strata. We apply the method to the METRIcAL trial, a pragmatic cluster randomized study evaluating a personalized music intervention for nursing home residents with dementia. The analysis illustrates how accounting for implementation heterogeneity and individual compliance can provide insights beyond standard ITT analyses.
Related papers
- Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization
- Strategic Causal Policy Learning with Manipulable Eligibility
- BayesJudge: Uncertainty-Aware Bayesian Meta-Evaluation of Human and LLM Judgments
- Automating structural reliability analysis with a multi-agent large language model framework
- Improving Minority Population Sampling with BISG Probabilities: Evidence from a Survey of Jewish Americans
- When Do Sports Leagues Stop Being Competitive? Bayesian Detection of Competitive-Balance Eras from Match-Level Data