Estimating the Complier Average Causal Effect in a Meta-Analysis of Randomized Clinical Trials With Binary Outcomes Accounting for Noncompliance: A Generalized Linear Latent and Mixed Model Approach

Ting Zhou, Jincheng Zhou, James S. Hodges, Lifeng Lin, Yong Chen, Stephen R. Cole, Haitao ChuAmerican 2021 · DOI

Abstract

Noncompliance is a pervasive challenge in randomized clinical trials (RCTs) that can bias standard intention-to-treat (ITT) analyses. While ITT maintains the benefits of randomization, it often underestimates the biological efficacy of a treatment. This paper proposes a frequentist framework using Generalized Linear Latent and Mixed Models (GLLAMM) to estimate the Complier Average Causal Effect (CACE) in meta-analyses. By accounting for between-study heterogeneity through random effects and utilizing a one-step maximum likelihood estimation, the approach provides a robust alternative to complex Bayesian models for evaluating treatment effects across multiple trials.

Introduction: The Challenge of Noncompliance

In clinical research, participants frequently fail to adhere to their assigned treatment protocols. Standard ITT analysis compares groups based on their initial assignment, which provides a conservative estimate of the policy effect but obscures the actual treatment benefit for those who receive it. The Complier Average Causal Effect (CACE) addresses this by estimating the effect specifically within the latent subpopulation of "compliers"—individuals who would follow their assignment regardless of the group to which they are randomized.

Performing a meta-analysis of CACE estimates is complicated by heterogeneity across studies, including varying compliance rates and baseline risks. Traditional two-step methods, which calculate effects for each study before averaging them, can be biased and struggle with "boundary cases" where compliance is 0% or 100%.

Methodology: The GLLAMM Framework

The authors propose a "one-step" approach using Generalized Linear Latent and Mixed Models (GLLAMM), implemented via the SAS NLMIXED procedure. This method offers several advantages:

  • Latent Class Modeling: It categorizes participants into unobserved groups: always-takers, never-takers, and compliers (assuming the absence of "defiers").
  • Exact Likelihood: By using the exact binomial likelihood of raw data rather than summary statistics, the model reduces bias and improves the precision of confidence intervals.
  • Random Effects: The model incorporates random effects to account for variation in compliance rates and outcome probabilities across different studies.
  • Model Selection: Researchers can use the Bayesian Information Criterion (BIC) and Akaike Information Criterion (AIC) to determine the optimal random-effects structure for a given dataset.
Comparison of random-effects models using BIC and AIC criteria.

Case Study: Epidural Analgesia and Cesarean Delivery

The methodology was applied to a meta-analysis of 10 RCTs investigating whether epidural analgesia during labor increases the risk of cesarean delivery. This dataset is characterized by high and highly variable noncompliance rates.

Traditional explicit formulas for CACE often fail in such datasets when studies have zero compliance in one arm, as they cannot calculate standard errors (e.g., the Clark and Nikkola studies in the dataset). The GLLAMM approach handles these boundary issues gracefully through maximum likelihood estimation.

Results

The GLLAMM meta-analysis yielded a CACE estimate of 0.022 (95% CI: -0.004, 0.048). While the point estimate is higher than the ITT estimate (0.010), the confidence interval includes zero, suggesting insufficient evidence to conclude that epidurals significantly increase cesarean risk among compliers.

Forest plot comparing ITT, BHM, and ML estimates for individual studies and overall meta-analysis effects.

Simulation studies further validated the model, showing that BIC was highly effective at identifying the correct model structure, especially as the number of studies increased. The one-step GLLAMM approach generally produced narrower confidence intervals than two-step methods, demonstrating superior statistical efficiency.

Discussion and Limitations

The GLLAMM approach provides a powerful frequentist alternative to Bayesian Hierarchical Models (BHM), making advanced CACE meta-analysis more accessible to researchers using standard statistical software like SAS. However, the model currently relies on several key assumptions:

  1. Binary Compliance: It assumes compliance is all-or-nothing and does not yet account for partial adherence.
  2. Exclusion Restriction: It assumes the treatment assignment only affects the outcome through the treatment actually received.
  3. Sample Size: Like all maximum likelihood methods, it may exhibit bias when the number of studies is very small (e.g., fewer than five).

Future work may extend this framework to accommodate partial compliance and network meta-analyses involving multiple treatment arms.

Summary updated

Abstract

Noncompliance, a common problem in randomized clinical trials (RCTs), can bias estimation of the effect of treatment receipt using a standard intention-to-treat analysis. The complier average causal effect (CACE) measures the effect of an intervention in the latent subpopulation that would comply with their assigned treatment. Although several methods have been developed to estimate the CACE in analyzing a single RCT, methods for estimating the CACE in a meta-analysis of RCTs with noncompliance await further development. This article reviews the assumptions needed to estimate the CACE in a single RCT and proposes a frequentist alternative for estimating the CACE in a meta-analysis, using a generalized linear latent and mixed model with SAS software (SAS Institute, Inc.). The method accounts for between-study heterogeneity using random effects. We implement the methods and describe an illustrative example of a meta-analysis of 10 RCTs evaluating the effect of receiving epidural analgesia in labor on cesarean delivery, where noncompliance varies dramatically between studies. Simulation studies are used to evaluate the performance of the proposed method.
BibTeX
@inproceedings{zhou2021estimating,
  author = {Ting Zhou and Jincheng Zhou and James S. Hodges and Lifeng Lin and Yong Chen and Stephen R. Cole and Haitao Chu},
  title = {{Estimating the Complier Average Causal Effect in a Meta-Analysis of Randomized Clinical Trials With Binary Outcomes Accounting for Noncompliance: A Generalized Linear Latent and Mixed Model Approach}},
  booktitle = {American},
  year = {2021},
  doi = {10.1093/aje/kwab238},
}
Figure