Instrumental variable approaches for estimating complier average causal\n effects on bivariate outcomes in randomised trials with non-compliance

Karla Díaz-Ordaz, A Franchini, R. A. F. Grieve

Abstract

In randomized controlled trials (RCTs) with treatment non-compliance, instrumental variable (IV) approaches are essential for estimating the Complier Average Causal Effect (CACE). This paper extends these methods to cost-effectiveness analyses (CEA), where the inherent correlation between cost and health outcomes must be addressed. We propose and evaluate two primary methods: a Bayesian Full Likelihood (BFL) approach and a Three-Stage Least Squares (3sls) method. Using simulations and the REFLUX study—a trial comparing surgery to medical management for gastro-oesophageal reflux disease—we demonstrate that failing to account for outcome correlation and treatment endogeneity leads to biased estimates and poor coverage. Both BFL and 3sls provide robust, unbiased alternatives for policy-relevant causal inference.

Introduction

Randomized Controlled Trials (RCTs) are the gold standard for evaluating clinical interventions. However, the integrity of randomization is often challenged by non-compliance, where participants do not receive the treatment to which they were assigned. While Intention-to-Treat (ITT) analysis maintains the benefits of randomization by comparing groups as assigned, it often underestimates the structural effect of the treatment itself. Conversely, Per Protocol (PP) analysis, which excludes non-compliers, frequently introduces selection bias because those who switch treatments often differ systematically from those who do not.

The Complier Average Causal Effect (CACE) offers a middle ground, identifying the treatment effect specifically among "compliers"—those who would follow their assignment regardless of the group. In the context of Cost-Effectiveness Analysis (CEA), estimating CACE is complicated by the bivariate nature of the outcomes: costs and health effects (e.g., Quality-Adjusted Life Years). Because these outcomes are typically correlated, univariate IV models can produce misleading results, particularly regarding the Incremental Net Benefit (INB), which is the primary metric for policy decisions.

Methodological Framework

This study proposes two robust IV strategies designed to handle bivariate outcomes and endogeneity simultaneously:

1. Three-Stage Least Squares (3sls)

The 3sls method is a frequentist approach that extends the standard Two-Stage Least Squares (2sls) framework. It treats cost and health outcomes as a system of simultaneous equations. By estimating these equations jointly, 3sls accounts for the correlation between the error terms of the cost and health models, ensuring that the uncertainty surrounding the joint distribution is correctly captured.

2. Bayesian Full Likelihood (BFL)

The BFL approach utilizes a joint model where treatment assignment acts as an instrument for treatment received. It simultaneously models the relationship between the instrument, the treatment, and the bivariate outcomes. This framework allows for "Bayesian feedback," where the uncertainty in the compliance model propagates through to the final causal estimates of cost and health.

3. Benchmarking Naive Estimators

To highlight the necessity of these complex models, the authors compared them against Naive Bayesian estimators (uBN/uBGN). These simpler models ignore the endogeneity of treatment received, effectively assuming that the reasons for non-compliance are unrelated to the outcomes—an assumption rarely met in clinical practice.

Evaluation and Results

The methods were tested using a comprehensive factorial simulation and a real-world application to the REFLUX study.

Simulation Study

The simulation varied sample sizes ($n=100$ to $1000$), non-compliance rates (30% to 70%), and cost distributions (Normal, Gamma, and Inverse Gaussian). The results indicated:

  • Bias and Coverage: Both 3sls and BFL provided unbiased estimates with excellent confidence/credible interval coverage across most scenarios.
  • The Naive Trap: Naive models exhibited significant bias, particularly when non-compliance was high (70%), proving that ignoring endogeneity is catastrophic for causal inference.
  • Distributional Robustness: 3sls proved highly robust to skewed cost data. BFL performed well but showed sensitivity if the cost distribution was significantly misspecified (e.g., using a Normal distribution for Inverse Gaussian data).
  • Small Sample Performance: In small datasets ($n=100$), BFL tended toward over-coverage (wider intervals), while 3sls remained more precise.

The REFLUX Case Study

The REFLUX trial compared laparoscopic surgery to medical management for Gastro-Oesophageal Reflux Disease (GORD). With a 28% non-compliance rate in the surgery arm, the CACE analysis revealed that surgery was more cost-effective for compliers than the ITT analysis suggested. Specifically, the CACE for surgery costs was £2030 compared to the ITT estimate of £1103, reflecting the true resource intensity of the surgical procedure for those who actually underwent it.

Discussion and Conclusion

The primary insight of this research is that correlation is a first-order concern in CEA. Estimating cost and health effects through separate IV models leads to incorrect uncertainty measures for the INB, potentially resulting in flawed policy recommendations.

While 3sls and BFL are superior to traditional methods, they rely on the standard IV assumptions of the Exclusion Restriction (assignment only affects outcomes via treatment) and Monotonicity (no "defiers"). Furthermore, the BFL model is computationally intensive and requires careful prior specification. Nevertheless, for researchers seeking to provide policy-makers with accurate estimates of treatment effects in the presence of non-compliance, these bivariate IV approaches represent a significant advancement in health economic methodology.

Summary updated

Abstract

In Randomised Controlled Trials (RCT) with treatment non-compliance,\ninstrumental variable approaches are used to estimate complier average causal\neffects. We extend these approaches to cost-effectiveness analyses, where\nmethods need to recognise the correlation between cost and health outcomes.\n We propose a Bayesian full likelihood (BFL) approach, which jointly models\nthe effects of random assignment on treatment received and the outcomes, and a\nthree-stage least squares (3sls) method, which acknowledges the correlation\nbetween the endpoints, and the endogeneity of the treatment received.\n This investigation is motivated by the REFLUX study, which exemplifies the\nsetting where compliance differs between the RCT and routine practice. A\nsimulation is used to compare the methods performance.\n We find that failure to model the correlation between the outcomes and\ntreatment received correctly can result in poor CI coverage and biased\nestimates. By contrast, BFL and 3sls methods provide unbiased estimates with\ngood coverage.\n
BibTeX
@article{diazordaz2017instrumental,
  author = {Karla Díaz-Ordaz and A Franchini and R. A. F. Grieve},
  title = {{Instrumental variable approaches for estimating complier average causal\textbackslash\{\}n effects on bivariate outcomes in randomised trials with non-compliance}},
  journal = {Journal of the Royal Statistical Society Series A (Statistics in Society)},
  year = {2017},
  doi = {10.1111/rssa.12294},
}
Figure