Cluster randomized trials (CRTs) examine the effectiveness of interventions offered to groups of patients, such as patients receiving care at one clinic. However, intermediate factors that occur after researchers have assigned treatment groups but before measuring treatment outcomes may lead to post-randomization confounding in estimates of intervention effectiveness. Current methods for designing and analyzing CRTs do not adequately address post-randomization confounding and do not capture heterogeneous treatment effect (HTE).
To develop and apply statistical methods and software for estimating treatment effects in CRTs in the presence of post-randomization confounding
|Theoretical development; empirical analysis
|Data Sources and Data Sets
Data from two CRTs:
- The PROUD trial tested the effectiveness of the PROUD intervention, which used medication to treat opioid use disorder within primary care settings to reduce acute care utilization
- ADAPTABLE compared risk of a major cardiovascular event with a high versus a low dose of aspirin taken preventively
- Develop formulas for estimating causal effects
- Develop methods to estimate time-to-event outcomes in CRTs
- Apply methods to data from two CRTs
|Average and heterogeneous causal effects
The research team developed principal stratification methods to estimate causal effects for different patient groups created based on intermediate factors. The separate effects represent HTE in the population. The methods addressed three types of post-randomization confounding:
- Identification bias. The intervention affects the identification of eligible patients, such as when diagnosis rates increase at clinics assigned to one treatment, resulting in more eligible patients.
- Recruitment bias. The intervention affects recruitment, such as patients selecting clinics because they offer a specific treatment.
- Noncompliance bias. Patients decide to stop their assigned treatment or switch to a clinic assigned to a different treatment.
The research team first developed methods for estimating treatment effectiveness in the presence of identification or recruitment bias. The team applied the methods to data from the PRimary Care Opioid Use Disorders Treatment (PROUD) trial, which examined acute care utilization among patients diagnosed with opioid use disorder.
Next, the research team developed one Bayesian method and one weighting method to address noncompliance bias in estimating time-to-event outcomes in CRTs. The team applied existing and new methods to data from Aspirin Dosing: A Patient-Centric Trial Assessing Benefits and Long-Term Effectiveness (ADAPTABLE), which compared risk for a major adverse cardiovascular event with varying aspirin doses.
The research team developed an R software package, PStrata, to support principal stratification analysis for CRTs.
A physician, a cardiologist, and a statistician provided input during the study.
The principal stratification methods addressed post-randomization confounding and identified HTE in the PROUD and ADAPTABLE trials.
Using the new methods to address identification or recruitment bias, the research team determined that the PROUD intervention had a nonsignificant, 7% increase in the probability of receiving acute care, indicating possible identification or recruitment bias.
Using the standard intention-to-event method to address noncompliance bias with data from ADAPTABLE, the research team found no clinical difference in the risk of a major adverse cardiovascular event among patients taking 81 milligrams (mg) of aspirin versus those taking 325 mg. Using the new methods, they found a significant reduction in risk for a major adverse cardiovascular event with 81 mg of aspirin among patients who adhered to their assigned dosage.
The research team tested the methods using cross-sectional CRTs; they may not work for longitudinal CRTs.
Conclusions and Relevance
The statistical methods and R package from this study can help address post-randomization confounding and measure HTE in CRTs.
Future Research Needs
Future research could extend the methods to longitudinal CRTs.