Clarifying selection bias in cluster randomized trials.

Area of Science:

• Causal inference
• Biostatistics
• Clinical trial design

Background:

• Cluster randomized trials (CRTs) often suffer from post-randomization selection bias due to unblinded recruitment.
• This bias leads to systematic differences in patient characteristics between treatment and control groups.
• Such imbalances can compromise the validity of standard analyses.

Purpose of the Study:

• To rigorously define causal estimands in CRTs with selection bias.
• To identify conditions for valid estimation using standard covariate adjustment.
• To explore data and assumptions needed when standard methods fail.

Main Methods:

• Utilized the principal stratification framework to define two average treatment effect (ATE) estimands: for the overall and recruited populations.
• Derived analytical formulas for these estimands based on principal-stratum-specific causal effects.
• Conducted simulation studies to evaluate multivariable regression adjustment under various selection bias scenarios.

Main Results:

• When treatment effects vary across principal strata, ATEs for overall and recruited populations differ.
• Naïve intention-to-treat analyses in recruited samples yield biased estimates for both ATEs.
• Estimating the recruited population ATE requires homogeneous treatment effects; the overall ATE is often not estimable without additional data.

Conclusions:

• Improved analysis strategies are needed for CRTs with post-randomization selection bias.
• Explicitly defining target populations and adopting appropriate design/estimation strategies is crucial.
• Investigators should assess treatment effect heterogeneity and collect data on non-recruited individuals.