Causal inference methods to assess safety upper bounds in randomized trials with noncompliance

Area of Science:

• Biostatistics
• Clinical Trials
• Pharmacovigilance

Background:

• Noncompliance with treatment assignment in randomized trials complicates causal inference.
• Intent-to-treat (ITT) analysis provides unbiased estimates of treatment assignment effects but may underestimate actual treatment effects.
• ITT analysis can also lead to underestimation of the upper confidence limit for treatment effects.

Purpose of the Study:

• To compare hazard ratio and confidence interval upper bound estimates from causal inference methods accounting for noncompliance with those from ITT analysis.
• To evaluate the performance of causal inference methods versus ITT in cardiovascular safety trials for diabetes drugs.
• To assess the impact of noncompliance on bias and variance in estimating treatment effects.

Main Methods:

• Simulations were conducted using parameters relevant to cardiovascular safety trials of diabetes drugs.
• A hypothetical trial of 10,000 subjects (1:1 randomization) was simulated with varying noncompliance rates (discontinuation, crossover).
• Assumed true hazard ratios were 0.9, 1.0, and 1.3, with evaluations of non-complier risks. Causal survival analysis and ITT methods were applied.

Main Results:

• Causal analysis demonstrated minimal bias in estimating the true hazard ratio across most settings.
• ITT analysis was unbiased only when the true hazard ratio was 1; otherwise, it underestimated both benefit and harm.
• When ITT upper bounds were ≥1.3, causal analysis upper bounds were also ≥1.3 in nearly all simulations. When ITT upper bounds were <1.3 and true hazard ratio was 1, causal analysis upper bounds reached ≥1.3 in up to 66% of simulations.

Conclusions:

• Causal survival analysis is superior to ITT for estimating true hazard ratios with noncompliance.
• The large variance of causal analysis necessitates careful consideration for safety upper bound exclusion, particularly when the true hazard ratio is 1.
• These simulations offer a reference for bias-variance trade-offs in noncompliance management for diabetes drug safety trials, highlighting the need for further research in causal inference for safety upper bounds.