Estimating causal effects from a randomized clinical trial when noncompliance is measured with error

Area of Science:

• Biostatistics
• Clinical Trials
• Epidemiology

Background:

• Noncompliance in randomized trials complicates the interpretation of intervention effects.
• Accurate estimation of the causal effect, assuming perfect adherence, is scientifically crucial.
• Existing methods like inverse probability of compliance weighted (IPCW) estimators fail when compliance is self-reported or imperfectly measured.

Purpose of the Study:

• To develop a novel statistical approach for estimating causal treatment effects when compliance is unknown or measured with error.
• To address limitations of current methods that assume error-free compliance data.
• To improve the accuracy and efficiency of causal effect estimation in the presence of noncompliance.

Main Methods:

• Utilized biomarker data (e.g., nicotine levels) to model compliance probability using mixture distributions.
• Developed a new re-weighting estimator incorporating the probability of compliance given observed data and confounders.
• Assessed the consistency, asymptotic normality, and efficiency of the proposed estimator.

Main Results:

• The proposed method directly estimates compliance probability from data, even with unknown compliance status.
• The new estimator is consistent and asymptotically normal.
• Simulations show the proposed approach offers improved efficiency and reduced bias compared to standard IPCW and ad hoc methods.
• The method was successfully applied to a randomized trial of very low nicotine content (VLNC) cigarettes.

Conclusions:

• The developed statistical method effectively estimates causal effects in randomized trials with imperfect compliance.
• This approach provides a more robust and efficient alternative to existing methods when compliance is difficult to ascertain accurately.
• The findings have significant implications for interpreting clinical trial data, particularly in public health interventions like smoking cessation.