Adjustment for time-dependent unmeasured confounders in marginal structural Cox models using validation sample data.

Area of Science:

• Epidemiology
• Biostatistics
• Observational Studies

Background:

• Large observational databases for drug safety studies often lack crucial confounder information, leading to unmeasured confounding bias.
• Existing methods using validation samples struggle with unmeasured time-varying confounders that can also mediate treatment effects.

Purpose of the Study:

• To propose and compare novel methods for controlling unmeasured time-varying confounders in marginal structural Cox models using validation samples.
• To enhance the accuracy of drug safety assessments by addressing limitations in large observational datasets.

Main Methods:

• Developed and evaluated regression calibration and multiple imputation techniques for propensity score adjustment.
• Proposed two martingale residual-based methods to correct inverse probability of treatment weights for imputed unmeasured confounders.
• Utilized simulation studies to compare the performance of proposed methods against naive approaches.

Main Results:

• Martingale residual-based methods demonstrated a systematic reduction in confounding bias compared to naive methods.
• Multiple imputation incorporating martingale residuals achieved the best overall accuracy on average.
• Applied martingale residual-based imputation to investigate drug-induced hypoglycemia risk in diabetic patients.

Conclusions:

• The proposed martingale residual-based imputation methods effectively address unmeasured time-varying confounding in drug safety studies.
• These methods improve the reliability of observational studies by leveraging validation sample data.
• The findings have implications for accurately assessing drug risks, particularly in complex patient populations like diabetics.