Testing causal effects in observational survival data using propensity score matching design

Area of Science:

• Biostatistics
• Epidemiology
• Causal Inference

Background:

• Observational studies frequently involve time-to-event data, which are susceptible to observed and unobserved confounding biases.
• Existing survival analysis methods like Cox PH models rely on proportional hazards assumptions and offer limited approaches for hidden bias.
• Addressing confounding is crucial for accurate causal inference in survival analysis.

Purpose of the Study:

• To propose a novel strategy for testing survival function differences in observational studies using matching designs.
• To develop a sensitivity analysis for assessing the impact of unmeasured confounding on causal survival analysis findings.
• To evaluate the performance of the proposed methods through simulations and a real-world chronic liver disease cohort.

Main Methods:

• Application of the paired Prentice-Wilcoxon (PPW) test and its modified version on propensity score matched data.
• Development of a sensitivity analysis framework based on matched pairs to quantify the impact of unmeasured confounders.
• Simulation studies to compare the power of PPW-type tests against traditional methods when proportional hazards assumptions are violated.

Main Results:

• The PPW-type test demonstrates higher statistical power in scenarios where the proportional hazards assumption is not met.
• Initial analysis of chronic liver disease data suggested a significant treatment effect.
• Sensitivity analysis revealed that the observed treatment effect was not robust and became nonsignificant under potential unmeasured confounding.

Conclusions:

• The proposed PPW-based approach offers a valuable tool for causal survival analysis in observational studies, particularly when proportional hazards assumptions are questionable.
• Sensitivity analysis is essential for evaluating the robustness of findings in the presence of potential unmeasured confounding.
• The study highlights the importance of considering hidden biases to avoid spurious conclusions in observational research.