Evaluation of instrumental variable method using Cox proportional hazard model in epidemiological studies

Area of Science:

• Epidemiology
• Biostatistics

Background:

• Instrumental variable (IV) methods are employed with Cox proportional hazard (PH) models to assess treatment effects in survival data.
• The precise performance and validity of these IV methods in epidemiological survival analyses require further investigation.

Purpose of the Study:

• To evaluate the performance and validity of instrumental variable (IV) methods when applied with a Cox proportional hazard (PH) model for survival data analysis.
• To assess how confounding factors influence the accuracy of treatment effect estimates derived from IV analyses.

Main Methods:

• Utilized two-stage instrumental variable (IV) models within a Cox PH framework.
• Conducted simulations with varying confounder strengths and baseline hazards, employing a sample size of 10,000 and 1,000 replications per simulation.
• Evaluated true treatment effects (Hazard Ratios) of 3, 2, and 1 (null effect), calculating 95% confidence intervals from simulation estimates.

Main Results:

• When observed confounders were not included in IV models and confounding was moderate, two-stage IV estimates closely approximated true treatment effects.
• Effect estimates diverged from true values when observed confounders were incorporated into the IV models.
• Estimates from unadjusted and adjusted two-stage IV models were close to the true value for a null treatment effect (Hazard Ratio = 1).

Conclusions:

• Treatment effect estimates from IV analyses using the Cox PH model are valid under specific conditions.
• Validity is maintained when using unadjusted IV models with moderate confounding effects.
• Alternatively, IV estimates remain valid if the treatment under investigation does not significantly impact the survival outcome.