Instrumental variable estimation in a survival context.

Area of Science:

• Biostatistics
• Epidemiology
• Causal Inference

Background:

• Unobserved confounding can bias causal effect estimates in nonrandomized studies.
• Instrumental variable (IV) methods can mitigate confounding bias without observing all confounders.
• Existing IV methods are less developed for censored survival outcomes.

Purpose of the Study:

• To develop and present novel instrumental variable (IV) approaches for causal inference in survival regression analysis.
• To address the limitations of current IV methods in the context of censored survival data.
• To provide valid estimation strategies for treatment effects under confounding.

Main Methods:

• Developed two IV methods for regression analysis under an additive hazards model.
• Method 1: A two-stage regression approach analogous to two-stage least squares.
• Method 2: A control function approach incorporating residuals from a first-stage exposure regression.

Main Results:

• The proposed methods provide valid estimates of causal treatment effects in survival analysis.
• Demonstrated application using Mendelian randomization to assess diabetes' effect on mortality.
• Established analogous strategies applicable to proportional hazards models for rare outcomes.

Conclusions:

• The novel IV methods effectively address confounding bias in survival data.
• These approaches enhance causal inference capabilities for censored outcomes.
• The study provides practical tools for analyzing complex epidemiological and biostatistical data.