Instrumental variable estimation of the marginal structural Cox model for time-varying treatments

Area of Science:

• Causal inference
• Longitudinal data analysis
• Biostatistics

Background:

• Marginal structural models (MSMs) are used for time-varying treatments with confounding.
• Sequential randomization assumption prevents unmeasured confounding but is often violated.
• Marginal structural Cox models (MSCMs) are popular for censored time-to-event outcomes.

Purpose of the Study:

• To develop methods for identifying MSCM parameters under unmeasured confounding.
• To leverage time-varying instrumental variables (TVIVs) when sequential randomization fails.
• To extend existing causal inference techniques to complex longitudinal settings.

Main Methods:

• Utilized a time-varying instrumental variable approach to address unmeasured confounding.
• Established identification conditions for MSCM parameters, generalizing prior work.
• Developed weighted estimating equations for consistent and asymptotically normal estimators.

Main Results:

• Sufficient conditions for identification of MSCM parameters using TVIVs were established.
• The proposed instrumental variable condition restricts confounder-instrument interaction.
• The method extends inverse probability of treatment weighted estimation to the TVIV setting.

Conclusions:

• The developed TVIV method enables causal effect estimation in MSCMs with unmeasured confounding.
• This approach provides a robust framework for analyzing complex longitudinal data.
• The methodology was validated through simulations and applied to HIV incidence data.