Instrumental variable approaches for estimating time-varying treatment effects in comparative effectiveness research

Area of Science:

• Epidemiology
• Biostatistics
• Health Services Research

Background:

• Large-scale longitudinal data enable causal inference for health interventions.
• Instrumental Variable (IV) methods can mitigate confounding bias.
• Limited IV approaches exist for time-varying instruments and confounders.

Purpose of the Study:

• To evaluate two instrumental variable approaches for time-varying treatments.
• To compare g-estimation with inverse probability weighting in simulations and a real-world case study.
• To assess the impact of Adalimumab on Rheumatoid Arthritis quality of life using these methods.

Main Methods:

• Extended g-estimation for time-fixed IVs and compared with inverse probability weighting for time-varying IVs.
• Conducted a simulation study to assess performance under various scenarios.
• Applied methods to a US National Databank for Rheumatic Diseases cohort, using physician preference as an instrument for Adalimumab.

Main Results:

• G-estimation yielded unbiased and precise treatment effect estimates, robust to weak IVs and confounding.
• Inverse probability weighting performed adequately with strong time-varying IVs but degraded with weak IVs.
• Both methods indicated no quality of life improvement from sustained Adalimumab versus other biologics; g-estimation provided narrower confidence intervals.

Conclusions:

• The IV-based g-estimation approach is reliable for time-varying treatments with valid time-varying IVs.
• Inverse probability weighting is an alternative but requires strong IVs, which are uncommon in practice.
• G-estimation is recommended for robust causal inference in complex longitudinal health studies.