Mendelian randomization mixed-scale treatment effect robust identification and estimation for causal inference.

Area of Science:

• Epidemiology
• Biostatistics
• Genetic Epidemiology

Background:

• Standard Mendelian randomization (MR) relies on instrumental variables (IVs) that must be independent of confounders and have no pleiotropic effects.
• Violations of these assumptions, such as confounding or horizontal pleiotropy, can bias causal effect estimates.

Purpose of the Study:

• To develop novel identification conditions for causal effects in the presence of unmeasured confounding using potentially invalid instrumental variables.
• To introduce the Mendelian randomization mixed-scale treatment effect robust identification (MR MiSTERI) approach.

Main Methods:

• MR MiSTERI leverages an invalid IV by assuming the treatment effect is constant on the additive scale and confounding bias is constant on the odds ratio scale.
• A novel assumption of heteroskedasticity in residual outcome variance with respect to the IV is introduced.
• A three-stage estimator is proposed, followed by an efficient one-step-update estimator, and an extension for multiple weak invalid instruments (MR MaWII MiSTERI).

Main Results:

• The conjunction of the three assumptions allows for causal effect identification even with an invalid IV.
• MR MiSTERI demonstrates advantages in situations with heterogeneous pleiotropic effects.
• Simulation studies and UK Biobank data analysis confirm the robustness of the proposed methods.

Conclusions:

• The MR MiSTERI approach provides a robust method for causal inference when standard MR assumptions are violated.
• The MR MaWII MiSTERI extension enhances identification and accuracy when dealing with multiple weak invalid instruments.
• These methods offer valuable tools for addressing complex challenges in genetic epidemiology and causal inference.