Stratification-based instrumental variable analysis framework for nonlinear effect analysis.

Area of Science:

• Causal inference
• Epidemiology
• Biostatistics

Background:

• Nonlinear causal effects are common with continuous exposures, often requiring instrumental variables (IVs) to address unmeasured confounding.
• Existing IV methods for nonlinear analysis, like IV regression and control-function methods, suffer from low statistical power or potentially biased results.

Purpose of the Study:

• To propose a novel instrumental variables (IVs) framework for robust nonlinear causal effect analysis.
• To overcome limitations of existing methods, enabling accurate estimation of effect functions and identification of causal thresholds.

Main Methods:

• Introduced a three-part IV framework: Stratification, Scalar-on-function/scalar models, and Sum-of-single-effects estimation.
• The framework constructs strata where IV assumptions hold, linking local stratum-specific estimates to global effect estimation.
• Validated through extensive simulations comparing performance against existing nonlinear IV methods, especially under weak instrument conditions.

Main Results:

• The proposed framework demonstrated superior performance in predicting effect shapes and accurately estimating effect functions compared to existing methods.
• Successfully identified change points and their values across various scenarios, outperforming other nonlinear IV approaches.
• Application to UK Biobank data revealed a threshold effect of alcohol consumption on systolic blood pressure, aligning with medical guidelines.

Conclusions:

• The novel IV framework provides a powerful and flexible tool for nonlinear causal effect analysis, particularly in genetic epidemiology.
• It effectively estimates complex effect functions and pinpoints critical thresholds, offering more reliable insights than traditional methods.
• The findings support the use of this framework for uncovering nuanced causal relationships in observational data.