Instrumental Variable Model Average With Applications in Nonlinear Causal Inference.

Area of Science:

• Statistics
• Biostatistics
• Econometrics

Background:

• Instrumental variable (IV) methods are crucial for causal inference but susceptible to bias from weak or invalid instruments.
• Nonlinear relationships and instrument characteristics pose challenges for traditional IV estimation.

Purpose of the Study:

• To propose a novel two-stage nonlinear causal effect estimation method using model averaging.
• To mitigate bias arising from weak and invalid instruments in nonlinear causal inference.

Main Methods:

• A two-stage approach employing sliced inverse regression for nonlinear transformation and model averaging.
• Adaptive Lasso penalty in the second stage for instrument selection and causal effect estimation.

Main Results:

• The proposed estimator demonstrates favorable asymptotic properties.
• Numerical studies confirm its effectiveness in identifying nonlinear causal effects, even with weak/invalid instruments.

Conclusions:

• The developed method offers a robust approach to nonlinear causal inference.
• It successfully handles challenges posed by instrument quality, applicable to real-world datasets like the Atherosclerosis Risk in Communities study.