IV-learner: learning conditional average treatment effects using instrumental variables

Area of Science:

• Biostatistics
• Causal Inference
• Health Services Research

Background:

• Estimating treatment effects is crucial for clinical decision-making.
• Unmeasured confounding poses a significant challenge in observational studies.
• Instrumental variables (IVs) offer a potential solution but require careful implementation.

Purpose of the Study:

• To develop a robust method for estimating conditional average treatment effects (CATE) in the presence of unmeasured confounding using IVs.
• To address limitations of existing Neyman-orthogonal learners for IV regression.
• To improve the accuracy and precision of CATE estimation in complex clinical settings.

Main Methods:

• Leveraging instrumental variables (IVs) within a causal inference framework.
• Employing infinite-dimensional targeted learning to tailor first-stage predictions.
• Developing a targeted Neyman-orthogonal learner adaptable to various data types and IVs/covariates.

Main Results:

• The proposed targeted Neyman-orthogonal learner demonstrates substantial performance enhancements compared to previous methods.
• Simulation studies confirm the improved accuracy and precision of the new IV-learner.
• Re-analysis of ICU transfer data validates the practical utility of the developed method.

Conclusions:

• The novel targeted learning approach effectively mitigates bias and estimation errors in CATE estimation with unmeasured confounding.
• This method provides a more reliable tool for identifying patient subgroups that benefit from specific interventions.
• The findings underscore the importance of advanced statistical techniques for improving causal inference in healthcare research.