Doubly robust machine learning-based estimation methods for instrumental variables with an application to surgical...

Area of Science:

• Epidemiology
• Health Services Research
• Biostatistics

Background:

• Randomized controlled trials (RCTs) are often infeasible in comparative effectiveness research.
• Instrumental variable (IV) designs are valuable alternatives when RCTs are not possible.
• Gallbladder inflammation (cholecystitis) treatment decisions present challenges for standard IV methods due to potential model misspecification.

Purpose of the Study:

• To investigate the comparative effectiveness of surgical versus non-surgical treatments for cholecystitis.
• To introduce advanced instrumental variable methods utilizing a doubly robust machine learning framework.
• To extend these methods for analyzing heterogeneous causal effects and conducting sensitivity analyses.

Main Methods:

• Employed a doubly robust machine learning framework for instrumental variable estimation.
• Utilized surgeon's operative tendency as an instrument for surgical treatment of cholecystitis.
• Developed and applied novel estimators for causal inference, including principal strata and sensitivity analyses.

Main Results:

• Operative care (surgery) demonstrates general effectiveness in treating cholecystitis.
• Machine learning-based IV methods provide consistent estimates and valid inference.
• Heterogeneous treatment effects were identified, with varying surgical benefits across patient subgroups.

Conclusions:

• Doubly robust machine learning enhances instrumental variable analysis for observational studies.
• Surgical intervention is typically more beneficial for cholecystitis, but patient-specific factors modify outcomes.
• Advanced causal inference methods are crucial for understanding treatment effectiveness in complex clinical scenarios.