Frameworks for estimating causal effects in observational settings: comparing confounder adjustment and instrumental...

Area of Science:

• Health research methodology
• Causal inference in observational studies
• Biostatistics

Background:

• Observational studies in health settings often face bias from confounding by indication.
• Common methods to address this bias include using confounders or instrumental variables (IVs).
• These methods rely on untestable assumptions, necessitating approaches that acknowledge potential imperfections.

Purpose of the Study:

• To formalize principles and heuristics for estimating causal effects using confounders and IVs when assumptions are potentially violated.
• To reframe observational studies as hypothesis testing across scenarios where one method's estimates are less inconsistent.
• To demonstrate these principles using donepezil's off-label use for mild cognitive impairment.

Main Methods:

• Discusses general principles for confounder and instrumental variable (IV) approaches under assumption violations.
• Explores linear and non-linear settings, including flexible methods like target minimum loss-based estimation and double machine learning.
• Applies principles to analyze donepezil's off-label use for mild cognitive impairment, comparing various methods.

Main Results:

• Compares and contrasts results from confounder and IV methods (traditional and flexible).
• Evaluates findings against a similar observational study and a clinical trial.
• Highlights the practical application of proposed principles in a real-world health scenario.

Conclusions:

• The study provides a framework for more robust causal effect estimation in observational health research.
• It emphasizes the importance of considering assumption violations and comparing different methodological approaches.
• The findings offer guidance for analysts navigating the complexities of confounding by indication.