Preference-based instrumental variable methods for the estimation of treatment effects: assessing validity and...

Area of Science:

• Biostatistics
• Pharmacoepidemiology
• Health Services Research

Background:

• Observational studies using administrative data are vital for regulatory and clinical decisions.
• Data limitations, lacking prognostic variables, challenge the validity of these studies.
• Instrumental variables (IV) methods, using provider-level practice variations, attempt causal inference but have complex assumptions.

Purpose of the Study:

• To propose a simple framework for evaluating violations of instrumental variables (IV) assumptions.
• To explore how treatment effect heterogeneity and assumption violations bias IV estimators.
• To provide methods for anticipating bias direction in IV analyses.

Main Methods:

• Developed a framework based on a single unobserved dichotomous variable.
• Analyzed potential bias in standard IV estimators relative to the average treatment effect.
• Applied the framework to a study on gastrointestinal bleeding risk from non-steroidal anti-inflammatory drugs (NSAIDs).

Main Results:

• The proposed framework clarifies how IV assumption violations and effect heterogeneity can bias results.
• It enables anticipation of bias direction using empirical data and subject matter knowledge (e.g., overuse/misuse of drugs).
• Demonstrated the framework's utility in a real-world NSAID safety comparison.

Conclusions:

• The framework offers a transparent approach to assessing the validity of IV analyses in pharmacoepidemiology.
• It aids researchers in understanding and mitigating potential biases in causal effect estimation.
• Facilitates more reliable evidence generation from observational health data for drug safety and effectiveness.