Propensity scores using missingness pattern information: a practical guide.

Area of Science:

• Epidemiology
• Biostatistics
• Health Informatics

Background:

• Electronic health records (EHRs) are rich data sources for health research.
• Propensity score analysis (PSA) is crucial for mitigating confounding bias in EHR studies.
• Missing data in EHRs, especially for confounders, poses significant analytical challenges.

Purpose of the Study:

• To explore and assess the underlying assumptions of the missingness pattern approach (MPA) for handling missing confounder data in PSA.
• To develop a practical framework for evaluating the plausibility of MPA assumptions in specific research contexts.
• To guide researchers on the appropriate application of MPA in observational studies using EHR data.

Main Methods:

• Utilized causal diagrams to systematically evaluate the plausibility of MPA assumptions across various scenarios.
• Developed a structured framework for assessing MPA assumption validity in real-world research settings.
• Applied the developed framework to a motivating study investigating renin-angiotensin system blockers and acute kidney injury risk using EHR data.

Main Results:

• Identified conditions under which MPA assumptions are likely to be violated.
• The proposed framework offers practical guidance for assessing MPA applicability.
• In the motivating study, MPA assumptions were found to be reasonable, facilitating its application.

Conclusions:

• The MPA is a viable method for handling partially observed confounders in PSA when its assumptions are met.
• The developed framework enhances the rigorous application of MPA in EHR research.
• This work provides critical tools for improving the validity of causal inference from observational health data.