Population-level and individual-level explainers for propensity score matching in observational studies.

Area of Science:

• Observational studies
• Causal inference
• Health services research

Background:

• Propensity score matching is widely used for cancer treatment survival analysis in large databases.
• Exclusion of observations is a common byproduct of propensity score matching schemes.
• These exclusions can be viewed as data-driven eligibility criteria, impacting generalizability.

Purpose of the Study:

• To develop methods for identifying characteristics of unmatched subpopulations in propensity score matching.
• To ascertain the population on which causal effects are estimated.
• To evaluate the representativeness of causal effects using machine learning in cancer registry data.

Main Methods:

• Utilized machine learning, specifically decision trees, to identify inclusion probabilities.
• Applied the method to two datasets from the National Cancer Database.
• Analyzed factors influencing the likelihood of observations being included in the matched sample.

Main Results:

• Decision trees showed younger patients had higher inclusion probabilities (≥0.90) compared to older patients (0.47-0.65) in one dataset.
• Older patients were least likely to be matched, with age being a key factor.
• In the second dataset, both age and surgery status influenced inclusion probability.

Conclusions:

• The study highlights the importance of addressing exclusions in propensity score matching.
• Machine learning can effectively characterize differences between matched and unmatched individuals.
• Complementing propensity score matching with other adjustment methods is recommended.