The exclusion of unmatched observations in propensity score matching has implications for the generalizability of causal effects. Machine learning methods can help to identify how the study population differs from the unmatched subpopulation.
There has been widespread use of propensity scores in evaluating the effect of cancer treatments on survival, particularly in administrative databases and cancer registries. A byproduct of certain matching schemes is the exclusion of observations. Borrowing an analogy from clinical trials, one can view these exclusions as subjects that do not satisfy eligibility criteria.
Developing identification rules for these “data-driven eligibility criteria” in observational studies on both population and individual levels helps to ascertain the population on which causal effects are being made. This article presents a machine learning method to determine the representativeness of causal effects in two different datasets from the National Cancer Database.
Decision trees reveal that groups with certain features have a higher probability of inclusion in the study population than older patients. In the first dataset, younger age categories had an inclusion probability of at least 0.90 in all models, while the probability for the older category ranged from 0.47 to 0.65. Most trees split once more on an even higher age at a lower node, suggesting that the oldest patients are the least likely to be matched. In the second set of data, both age and surgery status were associated with inclusion.
The methodology presented in this paper underscores the need to consider exclusions in propensity score matching procedures as well as complementing matching with other propensity score adjustments.