Comparing methods for estimation of heterogeneous treatment effects using observational data from health care...

Area of Science:

• Health Informatics
• Biostatistics
• Observational Studies

Background:

• Routinely collected health care data offer valuable real-world evidence complementing randomized controlled trials.
• Causal inference from observational health data is challenging due to noise, high dimensionality, and confounding.
• Existing methods for heterogeneous treatment effects often assume continuous outcomes, not typical binary/rare outcomes in health databases.

Purpose of the Study:

• Evaluate off-the-shelf causal inference methods for binary outcomes in health care database studies.
• Compare performance of Bayesian additive regression trees, causal forests, causal boosting, and causal multivariate adaptive regression splines.
• Assess methods' ability to estimate conditional risk differences in complex, real-world data structures.

Main Methods:

• Simulation studies mimicking comparative effectiveness research characteristics.
• Focus on estimating the conditional average effect of binary treatments on binary outcomes.
• Proposed a novel simulation design using real covariate/treatment data with simulated nonparametric outcomes.
• Applied the design to four published observational studies from major US health care databases.

Main Results:

• Bayesian additive regression trees and causal boosting demonstrated consistently low bias in conditional risk difference estimates.
• Performance varied across methods, highlighting the importance of method selection for specific data characteristics.
• The proposed simulation design effectively recapitulated key features of health care database studies.

Conclusions:

• Bayesian additive regression trees and causal boosting are recommended for causal inference in health care database studies with binary outcomes.
• These methods show robustness in estimating conditional risk differences in complex observational data.
• Further research should explore these methods in diverse real-world health data settings.