Integrating Machine Learning and Propensity Score-Based Methods for Making Causal Inferences From Real-World Data
Author(s)
Gallinaro J1, Shah V1, Lee H1, Kreif N2, Guerra I1
1IQVIA, London, LON, UK, 2University of Washington, Seattle, WA, USA
OBJECTIVES: Causal inference methods allow for comparative analyses using real world data (RWD) when trials are unfeasible. Propensity scores (PS) are routinely used in RWD applications for tasks such as mitigating confounding, evaluating treatment effectiveness, and constructing external comparator arms in the absence of a suitable control group. Machine learning (ML) has the potential to improve these analyses by handling complex relationships in large datasets. This review examined ML methods for PS estimation in causal inference analyses.
METHODS: We performed secondary research to identify (i) simulation papers comparing non-parametric ML with logistic regression for PS estimation, (ii) theoretical papers discussing the application of ML in doubly robust estimators for both PS and outcome model estimation, and (iii) applied papers that use ML for PS estimation in real-world comparative effectiveness research.
RESULTS: Out of 9 applied papers analyzed, 5 used the super learner, 2 used neural networks and 2 used random forests for PS estimation. However, 5 simulation studies showed mixed results when comparing logistic regression or nonparametric ML models for PS estimation in terms of estimated effect bias, covariate balance and PS overlap. Three theoretical papers discussed potential bias in treatment effect estimation from using non-parametric ML for PS estimation due to regularization bias and overfitting risk. These issues can be mitigated in doubly robust methods with cross-fitting, leveraging non-parametric ML for both PS and outcome models, providing estimators robust against model misspecification.
CONCLUSIONS: Under certain conditions, ML can improve the validity of PS in estimating treatment effects from RWD. However, the importance of the “human factor”, such as variable selection based on expert knowledge, is crucial. Doubly robust methods, particularly when used with cross-fitting, can produce less biased treatment effect estimates. Future research should focus on the less-explored topic of using ML for PS estimation in the context of external comparator arm studies.
Conference/Value in Health Info
Value in Health, Volume 27, Issue 12, S2 (December 2024)
Acceptance Code
P23
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics, Confounding, Selection Bias Correction, Causal Inference
Disease
no-additional-disease-conditions-specialized-treatment-areas