Machine learning and statistical models to address cluster-level confounding for the study of medical devices and surgical epidemiology using observational data

Surgeon and hospital-related features such as volume and experience can be associated with treatment choices and outcomes in observational studies related to medical devices and surgical epidemiology. This creates difficulty in estimating accurate treatment effects due to additional cluster-based confounding related to the surgeon and hospital characteristics. While methodologies exist to address patient-level confounding, current literature provides limited guidance on the best ways to tackle cluster-level confounding. In response to this gap in knowledge, I conducted a literature review on methods to address cluster-level confounding and used the findings to set the objectives for the subsequent chapters. I studied and compared the accuracy and precision of inverse probability weighting (IPW), propensity score matching, causal forests and cardinality matching in Monte Carlo simulation studies with many different cluster level confounding scenarios ranging from weak to strong confounding and different cluster structures and sizes. The identified methods were also applied to a real-world surgical safety study of pancreatic surgeries for cancer patients to illustrate the results of my simulations. I found that using a logistic regression-based propensity score model with both patient and cluster-level confounders is the best-performing method for IPW. However, IPW appeared less accurate and precise in scenarios with strong cluster-level confounding. Causal forest consistently outperforms IPW in these conditions, while the selection of confounders is data-driven. Cardinality matching performed similarly to propensity score matching but provided better covariate balance, except in scenarios with small cluster sizes. More research is needed to confirm the applicability of these findings, and to provide clearer guidance on the use of cardinality matching and causal forests in observational studies.