Abstract:
|
Propensity score methods are a common strategy for adjusting for observed confounders in non-experimental studies. Given the increasingly large size of datasets, such as electronic health records with the potential to control for thousands of observed covariates, machine learning methods have emerged as an effective strategy for estimating propensity score models in a flexible way, especially in the context of large numbers of covariates. This talk will present methods for estimating propensity scores using machine learning techniques such as generalized boosted models and present results from an in-depth simulation study showing their improved performance relative to logistic regression. The talk will also discuss the potential role of machine learning methods in the outcome model for improved estimation of causal effects. These methods have the potential to improve our ability to estimate causal effects in non-experimental settings in a range of fields, including mental health and drug abuse research.
|