|
Abstract:
|
Evidence from real world data (RWE) is increasingly valuable for medical decision making. With the potential for bias in RWE estimates of treatment effect and the numerous proposed estimation strategies to account for bias in these comparative non-randomized designs, researchers are beginning to use machine learning approaches to help optimize model selection and address model uncertainty. Model averaging (MA) is one of those approaches and is a well-accepted methodology that has now been adapted for comparative real-world analyses (Zagar et al. 2022). Within the proposed MA strategy, researchers specify a set of estimation strategies/models and use cross-validation to construct individual model weights in a data driven framework. We will review MA in the setting of comparative real-world analyses and provide a case study application of its use. This will include demonstrating the benefits of ensemble approaches relative to today’s standard of pre-selecting a single method. In addition, practical concerns such as selection of methods, pre-specification, appropriate sensitivity analyses, and comparisons to other machine learning based methods such as TMLE will be discussed.
|
