Interpretable Models in Regression Analysis

An interpretable model is preferred in practice due to several advantages, such as being robust to outliers, offering a better understanding of the underlying scientific question, and facilitating external communications. However, there is no standard approach to evaluate interpretation, and even its definition varies in the literature. In this chapter, we recommend defining an interpretation function with a multi-layer structure with base components based on domain knowledge of a specific problem. Additionally, to propose a meaningful model in practice, we also need to balance approximation, generalizability, and simplicity. In this framework, a DNN -based measure is utilized to select a proper model balancing the above aspects. We apply this approach to derive a sample size formula for an adaptive clinical trial design. The approximated formula greatly increases interpretation and facilitates discussion with cross-functional teams in the study design stage.