An overview of modern machine learning methods for effect measure modification analyses in high-dimensional settings

Abstract

A primary concern of public health researchers involves identifying and quantifying heterogeneous exposure effects across population subgroups. Understanding the magnitude and direction of these effects on a given scale provides researchers the ability to recommend policy prescriptions and assess the external validity of findings. Traditional methods for effect measure modification analyses require manual model specification that is often impractical or not feasible to conduct in high-dimensional settings. Recent developments in machine learning aim to solve this issue by utilizing data-driven approaches to estimate heterogeneous exposure effects. However, these methods do not directly identify effect modifiers and estimate corresponding subgroup effects. Consequently, additional analysis techniques are required to use these methods in the context of effect measure modification analyses. While no data-driven method or technique can identify effect modifiers and domain expertise is still required, they may serve an important role in the discovery of vulnerable subgroups when prior knowledge is not available. We summarize and provide the intuition behind these machine learning methods and discuss how they may be employed for effect measure modification analyses to serve as a reference for public health researchers. We discuss their implementation in R with annotated syntax and demonstrate their application by assessing the heterogeneous effects of drought on stunting among children in the Demographic and Health survey data set as a case study.