Modern Statistical Learning for Observational Data — Professional Development Continuing Education Course
ASA, Biometrics Section
While clinical trials provide the highest level of evidence to compare clinical treatments or public health interventions, they are often not feasible due to ethical, logistic or economic constraints. Observational studies provide an opportunity to learn about the effect of interventions for which little or no trial data are available. These studies constitute a potentially rich and relatively cheap source of information. However, in such studies, treatment or intervention allocation may be strongly confounded by other important patient characteristics and much care is needed to disentangle observed relationships and infer causal effects. In this course, we will provide an overview of modern techniques for analyzing observational data. We will focus primarily on recent advances in the field of targeted learning, which facilitates the use of state-of-the-art machine learning tools to flexibly adjust for confounding while yielding valid statistical inference. In contrast, conventional techniques for confounding adjustment rely on restrictive statistical models and may, therefore, lead to biased inference. We will discuss methods for inference on the effect of single time-point interventions, including their benefits and limitations. We will also introduce the multi time-point extension of these methods and discuss strategies for dealing with missing data. Methods will be illustrated using data from a recent observational study conducted using electronic medical records. At the conclusion of the course, attendees should be able to link scientific questions of interest to meaningful causal parameters, and perform estimation of those parameters using modern techniques.
Instructor(s): David Benkeser, Emory University, Rollins School of Public Health
