Causal inference is an emerging field in statistics. It has been rigorously studied under the counterfactual or potential outcome framework proposed by Donald Rubin (1978). Judea Pearl linked the causal inference with graph theory using directed acyclic graphs; James Robins further extend causal inference to the settings where exposure and confounding are time-varying. In addition to the theoretical and methodological development, causal inference analyses have also been widely used in biomedical research.
The first part of this course will introduce mechanisms that affect causal inference, including confounding, selection bias and interaction. Understanding the bias from confounding and selection helps studying causal inference. With knowledge about these biases, one can judge or even adjust for the bias using proper study designs or advanced statistical analyses. We will also introduce directed acyclic graph and study the biases using the graph theory.
The second part will introduce g-methods, including marginal structural models, G-formula/standardization and structural nested model/G-estimation. Based on the concept from the first part, we will learn how to construct statistical models that can account for the undue bias (confounding and/or selection bias).
The third part will introduce causal mediation model. Causal mediation analyses have received much attention recently. While the conventional causal inference concerns the causal relationship between an exposure (A) and the health outcome (Y), causal mediation model further characterizes the mechanism that mediates the causation. Specifically, causal mediation model studies the effect of A on Y mediated through a mediator (M), A->M->Y. Finally, we will introduce a relevant topic: instrumental variable analyses and discuss its required assumptions and how such an analyses provides a remedy for unmeasured confounding.