Regression adjustment for causal inference

Abstract

Causal inference identifies and quantifies the effect of an exposure (the hypothetical cause) to an outcome. While randomised controlled trial design is the preferred method for establishing and measuring causal effects, they are costly and not always feasible. Observational studies can be used for causal inference, but confounding is a key limitation. Typically, a confounder is a variable that causes the exposure and outcome. If left unaccounted for, confounders distort the estimate away from the true causal effect. Regression analysis is a common technique to analyse observational data, partly because of its ability to adjust for confounders by including those confounders as independent variables in the regression model. However, knowing when a variable should be adjusted is not always clear.