Fach- und Laienpublikum
Prof Dr. Sebastian Baumeister, Regensburg
Sebastian Baumeister from Regensburg University will talk about Covariate selection in observational studies with limited knowledge of the true causal structure.
Unconfoundedness is assumed to generate an unbiased estimate of a causal effect on an outcome. If there is sufficient knowledge on the underlying causal structure, established covariate selection criteria based on a causal directed acyclic graph (DAG) can be used to select subsets of observed pretreatment covariates. However, although variable selection approaches should be based on an understanding of the causal structure representing the common cause pathways between treatment and outcome, the true causal structure is rarely known. When knowledge of the causal structure is not sufficient for use of a DAG, alternative covariate selection strategies are needed. Common suggestions when the causal structure is only partially known include “all observed pretreatment covariates” (Rubin) or the “disjunctive cause criterion” (VanderWeele & Shpister). Data-driven procedures for selection of covariates have also been proposed (e.g., change-in-MSE, focused selection, CovSel). Defining subsets of covariates that are sufficient to control for confounding empirically is difficult in the absence of background knowledge on the causal data-generating mechanism because adjustment for covariates that are affected by the exposure or the outcome and adjustment for instruments can increase bias.
Prof Dr. Sebastian Baumeister, University of Regensburg
Berlin Epidemiological Methods Colloquium
Seminar room of the Neurology Clinic; Bonhoefferweg 3 entrance, 3rd floor, Charité – Campus Mitte