WL12: Teaching Causal Inference Using Observational Data

Music City Center 

Most students are more interested in causal inference than in population inference. Those in the health sciences, social epidemiology and the social sciences are more interested in using observational data as evidence for causal connections. . These students should be introduced to statistical methods to infer, disprove, or evaluate observational causation. Two methods (Rubin's propensity scores and Pearl's Directed A-cyclic Graphs) typically require software. This round-table discusses a third approach for an observational-causation course that doesn't require software and doesn't use Algebra. This introductory course uses (1) the Wainer graphical method of taking into account the influence of a measured binary confounder, (2) the Cornfield necessary conditions for a binary confounder to nullify or reverse an association involving a binary predictor, and (3) the Schield proposal for a standard distribution of confounders. The goal isn't to prove causation so much as teaching students to take care in using a crude (unadjusted) association as evidence for causation. Students learn how taking into account (controlling for) a related factor can change the size and direction of a two-group comparison. These three ideas are featured in the introductory Statistical Literacy course taught at the University of New Mexico (UNM) and at New College of Florida (NCF). The course at UNM satisfies a math requirement in their general education curriculum and is required by undergraduate majors in statistics. The course at NCF will be required of all new students in fall 2025.

observational causation

confounding

statistical literacy

general education 

Section on Teaching of Statistics in the Health Sciences