Introducing causal inference to the medical curriculum using temporal logic to draw directed acyclic graphs

Abstract

Directed acyclic graphs (DAGs) might yet transform the statistical modelling of observational data for causal inference. This is because they offer a principled approach to analytical design that draws on existing contextual, empirical and theoretical knowledge, but ultimately relies on temporality alone to objectively specify probabilistic causal relationships amongst measured (and unmeasured) covariates, and the associated exposure and outcome variables. While a working knowledge of phenomenology, critical realism and epistemology seem likely to be useful for mastering the application of DAGs, drawing a DAG appears to require limited technical expertise and might therefore be accessible to even inexperienced and novice analysts. The present study evaluated the inclusion of a novel four-task directed learning exercise for medical undergraduates, which culminated in temporality-driven covariate classification, followed by DAG specification itself. The exercise achieved high levels of student engagement, although the proportion of students completing each of the exercise’s four key tasks declined from close to 100% in tasks 1 and 2 (exposure and outcome specification; and covariate selection) to 83.5% and 77.6% in the third and fourth tasks, respectively. Fewer than 15% of the students successfully classified all of their covariates (as confounders, mediators or competing exposures) using temporality-driven classification, but this improved to more than 35% following DAG specification – an unexpected result given that all of the DAGs displayed at least one substantive technical error. These findings suggest that drawing a DAG, in and of itself, increases the utility of temporality-driven covariate classification for causal inference analysis; although further research is required to better understand: why even poorly specified DAGs might reduce covariate misclassification; how ‘wrong but useful’ DAGs might be identified; and how these marginal benefits might be enhanced with or without improvements in DAG specification.