Assuming in Public: Use DAGs to improve transparency and causal estimation

Abstract

Scientists routinely make causal inferences – whether implicit or explicit – about correlations generated from statistical analyses of experimental and observational data. However, while theorized causes are usually directionally specific, correlations are inherently symmetric or directionally ambiguous. Moreover, multiple causal structures can produce equivalent correlational results, posing significant threats to the validity of statistical inferences.

Fortunately, advances from the “causal revolution” in science and statistics have provided us with powerful tools, such as potential outcomes and directed acyclic graphs (DAGs), to better understand causes and effects. This talk will focus on how DAGs can help us “assume in public” more effectively. By introducing DAGs early in the research workflow and adhering to simple rules for their use, we can formalize the causal assumptions underlying our theories and statistical models, thereby enhancing transparency and reducing avoidable biases in causal estimation.

The presentation will cover the four foundational structures in causal systems, as represented in DAGs: complete independence, pipes, forks, and colliders. Real-world and simulated examples – drawn from the speaker’s blog posts – will illustrate key concepts, such as d-separation, “good and bad controls,” and adjustment sets. Finally, the talk will introduce tools and resources to help researchers more confidently and effectively navigate the assumptions and challenges of causal inference.