Causal inference involves using directed acyclic graphs (DAGs) and the potential outcomes model to assess effect modification (EM) from a causal perspective. DAGs can help identify confounding and selection bias. The potential outcomes model distinguishes statistical from causal interaction and allows estimating average risk or risk differences to determine if causal interaction is present on the additive scale. Studies should minimize bias, control confounding, and quantify uncertainty to best assess causal EM.