Summary
The episode works through four questions that students often find difficult in mediation and moderation analysis. It explains why randomizing an exposure leaves the relationship between the mediator and the outcome open to confounding, using the low-birth-weight paradox analysed by Hernández-Díaz, Schisterman and Hernán as a real example. It then works through simple slopes in a centred interaction model, the additive and multiplicative scales of interaction in the asbestos and smoking data of Hammond, Selikoff and Seidman, and a hypothetical caregiver study in which opposing direct and indirect paths produce a total association close to zero. The hosts close by discussing equivalent path models and whether mediation analysis should be applied to cross-sectional data, and they agree on a reporting rule for such analyses.