The effects of interventions targeting potentially contagious outcomes are often evaluated in clusters of interacting individuals. Contagion induces complex dependence: an individual’s outcome may be influenced by other cluster members’ treatments, outcomes, or both. A widely accepted approach to evaluate the efficacy of such interventions defines the “direct effect” as the contrast between the potential infection outcomes under treatment and no treatment, averaged over the conditional distribution of possible treatment assignments to other individuals. Theoretical and empirical studies have shown that the direct effect may change depending on the treatment allocation within clusters. At the same time, the direct effect is often interpreted individualistically, as if it summarizes the infection risk of an individual under treatment versus no treatment, with exposure to infection held constant. Using an agent-based model of contagion, we explain this discrepancy, and show how some randomization designs ensure that the direct effect cannot have an individualistic interpretation.