The drivers of variable disease risk in complex multi-host disease systems have proved very difficult to identify. Here we test a model that explains the entomological risk of Lyme disease (LD) in terms of host community composition. The model was parameterized in a continuous forest tract at the Cary Institute of Ecosystem Studies (formerly the Institute of Ecosystem Studies) in New York State, U.S.A. We report the results of continuing longitudinal observations (10 years) at the Cary Institute, and of a shorter-term study conducted in forest fragments in LD endemic areas of Connecticut, New Jersey, and New York, USA. Model predictions were significantly correlated with the observed nymphal infection prevalence (NIP) in both studies, although the relationship was stronger in the longer-term Cary Institute study. Species richness was negatively, albeit weakly, correlated with NIP (logistic regression), and there was no relationship between the Shannon diversity index (H') and NIP. Although these results suggest that LD risk is in fact dependent on host diversity, the relationship relies explicitly on the identities and frequencies of host species such that conventional uses of the term biodiversity (i.e., richness, evenness, H') are less appropriate than are metrics that include species identity. This underscores the importance of constructing interaction webs for vertebrates and exploring the direct and indirect effects of anthropogenic stressors on host community composition.