Lyme disease has been the subject of many studies due to increasing incidence rates year after year and the severe complications that can arise in later stages of the disease. Negative binomial models have been used to model Lyme disease in the past with some success. However, there has been little focus on the reliability and consistency of these models when they are used to study Lyme disease at multiple spatial scales. This study seeks to explore how sensitive/consistent negative binomial models are when they are used to study Lyme disease at different spatial scales (at the regional and sub-regional levels). The study area includes the thirteen states in the Northeastern United States with the highest Lyme disease incidence during the 2002–2006 period. Lyme disease incidence at county level for the period of 2002–2006 was linked with several previously identified key landscape and climatic variables in a negative binomial regression model for the Northeastern region and two smaller sub-regions (the New England sub-region and the Mid-Atlantic sub-region). This study found that negative binomial models, indeed, were sensitive/inconsistent when used at different spatial scales. We discuss various plausible explanations for such behavior of negative binomial models. Further investigation of the inconsistency and sensitivity of negative binomial models when used at different spatial scales is important for not only future Lyme disease studies and Lyme disease risk assessment/management but any study that requires use of this model type in a spatial context.