Traffic safety has increasingly become an important concern in developing long-term transportation planning strategies. Since transportation planning steps always involve some kinds of geographic entity, predicting crashes for those entities is not only a mere avenue of analytic methods in safety research, but also influential to practical application in road infrastructure design and management. However, the analyses using different spatial units are subjected to the modifiable areal unit problem (MAUP), which refers to the issue of inconsistent statistical results when dealing with geographic data of different aggregation configurations. Especially, a high-level of spatial aggregation of data could bring about the loss of detailed spatial information, also known as the scale effect. In this study, we propose Bayesian multi-scale models that are capable of accounting for the scale effect due to the high-level spatial aggregation of traffic and crash data. The performances of proposed models were assessed, as compared to the conventional (independent) model, using the crash data of two geographical scales, i.e. block groups (lower level) and census tracts (higher level) in Hillsborough County of Florida. The results indicate that the proposed multi-scale models could address the scale effects and enhance the model performance at the highly aggregated spatial units such as census tracts. This study sheds light on exploring the nature of scale effect in the macroscopic crash analysis.