In conventional safety analysis, traffic and crash data are often aggregated at the geographical units like census tracts, street blocks, and traffic analysis zones, which are often delineated by roads and other physical entities. A considerable proportion of crashes may occur at or near the boundary of geographical units. Such the crashes, also known as boundary crashes, can correlate with the explanatory variables of neighboring geographical units, regardless of the spatial proximity. This could then bias the parameter estimation of crash frequency model. In this study, a novel data-driven approach is developed for the allocation of boundary crashes. For example, crash severity and bicyclist characteristics are considered in the crash feature-based allocation. An illustrative case study based on built environment, population, traffic and bicycle crash data from 289 Lower Layer Super Output Areas (LSOAs) of London in the period 2017–2019 was conducted. Results indicate that high matching percentages of boundary crash allocation can be achieved. Furthermore, prediction performances, in terms of root mean square error (RMSE) and mean absolute error (MAE), of the crash frequency models based on the proposed crash feature-based allocation method is superior, compared to that based on conventional boundary crash allocation methods like half-and-half and iterative assignment approaches. Last but not least, more influencing factors that affect the bicycle crash frequency at macroscopic level can be identified. Findings should be indicative to the spatial safety analysis for different geographical configurations.