Although traffic speed is considered as one of the major contributory factors for crashes, research findings on the relationship of speed and crash frequency are not consistent. This is possibly because the overly aggregated data, used by segment-based crash modelling approaches, fail to describe the pre-crash traffic conditions that are related with crash occurrences. This paper presents an alternative data-aggregation method aiming at a more accurate representation of the pre-crash traffic and geometrical conditions that are identified based on their geo-coded crash locations. Crashes are classified into homogeneous pre-crash condition scenarios (defined by speed, volume and geometric characteristics) and are modelled by severity type (ie fatal, serious, slight) employing Bayesian multivariate Poisson log-normal regression. The main findings of the analyses suggest that speed is proportionally related with both crash frequency and severity and that this relationship is described better by a quadratic rather than a linear specification. Speed variance, that is indirectly represented by low volume scenarios, is also found to contribute to fatal and serious-injury crashes. From a methodological point of view, crash location accuracy is found to affect analyses significantly and improve the modelling results.