This paper is limited to the "size problem": it empirically and systematically examines to what extent the length of road segment influences the statistical description of accident counts and density. Univariate descriptive statistics are systematically measured for accident count and ratio, using different motorway segment lengths. A classification of the studied aggregation levels is proposed leading to three distinct groups of segments: (1) very small segments (1 hm), where observed accident counts are almost Poisson distributed, (2) large segments (more than 20 hm), where accident counts are almost normally distributed, and (3) medium size segments (3-20 hm), where accident counts have an intermediate empirical distribution. It is suggested that generalizations made at one level of spatial aggregation may not necessarily hold at another level. Conclusions derived at one scale may be invalid at another. The paper shows how important preliminary examination of the data can be, without any sophisticated statistical test. It also shows that one best or unique level of aggregation is not available: it depends upon the objective of the study. Results of any statistical analysis or modelling procedure should be seen as conditional on the scale of the analysis. (A)
Abstract