Essentially, this relates to the mathematical problem of naturally poor correlation between low value continuous values (e.g. 0.003) with discrete values (0, 1, 2, etc). Has anyone come up with a workable solution to this in road safety?
I attach a map of local road crashes in parts of Melbourne, Australia, to demonstrate. The local roads (within the squares defined by arterials) have hardly any crashes. Models come up with low values no matter how good the model is. Poor correlation/model fit is guaranteed. Solution?