Probabilistic Data Linking for Mapping Crashes in Large Networks

The limited quality of location data poses a major problem to those who want you use these data for research or safety management. A considerable proportion of crashes remain unassigned to specific road locations. This fact is sometimes overlooked because crash database queries return crashes assignable to locations but do not issue warnings about crashes that are not assignable. Without reliable location information, crash data are of limited use for more refined safety research and for even basic road safety management. Probabilistic linking techniques are frequently chosen to link medical, census, and other population records for mainly medical research but also for security concerns and market studies. This paper presents a first application of the probabilistic method to linking road crash records with road inventory records. Due to missing or incorrect data, multiple locations are possible candidates and probabilistic linking offers a solution by selecting the highest likelihood locations. This paper uses the Bayesian approach to link data in a version specialized to linking one-to-many records. Preliminary effort and its results are presented and discussed.