Spatio-temporal methods for the analysis of crime and traffic safety data

Resumen

Since physician John Snow analyzed the spatial distribution of cholera cases detected in the 1854 epidemic in London, many disciplines have benefited from the existence of spatio-temporal statistical methods: agriculture, astronomy, biology, epidemiology, geology, hydrology, meteorology, and remote sensing, among others. This thesis therefore focuses on the development and application of spatio-temporal methods in the context of two disciplines: traffic safety analysis and criminology. In particular, a capital objective has been to detect research gaps in the currently available literature. Thus, the investigation of several types of problems that usually arise in these two fields, which require a specific statistical approach, has led to the structuring of this thesis as follows. Firstly, after an introductory chapter, two studies in the context of traffic safety analysis where the use of a linear network structure is fundamental are shown. The first one contains a street-level multivariate analysis of the occurrence of traffic accidents accounting for the presence of intersection and non-intersection segments. Next, in Chapter 3, a method is presented and employed for the detection of differential risk "hotspots" along a network. Chapter 4 includes a spatio-temporal analysis of a burglary dataset focused on the phenomenon of near-repetition, which is capital in the field of criminology. The classic version of the Knox test is adapted to account for spatio-temporal burglary risk heterogeneity, which provides a more accurate representation of the magnitude of the phenomenon. Specifically, an adjustment is proposed that is suitable in a context of absence of spatial-temporal variation in both the exposure variable and the covariates. Chapter 5 includes a detailed study of the modifiable area unit problem (MAUP) in the context of traffic safety analysis. As a novelty compared to previous studies, the scale and zoning of the spatial structures considered are explicitly controlled. Furthermore, the analysis does not only focus on the final consequences in terms of estimation and precision of the models, but also on the alterations that occur in the different variables involved. Chapter 6 is dedicated to the comparison of several methodologies that can be selected to analyze how the proximity to certain places influences the incidence of an event of interest. Specifically, this comparison is made to assess the relationship between traffic accidents and the location of educational centers. Chapter 7 focuses on analyzing an issue that has been given great importance in quantitative criminology: the loss of reliability of analyses as a result of the presence of non-geocoded events. It has been estimated that reaching 85% geocoding success rate is enough to carry out further analysis of the data. In this thesis, this percentage is reestimated taking into account some factors and methods not taken into account in the initial estimation. It is concluded that reaching 85% success rate in the geocoding process may not be sufficient under certain conditions. Finally, Chapter 8 includes the description of two R packages that have been developed during this thesis: SpNetPrep, which allows the preprocessing and curation of a linear network, and DRHotNet, which implements the "hotspot" detection procedure described in Chapter 3.