The identification of traffic crash hot zones under the link-attribute and event-based approaches in a network-constrained environment Becky P.Y. Loo ⇑ , Shenjun Yao Department of Geography, The University of Hong Kong, Pokfulam, Hong Kong article info Article history: Received 6 April 2011 Received in revised form 4 July 2013 Accepted 4 July 2013 Keywords: Link-attribute Event-based Network Traffic crashes Hot zone abstract In the spatial analysis of road traffic crashes, a hot zone methodology explicitly uses the network conti- guity of more than one road segment as a criterion in identifying crash clusters. In this paper, 603 sim- ulated patterns of traffic crashes in three simplified hypothetical networks and the empirical crash pattern in Hong Kong from 2008 to 2010 (with a total of 30,490 traffic crashes on 1090 km of roads) are analyzed using the link-attribute approach and the network-constrained event-based approach. Pro- cedures for identifying hot zones using statistical thresholds are developed. This paper represents the first systematic comparison of hot zone results using these two different approaches. The results suggest that the link-attribute approach and network-constrained event-based approach are usually consistent but there are major differences between the two approaches. Ó 2013 Elsevier Ltd. All rights reserved. 1. Introduction Do spatial clusters of traffic crashes exist? If so, then where are they? How can we identify them in an efficient and scientific man- ner? Traffic crash hot zones (also called black zones) are defined as ‘‘a set of contiguous road segments taken together and character- ized by a high number’’ of traffic crashes (Flahaut, Mouchart, Mar- tin, & Thomas, 2003: 992). Traffic crash hot zones are distinct from traffic crash hotspots (also known as blacksites or blackspots) (Cheng & Washington, 2005; Elvik, 1997, 2006; Geurts, Wets, Brijs, & Vanhoof, 2004; Huang, Chin, & Haque, 2009; Miranda-Moreno, Labbe, & Fu, 2007) because of two distinct characteristics. First, hot zones must include more than one road segment (the distinc- tion does not lie in the length of the road segments). Second, those road segments with high crash numbers must be contiguous (Loo, 2009; Moons, Brijs, & Wets, 2009a). With the hotspot methodol- ogy, a junction or an individual road segment with a large number of crashes is identified as a hotspot. With the hot zone methodol- ogy, more than one individual road segment is taken together to become a single hot zone, which is a single spatial cluster of traffic crashes. In other words, network contiguity of more than one road segment is an essential criterion for identifying hot zones. This paper addresses the methodological challenges of detecting hot zones using the concept of spatial autocorrelation (Fotheringham, 2009; Getis & Ord, 1992; Goodchild, 1986) and the statistical methods developed under two commonly used approaches in geographic information systems (GIS) analysis. The first approach is the link-attribute approach (Yamada & Thill, 2007, 2010). Spatial events such as traffic crashes are not analyzed directly but are instead assigned to geographic features, such as areas or a road network. For the former, the focus is usually on visualizing and explaining the spatial variability of the crash inten- sity across areas (polygons), such as traffic zones, census tracts, districts, regions and provinces (Chen, Lin, & Loo, 2012; Erdogan, 2009; Levine,Kim, & Nitz, 1995b). For the latter, traffic crashes are assigned to line and point features, namely roads (links) and junctions (nodes). Links are, in turn, divided into shorter segments called basic spatial units (BSUs) for detailed spatial analysis. Traffic crash numbers or rates are treated as attribute values of these geo- graphic features. The second approach is to consider the physical locations of individual crashes (events) directly. This approach is often termed the event-based approach (Yamada & Thill, 2007). Much event-based research on traffic crashes aims to describe spa- tial patterns using ‘‘point process’’ tools (Kim & Yamashita, 2007; Levine, Kim, & Nitz, 1995a; Okabe, Satoh, & Sugihara, 2009). To illustrate, the K-means clustering algorithm is used to group crashes into cluster centroids that minimize the sum of thesquared distance from every point to the K centers (Kim & Yamashita, 2007). In other words, all crashes are assigned to one of the spatial clusters. Spatial statistics are used primarily as a tool for data reduction and grouping. When the event-based approach is used to identify local clusters, this goal is accomplished by directly mea- suring the (physical or network) distance or the degree of concen- tration among the traffic crashes. With reference to hot zones, the methodology is better developed under the link-attribute approach 0198-9715/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.compenvurbsys.2013.07.001 ⇑ Corresponding author. Tel.: +852 3917 7024; fax: +852 2559 8994. E-mail address: bpyloo@hku.hk (B.P.Y. Loo). Computers, Environment and Urban Systems 41 (2013) 249–261 Contents lists available at ScienceDirect Computers, Environment and Urban Systems journal homepage: www.elsevier.com/locate/compenvurbsys