priyanthawedagamainfluenceurbanlanduse2006.md 49 KB
category: literaturenote citekey: priyanthawedagamainfluenceurbanlanduse2006 title: The influence of urban land-use on non-motorised transport casualties authors: "Priyantha Wedagama, D.M.; Bird, Roger N.; Metcalfe, Andrew V." year: 2006 date: 2006-00-00 2006 doi: 10.1016/j.aap.2006.01.006 publication: "Accident Analysis & Prevention" url: "https://doi.org/10.1016/j.aap.2006.01.006" zotero_key: A3XXM4AX zotero_storage: KD756EME collections: transpordi_planeerimine folder: 001_artiklid firstAuthor: "Priyantha Wedagama, D.M."
status: converted
Accident Analysis and Prevention 38 (2006) 1049–1057
The influence of urban land-use on non-motorised transport casualties
D.M. Priyantha Wedagama a, Roger N. Bird a,∗, Andrew V. Metcalfe b
a University of Newcastle upon Tyne, Transport Operation Research Group, Claremont Tower, Claremont Road, Newcastle upon Tyne NE1 7RU, United Kingdom b School of Applied Mathematics, University of Adelaide, North Terrace, Adelaide, SA 5005, Australia Received 15 April 2005; received in revised form 5 January 2006; accepted 14 January 2006
Abstract
The relationship between non-motorised road traffic casualties and land-use was investigated in two zones of approximately 8 km2 in Newcastle upon Tyne, England. Road traffic accidents are, more usually, analysed in relation to traffic flow, on the assumption that the latter can be derived from land-use data. Here, a direct relationship between primary functional land-use and non-motorised casualties is estimated.
We review past work in this area. A shortcoming of casualty data is that it does not record the origin and destination of the journeys being undertaken when the accident occurred. A method was established to identify zones within which most accidents could reasonably be expected to be related to the land-uses within that zone.
Generalised linear models were developed using non-motorised casualties as the response variable, with primary functional land-use, population density and junction density as explanatory variables. Separate models were constructed for each combination of cyclists and pedestrians, adults and children, working and non-working hours in city centre and suburban analysis zones.
In general, the study found that pedestrian casualties in the city centre zone are particularly associated with an increase in retail and community land-use during working hours. In the city centre zone, out of working hours, an increase in retail land-use (almost certainly clubs and bars) is also associated with an increase in pedestrian casualties. An increase in cyclist casualties during working hours (in the non-pedestrianised area) is associated with an increase in retail land-use.
© 2006 Elsevier Ltd. All rights reserved.
Keywords: Accidents; Crashes; Pedestrian; Cyclist; Casualties; Urban land-use; Spatial analysis
1. Background
Pedestrians and cyclists make up the two major modes of nonmotorised transport, which are environmentally benign, cheap and convenient for short distance journeys (up to 3.5 km). However, researchers have given less attention to these modes of transport, possibly because they are low technology, require little investment and cause less congestion (Rietveld, 2000). This is despite the fact that, in most countries, between 20 and 40% of road accidents involve pedestrians and cyclists.
According to urban safety management theory, land-use policy is one of the strategies used to prevent and reduce accidents (The Institution of Highways and Transportation, 1997). Different types of land-use generate different numbers of trips and engender different driver behaviour, which in turn have the potential to cause road accidents. Taking a wider perspective, many other aspects of road safety analysis are also associated with different land-uses and the associated activities (Peled et al., 1996). In the US, pedestrians are the second largest population group to die in motor vehicle related crashes (LaScala et al., 2000). In the UK, The House of Commons Transport Committee (1996) examined ways to reduce risk to pedestrians and cyclists. In addition, the Department for Transport (DfT) is promoting a modal shift to walking and cycling for shorter journeys, and aims to make these modes safer.
An increasing degree of urbanisation is associated with an increase in non-motorised transport casualties through demographic factors (e.g. population density), and road and traffic environment factors such as road length, junction density and land-use. For instance, the risk of a pedestrian accident is up to five times greater for children living in urban areas than for those living in rural settlements (Petch and Henson, 2000).
In this study, we concentrate on the association between landuse and road traffic accidents. In typical road accident analyses, land-use is not a direct input to accident prediction models,
∗ Corresponding author.
instead the number of trips aggregated as a traffic flow is used. However, our rationale is that it is the land-use that generates and attracts the trips, and that it may be possible, therefore, to model the occurrence of accidents in terms of land-use without explicit reference to traffic flows. We devise a method of analysis that does not require traffic flow as an input to the model. This is a particularly useful approach for policy makers who can make decisions about land-use before it generates the traffic, rather than waiting for the land to be developed and then having to manage the traffic problems it has generated. The two approaches are shown schematically below:
This approach may also take into account some aspects of driver behaviour which would need to be modelled, and estimated, separately in a conventional approach. For example, motorists may exercise particular caution in the vicinity of schools or hospitals, and pedestrians or cyclists may be more at risk in the vicinity of bars due to an increased possibility of alcohol impaired drivers.
In summary, the main purpose of this study is to identify land-use types that have high association with number of casualties, and advise on the most efficient strategy for reducing these casualties. As this is a novel approach to non-motorised accident investigation, a methodology had to be developed and suitable data sources identified.
2. Literature review
A few previous studies have been published on the links between land-use and road traffic accidents. For example Levine et al. (1995)investigated the relationship between zonal land-use and road traffic accidents. They found that residential population density, manufacturing, retail trade and services industry were positively related to the number of road traffic accidents. A study of child road safety in Salford, England (Petch and Henson, 2000) examined child casualties by Enumeration District zones, covering an area of about 100 km2 over a 40-month period. Factors related to land-use, principally the number of trip attractors or trip generators, percentage of terraced housing and amount of open space, were found to be significant explanatory variables. Ivan et al. (2000) identified the number of driveways to the public highway on each road segment, as a significant predictor of single vehicle and multi vehicle crashes.
Researching pedestrian casualties, LaScala et al. (2000) suggested giving priority to efforts to prevent pedestrian alcohol impairment and reducing neighbourhood alcohol availability. This was based on the relationship between the spatial location of alcohol retail outlets, demographic factors and the road environment with pedestrian injuries. In relation to land-use, the number of bars, off-licences and restaurants per kilometre Bruinsma, 1998). Generally, the findings were that the spatial
and attitudes for predicting bicycle safety.
of roadway were found to be linked to pedestrian injuries. In terms of cyclist safety, little has been found related to land-use. In a study of bicycle commuter safety on-road, off-road and on footways, Aultman-Hall and Kaltenecker (1999) found that land-use characteristics were less significant than traffic levels
Other studies have attempted to relate land-use characteristics to travel patterns, particularly for walking and cycling (Badoe and Miller, 2000; Cervero and Kockelman, 1997; Rietveld and
location of activities and their spatial density have a significant role in the determining the relationship between land-use and transport. Land-use patterns tend to be due to local decisions to respond to market demand for housing, employment and services. In theory, land-use should be planned to minimise road traffic conflict, particularly between motorised and non-motorised transport. This can be achieved by reducing the need for motorised transport, for example, by locating shops and schools within walking distance of homes. In other words, locating trip generators and trip attractors close to each other could reduce the need to travel by car, thus encouraging walking and cycling.
Kim and Yamashita (2002) compared the vehicle-to-vehicle crashes with vehicle-to-pedestrian and vehicle-to-bicycle per acre land-use category for Honolulu over a 10-year period. They found that vehicle-to-vehicle crashes were highest in commercial and industrial (6.62/10 acre year), visitor lodging (5.15/ 10 acre year) and manufacturing and industry (3.67/10 acre year). The order was somewhat different for vehicle-topedestrian crashes with visitor lodging (0.43/10 acre year), commercial and industrial (0.30/10 acre year) and public services (0.20/10 acre year) being the three highest categories. The vehicle-to-bicycle crashes followed the similar pattern although the crash rates were typically less than a half (visitor lodging 0.22/10 acre year).
3. Urban land-use
Land-use classification in this study is based on the ODPM's (Office for the Deputy Prime Minister) classification (and abbreviations), viz.:
Residential (R), such as houses and flats, and roads or paths within such areas.
Institutional/communal accommodation (Q), such as hotels/ hostels, old people's homes, children's homes, monasteries and convents.
Highways and road transport (H), such as through routes and distributor roads in housing estates, bus stations and public car parks.
Non-highway transport routes and places (T), such as railways, airports and dockland.
Utilities (U), such as cemeteries and crematoria, power stations, water works, gas works, refuse disposal places (except those in landfill waste disposal (Y)), TV masts and electricity substations, or water works.
Industry (I), such as factories, refineries, shipbuilding yards, mills and other industrial sites.
Offices (J), such as local and central government offices, banks, building societies and other service industries.
Retailing (K), such as shops, garages, public houses, restaurants and post offices.
Storage and warehousing (S), such as depots, scrap and timber yards and warehousing.
Community buildings (C), such as health, educational, community and religious buildings and police stations, prisons, fire stations.
Leisure and recreational buildings (L), such as museums, cinemas, theatres, bowling alleys, sport halls, holiday camps, amusement arcades and buildings associated with outdoor recreation.
Vacant land previously developed (V), includes cleared sites used as temporary car parks or playgrounds.
This classification is used as a basis for data collection. However, to reduce the number of variables describing land-use in the model: residential (R) and Institutional/communal accommodation (Q) are grouped as residential (RQ); and industry (I) and storage (S) are grouped as industrial (IS).
It is recognised that some land-uses have different effects as trip attractors during working hours (Monday–Friday, 07:00 a.m. to 07:00 p.m.), non-working hours during the working week (Monday to Friday, 07:00 p.m. to 07:00 a.m.), and at weekends. For example, schools, offices and commercial premises attract people during working hours, while most pubs, bars and leisure facilities attract people after working hours. So, separate models were developed for casualties during working hours and during non-working hours.
A case study was undertaken in Newcastle upon Tyne, a city in north-east England. The spatial unit in this study was the Enumeration District (ED), the smallest census boundary in England. In the case study area, 185 EDs were used, with an average area of about 12 ha. Digital mapping of the area was obtained and analysed using a Geographical Information System (GIS). For each ED, the proportion of its area attributable to each land-use category was estimated. The land-use variables were defined as these proportions, and are therefore restricted to the range 0–1. Other variables calculated for each ED are length of road (km), population density (persons per km2) and junction density (junctions per km). The length of road in each ED is required as the analysis is in terms of casualties per kilometre, the road length being included as an offset in the generalised linear models. Population density may be linked to the number of nonmotorised trips generated for either working or non-working Table 1
Study variables for each ED
Response variables
• Pedestrian and cyclist casualties during working hours and non-working hours
Offset variable
• Road length (km)
Explanatory variables
- Population density (residents per km2)
- Junction density (number of junctions per km of road)
- Land-use proportions
purposes. Conflict between motorised and non-motorised transport frequently concentrates at junctions, so junction density was used to provide a measure of such locations within each ED. In summary, the variables measured for each ED and used as the basis for the analysis are listed in Table 1.
4. Data
4.1. Accident data
Statistics for casualties in the study area between 1998 and 2001 were obtained from the local government Traffic Accident and Data Unit. These data are supplied by police attending the scene of each injury accident who record a wide variety of relevant data. This study needed the location and the number of pedestrian and cyclist casualties. Throughout this paper we deal with numbers of casualties, rather than numbers of accidents, although the majority (702 out of 706) of accidents with any non-motorised road user casualties, in our study, involved only one such casualty. The following abbreviations are used to subdivide the casualty data:
Pwh: pedestrian during working hours; Pnwh: pedestrian during non-working hours; Cwh: cyclist during working hours; Cnwh: cyclist during non-working hours.
Fig. 1(a) shows the total number of non-motorised casualties in the case study area between 1998 and 2001 within these categories. Fig. 1(b) shows the total number of casualties classified as child (up to 16 years old) or adult (17 years and above).
4.2. Land-use data
Land-use can be classified in two main ways, either functional or physical. Functional land-use classification is based on the activity taking place on the land (e.g. services/offices, residential, industry or vacant land). Physical land-use classification depends on the shape and form of the land or the building on it (e.g. high rise industrial buildings, low-rise non-residential buildings, detached housing, semi-detached housing, open area). A justification for this distinction is that some changes in landuse that do not require substantial alteration of the building, such as the housing being used as offices, or a warehouse being
Fig. 1. (a) Non-motorised transport casualties (all days) and (b) child and adult casualties (all days).
used for retailing (Rhind and Hudson, 1980). The appropriate classification method will depend on the study objectives. As the objective of this study is the link between accidents involving non-motorised road users and land-use, it is human activity related to the land-use that has more relevance than the physical properties. Therefore, the functional land-use is used in this study.
However, there is a further complication in measuring and classifying functional land-use, particularly in urban areas. This is the multiple use of the same land due to differing use of the floors of the same building (e.g. offices above a shop). Unless analysis of each building's floor area is to be undertaken, a judgement has to be made about which is the predominant activity on the site. This was determined on the basis of which land-use was judged to be the most significant trip generator/attractor. In most cases this will be the activity taking place at ground level. For example, a busy shop on the ground floor of a building is likely to generate significantly more trips than several floors of offices above. In this way, the primary functional land-use area was recorded for each plot of land. These were aggregated by land-use category within each ED, and expressed as a proportion of the area of each ED.
Land-use data were collected using digital maps supplied by the Digimap service of Edinburgh University (Edina). Where necessary the land-use was clarified by site visits.
4.3. Junction data
The road length and number of junctions were found using the digital maps. Junction density was defined as the number of junctions divided by the road length within an ED, and calculated using the ARC/INFO system.
4.4. Population data
Population data were taken from the UK 1991 census data, the latest available at the time of the study). Population density was obtained by dividing the population of each ED by its area measured in square kilometres.
5. Methods
Each spatial analysis is based on about 90 EDs within a carefully defined analysis zones. The mean of the Poisson distribution, or if appropriate more highly dispersed distribution, of the number of accidents within an ED is assumed to be proportional to the length of road within the ED, possibly modulated by population density, junction density and the land-use proportions. The estimation of the modulation is the main aim of the spatial analysis.
5.1. Defining the analysis zones
The first step was to create two suitable zones for analysis. Each zone was then used as a basis for analysing the relationship between land-use and the non-motorised transport casualties. The size of the zone was set at a radius of about 3 km. This size was based on the assumption that pedestrians or cyclist trips are typically less than 3.5 km in length (Rietveld, 2000).
In creating an analysis zone, one ED would be selected as the core. The surrounding EDs, with centroids within the required radius create a cluster around the core. Criteria for selecting a core ED are a land-use proportion of trip attractors (retails, offices and industrial) greater than 50% and being surrounded by EDs which themselves had a wide range of trip attractors and housing. From the typical maximum length of non-motorised trips, it is likely that most trips to the core zone have origins within the surrounding cluster. It is assumed that most other nonmotorised trips will have destinations in the cluster, although it is inevitable that a few do not. The objective of choosing a cluster of EDs centred on a core is to keep the proportion of non-motorised through journeys to a minimum. The clusters define the analysis zones. Land-use within an analysis zone may be relevant to all non-motorised trips (and therefore casualties) within it, except in the case of trips with destination outside the analysis zone.
In the case study area, two EDs were found to have a proportion of trip attractors greater than 50%, and both were suitable as a core for a cluster. The two analysis zones are indicated on Fig. 2 and are hereafter referred to as zone NCL (dot pattern) with its core in Newcastle City Centre, and zone GSF (vertical line pattern) with its core in Gosforth. Zone NCL consists of 92 EDs, while zone GSF consists of 93 EDs.
The land-use distribution within each analysis zone (zone GSF and zone NCL) is shown in Table 2. From this it can be seen that the proportion of trip generators (R and Q) is greater in zone GSF (42%) than in zone NCL (23%). In addition, population density in zone GSF is greater than in zone NCL. In contrast,
Fig. 2. Two analysis zones.
the proportion of trip attractors (IS, J, K, C and L) is lower in zone GSF (21%) than in zone NCL (28%). The proportion of footways is the same (7%) in each zone, while the proportion of roads in zone NCL (15%) is somewhat greater than in zone GSF (10%).
Pedestrian and cyclist casualties are more concentrated within zone NCL than in zone GSF. Without detailed analysis, it might be inferred that the greater the proportion of trip attrac-
Table 2 Land-use proportion
| Buffer zone | Zone GSF | Zone NCL | | |---------------------------------|----------|----------|--| | Population density (person/km2) | 3660 | 3369 | | | Junction density (junction/km) | 7.00 | 6.00 | | | R (residential) | 0.42 | 0.21 | | | Q (communal accommodation) | 0.00 | 0.02 | | | H (roadway) | 0.10 | 0.15 | | | T (non-highway places) | 0.01 | 0.01 | | | U (utility) | 0.00 | 0.01 | | | I (industry) | 0.02 | 0.04 | | | J (offices) | 0.01 | 0.03 | | | K (retail) | 0.01 | 0.05 | | | S (storage) | 0.00 | 0.02 | | | C (community building) | 0.13 | 0.10 | | | L (leisure) | 0.04 | 0.04 | | | V (vacant land) | 0.17 | 0.24 | | | W (footway) | 0.07 | 0.07 | |
tors in a certain zone, the more casualties, regardless either of population density or of proportion of trip generators.
5.2. Correlation between explanatory variables
Correlations between the explanatory variables (population density, junction density and land-use proportions) were calculated separately for EDs in the two analysis zones (Tables 3a and 3b).
The highest correlations are those between population density and proportion of residential land-use. High correlations are inevitable, since the population resides in the residential area, but housing density can vary considerably between EDs. Negative correlations, such as that between RQ and K in NCL arise, partly at least, from the constraint that land-use proportions for an ED add to 1. Thus, a high proportion of residential use, 0.8 for example, restricts K to below 0.2. Generally, the other correlations between pairs of individual explanatory variables were not high enough to indicate obvious problems of multicollinearity when fitting GLIM models. However, linear combinations of explanatory variables may have higher correlations, and as discussed later, there were some advantages to limiting explanatory land-use proportions to trip attractors (IS, J, K, C and L) in the final models.
5.3. Distinction between working and non-working hours
The extent to which some land-use types generate trips differs between working hours and non-working hours. During nonworking hours industry (IS) and offices (J) will have little effect as trip attractors, because they are usually closed. During working hours retail (K) will be predominantly shops whereas during non-working hours it will include bars.
This study treats casualties during working hours (Monday– Friday, 07:00 a.m. to 07:00 p.m.) and non-working hours (Monday–Friday, 07.00 p.m. up to 07.00 a.m.) separately. Weekends are not included in the following analysis. The date and time of occurrence of casualties were needed to classify them into the working hours or non-working hours data set. This was carried out using the ARC/INFO and Arcview GIS.
5.4. GLIM models
It is likely that land-use affects cyclist and pedestrian casualties in different ways. Furthermore, land-use is likely to have a different effect on casualties during working hours than during non-working hours. The two analysis zones are also rather different, GSF being a suburb and NCL inner city. Therefore, eight separate models were developed initially: cyclist casualties and pedestrian casualties by working hours and non-working hours by analysis zone (GSF and NCL). Following this, models were fitted separately for children and adults making sixteen altogether. All the models had the general form that the logarithm of the hypothetical mean casualty rate in an ED (µ) was equal to a linear combination of the explanatory variables with the logarithm of road length (rdlen) as an offset. Nearly all the non-motorised casualties occurred singly. Only four accidents
Table 3a Correlation of land-uses in zone GSF
| Popd | Jund | RQ | IS | J | K | C | L | V | W | H | T | U | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Popd | 1.00 | ||||||||||||
| Jund | 0.04 | 1.00 | |||||||||||
| RQ | 0.62 | 0.02 | 1.00 | ||||||||||
| IS | −0.22 | 0.02 | −0.34 | 1.00 | |||||||||
| J | −0.07 | 0.01 | −0.25 | −0.07 | 1.00 | ||||||||
| K | −0.05 | 0.05 | −0.16 | 0.01 | 0.08 | 1.00 | |||||||
| C | −0.44 | −0.03 | −0.47 | −0.07 | 0.17 | −0.10 | 1.00 | ||||||
| L | −0.25 | −0.08 | −0.21 | −0.07 | −0.08 | −0.08 | 0.04 | 1.00 | |||||
| V | −0.33 | −0.04 | −0.42 | −0.07 | −0.09 | −0.09 | −0.14 | −0.09 | 1.00 | ||||
| W | 0.40 | 0.04 | 0.22 | −0.08 | 0.12 | 0.01 | −0.26 | −0.20 | −0.27 | 1.00 | |||
| H | 0.60 | 0.09 | 0.21 | −0.11 | 0.13 | 0.08 | −0.38 | −0.30 | −0.16 | 0.33 | 1.00 | ||
| T | −0.11 | 0.02 | −0.21 | 0.10 | −0.04 | −0.04 | −0.06 | −0.05 | 0.00 | −0.10 | 0.00 | 1.00 | |
| U | −0.08 | 0.05 | −0.03 | −0.03 | −0.05 | 0.46 | −0.07 | −0.04 | −0.04 | −0.05 | −0.12 | −0.04 | 1.00 |
Table 3b Correlation of land-uses in zone NCL
| Popd | Jund | RQ | IS | J | K | C | L | V | W | H | T | U | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Popd | 1.00 | ||||||||||||
| Jund | −0.01 | 1.00 | |||||||||||
| RQ | 0.46 | 0.01 | 1.00 | ||||||||||
| IS | −0.27 | 0.00 | −0.32 | 1.00 | |||||||||
| J | −0.26 | 0.05 | −0.39 | 0.37 | 1.00 | ||||||||
| K | −0.02 | 0.02 | −0.44 | 0.04 | 0.32 | 1.00 | |||||||
| C | −0.27 | 0.07 | −0.38 | −0.05 | 0.00 | −0.01 | 1.00 | ||||||
| L | −0.19 | −0.06 | −0.22 | −0.02 | 0.00 | 0.04 | 0.00 | 1.00 | |||||
| V | −0.37 | −0.09 | −0.45 | −0.01 | −0.12 | −0.16 | −0.07 | 0.00 | 1.00 | ||||
| W | 0.41 | 0.00 | −0.08 | 0.09 | 0.10 | 0.04 | −0.13 | −0.16 | −0.18 | 1.00 | |||
| H | 0.45 | 0.07 | −0.14 | 0.13 | 0.24 | 0.40 | −0.27 | −0.06 | −0.35 | 0.56 | 1.00 | ||
| T | −0.24 | 0.04 | −0.26 | 0.10 | 0.10 | 0.39 | −0.11 | −0.04 | 0.02 | 0.11 | 0.23 | 1.00 | |
| U | −0.20 | −0.04 | −0.19 | −0.02 | −0.06 | −0.04 | 0.07 | 0.00 | −0.02 | −0.19 | −0.19 | −0.03 | 1.00 |
gave rise to two pedestrian casualties. This might account for some slight overdispersion. The explanatory variables included population density and junction density as well as land-use proportions. Typically:
$$\begin{split} \ln(\mu) &= \ln(\text{rdlen}) + \beta_0 + \beta_1 \text{Popd} + \beta_2 \text{H} + \beta_3 \text{K} + \beta_4 \text{C} \ &+ \beta_5 \text{L} + \beta_6 \text{IS} + \beta_7 \text{J} + \beta_8 \text{V} + \beta9 \text{T} + \beta{10} \text{W} \ &+ \beta_{11} \text{Jund} \end{split}$$
where the βi are coefficients to be estimated from the data.
The proportion of residential land-use (RQ) was not included in any models because it is highly correlated with population density and intuitively measures a similar feature of an ED. Industry (IS) and offices (J) were excluded from models for casualties during non-working hours because they are usually closed. With these exceptions, the models were developed first using all land-use proportions as potential explanatory variables and then by restricting land-use proportions to trip attractors (IS, J, K, C and L in working hours; K, C and L in non-working hours), referred to as essential land-use. Some correlation between explanatory variables is inevitable in survey work, but this has less influence on the estimates of coefficients of essential land-use, and their standard errors, if the potential explanatory variables are restricted to essential land-use rather than being selected from all land-use. The best fitting models are summarised in Tables 4a–4d.
5.5. Models for overdispersion
If the number of casualties has a Poisson distribution its mean equals its variance, and the sample deviance has an approximate chi-square distribution with degrees of freedom equal to the number of data less the number of parameters estimated from the data. However, the approximation is poor if the mean of the Poisson distribution is small, when the deviance is likely to be considerably smaller.
A ratio of deviance to degrees of freedom that is significantly larger than 1 indicates that a distribution with a variance greater than its mean would be a better model than the Poisson distribution. The negative binomial (NB) distribution is such a distribution, which has been used for the analysis of road traffic accidents by Miaou (1994), for example. The NB distribution arises from a mixture of Poisson distributions. So, if casualties within an ED have a Poisson distribution, but there are differences between EDs that are not accounted for by the recorded explanatory variables, the NB distribution will be appropriate. Zero inflated distributions are alternative models for probability distributions with variances that are greater than the mean. The zero inflated Poisson (ZIP) and zero inflated negative binomial
Table 4a Significant land-use variables, and their coefficients (standard error in parentheses underneath) for child/adult by working hours/non-working hours categorisation pedestrian casualties in Newcastle upon Tyne zone
| Variable | Child working hours | Adult working hours | Child non-working hours | Adult non-working hours |
|---|---|---|---|---|
| 1 | −2.45 (0.52) | −1.47 (0.37) | −3.77 (1.01) | −3.12 (0.61) |
| Population density | 0.354e−4 (0.317e−4) | 0.492e−4 (0.240e−4)* | −0.233e−5 (0.765e−4) | 0.477e−4 (0.370e−4) |
| Junction density | 0.111 (0.036)* | 0.080 (0.029)* | 0.158 (0.082)+ | 0.126 (0.050)* |
| IS (industrial) | 2.32 (0.99)* | 1.02 (0.83) | ||
| J (offices) | 0.45 (1.85) | 0.62 (1.53) | ||
| K (retail) | 7.12 (0.99)* | 5.14 (0.80)* | 9.32 (1.85)* | 7.12 (1.12)* |
| C (community) | 2.77 (0.88)* | 1.92 (0.70)* | −6.45 (5.95) | 1.54 (1.35) |
| L (leisure) | 1.09 (2.47) | −0.01 (2.04) | −9.31 (9.39) | −0.92 (4.40) |
| Deviance; d.f. | 82.2; 84 | 94.2; 84 | 26.1; 86 | 48.8; 86 |
Table 4b Significant land-use variables, and their coefficients (standard error in parentheses underneath) for child/adult by working hours/non-working hours categorisation pedestrian casualties in Gosforth zone
| Variable | Child working hours | Adult working hours | Child non-working hours | Adult non-working hours |
|---|---|---|---|---|
| 1 | −3.86 (0.86) | −1.82 (0.49) | −5.15 (2.51) | −2.38 (2.65) |
| Population density | 0.213e−3 (0.104e−4)* | 0.943e−4 (0.674e−4) | −0.466e−3 (0.215e−3)* | −0.633e−3 (0.762e−3) |
| Junction density | 0.044 (0.013)* | 0.025 (0.010)* | −0.250 (0.283) | −0.188 (0.502) |
| IS (industrial) | −15.5 (11.6) | −1.82 (3.02) | ||
| J (offices) | 3.52 (4.20) | 0.72 (3.46) | ||
| K (retail) | 12.8 (2.7)* | 6.92 (1.76)* | 2.32 (12.45) | −9.13 (84.1) |
| C (community) | 4.89 (1.32)* | 2.66 (0.81)* | −11.5 (23.8) | 5.33 (3.54) |
| L (leisure) | −4.95 (5.91) | −0.39 (2.00) | ||
| Deviance; d.f. | 59.5; 85 | 84.4; 85 | 15.1; 88 | 9.2; 88 |
Table 4c Significant land-use variables, and their coefficients (standard error in parentheses underneath) for child/adult by working hours/non-working hours categorisation cyclist casualties in Newcastle upon Tyne zone
| Variable | Child working hours | Adult working hours | Child non-working hours | Adult non-working hours |
|---|---|---|---|---|
| 1 | −0.71 (0.64) | −0.76 (0.44) | −4.18 (1.81) | −4.07 (1.62) |
| Population density | 0.106e−3 (0.738e−4) | −0.112e−4 (0.406e−4) | −0.292e−3 (0.344e−3) | −0.845e−4 (0.166e−3) |
| Junction density | −0.08 (0.09) | 0.025 (0.052) | −0.03 (0.20) | 0.179 (0.095)+ |
| IS (industrial) | −1.54 (1.90) | −1.85 (1.43) | ||
| J (offices) | −0.11 (4.62) | 3.25 (2.31) | ||
| K (retail) | −1.59 (2.73) | −4.77 (2.02)* | 4.73 (3.91) | 3.28 (3.59) |
| C (community) | −0.75 (1.93) | −0.37 (1.17) | 5.69 (2.74)* | 0.31 (3.89) |
| L (leisure) | 0.63 (2.78) | −0.22 (2.28) | 7.28 (4.32)+ | −38.7 (47.3) |
| Deviance; d.f. | 42.6; 84 | 61.1; 84 | 12.0; 86 | 13.9; 86 |
Table 4d Significant land-use variables, and their coefficients (standard error in parentheses underneath) for child/adult by working hours/non-working hours categorisation cyclist casualties in Gosforth zone
| Variable | Child working hours | Adult working hours | Child non-working hours | Adult non-working hours |
|---|---|---|---|---|
| 1 | −2.05 (0.98) | −1.09 (0.50) | −2.28 (1.49) | −1.65 (0.91) |
| Population density | 0.232e−3 (0.182e−3) | −0.475e−4 (0.790e−4) | 0.143e−3 (0.179e−3) | 0.274e−3 (0.200e−3) |
| Junction density | −0.0095 (0.060) | 0.011 (0.017) | −0.347 (0.261) | 0.005 (0.055) |
| IS (industrial) | −23.4 (13.6)+ | −8.49 (8.04) | ||
| J (offices) | −8.57 (12.36) | −4.98 (5.70) | ||
| K (retail) | 16.4 (3.2)* | 8.22 (2.09)* | 0.83 (9.83) | 0.83 (7.40) |
| C (community) | 2.94 (1.60)+ | 1.14 (1.00) | 1.86 (2.76) | −2.49 (2.94) |
| L (leisure) | −3.50 (3.17) | |||
| Deviance; d.f. | 34.1; 86 | 65.7; 85 | 26.4; 88 | 48.8; 86 |
distribution (ZINB) have been found to give a good empirical fit to road traffic casualty data (Shankar et al., 1997), although the physical interpretation of EDs having a probability of being completely safe is implausible. Both models make allowance for overdispersion when estimating the standard errors of coefficients. If this is not done results may appear to be statistically significant when they are not.
If child and adult pedestrian casualties are analysed together, the casualties during working hours are overdispersed. In zone NCL, the Poisson deviance for the best fitting model with seven explanatory variables is 156.2 with 84 degrees of freedom. The NB distribution was a significantly better fit (LR test P = 0.000). Also, the likelihood for the NB distribution was only slightly smaller than that for the ZIP distribution, which has additional parameters relating the logit of the probability of zero accidents to the explanatory variables. In zone GSF, the Poisson deviance for the best fitting model, again with seven explanatory variables, was 127.8 on 85 degrees of freedom. The ZIP likelihood was noticeably higher than that for the NB, and it was a significantly better fit than the Poisson (Vuong test, P = 0.034). However, following a referee's suggestion, the overdispersion with the Poisson distribution can be removed by analysing the child and adult casualties separately, and these are the results presented in Table 4.
6. Results and discussion
Zone GSF and zone NCL have different patterns of landuse. More accidents occurred in zone NCL than in zone GSF. Within zone NCL, pedestrian casualties are concentrated around the city centre whereas cyclist casualties are distributed more evenly through the area. Both pedestrian and cyclist casualties are distributed relatively evenly across the area of zone GSF.
Generally, using all land-use categories (except RQ) leads to a degree of multicollinearity. Consequently, the estimated coefficients of land-use proportions can have large standard errors and few are individually statistically significant. This leads to difficulties in interpretation. Restricting the analysis to essential land-use (IS, J, K, C, L during working hours, K, C, L during non-working hours) tends to give more reasonable models and more precise results. The Poisson regressions summarised in Table 4 have the natural logarithm of road length (km) as an offset. The scale parameter is taken as 1 and the standard errors are shown in brackets underneath the estimated coefficients. Coefficients that are statistically significantly different from 0 at an approximate 5% level (ratio of coefficient to its standard error exceeds 2 in absolute value), or 10% level (ratio of coefficient to its standard error exceeds 1.7 in absolute value) are marked with (*) or (+), respectively.
In addition to logarithm of road length included as an offset, population density and junction density are included in all the models because they are variables that are expected to be positively associated with casulalties. There was evidence that increasing population density was associated with pedestrian accidents, for most categories, but no convincing evidence of such an association with cyclist casualties. One explanation is that cycle trips take longer and are likely to pass through several EDs. As expected, there is a statistically significant positive association between junction density and pedestrian casualties for all categories except during non-working hours in the Gosforth zone. The only statistically significant association between junction density and cyclist casualties was for adults in the city centre zone during non-working hours. The lack of significance may be due to small numbers but combining the adult and child casualties does not change the conclusions.
During working hours pedestrian casualties in both zones increase significantly with: population density, junction density, proportion of community building (C) and proportion of retail (K). The trends are similar for adults and children, although the estimated coefficients are somewhat higher for the children, most noticeably for retail in the Gosforth zone. In addition, the coefficient for industrial (IS) in the city centre zone is statistically significant and positive for children.
During working hours, retail has a statistically significant positive association with cyclist casualties in zone GSF, for both children and adults, but a negative association, statistically significant for adults, in the city centre (zone NCL). The latter may be explained by pedestrianisation around the shopping area in the city centre, which means shops are not well positioned to attract cyclists. Cycling is perhaps more suited to commuting journeys than to city centre style shopping. It is also worth noting that while the main shopping street in zone NCL is pedestrianised and cyclists are banned from it; the main shopping street in zone GSF is open to both motorised traffic and cyclists.
Retail (K), during non-working hours is more related to landuses such as bars, public houses and restaurants, than shops. This is a convincing explanation for the highly statistically significant positive association between retail and both adult and child pedestrian casualties during non-working hours in the city centre zone, where bars are open from 05:00 p.m. to midnight.
Fortunately, there were few casualties in the non-working hour cyclist category. Eleven casualties in 10 out of 93 Enumeration Districts in zone GSF and 10 casualties in 9 out of 92 Enumeration Districts in zone NCL, so there were few statistically significant results. The exceptions are the positive coefficients for community and leisure for child casualties in the city centre zone.
7. Conclusion
During working hours pedestrian casualties are particularly associated with an increase in retail and community land-use. Although the results are qualitatively similar for children and adults, some of the estimated coefficients are significantly higher for children. In the NCL zone an increase of 0.10 in the proportion of retail land-use is estimated to increase the mean number of child casualties by a factor of 2.04 (exponential of the product of 0.10 with the estimated coefficient 7.12) and adult casualties by a factor of 1.67. A similar increase in the proportion of community land-use gives a factor of 1.32 and 1.21 for children and adults, respectively. Corresponding factors for the GSF zone are: 3.60 and 2.00 for an increase of 0.10 in the proportion of retail land-use for children and adults respectively; and 1.63 and 1.30 for an increase in the proportion of community land-use for children and adults, respectively. As might be expected, the estimated increase in the mean number of casualties with increases in population density and junction density were, except for the latter for children in the NCL zone, statistically significant at the 5% level.
The two zones differ in terms of pedestrian casualties during non-working hours. In the GSF zone there is no statistically significant evidence of an association between retail land-use and casualties. In contrast, an increase of 0.10 in the proportion of retail land-use in the NCL zone is estimated to increase the mean number of casualties by factors of 2.54 and 2.04 for children and adults respectively. In the GSF zone, an increase of 0.10 in the proportion of community land-use is estimated to increase the mean number of adult casualties by a factor of 1.17, which is not statistically significantly different from 1 at the 10% level. The corresponding estimate for children is less than 1, but it is not statistically significant from 1. Community land-use within the NCL zone is mainly restricted to working hours, and any increase is unlikely to affect casualties out of working hours. The estimated factors are less than 1, but again, they are not statistically significantly different from 1.
The analyses lead to a general conclusion that priority should be given to reducing pedestrian accidents associated with retail outlets, probably shops, during working hours, with retail outlets, almost certainly clubs and bars, in city centres during non-working hours, and with community land-use.
The analysis of cyclist casualties within the GSF zone leads to the conclusion that the mean number of casualties during working hours increases significantly with the proportion of retail land-use and, to a lesser extent, with the proportion of community land-use. An increase of 0.10 in the proportion of retail land-use is estimated to increase the mean number of child casualties by a factor of 5.16 and adult casualties by a factor of 2.28. A similar increase in the proportion of community land-use gives factors of 1.34 and 1.12 for children and adults, respectively. Such increases are not apparent in the City Centre, possibly due to the ban on cyclists in the pedestrianised main shopping street making the shops unattractive to cyclists.
Acknowledgements
The authors would like to thank Mr. Graham Robinson of Gateshead MBC for providing accident data, Edina, and an anonymous referee, for helpful suggestions.
References
- Aultman-Hall, L., Kaltenecker, M.G., 1999. Toronto bicycle commuter safety rates. Accid. Anal. Prev. 31, 675–686.
- Badoe, D.A., Miller, E.J., 2000. Transportation-land-use interaction: empirical findings in North America and their implications for modelling. Transportation Res. Part D 5, 235–263.
- Cervero, R., Kockelman, K., 1997. Travel demand and the 3Ds: density, diversity and design. Transportation Res. Part D 2 (3), 199–219.
- Ivan, J.N., Wang, C., Bernardo, N.R., 2000. Explaining two lane highway crash rates using land-use and hourly exposure. Accid. Anal. Prev. 32, 787– 795.
- Kim, K., Yamashita, E., 2002. Motor vehicle crashes and land use—empirical analysis from Hawaii. Transportation Research Record 1784, TRB, National Research Council, Washington, DC, pp. 73–79.
- LaScala, E.A., Gerber, D., Gruenewald, P.J., 2000. Demographic and environmental correlates of pedestrian injury collisions: a spatial analysis. Accid. Anal. Prev. 32, 651–658.
- Levine, N., Kim, K.E., Nitz, L.H., 1995. Spatial analysis of Honolulu motor vehicle crashes: II zonal generators. Accid. Anal. Prev. 27 (5), 675– 685.
- Miaou, S.P., 1994. The relationship between truck accidents and geometric design of road sections: Poisson versus negative binomial regressions. Accid. Anal. Prev. 26 (4), 471–482.
- Peled, A., Haj-Yehia, B., Hakkert, A.S., 1996. ArcInfo-Based Geographical Information System For Road Safety Analyses and Improvement, ESRI User Conference.
- Petch, R.O., Henson, R.R., 2000. Child road safety in the urban environment. J. Transport Geogr. 8, 197–211.
- Rhind, D., Hudson, R., 1980. Land-use. Methuen and Co. Ltd., New York.
- Rietveld, P., Bruinsma, F., 1998. Is Transport Infrastructure Effective? Transport Infrastructure and Accessibility Impacts in the Space Economy. Springer, Germany.
- Rietveld, P., 2000. Non-motorised modes in transport systems; a multimodal chain perspective for the Netherlands. Transportation Res. Part D 5, 31–36.
- Shankar, V., Milton, J., Mannering, F., 1997. Modelling accident frequencies as zero altered probability processes: an empirical inquiry. Accid. Anal. Prev. 29 (6), 829–837.
- The Institution of Highways and Transportation. Transport in The Urban Environment. The IHT, London, 1997.