Applied Geography 31 (2011) 65e75
Contents lists available at ScienceDirect
Applied Geography journal homepage: www.elsevier.com/locate/apgeog
Fire incidence in metropolitan areas: A comparative study of Brisbane (Australia) and Cardiff (United Kingdom) Jonathan Corcoran a, *, Gary Higgs b, Angela Higginson c a
The University of Queensland, School of Geography, Planning and Environmental Management, Queensland 4072, Australia University of Glamorgan, GIS Research Centre, Wales Institute of Social and Economic Research, Data and Methods, Faculty of Advanced Technology, Pontypridd, Wales CF37 1DL, United Kingdom c The University of Queensland, Institute of Social Science Research (ISSR), Queensland 4072, Australia b
a b s t r a c t Keywords: Fire incidence Socio-economic factors GIS Risk analysis Cardiff Brisbane
In our previous research we applied spatial statistics and regression analysis to explore the relationships between the types of socio-economic factors that are associated with different fire incident types for an area of South Wales, UK. In this paper, we extend this analysis by using a comparative approach applying regression analysis to examine intra-urban trends in fire incidence using the case studies of Brisbane (Australia) and Cardiff (United Kingdom). Whilst drawing attention to the problems faced by researchers using spatial data in comparative contexts, this has revealed some important similarities and differences in associations, for example in relation to the residential patterns of the two cities which is reflective of their respective wider urban geography. We conclude by demonstrating the differences in trends between Cardiff and Brisbane and by highlighting outcomes from this research that are of relevance to policy makers in urban contexts. The latter could include for example those charged with identifying high risk communities, designing possible intervention strategies such as safety campaigns or with implementing educational programmes. Ó 2010 Elsevier Ltd. All rights reserved.
Introduction Nationally the UK government has set targets to reduce by 2010 the number of accidental fire-related deaths in the home by 20%, to reduce the number of deliberate fires by 10% and to reduce the number of fires that occur in buildings by 3% (ODPM, 2005). In order to support such targets, emergency services such as fire and rescue services are increasingly collecting detailed operational incident data using spatial collection and analytical technologies. Geographical Information Systems (GIS) are being used in risk assessment, road routing (response) analysis and thematic and ‘hot-spot’ mapping applications as well as in areas such as performance monitoring in relation to central government targets. The key aims include the use of such systems to support risk reduction, manage existing resources, improve response times to incidents, model the consequences of resource deployments and aid the prevention of emergency incidents through safety campaigns (Hooper, 2006). In relation to the former, for example, risk
* Corresponding author. Tel.: þ61 7 3365 6517; fax: þ61 7 3365 6899. E-mail addresses:
[email protected] (J. Corcoran),
[email protected] (G. Higgs),
[email protected] (A. Higginson). 0143-6228/$ e see front matter Ó 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.apgeog.2010.02.003
assessments are calculated for different types of dwellings in relation to historical patterns of fire data and local population data and for non-fire incidents such as flooding and road vehicle accidents based on spatial patterns of incidents. Previous studies have drawn attention to the importance of determining potential influences on trends in fire incidence in relation to fire-related mortality and injury rates and the economical consequences of fire events. Zhang, Lee, Lee, and Clinton (2006) quote global figures at the turn of the century of annual deaths caused by fires, mostly in the home, of approximately 300,000. In Australia alone, the same authors quote Australian Bureau of Statistics figures for 2000 that indicate there were over 10,000 residential fires annually and that these led to 1500 injuries and 70 deaths. Hidden within such figures are likely to be spatial variations in mortality and injuries related to socioeconomic trends within for example urban areas. A review of early research in this field found that in the United States areas with high numbers of disadvantaged residents were more vulnerable to residential fires but also drew attention to the impact of household level factors such as family composition (FEMA, 1997). Such reviews have drawn attention to the association between demographic characteristics and population density and risk factors such as building types, smoking patterns, variations in smoke detector
66
J. Corcoran et al. / Applied Geography 31 (2011) 65e75
installations and levels of educational attainment (Jennings, 1999). Our earlier research (reviewed in the next section) highlights for example those household and area characteristics associated with call-outs in South Wales (Corcoran, Higgs, Brunsdon, Ware, & Norman, 2007). In that paper, we used data sets for an operational fire brigade area in order to use South Wales as a “test-bed” for our methods of analysing potential associations of fire incidence with socio-economic data based on the 2001 UK Census of Population. The primary aim of this paper is to build on that body of research and draw on a sub-set of the data we used for our South Wales case study area (in this paper clipped to the urban area of Cardiff, United Kingdom) and a comparable data set for Brisbane (Australia) in order to examine if there is any commonality in the types of factors that are associated with fire incident types in both cities. In so-doing we aim to establish if the methodologies we used in our previous approach are transferable to a different geographical, environmental and social context and comment on any differences in data collection procedures for example which may account for such trends. In the absence of disaggregate data at the level of individual households, a spatial (ecological) approach is adopted in order to conduct a preliminary analysis of the characteristics of areas associated with high and low levels of fire incidence. Detailed regression techniques are used to examine how the incidence of call-outs is associated with census variables at an intra-urban scale of analysis to identify the types of factors that may be influencing such trends. The analyses of spatial patterns of crime incidence and health data have formed the basis for a number of studies (for examples of such work see Chainey & Radcliffe, 2005; Cromley & McLafferty, 2002). However, there is relatively little published research that has explored the potential association of census data and fire incidence at the intra-urban scale (see the following section). The use of GIS-based techniques to analyse spatial trends has only relatively recently been made possible through the availability to academics of detailed geo-coded data increasingly being collected by fire and rescue services as part of their day-today operational duties. We further suggest that the type of comparative research conducted as part of our study can highlight the extent to which such patterns are an artefact of the data collection procedures in place in differing contexts or alternatively represent ‘real’ trends that are replicated in contrasting environments. If the latter is the case, such local factors may be worthy of further investigation in order to explain patterns and to identify root causes of trends through, for example, multi-method approaches. Finally, from the operational perspectives of the fire services which serve these two cities, we contend that the spatial analysis of fire incidence patterns can be used for example to proactively target areas which have higher than expected incidence of call-outs through such devices as educational and awareness campaigns as well as having wider strategic benefits in areas such as performance management and joint working with other local and central government services. However, as we demonstrate, given the contrasting geographical and attribute characteristics of socio-economic data and the nature (for example terminologies, spatial and temporal referencing) of the incident data collected by fire services in the UK and Australia, comparative studies of this nature are by no means straightforward. In addition controlling for all variables in such studies is problematic; Brisbane and Cardiff, for example, have very different climatic regimes which may impact on such trends and we aim to explore such influences in our further research. In the next section, we briefly review previous studies that have examined the association between socio-economic factors and fire incidence before detailing our methodological approach in section three.
Socio-economic factors and fire incidence: previous research In a review of existing research on socio-economic variables and their association to fire incidence, Jennings (1999) detailed a series of primarily US-based ecological studies. In this review he states that “ecological approaches to understanding urban life provide a valuable framework for discussing and enumerating the factors contributing to the fire problem” (p. 27). Jennings reports that socio-economic associations were found to be most prevalent when analysed at the census tract level including factors such as overcrowding and substandard housing and that these factors were associated with heightened fire risk (Wallace & Wallace, 1984). That review also highlighted the types of socio-economic factors that could be important in other (non-US) contexts referring to a UKbased study in particular (Chandler, Chapman, & Hallington, 1984) where factors including the age of housing, housing tenure, and socio-economic status (social class, unemployment status, ethnicity) were found to correlate with fire incidence. In addition to investigations of fire incidence (largely based on domestic fires) there has been a particular focus on the examination of salient socio-economic factors to try to account for patterns in fatal fire incidents. In one piece of work Duncanson, Woodward, and Reid (2002) found that fatal domestic fires were more prevalent in socio-economically deprived mesh blocks. Other work has explored associations between population density, demographic characteristics, and risk factors such as building types, smoking patterns, variations in smoke detector installations and levels of educational attainment; all of which have been found to correlate with such variations (Gunther, 1981; Holborn, Nolan, & Golt, 2003; Krisp, Virrantaus, & Jolma, 2005; Runyan, Bangdiwala, Linzer, Sacks, & Butts, 1992; Shai & Lupinacci, 2003). A more recent investigation concerned with mapping injuries in children from fires and burns in New South Wales, Australia has highlighted that children living in rural and remote areas, as well as in some metropolitan areas, were at greater risk (Poulos et al., 2009). More generally such studies have demonstrated the potential of spatio-temporal analytical mapping techniques to identify communities at risk (Asgary, Ghaffari, & Levy, 2010; Corcoran, Higgs, Brunsdon, & Ware, 2007; Corcoran, Higgs, Brunsdon, Ware, & Norman, 2007). In this paper we extend the analysis presented in our previous research to an intra-urban comparison of fire incident trends for two cities e namely Brisbane (Australia) and Cardiff (United Kingdom) e drawing on potentially comparable sources of both fire incident data and detailed resolution census variables. We argue that there is a real need to undertake comparative analysis (as much as the respective data permits) of the underlying trends in fire incidences in relation to socio-economic factors in different contextual situations. Such a study is particularly timely given the widespread adoption of GIS technologies and improvements in the collection of more positionally accurate operational data related to fire incidents by the emergency services that permits such detailed disaggregate analysis. An AustralianeUK study will also permit a wider comparison to the trends found for residential fires in urban United States by Gunther (1981) and Jennings (1999). This in turn had built on research conducted in the UK context by Chandler et al. (1984) who demonstrated the importance of such factors as the age of housing, housing tenure and socio-economic status in explaining patterns of fire rates for three urban areas. Nicolopoulos, Murphy, and Sandinata (1997) examined the relationships between fire incidence, socio-economics and demographic composition within the greater Sydney area. Their study highlighted relationships between fire incidence (total fires, house fires, structure fires, arson, bush and grassland fires) and total population, age, ethnicity, language spoken at home, school leaving age, socio-economic wellbeing, tertiary qualification attainment, income, unemployment
J. Corcoran et al. / Applied Geography 31 (2011) 65e75
type and housing tenure. Their study used firstly a correlation analysis and then factor analysis to analyse the interrelationships between twenty three variables identified from the correlation stage from which two significant factors were identified namely demographic and social structure such as education, housing tenure, population composition (Factor 1), and a measure of the socio-economic/income structure (Factor 2). Using these two factors multiple regression analyses were conducted for each fire incident type. Results suggested a significant relationship between fire incidence and the socio-economic composition of the community and home ownership. The number of elderly people was found to be weakly related to fire incidence and income was demonstrated to provide greater explanation than ethnicity. Such studies have the potential to provide contextual information for more detailed investigations and to design and monitor fire reduction and safety initiatives. Allareddy, Peek-Asa, Yang, and Zwerling (2007), for example, used a survey of just over a thousand households in a rural county in Iowa to examine the types of occupant and their characteristics (e.g. income, educational levels) and household factors (e.g. residence type, ownership, structural and maintenance information, fire safety equipment) associated with self-reported fires. Just over 11% of these households reported a residential fire in the time period under investigation and the researchers found significant relationships with occupant characteristics (e.g. presence of a person in the household with alcohol problems) and factors such as presence of a fire extinguisher. Drawing upon the body of existing literature on socio-economic research, our previous research (Corcoran, Higgs, Brunsdon, & Ware, 2007; Corcoran, Higgs, Brunsdon, Ware, & Norman, 2007) used a combination of spatial statistical techniques and regression analysis in order to explore the association between socioeconomic factors (derived from census-based variables) and different types of fire incidence for a fire and rescue service in South Wales, UK. That research demonstrated a correlation between such incidence and deprivation, with less deprived areas being associated with the ‘lower risk’ categories for each of the incident type studied (household fires, vehicle fires, secondary fires and malicious false alarms). However, there were interesting differences between the levels of association with socio-economic variables and incident types which, we suggested, warranted further study. The area under investigation for that study included a mixture of some of the most deprived and affluent communities in Wales, and noticeable differences in trends were found for example between rural and urban areas which had divergent deprivation profiles (based on the Townsend Index e a standard measure of deprivation in the UK e Townsend, Phillimore, & Beattie, 1988). Further study of “unconstrained” variables from the 2001 census (the selection of which was informed by a review of the studies referred to above) using binomial regression associations between different incident types revealed the importance of factors such as areas with lower proportions of white residents and with residents that had lower levels of educational attainment (in relation to property fire call-outs), census wards with higher than average numbers of overcrowded households (in relation to vehicle fire call-outs) and those with lower proportions of car owners (in relation to false alarm call-outs). Here, an overcrowded household was defined as households with more than one person per room (after Duncanson et al., 2002). Many of the factors identified in that particular study (such as the importance of the levels of educational attainment of people living in those census tracts) mirrored those found in a larger (2000e2001) national study of 24 fire and rescue services in England and Wales (ODPM, 2004). In our previous research we suggest that the lack of a common spatial scale of analysis in such studies limits the transferability of many of these findings (Corcoran, Higgs, Brunsdon, Ware, &
67
Norman, 2007). Jennings (1999) cited research that suggests that the relationship between census population and housing variables and fire incidence is scale dependent and that this needs to be taken into account in comparative research. In this paper we aim to address this by adopting the smallest units for which socioeconomic data is available from the respective censuses whilst recognising that there are some factors that may be associated with fire incidence that are not currently captured in the respective censuses. Specifically we build on such studies using regression approaches to investigate spatial patterns through the identification of potential relationships of census-derived variables and fire incident data for the cities of Cardiff and Brisbane. The methodology we have adopted is described in more detail in the next section. Methodology Study areas Fig. 1 shows the socio-economic variation across each of our study areas using the Townsend index for Cardiff and the SocioEconomic Indexes for Areas (SEIFA) for Brisbane. The SEIFA index of relative socio-economic disadvantage e one of the four SEIFA indexes compiled by the Australian Bureau of Statistics (ABS) e was used to represent the degree of relative advantage to disadvantage (ABS, 2001b). Both the Townsend and SEIFA indices attempt to capture socio-economic characteristics across census regions and thus are used here by way of comparison. Whilst it is recognised that each index is founded upon discrete methodologies and incorporates different census variables, the use of these indices does provide a contextual background to the analysis of residential fires in both cities and both have broader policy relevance. In Cardiff, for example, the principal pockets of deprivation are in the some of the suburbs of the east and west of the city but even within the areas redeveloped through regeneration schemes in the docklands of the city, where the National Assembly for Wales is situated, there are some of the most deprived census areas in the whole of Wales. The northern suburbs in contrast tend to be dominated by more affluent census areas which account for some of the most affluent in Wales on a range of Government-based deprivation indicators. The heterogeneous nature of urban form is mirrored in Brisbane where areas of greatest disadvantage exist south west of the city centre extending linearly in addition to pockets of disadvantage located east of the city centre. In contrast, the most socioeconomically advantaged areas in Brisbane are largely located in the western suburbs. Detailed description of the database Fire incidence databases Disaggregate databases of call-out types were examined for building fires, vehicle fires, secondary fires and malicious false alarms (hoax calls). Building fires include all fires involving property for example, dwellings, public buildings and workplaces. Secondary fires include fires in derelict buildings, refuse and refuse containers, outdoor structures, for example, fence, gate and road signs and grassland fires. The final incident type investigated, malicious false alarms categorises false alarms deemed malicious or deliberate. Incident types matching these descriptions were extracted for both Cardiff and Brisbane as these could be best matched across the two formerly disparate command and control databases. However, it is recognised that some differences in reporting may result in differences in the summary statistics presented in Table 1. For example, fires in derelict buildings are classified as secondary fires under the categorisation used by the fire
68
J. Corcoran et al. / Applied Geography 31 (2011) 65e75
Cardiff
#
0
2
4
6 Kilometers
Cardiff Docks
Townsend index Most Deprived Next Deprived Median Next Affluent Most Affluent No value Central Business District Main road
Brisbane Port of Brisbane #
N 0
5
10
15
20 Kilometers
SEIFA Index Most Disadvantaged Next Disadvantaged Median Next Advantaged Most Advantaged No value
Fig. 1. Socio-economic variations in Brisbane (SEIFA disadvantage measure) and Cardiff (Townsend deprivation index) at CD and OA level respectively using quintiles.
service operational in the Cardiff case study area. Under the categorisation used by the Queensland Fire and Rescue Service (QFRS), the occupancy or structural soundness of the building is not used for incident classification. Consequently any fires in derelict buildings in Brisbane are classified as structural fires. Organisational practice can also influence coding, even when the database options are the same across the two jurisdictions. For example, if an alarm is activated, yet on arrival there is no fire and no witness to the activation, it can be extremely difficult to accurately ascertain the cause of the activation. In such circumstances, organisational policy may influence an officer’s coding decision, particularly with regard to incidents where the public may be charged for call-outs. In Queensland, false alarms can be billable calls when the alarm is activated due to certain circumstances under the occupier’s control. This can lead fire officers to err on the
side of caution and code a false alarm as malicious or as accidental, in cases where they cannot pinpoint the actual cause of activation. Both the Cardiff and Brisbane call-outs cover the period 1st January 2000e31st December 2004 with a date, time, grid reference and incident type recorded for each incident. There are important differences between the two sources of fire incident data. For example, with regard to the geographical referencing of incidents, hand-held GPS have been used since 2002 to create the database of fire call outs in Cardiff so there have been improvements in later years in the accuracy of the databases (there are still acknowledged problems with some locational referencing however; see, Corcoran, Higgs, Brunsdon, Ware, & Norman, 2007 for further details). However, for the Brisbane area, fire officers are required to interpret a map and create a thirteen-digit geo-code manually. Naturally, this leaves the geo-coding of the Brisbane data
J. Corcoran et al. / Applied Geography 31 (2011) 65e75 Table 1 Comparison between Brisbane and Cardiff (census and fire incidence variables). Brisbane
Cardiff
Population (2001 census) Area (hectares) Mean area of census areas (hectares) Number of census areas with valid population Mean population per census area
888,252 132,350 78.82 1667 533
305,164 14,030 14.16 990 308
Annual building fires (mean rate per 1000 population) Median annual building fires per 1000 population Range of annual building fire rates
0.75 0.40 0e53.33
1.99 1.09 0e31.93
Annual vehicle fires (mean rate per 1000 population) Median annual vehicle fires per 1000 population Range of annual vehicle fire rates
0.85 0 0e238.46
2.76 0.69 0e82.63
Annual secondary fires (mean rate per 1000 population) Median annual secondary fires per 1000 population Range of annual secondary fire rates
4.36
7.17
0.91 0e906.67
2.18 0e170.28
Annual malicious false alarm (mean rate per 1000 population) Median annual malicious false alarm per 1000 population Range of annual malicious false alarm rates
0.43
1.19
0
0
0e79.11
0e91.08
open to potentially more error that could vary over time. For both data sets there are a portion of records that are either missing or have an incorrect geo-code, which varies by incident type. For the Cardiff data the largest is for hoax calls equating to 2.57% (n ¼ 31) of call-outs and 8.17% (n ¼ 333) for building fires in the Brisbane data. In this study we have not attempted to fully evaluate the implications of records that are missing a geo-code; the volume of those records missing a spatial identifier was not deemed to significantly affect the analysis presented in this paper. Table 1 compares the fire incident rates for the two cities. Overall, Cardiff has a significantly higher annual rate of building fires than Brisbane which has over twice the population although the figures do hide important spatial variations within the respective cities. Similarly, the annual vehicle fires rate for Cardiff is over three times that of Brisbane, secondary fires almost twice as large and malicious false alarms over twice as large for Cardiff. However, as the range of annual rates illustrates, these ‘global’ figures need to be examined in the context of intra-urban socioeconomic characteristics and in the next section we outline how census data can be used to contextualise patterns of fire incidents in both cities. Census data An important element of the study is the use of census-based variables in order to examine potential relationships with the numbers of fire incidents. This was guided to a certain extent by previous research that had identified census-based variables that were associated with levels of fire incidence but also involved the analysis of policy-based deprivation measures that have been used in wider contexts within the UK and Australia (Chandler et al., 1984; Corcoran, Higgs, Brunsdon, Ware, & Norman, 2007; Duncanson et al., 2002; Jennings, 1999). This is not the place to describe in detail the characteristics of the two census collection procedures (see for example, Rees et al. (2002) for the UK Census of Population and Australian Bureau of Statistics (ABS, 2001a) for details on the Australian Census of Population and Housing). The latest Australian Census was conducted in August 2006; however, to try to ensure compatibility, data were collated from both the 2001 UK and Australian Censuses. Throughout this analysis we have sought to ensure compatibility both in terms of spatial units and census variables taken from
69
the 2001 censuses held in the UK and Australia. In order to try to ensure compatibility we have used the smallest spatial units for which information is available from the 2001 census conducted in both countries. In the UK, output areas (OAs) in the UK constitute the lowest-level spatial unit, roughly approximating to Census Blocks in the US census hierarchy, and have a target household count of 125 and an average population of 300. In Australia data is available at the Census Collection District (CD) level which on average covers approximately 220 dwellings; although from 2011 the smallest units will be SA2s (Statistical Area 2) approximating to between 30 and 60 dwellings (Hugo, 2007). The relationship between socio-economic characteristics of the resident population, and the annual rate of call-outs per 1000 population was examined at the CD and OA levels for both Brisbane and Cardiff respectively. Each call-out was matched to a CD or OA using the geo-coded location from the original fire report and the annual rate of call-outs was calculated by averaging the number of call-outs responded to by fire agencies in the 2000e2004 calendar years, and dividing by the resident population as at the time of the 2001 census. Given access to different time frames for the fire incident data would permit a better alignment to the census data, for example incident data spanning 1999e2003 matched to the 2001 census and 2001e2006 matched to the 2006 census. However, the data provided by each of the fire services represents the best possible match to census data given the UK census is conducted every ten years (the most recent census being 2001) in comparison to the five yearly census in Australia. The effect of socio-economic disadvantage on the rate of all the analysed call-out types can be firstly demonstrated by the use of two composite census variables, the Socio-Economic Indexes for Areas (SEIFA) Index (Brisbane) and the Townsend Index of Multiple Deprivation (Cardiff). The SEIFA index consists of four indexes developed by the Australian Bureau of Statistics and derived from the Census of Population and Housing (ABS, 2001b) and has been used as a measure of advantage/disadvantage in a number of areas, particularly related to health applications (Armfield, 2007; Cass, Cunningham, Wang, & Hoy, 2007; Walker & Becker, 2005). The SEIFA index was initially developed in 1990 and configured to reflect the general socio-economic well-being of a neighbourhood that includes variables such as income and occupation to categorise areas as relatively advantaged or disadvantaged. The Townsend index of multiple deprivation (Townsend et al., 1988) was first developed using data from the 1991 UK census with a focus on those aspects of material deprivation postulated to be linked to health variations. The Index has similarly been used in the UK in a range of application areas (see for example, Morgan, Ahmed & Kerr, 2000; Prendergast, Beal, & Williams, 1997; Watson, Cowen, & Lewis, 1996) and is constructed from small area census data on unemployment, non-home ownership, non-car ownership and household overcrowding (Senior, 2002). We have used both the SEIFA and Townsend measures at the lowest level of census geography in a preliminary analysis to describe patterns of fire incidence in relation to patterns of socio-economic deprivation/disadvantage in the two study areas. The disaggregate census variables used to create each of the indexes are listed in Table 2. In addition, to these disadvantage/deprivation measures, ‘unconstrained’ socio-economic predictor variables were sourced from the UK and Australian censuses, both conducted in 2001. Due to the large number of variables and the high levels of correlation between them, principal components analysis was used to reduce the number of variables in the model. For each of the six socioeconomic areas of interest chosen as potential predictors (family structure, educational qualifications, number of cars, home ownership status, property type and ethnicity), the censuses provide between four and seven individual variables with which to
70
J. Corcoran et al. / Applied Geography 31 (2011) 65e75
Table 2 Variables used in the construction of the SEIFA and Townsend indexes. SEIFA % People aged 15 years and over with no post-school qualifications % Occupied private dwellings with no internet connection % People with stated annual household equivalised income between $13,000 and $20,799 (approx. 2nd and 3rd deciles) % Employed people classified as Labourers % Households paying rent less than $120 per week (excluding $0 per week) % People aged under 70 who have a long-term health condition or disability and need assistance with core activities % Employed people classified as Machinery Operators and Drivers % People (in the labour force) unemployed % One parent families with dependent offspring only % Households renting from Government or Community organisation % Employed people classified as Low Skill Community and Personal Service Workers % Occupied private dwellings requiring one or more extra bedrooms (based on Canadian National Occupancy Standard) % Occupied private dwellings with no car % Occupied private dwellings with four or more bedrooms % People aged 15 years and over at university or other tertiary institution % Households paying mortgage greater than $2120 per month % Households paying rent greater than $290 per week % People aged 15 years and over with an advanced diploma or diploma qualification % Employed people classified as Professionals % Occupied private dwellings with a broadband internet connection % People with stated annual household equivalised income greater than $52,000 (approx 9th and 10th deciles)
education (final year schooling and post-school study) loaded most highly. The differences are reflected in the naming of the variable descriptors. When attempting to create a composite variable for ethnicity, the difficulties in data matching were more pronounced, due to the census questions asked. The UK census asked questions about ethnicity which related to ‘race’ ethnicity, whilst the Australian census questions related to ‘country of origin’ ethnicity. For Brisbane, Australian, NZ or European born persons (very broadly generalised as ‘white’) loaded most highly on the variable ethnic. In contrast, persons identifying their racial ethnicity as white had a strong negative loading on Cardiff’s ethnic variable. Statistical approaches The aim of the regression analysis is to establish if the socioeconomic factors associated with fire incidents are different in the Australian context than in our previous study of South Wales (Corcoran, Higgs, Brunsdon, Ware, & Norman, 2007) and for a subset of that database, i.e. for the city of Cardiff. Being count data, it
Table 3 Variables entered into principal components analysis to create composite variables. Composite variable
Brisbane variables
Cardiff variables
Family
Couple family, no children Single parent w/children under 15 Single parent w/dependent students (15e24) Couple family w/children under 15 Couple family w/dependent students (15e24) No vehicles in household 1 vehicle in household 2 vehicles in household 3 or more vehicles in household No formal qualifications Year 10 or equivalent Year 12 or equivalent Certificate or advanced diploma level Degree or postgraduate level
Couple family, no children Couple family, dependent children Couple family, no dependent children Single parent, dependent children Single parent, no dependent children No vehicles in household 1 vehicle in household 2 vehicles in household 3 or more vehicles in household No formal qualifications Level e.g. 1 þ O-level passes (but less than 5) Level 2 e.g. 5 þ O-level passes Level 3 e.g. 2 þ A-levels, higher school cert Level 4 or 5 e.g. degree, higher degree, nurse Owned outright Purchasing (incl rent-to-buy & shared) Rent from council/state housing scheme Non-govt renting (includes private & social)
Townsend % Economically active people unemployed % Households with more than one person per room % Households with no car % Households not owner-occupied Cars
describe a household. For example, the Australian census provides four different variables which measure the number of cars in a household (no vehicle, one vehicle, two vehicles, three or more vehicles), and within a given geography, household responses to these variables tend to be correlated. A composite variable was generated by simply taking a single principal component from each group using the weights of the first component created from the principal components analysis. Using this approach has the effect of overcoming the within group correlation e for example, increases in the percentage of two car households must then be associated with an overall reduction in the percentages in other categories of cars per household (Corcoran, Higgs, Brunsdon, Ware, & Norman, 2007). Table 3 lists each the variables entered into the principal components analysis for each composite variable. Using this process on each of the six sets of socio-economic variables, we created six composite variables for both Brisbane and Cardiff, which were then used as predictors in the regression models. The census variables selected were matched between Brisbane and Cardiff as closely as possible; however, in order to do so, the composite variables used in this analysis differ from those constructed in our previous paper (Corcoran, Higgs, Brunsdon, Ware, & Norman, 2007), and are not strictly comparable. Tables 4 and 5 show the loading structure of the most important census variables onto the composite variables. Four of the six composite variables showed similar loading structures between Brisbane and Cardiff, although the levels of variance explained by the first factor differed between the data sets. Two variables (quals and ethnic) did not match between the two data sets as well as the other four. In the variable quals, the loading structure differed between locations; for Brisbane, all levels of formal education loaded highly on quals, whilst for Cardiff higher
Quals
Tenure
Accom
Ethnicity
Owned outright Purchasing (incl rent-to-buy & shared) Rent from council/state housing scheme Non-govt renting (includes private & social) Other tenure type Unoccupied premises Detached house Semidetached (incl townhouse, terrace) Flat, unit or apartment Other (incl vans, mobile or temporary) In or attached to commercial building Born in Australia Born in New Zealand Born in NW or SE Europe Born in North Africa or Mediterranean Born in NE, SE or Southern & Central Asia Born in the Americas Born in sub-Saharan Africa
Unoccupied premises Detached house Semidetached (incl. townhouse, terrace) Flat, unit or apartment Other (incl. vans, mobile or temporary) In or attached to commercial building White Mixed race Indian, Pakistani, Bangladeshi, other Asian Caribbean, African, other Black Chinese, other ethnic group
J. Corcoran et al. / Applied Geography 31 (2011) 65e75 Table 4 Brisbane e results of principal components analysis (Three highest loadings on first component reported). Variable group name (% of variance explained) Family Couple families (55.71%)
Cars High car ownership (50.47%)
Quals Formal education (52.36%)
Tenure High owner occupier rate (33.78%)
Accom Flats & unoccupied (32.73%) Ethnicity White (43.48%)
Variable name
Variable description
Loading
cf_u15 cf_stu couple_noc
Couple family with dependent children under 15 Couple family with dependent students 15e24 Couple family no children
0.5114 0.4582 0.4828
no_vehicle 2_vehicles 3_or_more
No vehicles in household 2 vehicles per household 3 or more vehicles per household
0.2185 0.6758 0.6466
yr_10 yr_12 cert_or_ adv_dip degree
Achieved Year 10 or equivalent education Achieved Year 12 or equivalent education Achieved Certificate 1e4 or Advanced Diploma Degree
0.4500 0.5287 0.5678 0.4312
owned purchasing other
Fully owned Being purchased (including under rent/buy scheme) Other tenure type
0.6767 0.6137 0.2910
unoccupied detached flat
Unoccupied private dwelling Separate house Flat, unit or apartment
aus nz euro
Born in Australia Born in New Zealand Born in NW or SE Europe
0.5246 0.4689 0.6442 0.4672 0.3769 0.4911
was assumed that the best fitting distribution for both the Brisbane and Cardiff building fire rate data would be either the Poisson or the Negative Binomial distribution. For each of the four call-out types (building fires, vehicle fires, secondary fires and malicious false alarms) a regression model was fitted using the six composite census variables as predictors of the average number of incidents attended annually. A Negative Binomial regression was conducted, and the data also assessed against a Poisson distribution for comparison of the dispersion parameter alpha against zero. When alpha equals zero, the Negative Binomial reduces to the Poisson distribution. For both locations, the likelihood-ratio test of alpha ¼ 0 was significant (p < .0001), indicating that the data were over-dispersed, and better described by a Negative Binomial rather than a Poisson distribution. The Corcoran, Higgs, Brunsdon, Ware, and Norman (2007) study demonstrated the findings from the negative binomial model to be more reliable in terms of modelling the spatially aggregated count data. As such we apply a negative binomial to the data described in the comparative study described here. Results
resident population in Brisbane. Fig. 2 shows the association between higher levels of disadvantage and increased number of call-outs. The indices have been divided into quintiles, and the direction of the Advantage Disadvantage Index has been changed so that, consistent with the Townsend Index, lower scores indicate lower rates of disadvantage. Additionally, the outer values have not been graphed, and the scales have been rendered the same, for clarity. However, the trends are common to both cities, i.e. there is an increase in incidents per 1000 population with deprivation for all types with the SEIFA (Brisbane) and Townsend (Cardiff) measures. Regression analysis The regression model was fitted for four of the incident types. The robust Negative Binomial regression models predicting the rate of each of the incident types per 1000 population from the six composite variables from Tables 4 and 5 were statistically significant for both the Brisbane and Cardiff datasets (Table 6). A robust regression was chosen to account for correlation between predictor variables. To assess the strength of the relationships between the set of predictor variables and the response variable, when using ordinary least squares (OLS) regression, one would usually interpret R-squared. Negative Binomial regression does not provide estimates for R2 in the same way as OLS regression, rather it provides McFadden’s Pseudo R2, which is calculated as:
Table 5 Cardiff e results of principal components analysis (three highest loadings on first component reported). Variable group name (% of variance explained) Family Couple families (35.23%)
Cars High car ownership (60.07%)
Quals High Education levels (42.78%)
Variable name
Variable description
sp_dep cf_dep couple_noc
Single parent family with dependent children Couple family with dependent children Couple family no children
0.4135 0.5182 0.5779
no_vehicle 2_vehicles 3_or_more
No vehicles in household 2 vehicles per household 3 or more vehicles per household
0.5993 0.6027 0.5099
no_quals lvl_1 lvl_4_or_5
No formal qualifications Level 1 e.g. 1 þ O-level passes (but less than 5) Level 4 or 5 e.g. degree, higher degree, nurse
0.5707 0.3723 0.6196
Fully owned Being purchased (including shared ownership) Rent from private landlord or housing association
0.5400 0.5425 0.5169
Tenure High Owner Occupier owned rate (49.51%) purchasing rent
Accom Flats & unoccupied (31.93%)
Exploratory analysis A preliminary investigation involved an exploratory analysis of spatial trends of fire incidence and socio-economic characteristics. One multivariate outlier was dropped from the analysis for each of the Brisbane and Cardiff study areas, along with 11 CDs with zero
71
Ethnicity Non-white (43.39%)
unoccupied semidet flat
Loading
Unoccupied private dwelling 0.5179 Semidetached (including 0.4836 terrace housing) 0.6200 Flat, unit or apartment
Asian Indian, Pakistani, Bangladeshi, 0.4756 Black other Asian 0.5076 Chinese_other Caribbean, African, other Black 0.4813 Chinese, other ethnic group
72
J. Corcoran et al. / Applied Geography 31 (2011) 65e75 Brisbane - Building Fire rate by Disadvantage
Cardiff - Building Fire rate by Disadvantage
10
10
8
8
6
6
4
4
2
2
0
0 1
2
3
4
5
1
SEIFA Index of Advantage Disadvantage quintiles: lower score indicates less disadvantage
2
3
4
5
Townsend Index of Multiple Deprivation quintiles: lower score indicates less deprivation
Brisbane - Secondary Fire rate by Disadvantage
Cardiff - Secondary Fire rate by Disadvantage
40
40
35
35
30
30
25
25
20
20
15
15
10
10
5
5
0
0 1
2
3
4
5
1 2 3 4 Townsend Index of Multiple Deprivation quintiles: lower score indicates less deprivation
SEIFA Index of Advantage Disadvantage quintiles: lower score indicates less disadvantage
5
Cardiff - Vehicle Fire rate by Disadvantage
Brisbane - Vehicle Fire rate by Disadvantage 15
15
10
10
5
5
0
0 1
2
3
4
1
5
SEIFA Index of Advantage Disadvantage quintiles: lower score indicates less disadvantage
2
3
4
5
Townsend Index of Multiple Deprivation quintiles: lower score indicates less deprivation
Brisbane - Malicious False Alarm rate by Disadvantage
Cardiff - Malicious False Alarm rate by Disadvantage
8
8
6
6
4
4
2
2
0
0 1
2
3
4
5
SEIFA Index of Advantage Disadvantage quintiles: lower score indicates less disadvantage
1
2
3
4
5
Townsend Index of Multiple Deprivation quintiles: lower score indicates less deprivation
Fig. 2. Correlation between fire indices and indices of deprivation (SEIFA for Brisbane and Townsend for UK).
(1 e log-likelihood (model)/log-likelihood (constant only))
(1)
Pseudo R2 does not have the same meaning as OLS regression R2, and should not be interpreted as the proportion of variance in the response variable explained by the predictor variables. Pseudo R2 values are considerably lower than their OLS R2 counterpart, and values between 0.2 and 0.4 would indicate an excellent fit
(McFadden, 1979). Table 6 shows the Pseudo R2 statistic for each analysis, which is used to demonstrate the relative predictive strength of the models for the two locations. Not all predictor variables in the regression model were individually statistically significant. The Chi2 and Pseudo R2 results for the final robust negative binomial regression models using only the individually significant variables, are shown in Table 7. All Cardiff
J. Corcoran et al. / Applied Geography 31 (2011) 65e75 Table 6 Results of robust negative binomial regression predicting rate of fires from the six composite variables presented in Tables 4 and 5. Brisbane Building fire rate
Secondary fire rate
Vehicle fire rate
Malicious false alarm rate
Chi2 (6) ¼ p < .0001 Pseudo R2 Chi2 (6) ¼ p < .0001 Pseudo R2 Chi2 (6) ¼ p < .0001 Pseudo R2 Chi2 (6) ¼ p < .0001 Pseudo R2
Cardiff 102.70 ¼ 0.0712 106.18 ¼ 0.0940 96.61 ¼ 0.1405 95.01 ¼ 0.0947
Chi2 (6) ¼ p < .0001 Pseudo R2 Chi2 (6) ¼ p < .0001 Pseudo R2 Chi2 (6) ¼ p < .0001 Pseudo R2 Chi2 (6) ¼ p < .0001 Pseudo R2
225.25 ¼ 0.0671 327.50 ¼ 0.0613 189.14 ¼ 0.0701 93.13 ¼ 0.0710
Table 7 Results of negative binomial regression (Chi2 and pseudo R2statistics) using individually significant variables only.
Building fire rate
Secondary fire rate
Vehicle fire rate
Malicious false alarm rate
Brisbane
Cardiff
Chi2 (3) ¼ 93.51 p < .0001 Pseudo R2 ¼ 0.0699 Chi2 (3) ¼ 101.22 p < .0001 Pseudo R2 ¼ 0.0908 Chi2 (5) ¼ 92.46 p < .0001 Pseudo R2 ¼ 0.1400 Chi2 (3) ¼ 73.67 p < .0001 Pseudo R2 ¼ 0.0916
Chi2 (4) ¼ p < .0001 Pseudo R2 Chi2 (3) ¼ p < .0001 Pseudo R2 Chi2 (4) ¼ p < .0001 Pseudo R2 Chi2 (4) ¼ p < .0001 Pseudo R2
210.59 ¼ 0.0665 330.82 ¼ 0.0602 174.84 ¼ 0.0697 68.35 ¼ 0.0687
incident types are predicted by quals, tenure and accom, with cars being significant also for rates of building fires and hoax calls. Family and ethnic are not significant individual predictors for any Cardiff incidents. Results were not as consistent for Brisbane, where no one variable was significant for all incident types, although cars, quals and accom were each significant for three out of four incident types. Table 8 shows the regression coefficients for each of the significant variables in the analyses. The accom variable behaves very differently between locations. When it is a significant predictor for Brisbane incident types, it is always a negative coefficient, indicating that higher rates of incidents are associated with lower rates of flats and unoccupied dwellings (as compared to detached and semidetached dwellings). Accom is a significant predictor for each of the Cardiff incident types, but here, higher rates of incidents are associated with higher rates of flats and unoccupied dwellings. This it is postulated may reflect the different patterns of urbanisation and housing stock between Brisbane and Cardiff.
73
Some interesting comparisons can be made in terms of different fire incident types. Firstly in terms of building fires, the individual predictors cars, tenure (p < .05 for each) and quals (p < .005) were significant for Brisbane. Higher rates of building fires were therefore associated with lower rates of car ownership, lower rates of formal education, and lower rates of owner occupation. Thus CDs with lower levels of engagement with education were more prone to building fires, a finding that mirrors trends in our previous work in South Wales and with other papers reviewed in Section Two of the paper. For Cardiff, the individual predictors cars, tenure (p < .05 for each), quals (p < .005) and accom (p < .001) were significant and higher rates of building fires were associated with lower rates of car ownership, lower rates of higher education, lower rates of owner occupation, and higher rates of flats and unoccupied dwellings. In terms of secondary fires, the individual predictors quals, accom (p < .001 for each) and ethnic (p < .05) were significant for Brisbane. Higher rates of secondary fires were associated with lower rates of formal education, lower rates of flats and unoccupied dwellings, and higher rates of white population. For Cardiff, the individual predictors quals, tenure and accom (p < .001 for each) were significant and higher rates of secondary fires were associated with lower rates of higher education, higher rates of flats and unoccupied dwellings, and lower rates of owner occupation. There are some interesting contrasts in the trends for vehicle fires for the two cities, in that the individual predictors quals (p < .001), family (p < .005), cars, tenure and accom (p < .05 for each) were significant for Brisbane with higher rates of vehicle fires associated with lower rates of couple families, higher rates of car ownership, lower rates of formal education, lower rates of home ownership, and lower rates of flats and unoccupied dwellings. In contrast, for the Cardiff data, the individual predictors cars (p < .005), quals, tenure and accom (p < .001 for each) were significant with higher rates of vehicle fires associated with higher rates of car ownership, lower rates of higher education, lower rates of owner occupation, and higher rates of flats and unoccupied dwellings. These trends confirm our findings from previous analysis of the wider South Wales area in which we found that wards (larger spatial units of approximately 1500 people) that had higher car ownership rates were more prone to these types of call-outs. It also confirms the importance of the educational variable in associations with vehicle fire call-outs. Finally, in terms of malicious false alarms, for Brisbane, the individual predictors cars (p < .001), accom (p < .05) and ethnic (p < .005 for each) were significant with higher rates of hoax calls associated with lower rates of car ownership, lower rates of flats and unoccupied dwellings, and higher rates of white population. In contrast, the individual predictors cars (p < .005), quals, tenure and accom (p < .001 for each) were significant for Cardiff with higher rates of hoax calls were associated with lower rates of car ownership, lower rates of higher education, lower rates of owner occupation, and higher rates of flats and unoccupied dwellings. Again,
Table 8 Regression coefficients and standard errors (for significant variables). Brisbane Building Family Cars Quals Tenure Accom Ethnic Constant
0.1439 (.0585) .1874 (.0648) 0.1119 (.0470)
Cardiff Secondary
Vehicle
0.6375 (.1008)
0.2918 0.4139 0.5061 0.2646 0.1691
0.2492 (.0412) .1744 (.0851) 0.4844 (.0324)
0.8959 (.0634)
(.0964) (.1762) (.1132) (.1279) (.0777)
0.8889 (.0614)
Hoax
Building
0.9129 (.1170)
0.1438 0.1159 0.1196 0.2099
0.1473 (.0605) 0.2083 (.0697) 1.3120 (.0933)
Secondary
Vehicle
(.0619) (.0399) (.0562) (.0314)
0.4932 (.0405) 0.2340 (.0482) 0.2179 (.0421)
0.2548 0.5992 0.4319 0.2207
0.5057 (.0423)
1.6103 (.0510)
Hoax (.0893) (.0698) (.0885) (.0507)
0.6268 (.0607)
0.2093 0.2029 0.2095 0.1979
(.1038) (.0784) (.0911) (.0694)
0.1733 (.1009)
74
J. Corcoran et al. / Applied Geography 31 (2011) 65e75
many of these variables may act as a more general proxy for deprivation with, for example, OAs with lower educational attainment and low car ownership more prone to this type of crime in common with some of the trends observed for Brisbane. The contrast, however, between the relationship with flats and unoccupied dwellings is an interesting one which possibly reflects the respective residential geographies of the two cities and possibly that of the chosen housing preferences of residents in Brisbane and Cardiff. Discussion In our previous research we presented an exploratory analysis of the incidence of call outs for the fire service for an area of South Wales (Corcoran, Higgs, Brunsdon, Ware, & Norman, 2007). Many of the patterns that emerged reflected both those of findings from previous studies and a wider national study conducted in the UK for 2000 and 2001 (ODPM, 2004). In our earlier paper we called for more comparative research with other cities/regions to see if the methodologies and findings related to the association between socio-economic variables and fire incidence were transferable to other contexts. We also suggested that such research should ideally be conducted at the most detailed spatial scales possible. In this study, we have attempted to address both strands of these research aims by providing a comparison of trends in the spatial distribution of fire incidence by socio-economic circumstances for Cardiff (UK) and Brisbane (Australia). This has involved detailed point level analysis of fire incidence data and the use of GIS and regression techniques to examine potential association with the socioeconomic geography of both cities. This in turn has revealed some important similarities as well as differences in associations, for example in relation to the residential patterns of the two cities which we posit is reflective of the respective wider urban geography of Cardiff and Brisbane. Whilst the links with measures of deprivation (both policy-based) and census proxies was perhaps to be expected, the study of spatial trends revealed differences in associations between call-outs and socio-economic factors which could provide the focus for more detailed contextual studies taking into account a wider range of non-census based variables and detailed case histories of the fire events themselves such as the cause of the fire. The comparative study has also highlighted a number of significant methodological concerns. Whilst some were evident in our previous study and were discussed in that research, this investigation has drawn attention to the problems faced by researchers using secondary spatial data in comparative contexts. Such factors are almost inevitable given that these are datasets based on the operational data collection procedures in place within these respective fire services. In addition, a limitation of the current analysis of building fires is that the response variable is the rate of all building fires per 1000 people, not just residential building fires, whilst all of the predictor variables relate to characteristics of the resident population. Given that industrial and commercial premises tend to cluster within cities, the rate of non-residential building fires would also be expected to vary by census or output area. The influence of the socio-economic characteristics of the resident population, on the rate of non-residential fires, is likely to be small at best. Therefore, it is reasonable to expect that the model for building fires would be strengthened if fires in non-residential premises were excluded from the analysis. This in turn highlights the problems of using exactly comparable fire incidence data (both in terms of the respective accuracy of geo-coding of events and the classification of fire types) but also the difficulties faced in using census data for the two countries which are not directly comparable. Another area of concern is the issue relating to the organisational differences in reporting between the two fire services. In an
attempt to reduce the effect we selected incident categories that permitted the best match, however it is recognised that some differences in incident volumes are inevitable. Both the identification and quantification of these differences are beyond the scope of the paper and their investigation reliant on obtaining further more detailed incident data describing the specific nature of each incident to more fully address this issue. In this paper we have attempted to use broadly comparable census variables from the two censuses for 2001 but this has not always proven to be possible, as the individual census questions do vary. For example, different educational structures in each location mean that we cannot strictly map the two census variables, as the high school certification processes are measured in number of A and GCSE levels in Cardiff, and number of grades completed in Brisbane. Under Principal Components Analysis, the loading structure of the quals variable also differed between locations; for Brisbane, the loading was highest on all formal education, whereas for Cardiff quals loaded most highly on higher education. These differences are a reflection not only of the different educational structures in place, but also potentially, of broader socio-economic differences. We have used the most detailed geography from both but again these are not exactly comparable. We note also that the two countries have different rules regarding the conduct of the two censuses with, for example, concerns expressed in some contexts of the under-count in urban areas in particular in the UK at the time of the 2001 census and the compulsory nature of participation in the census taking process in Australia. Nevertheless many of the issues identified by Hugo (2007) are common to both countries, for example the need to calculate daytime populations (arguably even more important in assessing population ‘at risk’ in the case of daytime fire events), the need to take into account temporary or migrant workers (again important in the context of the present study in terms of areas of possible spatial segregation) and the use of arbitrary administrative boundaries in the census collection and reporting process. We should at this stage also repeat that these are two urban conurbations and that more research would be needed to establish if such factors are also important in rural scenarios. Despite the problems, there are important questions that are raised that could be the focus for future research. For example, whilst the associations between rate of incidents and indices of deprivation/disadvantage is consistent across call-out types, the regression model does not explain why Cardiff has a mean incident rate which is more than twice that of Brisbane. The difference may be due to a difference in the actual number of fires occurring which require fire service attendance, or due to a reporting bias. Possible explanations to be explored in future research include differences in: age, design or construction of buildings; climate and the need for heating; smoke alarm penetration rates; coding classifications by fire services; community fire safety awareness. The construction of a common deprivation index could also help to determine whether the difference in call-out rates is partly due to differences in degrees of socio-economic deprivation/disadvantage between Brisbane and Cardiff. As discussed previously, it is also reasonable that the building fire rates may differ due to the influence of non-residential fires in the model. Finally it would be useful if such trends could be examined in the light of spatial and temporal patterns of initiatives such as safety campaigns, educational programmes and access to smoke alarms should such data become available in the future at the small area level (see for example Zhang et al., 2006). Conclusion This study is part of our on-going research into the association of fire incidence and socio-economic conditions in urban areas
J. Corcoran et al. / Applied Geography 31 (2011) 65e75
using spatial and statistical techniques. This research should have benefits for policy makers charged with identifying populations at risk, designing possible intervention strategies such as safety campaigns and in implementing educational programmes. Whilst it is difficult to establish the root causes through ecological studies of this nature we contend that the types of analysis reported in this paper have the potential to contribute to these overall aims through investigations of the relationships between the types of fire incidents collected by fire and rescue services and both census (and non-census based) measures of socio-economic circumstances. Mapping can be used to help spatially target interventions through the visualisation of salient area-based characteristics that have been established through statistical modelling to be associated with higher than average levels of fire incidence. The production of these maps may also be of use to strategic planners as a tool to help allocate finite resources to geographically specific areas. However, such comparative research is not without its problems as we highlight above; data incompatibility has prevented a complete like-for-like analysis and avenues for further research in this area are highlighted. Nevertheless, within these caveats, the research conducted to date has revealed variations in the types of factors associated with fire incidence in urban areas, identified important differences in such correlations in relation to different fire incident types and suggested possible reasons for such trends based on the respective residential patterns of Cardiff and Brisbane and the prevailing patterns of deprivation within the two cities. Acknowledgements This paper is based on research conducted on a project funded by the Australian Research Council Linkage program grant #LP0883861. We would like to thank the South Wales Fire and Rescue Service and the Queensland Fire and Rescue Service for access to the data on which the paper is based. However, the interpretations of the analysis are solely those of the authors and do not necessarily reflect the views and opinions of the Services or any of their employees. The work uses UK 2001 Census of Population data and GIS boundary data obtained via MIMAS CASWEB and EDINA UKBORDERS; academic services both supported by ESRC and JISC. These data are copyright of the Crown and are reproduced with permission of the Controller of HMSO. References Allareddy, V., Peek-Asa, C., Yang, J., & Zwerling, C. (2007). Risk factors for rural residential fires. Journal of Rural Health, 23(3), 264e269. Armfield, J. M. (2007). Public water fluoridation and dental health in New South Wales. Australian and New Zealand Journal of Public Health, 29(5), 477e483. Asgary, A., Ghaffari, A., & Levy, J. (2010). Spatial and temporal analyses of structural fire incidents and their causes: a case of Toronto, Canada. Fire Safety Journal, 45 (1), 44e57. Australian Bureau of Statistics (ABS). (2001). Census of population and housing 2001. Census products. Canberra: ABS. Australian Bureau of Statistics (ABS). (2001). Information paper: An introduction to socio-economic indexes for areas (SEIFA). Canberra: ABS. Cass, A., Cunningham, J., Wang, Z., & Hoy, W. (2007). Social disadvantage and variation in the incidence of end-stage renal disease in Australian capital cities. Australian and New Zealand Journal of Public Health, 25(4), 322e326. Chainey, S., & Radcliffe, J. H. (2005). GIS and crime mapping. London: Wiley.
75
Chandler, S. E., Chapman, A., & Hallington, S. J. (1984). Fire incidence, housing and social conditions e the urban situation in Britain. Fire Prevention, 172, 15e20. Corcoran, J., Higgs, G., Brunsdon, C., & Ware, A. (2007). The use of co-maps to examine the spatial and temporal dynamics of fire incidents: a case study in South Wales, UK. Professional Geographer, 59(4), 522e537. Corcoran, J., Higgs, G., Brunsdon, C., Ware, A., & Norman, P. (2007). The use of spatial analytical techniques to explore patterns of fire incidence: a South Wales case study. Computers, Environment and Urban Systems, 31(6), 623e647. Cromley, E. K., & McLafferty, S. L. (2002). GIS and public health. New York: Guilford Press. Duncanson, M., Woodward, A., & Reid, P. (2002). Socioeconomic deprivation and fatal unintentional domestic fire incidents in New Zealand 1993e1998. Fire Safety Journal, 37(2), 165e179. Federal Emergency Management Agency (FEMA). (1997). Socio-economic factors and the incidence of fire, report no. FA170. National Fire Data Centre, United States Fire Administration. Gunther, P. (1981). Fire cause patterns for different socio-economic neighbourhoods in Toledo, Ohio. Fire Journal, 75(3), 52e58. Holborn, P. G., Nolan, P. F., & Golt, J. (2003). An analysis of fatal unintentional dwelling fires investigated by London Fire Brigade between 1996 and 2000. Fire Safety Journal, 38, 1e42. Hooper, R. (2006). Deploying GIS-based risk planning, operational planning and resource management at the London Fire Brigade. Paper presented at the Association for Geographic Information Annual Conference, London, UK. Hugo, G. (2007). Space, place, population and census analysis in Australia. Australian Geographer, 38(3), 335e357. Jennings, C. R. (1999). Socioeconomic characteristics and their relationship to fire incidence: a review of the literature. Fire Technology, 35(1), 7e34. Krisp, J. M., Virrantaus, K., & Jolma, A. (2005). Using explorative spatial analysis to improve fire and rescue services. In P. Van Oosterom, S. Zlatanova, & E. M. Fendel (Eds.), GeoInformation for disaster management (pp. 1283e1296). Springer. McFadden, D. (1979). Quantitative methods for analysing travel behaviour of individuals: some recent developments. In D. A. Hensher, & P. R. Stopher (Eds.), Behavioural travel modelling (pp. 279e319). London: Croom Helm. Morgan, C. L. I., Ahmed, Z., & Kerr, M. P. (2000). Social deprivation and prevalence of epilepsy and associated health usage. Journal of Neurology, Neurosurgery and Psychiatry, 69, 13e17. Nicolopoulos, N., Murphy, M., & Sandinata, V. (1997). Socio-economic characteristics of communities and fires. In Statistical research. Sydney: NSW Fire Brigades. Office of the Deputy Prime Minister (ODPM). (2004). Arson control forum: Social exclusion and the risk of fire. Research Bulletin No. 4. London: ODPM. Office of the Deputy Prime Minister (ODPM). (2005). The fire and rescue services national framework document. London: ODPM. www.communities.gov.uk/ documents/fire/pdf/128923.pdf Last Accessed 12.02.10. Poulos, R. G., Hayen, A., Chong, S. S. S., & Finch, C. F. (2009). Geographic mapping as a tool for identifying communities at high risk of fire and burn injuries in children. Burns, 35(3), 417e424. Prendergast, M. J., Beal, J. F., & Williams, S. A. (1997). The relationship between deprivation, ethnicity and dental health in 5-year-old children in Leeds, UK. Community Dental Health, 14(1), 18e21. Rees, P., Martin, D., & Williamson, P. (Eds.). (2002). The census data system. Chichester: Wiley. Runyan, C. W., Bangdiwala, S. I., Linzer, M. A., Sacks, J. J., & Butts, J. (1992). Risk factors for fatal residential fires. New England Journal of Medicine, 327(12), 859e863. Senior, M. (2002). Deprivation indicators. In P. Rees, D. Martin, & P. Williamson (Eds.), The census data system (pp. 123e139). Chichester: Wiley. Shai, D., & Lupinacci, P. (2003). Fire fatalities among children: an analysis across Philadelphia’s census tracts. Public Health Reports, 118, 115e126. Townsend, P., Phillimore, P., & Beattie, A. (1988). Health and deprivation: Inequality and the North. London: Croom Helm. Walker, A. E., & Becker, N. G. (2005). Health inequalities across socio-economic groups: comparing geographic-area-based and individual-based indicators. Public Health, 119(12), 1097e1104. Wallace, D., & Wallace, R. (1984). Structural fire as an urban parasite: density dependence of structural fire in New York City, and its implications. Environment and Planning A, 16, 249e260. Watson, J. P., Cowen, P., & Lewis, R. A. (1996). The relationship between asthma admission rates, routes of admission, and socioeconomic deprivation. European Respiratory Journal, 9, 2087e2093. Zhang, G., Lee, A. H., Lee, H. C., & Clinton, M. E. (2006). Fire safety among the elderly in Western Australia. Fire Safety Journal, 41, 57e61.