Article Text
Abstract
Background Fires and burns are a leading cause of unintentional injury death in the USA. Although it has been anecdotally reported that vacant dwellings are at a higher risk for fire, the association between vacancy and fire risk at the individual household level has not been empirically measured.
Methods In this cross-sectional study, geocoded residential vacant properties (VP) and fire events are analysed in Baltimore City at the census tract level and the individual household level.
Results On average, a 10% increase in the proportion of vacancies in a census tract was associated with a 9.9% increase in fires (95% CI: 5% to 15%). Random-effects Poisson models, controlling for housing and neighbourhood conditions, found contagion effects. The risk of fire in an occupied dwelling increased by 8% (95% CI: 1% to 10%) for every vacant structure within 10 m, and the risk of fire decreased by half (95% CI: 45% to 62%) for every km between an occupied dwelling and vacant building. Close proximity to VP was associated with trash fires within dwellings (p=0.039) and structure fires (p=0.012).
Conclusions We believe that this is the first study to demonstrate increased risk posed by nearby VP at the household level, confirming earlier ecological analyses of the role of VP as strong correlates of home fires. Measurement of this risk can motivate property owners, policy makers and insurers to invest in risk reduction measures that include building maintenance and trash removal.
- Burn
- Clustered analyses
- epidemiology
- geographical/spatial analysis
- poverty
- health Services
- alcohol/drugs
- international
- methods
- economics
- MVTC
- Counselling
- behavioural
- evaluation
- psychological
- violence
- poverty
- community
- child
- public health
- education
Statistics from Altmetric.com
- Burn
- Clustered analyses
- epidemiology
- geographical/spatial analysis
- poverty
- health Services
- alcohol/drugs
- international
- methods
- economics
- MVTC
- Counselling
- behavioural
- evaluation
- psychological
- violence
- poverty
- community
- child
- public health
- education
Background
Fires and burns remain the seventh leading cause of unintentional injury death in the USA,1 with residential fires killing approximately 2780 civilians in 2008. In the same year, residential fires caused 13 560 civilian injuries.2 These fires carry a heavy economic burden as well—an estimated US$42.5 billion in human losses3 and US$8.5 billion in property damage annually.2
Although it has long been assumed that vacant properties (VP) are at a higher risk of catching fire than occupied properties, few studies have quantified the relationship between VP and fire risk. Two studies of fire rates in the 1970s found VP to be over-represented among structural fires in the cities of Newark, New Jersey,4 and Highland Park, Michigan,5 but since then, only a small body of research on community-level predictors of fire risk has been assembled. Several studies have found vacancy to be a predictor of neighbourhood or census tract level fire incidence along with other socioeconomic indicators such as income level, rates of home ownership and average age of housing structures.2 6–9
A more recent report by Ahrens described the patterns of fires in vacant structures. The leading cause of such fires is intentional fire setting, followed by unintentional causes related to cooking, heating and lighting equipment. Eight percent of VP fires are caused by exposure to fires in nearby structures.10 Fires in VP spread to other structures 11% of the time, as opposed to only 3% of occupied structure fires. Unsecured VP are more likely than secured VP to burn due to intentionally set fires, and fires in such VP are also more likely to spread to other structures.10
Though it has been established that neighbourhoods with a large proportion of vacant structures are more likely to have a higher incidence of fire, we could find no previous research looking at the effects of vacancies on fire risk at the individual property level. An analysis at the individual level can inform homeowners and city officials about the magnitude of the additional fire risk residents incur by residing in a home adjacent to a vacant building and can guard against biases induced by aggregating spatial data (sometimes referred to as ‘ecological fallacies’11). Understanding the risks that vacant buildings pose to their neighbours can help policy makers justify investments in vacancy reduction or bricking and boarding. The purpose of this study is to examine the risk of residential fire due to proximity to vacant property in Baltimore City at the census tracti and individual household levels.
Methods
A listing of VP and their addresses was obtained from Baltimore City's Mayor's Office of Information Technology in July of 2008. This office maintains the most updated listing of vacant dwellings and lots in Baltimore City. Data on all buildings in Baltimore were obtained from MD Property View 2007, a database developed by the MD Department of Planning that includes a jurisdiction's property map and parcel information.12 Of all 235 862 properties in the MD Property View 2007, 214 787 of those represented residential dwellings were included. Properties (21 075) were removed because they represented non-residential structures such as places of worship, schools, commercial spaces, hospitals and State or Federally owned buildings. A list of addresses to which the fire department was called for an emergency visit in 2004–2007 was obtained from the Baltimore City Fire Department. The fire department had coded whether fire fighters had found fires at each of the fire calls in the data set. Non-fire-related calls such as scalds, gas leaks, smelled smoke but no fire found and other false alarms were excluded. VP were restricted to those that had previously been residential dwellings. Vacant or burned warehouses, shops and other businesses were excluded. Census tract boundaries from 2000 US Census were obtained at http://www.census.gov/. The Property View database provided data on the price of the dwelling at its most recent sale and the date the dwelling was built, and indicated whether the owner of the building lived in the building (owner occupancy).
Fire events and VP were geocoded to match the spatial relationships in MD Property View with ArcMap 9.2 GIS software (Redlands, California: Environmental Systems Research Institute, 1992–2008). Using census tract boundaries, vacant buildings per total dwellings, average sale price, percent owner occupancy, average year of construction and homes per km2 for each census tract were calculated in ArcMap. ArcMap also was used to calculate the numbers of vacant residences within a radius (ie, buffers) of 5, 10, 25, 50, 100, 150 and 350 m of an occupied dwelling and the distance from each dwelling to the nearest VP. Because we had no a priori means of determining how fire risk scaled with a dwelling's distance to a vacancy, we selected a range of cut points to allow us to identify the upper bound beyond which a vacant building would not convey additional risk. In addition to calculating the number of VP within buffers of varying radii, the distance between each home and the nearest VP was also calculated. ArcMap also was used to calculate the number of residential dwellings within a 10 m radius as a measure of housing density.
The relationship between vacant buildings and fire rates was initially explored at the census tract level by examining overlap between the rate of fires per census tract and the proportion of vacant buildings per census tract. Thematic maps of each variable were prepared and compared to determine the degree of overlap of high vacancy rates and high rates of fires (for fires and VP, see figure 1). An ecological analysis used negative binomial regression to estimate the association between vacancy rates and incident fire rates per dwelling at the level of the census tract.
After the ecological analysis, the relationship between vacancy and fires was examined at the level of individual dwelling with a Poisson multilevel model. The association between fires and increasing numbers of vacancies with the buffer zones was estimated. Numbers of fires with radii of 5, 10, 25, 50, 100, 150 and 350 m were treated as independent variables in separate models comparing VP exposure to numbers of fires. An additional model examined the association between fires and the distance to the nearest vacancy (km). The Poisson models included random intercepts at census tracts to adjust standard errors for correlations in fire risk within census tracts. Finally, the association between vacancies and the top 11 causes of fire was estimated in logistic models where an individual cause of fire. Each of the 11 causes of fire was modelled as a dependent variable in separate logistic regressions, where y=1 represented one cause of interest and y=0 represented all 10 other causes of fire. For example, indoor trash fires (y=1) where compared to fires from gas leaks + cooking fires + boiler malfunction, etc (y=0). The 11 causes of fire represented all the causes of fire that were over 1% of all fires. The random-effects Poisson and logistic models controlled for year the property was built, owner occupancy (yes/no), price of the dwelling at last sale and the number of occupied properties within a 10 m radius of the property of interest to control for housing density. All analyses were done with the R statistical package13 and the ‘lme4’ package.14 Residual spatial autocorrelation was examined with the ‘gstat’ package in R.15
Results
Frequencies of fires and dwelling status
The fire department responded to 11 464 fires in occupied dwellings from 2004 to 2007, of which 7864 were matched to dwellings from the Property View (69%). Thirty-one per cent of records on fires in homes were incomplete or did not match standard spellings or abbreviations for street names and could not be matched to Property View records. Of the city's total of 214 787 residential dwellings, 16 409 (7.6%) were vacant in 2008.
Census tract level results
At the aggregate level, census tracts with high proportions of vacant dwellings and census tracts with high proportions of fires may tend to be in the centre and northwest of Baltimore City by visual inspection (figure 1). In the negative binomial regression at the level of the census tract, a 10% increase in proportion of vacancies in a census tract was associated with a 9.9% increase in fires in that tract (95% CI: 5% to 14%) on average (table 1).
Individual level results
At the level of individual dwellings rather than census tracts, vacant buildings in Baltimore were 28% more likely to have a fire than an occupied dwelling (95% CI=1.21 to 1.36) in a univariate analysis. The average number of fires in dwellings increases as the number of vacant dwellings within 10 m increases.
Random-effects results for fires of all causes
A multivariate regression model with random intercepts at census tracts estimated that dwellings have 11.9% (95% CI: 8% to 16%) more fires for every additional VP within 5 m of a dwelling and 7.6% (95% CI:5% to 10%) more fires for every additional VP within 10 m of a dwelling. As the distance between a dwelling and the vacant buildings expands, the association between vacant buildings and fires becomes weaker. At distances of 25 m, an additional vacant building is associated with 3.8% (95% CI:3% to 5%) more fires, and at 50 m, an additional vacant building is associated with 2% (95% CI:2% to 3%) more fires. When distance to the nearest VP is considered, rather than number of VP in a buffer around an occupied property, exposure to VP remains significant; dwellings are 55% (95% CI:45% to 62%) less likely to experience fires with every km from the nearest vacant home (table 2). Owner occupancy and housing density are also significantly related to numbers of fires (table 2). Variograms of standardised residuals from the Poisson models showed little residual spatial autocorrelation (figure 2).
Random-effects results for fires of specific causes
With regard to specific causes of fires, an increase of 1 km in the distance between an occupied dwelling and the nearest vacant building is significantly protective from indoor trash fires (OR=0.17%, 95% CI: 0.03% to 0.91%) and building/structure fires (OR=52%, 95% CI: 31% to 87%) in a multi-level logistic regression controlling for price, owner occupancy, year of construction and housing density. Increasing distance from home to the nearest vacant building is associated with fires where the fire department is alerted by alarm systems (OR=2.44, 95% CI: 1.42 to 4.19, table 3).
Discussion
This study demonstrates that VP are associated with more fires at a community level and among immediately adjacent dwellings. To date, research has focused on census tract level analyses,2 6–9 and our results confirm these earlier findings. In our work, exposure to VP is measured in terms of numbers of VP within varying distances of dwellings and as the distance to the nearest VP. Numbers of VP within 100 m of a dwelling and distance to the nearest VP are significant risk factors for fire. We found that nearby VP continue to convey a significant risk of fire even when we controlled for other significant risk factors like owner occupancy and cost and age of the dwelling, and when a random effect at census tract level was adjusted for residual spatial autocorrelation. We believe that the multivariate random-effects model yields a conservative measure of the VP effect, and we present several measures of VP exposure to underscore the strength of the VP fire association. Moreover, we observed a gradient of effect where VP effect measures increase as the radius between homes and VP decreases; the gradients of effect may represent a ‘dose–response’ which bolsters our confidence in the finding. This study adds new information on the relationship between VP and house fires at the individual household level, showing that proximity to a vacant building is associated with heightened risk of fires.
It is well known that VP have higher rates of intentional and unintentional fires.10 What was not previously understood was that occupied properties neighbouring VP were associated with increased risk simply by virtue of their proximity. This increased risk was not explained by other known fire risk factors such as age of the property. These results increase the importance of addressing VP because they are at increased risk for fires themselves and because their increased risk jeopardises other properties and residents in their neighbourhood.
One limitation of the study concerns the data that are available for these types of analyses. The number of VP changes over time, and although the responsible government agency attempts to keep an accurate and updated list, it is possible that vacancies go unidentified. In this case, our data would underestimate the number of VP. On the other hand, when properties are reoccupied, there may be a time lag for having the home moved off the list of VP, which would overestimate the number. Nevertheless, the data set we used is comparable to those used in similar previous research, and without extensive resources, it is not possible to collect original data to answer our research questions. A second limitation is that the analysis does not establish a causal relationship between VP and fires. Although fires may start in VP and spread, VP may be an indicator of larger socio-economic declines in a neighbourhood, and these changes may lead to fires. Also, our vacancy data are from 2008 and the fires data are from 2004 to 2007, so the temporal misalignment hinders causal inference. On the other hand, we do not believe that the spatial distribution of VP changed appreciably over this period. We failed to geocode 31% of the fires that were reported by the fire department. The large majority of the fires that we could not geocode had addresses that were either incomplete or spelt in a non-standard manner. We do not believe that incomplete or misspelled addresses would vary in relation to proximity to VP, and thus, data missing from the analysis would not bias the VP effect estimate.
Ownership of a housing unit was associated with a 35% decrease in fire risk, which may be attributed to live-in owners being more likely to take fire prevention steps that external landlords may ignore or renters may not undertake. Increases in the number of homes within 10 m of a dwelling were associated with protection from fires, which reflects greater fire awareness in high-density living conditions.
Taken together, the new findings of this research can be used to target fire prevention efforts and risk communication messages to individuals residing near vacant buildings. In particular, our data suggest that cleaning out trash and maintaining buildings structurally can prevent fires in high-risk homes surrounded by vacant buildings. The results suggest potential benefits from boarding up VP to reduce the potential for fire-causing human activity. Community residents should be informed of their individually increased risk if they live near a vacant property. At a minimum, such information can raise the salience of fire risk among individual residents and offer an opportunity to provide prevention education as well as test and install smoke alarms. In addition, such information could spur community action to deal with VP in a timely manner to protect entire neighbourhoods. Given the societal burden imposed by fires, especially in urban areas, such efforts should be given high priority.
What is already known
Previous work focused on rates of fire at aggregate geographic areas, and found links between proportions of vacant buildings and fire rates.
What this study adds
We believe that our paper is the first to examine how decreasing proximity to vacant buildings can increase the risk of home fire on the level of individual homes.
Footnotes
Funding This study was supported through a grant from the Centers for Disease Control and Prevention (5R18CE001339) and the National Institute of Child Health and Human Development (R01 HD059216-04).
Competing interests None.
Ethics approval The study was approved by the Johns Hopkins Institutional Review Board.
Provenance and Peer review Not commissioned; externally peer reviewed
↵i Census tracts are delineated by the federal government of the USA for purposes of the decennial census. Census tracts are intended to be homogenous with respect to socio-economic and demographic variables (http://www.census.gov/).