Do fires discriminate? Socio-economic disadvantage, wildfire hazard exposure and the Australian 2019–20 ‘Black Summer’ fires

We examine the relationship between socio-economic disadvantage and exposure to environmental hazard with data from the catastrophic 2019–2020 Australian wildfires (Black Summer) that burnt at least 19 million hectares, thousands of buildings and was responsible for the deaths of 34 people and more than one billion animals. Combining data from the National Indicative Aggregated Fire Extent (NIAFE) and 2016 Socio-Economic Indexes for Areas (SEIFA), we estimate the correlation between wildfire hazard exposure and an index of community-level socio-economic disadvantage. Wildfire hazard exposure is measured as the interaction between the percentage of area burnt and proximity of the fire to settlements. The results reveal a significant positive relationship between fire hazard exposure and socio-economic disadvantage, such that the most socio-economically disadvantaged communities bore a disproportionately higher hazard exposure in the Black Summer than relatively advantaged communities. Our spatial analysis shows that the socio-economic disadvantage and wildfire hazard exposure relationship exists in inner regional, outer regional and remote areas of New South Wales and Victoria, the two worst-hit states of the Black Summer catastrophe. Our spatial analysis also finds that wildfire hazard exposure, even within a small geographical area, vary substantially depending on the socio-economic profiles of communities. A possible explanation for our findings is resource gaps for fire suppression and hazard reduction that favour communities with a greater level of socio-economic advantage.


Introduction
As the planet warms, the number of catastrophic wild res is growing worldwide at an alarming rate (Jolly et al. 2015;Lindenmayer and Taylor 2020). The unprecedented scale of devastation caused by wild res in the United States, Australia and Brazil in 2019-2020 foreshadows the shifting ground of global wild re risk in a changing climate. Spatio-temporal trends from 1979 to 2013 reveal an 18.7% increase in global reseason length and 25% increase in re susceptibility of the world's vegetated surface (Jolly et al. 2015). The frequency and intensity of large-scale wild res have increased in the past two decades and is expected to increase further in the coming decades (Sharples et al. 2016;Abram et al. 2021;Lindenmayer and Taylor 2020).
In terms of the historical record, 2019 was both the warmest and the driest ever recorded in Australia (Filkov et al. 2020). An increasing temperature trend is observed over much of Australia in the past century with the average surface air temperature increasing by 1°C since 1910 (BoM andCSIRO 2018). The risks of higher temperatures are compounded in years where there is below average rainfall (Grose et al. 2014;BoM and CSIRO 2018).
Notably, south-eastern and south-western Australia are becoming drier, with May to July rainfall in south-eastern Australia seeing the largest decrease, by around 20% since 1970, and an 11% decrease in April to October rainfall since the late 1990s (BoM and CSIRO 2018).
High temperatures and low rainfall combined led to low soil moisture levels and a cumulative Forest Fire Danger index (FFDI) in 2019 that was the highest on record for 60% of Australia's land area. These very high-risk conditions were the trigger for the largest series of wild res, in terms of area and over one summer, ever experienced in Australia. These Black Summer res resulted in the burning of at least 19 million hectares, destruction of more than 2,400 buildings, and loss of 34 human lives (Filkov et al. 2020) and more than one billion animals (Richards et al. 2020). The res, even by a global comparison, burned the highest percentage of any continental forest biome (21%) from 2000-2019 (Boer and de Dios 2020).
Socio-economic disadvantage encompasses the relative inability of individual and communities to access economic and social resources and to participate in society (Australian Bureau of Statistics 2018a). Socio-economic disadvantage is a broader concept than poverty as it goes beyond nancial resources to encompass a wide array of non-nancial, social and community resources that account for ethno-linguistic and religious differences, and the availability, accessibility and affordability of public goods and services. All else equal, those who suffer from the greatest socioeconomic disadvantage have the fewest resources of their own to cope with disasters. Thus, disadvantage is key to understanding how climate change risks (Leichenko and Silva 2014;Hallegatte et al. 2020) and other 'natural' disasters, including wild res (Cottrell, 2005), affect heterogenous populations differentially.
Our work builds on the existing literature that has examined the link between poverty or socio-economic disadvantage and environmental risks. This link varies across countries, communities and hazard types. Studies on ooding and cyclone or hurricane show that socio-economic disadvantage is frequently associated with a higher likelihood of being affected by a hazard (e.g., living on a ood-plain near a river susceptible to ooding) and fewer resources to mitigate its impact (e.g., insurance may be unavailable or unaffordable) (Brouwer et al. 2007;Akter and Mallick 2013;Elliott and Pais 2006;McDougall 2007).
The link between socio-economic disadvantage and wild re hazard is not so straightforward and may vary depending on land use pattern, characteristics of the population that live in wild re prone locations (regional, rural and peri-and rural-urban interfaces) and socio-economic and institutional arrangements surrounding wild re risk reduction. For example, earlier studies in the United States nd that low-income rural populations with less formal education are more at risk of wild res or bear a disproportionately high burden of its direct and indirect impacts (Lynn and Gerlitz 2005;Ojerio et al. 2011). More recent studies from the United States reveal a negative association between socio-economic disadvantage and wild re sensitivity or overall wild re risk (Wigtil et al. 2016;Davies et al. 2018;Paveglio et al. 2016;Paveglio et al. 2018). Wigtil et al. (2016) and Davies et al. (2018), for instance, show that socio-economically advantaged population (e.g., white and rich) inhabit highly wild re-prone locations due to these locations' high environmental amenities and corresponding property values. Likewise, Paveglio et al. (2016Paveglio et al. ( , 2018 nd a positive relationship between income, education and sensitivity to potential wild re losses in Montana and Idaho. The interlinkage between socio-economic disadvantage and wild re hazard exposure has not been studied in Australia using large-scale data. The only study that comes closest to exploring this link in Australia uses qualitative data from a farming district (Wulgulmerang) of East Gippsland, Victoria (Whittaker et al. 2012). The study nds that declining farm incomes, depopulation, and low access to essential services increase residents' exposure to wild res. Given the growing frequency and intensity of extreme wild re events in Australia and rapidly expanding settlements in periand rural-urban sprawls (Koksal et al. 2019;Foster et al. 2013), an in depth understanding of the interlinkage between socio-economic disadvantage and wild re hazard exposure is important for informing Australia's wild re hazard risk reduction strategies (Royal Commission into National Natural Disaster Arrangements 2020). Unlike in the United States, wild re prone locations of Australia, particularly the peri-and rural-urban fringes, are inhabited by a larger percentage of socio-economically disadvantaged communities compared with inner city locations (Alexandra 2020). The relatively higher value properties in Australia are concentrated around the central or near central locations in the capital cities where wild re hazard exposure is the lowest. Hence, the interlinkage between wild re exposure and socio-economic disadvantage in Australia is likely to be different to the United States.
Our study contributes to understanding disaster risk which is a priority area (Priority 1) for action identi ed by the Sendai Framework for Disaster Risk Reduction 2015−2030(United Nations 2015. The four speci c ways we enhance understanding of disaster risk is as follows. First, we examine the link between socio-economic disadvantage and wild re hazard exposure in Australia utilising the broad spatial coverage of the Black Summer res at the Statistical Area 1 (SA1) level, which is the second smallest geographical unit designated in the Australian Statistical Geography Standard (ASGS) (Australian Bureau of Statistics 2016). This is the rst large-scale quantitative study in Australia in the context of an unprecedented extreme wild re event. Second, in addition to exploring the link at the national level, we examine whether wild re hazard exposure is correlated with remoteness of the settlements. Wigtil et al. (2016) examined spatial variation of the link but only at two levels, i.e., wildland-urban interface and non-wildland-urban interface. We examine this relationship at ve levels of remoteness (i.e., major cities, inner regional, outer regional, remote and very remote areas). Third, it uses a novel approach to measure wild re exposure. The most used indicator for wild re exposure is re extent, i.e., the spatial coverage of a re (Lindenmayer and Taylor 2020;Jolly et al. 2015) or settlement proximity to a wild re (Hazlett and Mildenberger 2020). We use a multidimensional index to measure wild re hazard exposure that accounts for both re extent and re proximity to settlements. Finally, our study uses actual re hazard exposure (instead of potential or simulated hazard exposure) of a large spatial variation covering a combined area of approximately 12 million hectare that was burned at different times. Thus, while there was similarity in terms of high re hazard conditions in different locations, the res that occur were multiple and separate events.

Theoretical Framework
The disaster literature de nes risk (expectation of loss) as a function of hazard (potential for loss or harm from adverse events), vulnerability (susceptibility to losses from adverse events) and exposure (the presence of population or other elements of interest in a location which is susceptible to hazard) (IPCC 2014). Hazard is the future or possible occurrence of a natural or human induced event which has the potential to in ict harm on vulnerable and exposed elements (IPCC 2014). Vulnerability is the propensity of a system to suffer negative consequences when impacted by adverse events (IPCC 2014). The degree of vulnerability depends on a system's sensitivity (e.g., a house built with highly ammable materials) and adaptive capacity (e.g., personal re ghting equipment, purchase of adequate insurance) (IPCC 2014). According to these conceptualizations, vulnerability is a pre-existing and internal property of a system while exposure represents an external property. Both vulnerability and exposure are necessary elements but, by themselves, are not su cient determinants of risk.
A key nding of the existing disaster literature is that the biophysical properties of an area alone do not fully determine hazard exposure and vulnerability to climate related natural disasters. Hazard exposure is also determined by a broad set of socio-economic and demographic factors such as wealth, social status, age, gender, ethnicity (IPCC 2014). In this perspective, and according to Cutter's Hazards-of-Place (HOP) Model (Cutter 1996), exposure to hazard is a complex interaction of socio-economic and biophysical conditions of a place. For instance, in the context of wild res, the frequency, severity and nature of wild re in a location are, in part, determined by land use (e.g., area in forest or grassland), settlement patterns (e.g., along a river or on a ridge), vegetation management (e.g., degree of hazard reduction) and compliance with re warnings (Hawbaker et al. 2013;Syphard et al. 2013).
Socio-economic factors may in uence wild re hazard exposure in four ways. First, socio-economically disadvantaged communities may live in more wild re prone locations due to housing affordability (Alexandra 2020). Second, allocation of public resources for preparedness, hazard reduction and suppression may disproportionately favour the most advantaged communities (Royal Commission into National Natural Disaster Arrangements 2020). For instance, evidence from the United States shows that poor households are less likely than non-poor households to bene t from federal programs designed to reduce wild re risk (Lynn and Gerlitz 2005;Brunet et al. 2001). Third, socio-economically advantaged communities may be able to commit greater volunteer time to re suppression (McLennan and Birch 2005) and, during a re season that continues for a long time, volunteer re ghter numbers may decline proportionally more in areas where people's incomes are more reliant on small businesses or casual or hourly employment (Volunteering Australia 2020). Finally, ignition risk caused by malicious intent and/or negligence may be lower in socially advantaged communities. For instance, poor academic achievements, substance abuse, unemployment, the absence of a mother or father gure through childhood, are common characteristics of arsonists in Australia (Bell et al. 2018;Ellis-Smith et al. 2019). Such characteristics are also relatively more prevalent in socio-economically disadvantaged than advantaged communities.

Measuring wild re exposure
Fire data were collected from the National Indicative Aggregated Fire Extent (NIAFE) dataset published by the Department of Agriculture, Water and the Environment (2020) and was developed using inputs from the national Emergency Management Spatial Information Network Australia (EMSINA) and other supplementary sources, such as relevant state and territory agencies and the Northern Australian Fire Information website. These data present a record of total re extent starting from 1 July 2019 to 25 May 2020.
Our outcome variable of interest is wild re hazard exposure. Wild re hazard exposure is operationalised as the interaction between re extent and proximity of re to settlements. Previous studies operationalised this construct using re extent (i.e., area burnt) (Ager et al. 2018) or residents' proximity to a wild re (Hazlett and Mildenberger 2020). The reason for combining these two indicators is that re extent alone does not fully capture wild re hazard. For instance, it is possible that a large share of an area was burnt, but the hazard exposure was low because the re was far away from settlements. We undertake separate analyses using re extent and re proximity to settlements as dependent variables to test the robustness of our ndings.
Our focus is on the immediate effects of wild res (extent and proximity) rather than on either the short or long-term health consequences of smoke and which may exceed the direct loss of life from re. For example, the Black Summer wild res are estimated to have caused more than 400 excess deaths (any smoke-related cause), more than 1,100 hospitalisations for cardiovascular problems and more than 1,300 presentations to emergency departments with asthma over the period 1 October 2019 to 10 February 2020 (Arriagada et al. 2020).
Fire extent is a continuous variable that is calculated as the percentage of burnt area in a SA1. Proximity is an ordinal variable comprising eight integer values ranging from 1 to 8. A high value of proximity indicates low distance between a settlement and a re. Wild re hazard exposure is an index generated by interacting re extent and re proximity. The index takes a value between 0 and 800. A high value of wild re exposure means both re extent and re proximity are high and vice versa.

Index of Relative Socio-Economic Disadvantage (IRSED)
Our key independent variable is the Index of Relative Socio-Economic Disadvantage (IRSED). The IRSED is one of the four indices of Socio-Economic Indexes for Areas (SEIFA) prepared by the Australian Bureau of Statistics. Using information from Census data, SEIFA assigns scores and classi es each SA1 based on the collective socio-economic advantage and disadvantage of its population. SEIFA scores are not available for areas that do not have su cient population or have a high rate of missing responses for the key variables. SEIFA include the following four indices: Given that the focus of this study is to explore the link between socio-economic disadvantage and wild re hazard exposure, we use the IRSED for our main analysis. The IRSED is presented as raw scores and deciles. Deciles divide SA1s into ten groups based on their raw scores. We use deciles for our main analysis because it allows us to test for non-linearity in the relationship. As per the original coding, a low value of the IRSED means a high proportion of relatively disadvantaged population in an area. For example, a low value of the IRSED means a high proportion of the population earn a low income, live in a private dwelling with one or no bedrooms and/or a higher percentage of household with children and a jobless parent and/or a high incidence of no internet access and/or poor English language pro ciency and so on (Australian Bureau of Statistics, 2018a). For ease of interpretation and consistency with the concept of disadvantage, we recode IRSED deciles in reverse order as IRSED_r. For example, IRSED decile 10 is coded as IRSED_r decile 1, decile 9 is coded as 2 and so on.

Empirical Framework
The empirical analyses examined the correlation between wild re hazard exposure and the recoded deciles of the socio-economic disadvantage of communities (i.e., the IRSED_r). Our null and alternative hypotheses were: H 0 : Wild re hazard exposure is uncorrelated with IRSED_r.
H A : Wild re hazard exposure is correlated with IRSED_r.
The following reduced-form xed-effects regression model is used to test this hypothesis: In equation (1), i stands for SA1. Wild re Exposure is the interaction of re extent (i.e., total burnt area as a percentage of the total size of a SA1) and proximity of re to settlements. IRSED_r is recoded deciles (i.e., a higher value means higher disadvantage). X is a vector of SA1-level control variables including population density, population density squared, average re extent during 2011-2016, remoteness, forested area as a percentage of the total area, distance from the coast, and presence of a re station and number of business. The term is SA3 xed effects, which is the most parsimonious speci cation possible given the nature of the data[1]. It captures a wide array of unobserved characteristics at the SA3-level that are likely to be correlated with IRSED_r and wild re hazard exposure. For example, they capture variation in climatic pattern, drought condition, vegetation mosaic, infrastructure condition and other factors related to location. The robust error term is may include SA1 level unobserved factors; we assume that the variables included in are uncorrelated with IRSED_r.

Robustness tests
First, we re-ran our original speci cation keeping wild re hazard exposure as the dependent variable but replacing our key independent variable, i.e., IRSED_r, by alternative measures of socio-economic disadvantage, being IRSEAD_r, IER_r and IEO_r. These models are estimated to check whether the possible socio-economic disadvantage and wild re hazard exposure association is sensitive to the way socio-economic disadvantage is measured. Second, we re-ran the speci cation outlined in equation 1 using wild re extent and proximity separately as indicator of wild re exposure.
These two models test whether the hypothesised relationship between wild re hazard exposure and socio-economic disadvantage holds when wild re hazard exposure is operationalised using re extent or re proximity to settlement.

Regression results
Our analysis sample includes all SA1s that experienced re incidents during 2019−2020 re season and for which SEIFA data are available. The analysis sample excludes observations from the Northern Territory and northern areas of Western Australia, South Australia and Queensland[1] as burned area in northern Australia represents controlled res that are part of the natural landscape management dynamics (Department of Agriculture, Water and the Environment 2020). Table 1 presents the descriptive statistics of key variables for the analysis sample.
The main results from Tobit regression models are presented in Table 2. We use Tobit models because our dependent variable is censored. More speci cally, the Wild re Exposure index cannot take negative value or a value greater than 800. Columns 1 and 2 present the estimated coe cients obtained, respectively, from a bivariate and a full model.
In column 1, all coe cients for the nine IRSED_r deciles are positive and signi cant at the 1% level. In column 2, the coe cients remain statistically signi cant at least at the 5% level, except for the coe cient of decile 2 when control variables and SA3 xed effects are accounted for. Each coe cient of the IRSED_r decile estimates the difference in Wild re Exposure index value between decile 1 (the most advantaged group) and the corresponding decile. For example, in column 2, the estimated coe cient of decile 3 is 77.300 (p<0.05). This coe cient implies that, on average, the SA1s that are classi ed as decile 3 experienced 77.310 points higher wild re exposure relative to the SA1s that were classi ed as decile 1, all else constant. The Wild re Exposure index scores of the SA1s classi ed as decile 8, 9 and 10 (most disadvantaged groups), on average and all else being constant, are 106.500, 96.220 and 144.400 points higher than the SA1s of decile 1 (the most advantaged SA1s), respectively. The remaining coe cients can be interpreted in a similar way.
To visualise the difference in wild re exposure across IRSED_r deciles, we plotted the mean and 95% con dence intervals of the predicted values (evaluated at the mean values) of wild re hazard exposure across IRSED_r deciles (Figure 1). The gure shows that the predicted value of the wild re exposure index is lowest for decile 1 (the least disadvantaged SA1s). The average predicted wild re exposure index signi cantly (p<0.01) increases for decile 2. The predicted wild re exposure index score does not vary signi cantly among deciles 2 to 7 which implies that the SA1s of these deciles were more or less equally exposed to wild res. The wild re exposure index scores shift up signi cantly (p<0.001) for SA1s that are classi ed as IRSED_r deciles 8, 9 and 10 implying that the most disadvantaged SA1s were most exposed to Black Summer res. Note that the differences in predicted wild re hazard exposure scores between these three most disadvantaged deciles and the rest of the seven deciles are statistically insigni cant. These results imply the presence of non-linearity in the interlinkage of wild re hazard exposure and socio-economic disadvantage.
Turning our attention to the control variables (Table 2), wild re hazard exposure has a signi cant positive correlation with forest cover. This is because forested areas have more fuel load and are, hence, more prone to large re extent during a wild re. The history of wild re exposure (i.e., average percentage of area burnt from 2011-2016) also has a signi cant positive correlation with 2019-2020 wild re exposure. This implies that areas prone to wild res in a regular year were more likely to experience a high wild re hazard exposure in 2019-2020. The coe cient for distance from the coast has a signi cant negative correlation with wild re hazard exposure at the 5% level, i.e., all else being constant, wild re hazard exposure in inland areas was signi cantly lower than coastal areas.
Among other control variables, the coe cient for population density has a signi cant positive correlation with wild re hazard at the 10% level in column 2, i.e., densely populated areas were signi cantly more exposed to wild res. The number of re stations is not a signi cant covariate of wild re hazard exposure but the coe cient of the number of businesses in an area is negative and signi cant at the one percent level in column 2.
This means that the areas that had high business activities were signi cantly less exposed to wild re hazard.
The coe cients for remoteness are positive and signi cant at the 1% and 5% level in column 2. Here, the base level for remoteness is 'major cities'. Our nding from the estimated coe cients is that, as expected, major city areas experienced signi cantly lower wild re hazard exposure than inner and outer regional areas and remote and very remote areas. Note also that the SA1s with less wild re hazard exposure may still incur a substantial smoke-related health burden associated with wild res that may be located a large distance away from them (Arriagada et al. 2020).

Robustness tests
SM3 presents the results of three Tobit regression models in columns 1, 2 and 3 that use IRSEAD_r, IER_r or IEO_r deciles as the key independent variable, respectively. The coe cients of IRSEAD_r (SM3, column 1) are all positive (except for the coe cient of decile 2) but only the coe cient for decile 10 (most disadvantaged SA1s) is signi cant (p<0.05). All coe cients of IER_r (SM3, column 2) are positive and signi cant at least at the 10% level (except for the coe cient of decile 2). All coe cients of IEO_r (SM3, column 3) are positive and signi cant at least at the 10% level (except for the coe cient of decile 2). These ndings suggest that the interlinkage between socio-economic disadvantage and wild re exposure is, in general, consistent across various measures of socio-economic disadvantage. The strength of the positive correlation may vary depending on the way socio-economic disadvantage is operationalised.
In SM4, column 1, we present the results of a Tobit regression model that uses wild re extent as the dependent variable instead of the interaction of wild re extent and proximity. IRSED_r deciles are used as the key independent variable. Consistent with our main ndings (presented in Table 2), the results presented in SM4 and column 1 show wild re extent is signi cantly higher for all larger IRSED_r deciles relative to IRSED decile 1, all else being constant. In SM4, column 2, we present the results of an ordered probit regression model with wild re proximity to settlements as the dependent variable. We used an ordered probit regression model because proximity is an ordered variable varying between 1 (furthest) and 8 (closest). The coe cients for IRSED_r are positive and signi cant at least at the 10% level for the top three deciles (most disadvantaged SA1s) implying that the most disadvantaged SA1s were located in closest proximity to wild res compared to less disadvantaged SA1s.

Results of the spatial analysis
Figures 1, 2 and 3 present the spatial distribution for the interlinkage of IRSED_r scores and wild re hazard exposure in Australia, New South Wales (NSW) and Victoria, respectively. This analysis explores whether the interlinkage between socio-economic disadvantage and wild re hazard exposure is speci c to remoteness of a location. We categorise the quantitative wild re hazard exposure into three levels for ease of mapping [low (=1), medium (=2) and high (=3)] and IRSED_r deciles into three levels as IRSED_r3 [least disadvantaged (=1), disadvantaged (=2) and most disadvantaged (=3)].
The qualitative levels are based on the quantitative values for wild re hazard exposure index. Low wild re hazard exposure is when the wild re hazard exposure index value for an area is less than 100. Wild re hazard exposure index values greater than 100 but less than 300 are categorised as a medium and those greater than 300 are categorised as a high. IRSED_r deciles 1-4 are classi ed as the least disadvantaged group, deciles 4-6 are grouped as disadvantaged, and deciles 7-10 are categorised as most disadvantaged.
According to Figure 2, the wild re hazard exposure and IRSED_r3 interlinkage is most pronounced in the eastern states of Australia, speci cally NSW and Victoria. Wild re hazard exposure in these two states was signi cantly higher than Western and South Australia which is consistent with previous ndings of spatial distribution of wild re hazard exposure in Australia (Lindenmayer and Taylor 2020). See SM5 and SM6 for the spatial distribution of wild re hazard exposure across states.
Focusing on NSW (Figure 3), the 2019-2020 wild res were a major city, inner and outer regional phenomenon. Remote and very remote areas of NSW were not severely hit. In Victoria (Figure 4), the 2019-2020 wild re hazard was more prominent in outer regional and remote areas, with the most pronounced wild re hazard and IRSED_r3 interlinkage.
To understand how remoteness[2] intersects with the wild re hazard exposure and IRSED interlinkage, we estimated the mean value of wild re hazard exposure for IRSED_r3 categories across ve regional strata in NSW and Victoria. The results of the analysis are presented in Table 3. Column 1 presents the results for the NSW sample. In major city areas in NSW, wild re hazard did not vary signi cantly across IRSED_r3 levels. In inner regional areas, disadvantaged (p<0.001) and the most disadvantaged (p<0.01) communities experienced signi cantly higher wild re hazard exposure than the least disadvantaged areas. However, the wild re exposure level of disadvantaged and most disadvantaged communities of inner regional NSW did not vary signi cantly (p=0.871). The largest difference in wild re hazard exposure across IRSED_r3 is observed for the outer regional areas of NSW. The most disadvantaged areas in the outer regional NSW were exposed to the highest wild re hazard among all remoteness-IRSED_r3 combinations. The difference of this group's wild re exposure with the rest of the groups is signi cant at the 1% level.
Next, we examined the spatial distribution of the wild re hazard exposure and IRSED_r3 interlinkage in Victoria. Like NSW, the most disadvantaged SA1s experienced signi cantly higher wild re hazard exposure in outer regional and remote areas, relative to the least disadvantaged and disadvantaged SA1s in those locations.

Possible Mechanisms
Wild re risk management in Australia is comprised of three components: (a) prevention, i.e., reduction of fuel load (at lower danger re weather conditions) and also ignitions; (b) physical preparation to minimise spread, i.e., creating rebreaks, access roads, and water supply infrastructure; (c) use of early detection systems and rapid deployment of re ghters and equipment.
Prevention and physical preparation are classi ed as hazard reduction activities. Australia's disaster resilience framework is centred on the idea of shared responsibility that requires local community and emergency service agencies to work together for both hazard reduction and re suppression activities (McLennan and Handmer 2012). This approach shifts the responsibility from the central authority to individuals and the local community with the assumption that communities and individuals at risk understand the risk and have the capacity to prevent, prepare and respond to it.
The rst possible causal mechanism for why socio-economically disadvantaged communities suffered greater wild re hazard exposure in the Black Summer res could be differences in re suppression capability among communities and the level of support received from the government. This difference may arise because re ghting service in Australia is run by professional and volunteer re ghters. In general, professional re ghters who are trained to ght structural res are deployed in higher density areas with more valuable buildings. For instance, it has been observed in the US that where residents have a greater willingness to pay (in terms of property taxes) for these services and demand higher levels of re protection because their properties are, in general, more valuable (Brunet et al. 2001).
Professional re ghters in NSW, one of the worst-affected states in the Black Summer res, number some 7,000, and are paid for by the state government and operate under a single state structure (Fire and Rescue NSW). Similar arrangements exist in other Australian states. Professional re ghting services are supplemented by state-supported rural part-time re services based in rural communities that comprise, in NSW alone, more than 70,000 volunteers (Langford 2019). Australian states provide equipment and training to the volunteers as part of the rural re services in each state, and there is a longstanding and proud tradition of volunteers in rural communities providing much of the human power for rural re prevention. Rural re ghters also have a comprehensive understanding of local weather condition and they deal with seasonally occurring largescale landscape res in di cult conditions. Professional re ghters can also be deployed in rural and remote areas if rural re ghting capability is deemed inadequate. Professional re ghters deliver a higher-quality service, as measured by response and losses per reported re relative to volunteer re ghters for whom a national standard of competency requirements and accreditation protocol are currently lacking (Brunet et al. 2001; Royal Commission into National Natural Disaster Arrangements 2020). Thus, residents of peri-urban and rural-urban interface, with professional re ghting services and higher median incomes, should be expected to receive a better re ghting service than persons residing in regional or rural areas (who, typically, have a rural re ghting service and lower median incomes). Indeed, using US data, Brunet et al. (2001) nd a signi cant negative relationship between median household income, average years of education and population density and the proportion of residents in state counties that are serviced by volunteer re ghters.
It is possible that the relative advantage of a professional re ghting service may be represented in terms of a gradient of resources or quality in terms of volunteer re ghting services in rural services without professional re ghters. In other words, more disadvantaged and smaller rural communities may receive a lower level of rural re ghting service, in general, than larger, more a uent rural communities. If this is the case, then the disadvantage may arise from either the allocation of equipment at a state level, prioritisation of re mitigation services such that larger communities are prioritised over smaller communities when resources are scarce relative to need, or other factors concerning volunteering.
Another possible way to explain the socio-economic disadvantage and wild re hazard exposure interlinkage is that disadvantaged communities live in hazardous locations because of their low property values. However, our spatial analysis refutes this hypothesis for Australia. Our model controls for the history of wild re hazard exposure and the percentage of forest cover in an area. This means that the poverty and wild re hazard association estimated by our regression model is independent of the history of wild re hazard exposure and forest density. Thus, location alone is unlikely to be a potential driver of the socio-economic disadvantage and wild re hazard interlinkage.
A third pathway through which socio-economically disadvantaged communities might face a greater wild re hazard exposure is arson risk. Various statistics suggest that 40-50% of the res in Australia are deliberately lit, and another 47% are accidental, i.e., failure to comply with guidelines (Beale and Jones 2011). Arsonists are predominantly male youth and unemployed middle-aged men with a history of mental illness, substance abuse and criminal records (Bell et al. 2018;Ellis-Smith et al. 2019). Disadvantaged communities are more likely to have a higher proportion of such individuals than relatively less-disadvantaged communities, increasing their exposure to wild re hazard. However, re service authorities in NSW and Victoria con rmed that only a small proportion (<1%) of the 2019-2020 wild res were deliberately lit, with most started by lightning (Nguyen et al. 2020). Thus, in our view, re source is not a likely explanation of the positive relationship between socio-economic disadvantage and wild re hazard exposure.
A nal possible pathway could be heterogeneity in hazard reduction capabilities across communities. The report of the Royal Commission into National Natural Disaster Arrangements (2020) echoes a similar concern. The Royal Commission nd that, although disaster preparedness, response and recovery are responsibilities of state and territory governments, there is a practice of delegating these responsibilities to local governments who are not necessarily adequately resourced to meet those responsibilities. The Royal Commission also noted that resource sharing arrangements between local and state/territory governments and among local governments to support surge capacity are in many cases informal, ad-hoc and inadequate. Knowledge about local governments' capacity and capabilities to manage natural disasters is also insu cient at the state/territory government level which impedes e cient resource allocation for disaster risk reduction and response.

Discussion
The objective of this study was to examine the interlinkage between socio-economic disadvantage and wild re hazard exposure in the context of the 2019−2020 Black Summer res of Australia. Consistent with Cutter's Hazards-of-Place (HOP) Model, our analyses reveal the actual hazard exposure varies signi cantly within geographical areas that have the similar biophysical conditions but vary widely in terms of socio-economic characteristics. Our ndings are consistent with studies that investigated a similar link between hazard exposure and poverty or socio-economic disadvantage for oods, cyclones and hurricanes (Brouwer et al. 2007;Akter and Mallick 2013;Elliott and Pais 2006;McDougall 2007) but our ndings differ to the recent studies from the United States (Davies et al. 2018;Wigtil et al. 2016).
Our results reveal considerable non-linearity in the correlation between wild re hazard exposure and socio-economic disadvantage. This is a novel contribution of our study. Previous studies that examined the relationship between socio-economic disadvantage and environmental risk exposure, rarely tested the non-linear pattern of the interlinkage. We nd that the wild re hazard exposure did not increase linearly with an increase in socioeconomic disadvantage. The difference in wild re hazard exposure was most prominent between the SA1s of the top three disadvantaged deciles and the remaining seven deciles. In particular, the variation in wild re hazard exposure signi cantly differs between the most advantaged SA1s (decile 1) and relatively less advantaged SA1s (deciles 2−7). The difference in wild re hazard exposure between most disadvantaged SA1s and the rest remains consistent across all measures of socio-economic disadvantage tested in our analyses although the magnitude of the correlations varies across measures of socio-economic disadvantage.
As expected, forest and population density are signi cantly positively correlated with 2019−2020 wild re hazard exposure. High forest density means high fuel load and high population density increases ignition risks due to high level of human activities in an area. We also nd a history of hazard exposure is signi cantly and positively correlated with Black Summer re hazard exposure. After controlling for all these confounders, socioeconomic disadvantage remains a signi cant correlate of the wild re hazard exposure.
Our spatial analyses reveals that the interlinkage between wild re hazard exposure and socio-economic disadvantage is not a major city phenomenon. This interlinkage is most prominent in inner regional, outer regional, and remote Australia. We hypothesise this is because of the availability of professional re ghting services in major city areas compared to the rest of the geographical areas where re suppression is done by volunteer re ghters in a large landscape under highly di cult weather conditions. This is further evidenced by our regression result that shows areas with a high-level business activity are less exposed to wild re hazard.
We used a novel wild re hazard exposure index to capture two dimensions of wild re hazard exposure, namely, re extent and re proximity to settlements. This operationalisation of wild re hazard can be used, where applicable, in future analyses for a more comprehensive understanding of wild re hazard exposure of communities. We also nd that the interlinkage between wild re hazard exposure and socio-economic disadvantage remains positive and signi cant even when the wild re hazard exposure is operationalised using either wild re extent or wild re proximity.

Conclusions
Our ndings make an important contribution to the understanding disaster risk, the rst priority area for action of the Sendai Framework for Disaster Risk Reduction 2015-2030 which urges for a greater integration of scienti c knowledge and understanding on all dimensions of disaster risk to design appropriate preparedness and effective response strategies. Contrary to the ndings reported by the studies from the United States, our results reveal that socio-economically disadvantaged communities experienced highest wild re exposure during 2019−2020 re season in Australia.
Our ndings establish that this link between socio-economic disadvantage and wild re exposure was least prominent in major cities but was prevalent in the re affected locations of inner and outer regional locations and in rural Australia.
In disaster risk management and planning contexts, the built environment receives the highest priority when it comes to exposure analysis. Socioeconomic disadvantage is taken into account primarily for vulnerability analysis (in terms of sensitivity and adaptive capacity) (Royal Commission into National Natural Disaster Arrangements 2020). Our study reveals that socio-economic disadvantage of communities are strong determinants of disaster exposure; hence, they should also be taken into careful consideration when developing strategies to reduce disaster risks. Additionally, as noted by others , the spatial mapping of wild re extent, wild re proximity to settlements in a national database is another important way to improve wild re prevention and planning.
Future studies should test the potential pathways through which the hazard exposure and socio-economic disadvantage link is determined. Future studies should also explore the interlinkages among the other components of wild re risk, namely, hazard and vulnerability (i.e., sensitivity and adaptive capacity). This should offer a more holistic understanding of the spatial distribution of wild re risk and socio-economic vulnerability in Australia. Additionally, our wild re hazard exposure model does not account for smoke exposure. Smoke exposure not only leads to health risks but also in icts economic costs due to a reduction in demand for services in areas that are not directly impacted by the re. Thus, future analyses should also account for smoke pollution to capture a broader understanding of the link between wild re hazard exposure and socio-economic disadvantage.  Notes:

Tables
a The interaction of re extent (i.e., total burnt area as a percentage of the total size of a Statistical Area (SA)1) and proximity of re to settlements.