Regional participation trends for community wildfire preparedness program Firewise USA

Community-wide wildfire mitigation can effectively protect homes from structure ignition. The Firewise USA program provides a framework for grassroots wildfire preparedness. Here, we examine the 500 Firewise USA sites in California to understand participation and demographic trends. We find important regional differences regarding the influence of underlying fire hazard, fire history, and other Firewise sites on new site formation. Sites in the Bay Area and Sierras respond strongly to fire history and proximity to other Firewise sites, while Northern and Southern California have few Firewise sites despite underlying hazardous conditions and large fire history. Firewise sites are often whiter, older, and more well-educated than California’s median population, potentially leaving out many communities that do not meet this demographic profile but face severe risks from wildfires. These findings offer important insights into the factors motivating communities to pursue wildfire protection, particularly important given recent severe and destructive wildfire seasons.


Introduction
Climate change has contributed to the record-breaking large and destructive wildfires that have recently burned throughout communities in the Western United States and around the world [1]. Adaptation efforts to improve community wildfire protection typically entail fuel treatments or structure protection. Fuel treatments involve reducing fuel concentrations, often through mechanical thinning or prescribed burns, to remove combustible material and decrease the likelihood of severe wildfires [2]. Structure protection refers to both defensible space and home hardening. Defensible space is a buffer zone extending up to 100 feet (30 m) from a structure where vegetation and other combustible materials have been cleared and managed to prevent structure ignition [3]. Home hardening involves featuring or upgrading materials like roofs, windows, and vents with wildfire-resistant versions [4]. Homes in the wildland-urban interface (WUI) are at particular risk, both in the intermix WUI (where homes and fuels are mixed) and interface WUI (where homes abut fuels) [5].
Although individual structure protection can improve structure survivability during a wildfire, community-wide structure protection is more effective in protecting homes [6]. Suburban homes rarely have 100 feet (30 m) between them, exacerbating risk from wind-driven embers capable of igniting multiple structures. In higher-density neighborhoods, wildfire-related losses can cover the entire community, while a community of well-protected homes with low ignitability can survive without significant destruction [4]. Decisions to invest in wildfire protection often depend on neighbors' decisions. Homeowners who mitigate their wildfire risk are more likely to have neighbors who also mitigate [7,8]. However, wildfire mitigation can quickly become expensive. Costs can become a significant challenge for homeowners interested in reducing their wildfire risk [9]. Upgrading a roof to wildfire-resistant standards can cost up to $22 000 USD, while replacing exterior walls (including windows and doors) can exceed $40 000 USD [10].
Prior research has explored how and why communities pursue wildfire mitigation, though most of these studies examine individual resident activities rather than organized community programs. Residents are motivated to pursue wildfire mitigation activities due to their prior exposure to wildfires [11][12][13], disposable income [14], regulations based on underlying hazards or zoning [15], knowledge of structural risks [13], and belief that mitigation improves structure protection [16]. Residents who are older or retired, have higher household income, and have a bachelor's degree or higher complete more mitigation activities than younger [13,17,18], less wealthy [13,[17][18][19], and less educated residents [17,19]. Census tracts with predominantly Black, Hispanic, or Native American populations are disproportionately vulnerable to wildfires [20], but predominantly white neighborhoods are more likely to participate in community-wide wildfire mitigation efforts [21].
The National Fire Protection Association (NFPA) oversees a volunteer, grassroots, community-based wildfire preparedness program called Firewise USA [22]. As of 2022, more than 1,800 communities in the United States are recognized Firewise sites [23]. Forming a Firewise site involves five steps: (1) create a committee and identify the proposed site's boundaries (8-2500 single-family dwelling units); (2) obtain a written wildfire risk assessment that identifies priority areas for wildfire risk reduction; (3) develop a risk reduction and education action plan; (4) host an outreach event and implement risk reduction activities; and (5) track hours and financial investments in mitigation [24]. To renew their status and remain in good standing, sites must host an outreach event and dedicate at least one volunteer hour or its monetary equivalent ($28.54 USD, adjusted for inflation in 2021 dollars) per dwelling unit toward mitigation per year [23][24][25]. Potential mitigation activities include consultations with professionals, project coordination, attendance at educational events, and resident or contractor labor on structure protection or fuels management [25]. Homes in active Firewise sites are eligible for discounted insurance prices in certain states, including California [26], where insurance companies have recently issued policy non-renewals or cancellations given wildfire risks [27].
Firewise USA recognition is part of an ongoing and flexible process toward improved wildfire protection, 'not the end-all-and-be-all of wildfire safety' [28]. Each Firewise site chooses its mitigation activities, such as evacuation training, fuel treatments, defensible space, or home hardening [28]. Firewise recognition does not automatically equate to reduced wildfire risk, but dedicated Firewise sites that actively reduce their risk are better protected during an actual wildfire [29]. When the 416 Wildfire burned near the highly-engaged Falls Creek Firewise site in Durango, Colorado, firefighters used the site as a base and successfully protected the community's homes [30]. Successful sites typically have motivated leaders who engage their neighbors in recognizing and addressing their collective wildfire risks [31].
However, there is limited research on demographic makeup of communities in Firewise USA or on what factors may have influenced participation in the program, with most prior research focused on sites in New Mexico. Interviews across 16 Firewise sites in New Mexico revealed that participants recognized the value of Firewise in improving wildfire protection and documenting investments and progress to share success stories and retain institutional knowledge [32]. Participants in New Mexico's Firewise sites were generally aware of local fire dangers and local fire histories, and were often inspired to participate because of a nearby community's engagement in Firewise USA [33]. Firewise site formation is often reactive following wildfires that threaten or destroy nearby buildings, rather than proactive [29]. Here, we seek to identify trends in community participation in the Firewise program in California and to place these findings in context with prior research exploring how and why residents pursue wildfire mitigation activities [11][12][13][14][15][16][17][18][19][20][21]. As of January 2022, California had 500 Firewise sites established between 2005 and 2021, more than any other state [34]. Firewise has recently grown in popularity in California, particularly following recent wildfire seasons. Thirteen of the most destructive wildfires and twelve of the largest wildfires in California's history occurred between 2017 and 2021 alone [35,36]. Exploring the factors associated with the increase in engagement in the Firewise program may improve our broader understanding of why communities pursue wildfire mitigation activities.
Here, we explore three research questions intended to understand participation and demographic trends among Firewise sites in California. First, how has participation in Firewise USA changed between 2005 and 2021 across California based on self-reported data from Firewise sites? Second, what are the demographic and socioeconomic characteristics of Firewise sites, and how have those characteristics changed over time? Finally, what is the influence of underlying fire hazard, proximity to large wildfires, proximity to the WUI, and proximity to other sites on new Firewise sites? We hypothesize that Firewise communities will follow similar demographic and participation patterns as those from other studies that examine why individuals or communities pursue wildfire mitigation activities. Most of these prior studies examining factors influencing community wildfire mitigation activities rely on survey data supplemented by statistical analyses [11][12][13][14][16][17][18][19]. By contrast, we rely on data from the NFPA, California Department of Forestry and Fire Protection (CAL FIRE), and the U.S. Census Bureau to examine patterns across Firewise sites through GIS-based statistical analysis. We also consider regional differences in fire hazard and wildfire history [37,38], and use spatial clustering to organize Firewise sites into four regions to analyze the sites collectively. Recognizing regional differences may support better and more targeted efforts to improve community-wide wildfire mitigation activities. Our results reveal that underlying fire hazard and socioeconomic and demographic characteristics influence Firewise participation. Improved understanding of how and why communities participate in Firewise may better inform regional community-based wildfire preparedness and protection, particularly important given recent climate-exacerbated severe wildfire seasons across California and the Western United States.

Data sources 2.1.1. NFPA
We requested and received data on all Firewise communities in California from the NFPA in January 2022. We received two datasets from the NFPA. The first dataset included information on Firewise sites in good standing as of January 2022, referring to actively participating sites: site name, status (in good standing), state, county, city, latitude, longitude, total investment, approval date, resident count, dwelling units, and how the leaders heard about Firewise. The second dataset included information on all Firewise sites in good standing established in California between 2008 and 2021: site name and renewal status, site name, status (in good standing), year, city, county, state, activities completed in 2021, and an optional narrative. Data stem from annual self-reported information by leaders of each active Firewise site collated by the NFPA. Data were cleaned and organized and are summarized in table S1. For the purposes of this study, 'new' sites refer to Firewise sites that were established in 2021, while 'renewing' sites refer to Firewise sites that formed prior to 2021 and renewed their status in 2021. We used the latitude and longitude coordinates presented in the data to spatially align Firewise sites with other spatially indexed data outlined in the subsequent sections. We used K-means clustering to place the sites into four distinct site-clusters based on their geography; we selected four site clusters as the optimal number of sites based on silhouette width and within sums of squares (figure S1). This approach enabled us to recognize regional differences in wildfire risk and community risk reduction activities [37,38].

CAL FIRE
In order to assess the relationship between fire hazard and prior fire history to Firewise site establishment, we considered fire hazard, WUI, and fire perimeter data. We also received geospatial data on underlying fire hazard in California from CAL FIRE's Fire and Resource Assessment Program. The dataset detailed the three fire hazard severity zone (FHSZ) levels in California (moderate, high, and very high) across federal, state, and local responsibility areas from 2020 (figure S2). The dataset included an additional layer of 'Not Mapped,' which refers to areas for which CAL FIRE has not released FHSZ maps. FHSZ maps identify underlying fire hazard based on prior fire history, terrain, fuels, and typical fire weather [39]. The federal government is responsible for wildfire protection in Federal Responsibility Areas, referring to federal lands. CAL FIRE is responsible for wildfire protection in State Responsibility Areas, including state and private lands. Local government agencies are responsible for wildfire protection in local responsibility areas, referring to local jurisdictions [40]. In addition, we accessed fire perimeter data from CAL FIRE between 1990 and 2020. We excluded wildfires under 5,000 acres in our analysis to exclude small fires based on CAL FIRE's definition of 'large fires' [41].
WUI and housing density shapefiles and definitions were pulled from Radeloff, Helmers [42]. For each site, we created a buffer radius at 1 km intervals from 1 km to 50 km. We calculated the proportion of area within each buffer classified as moderate, high, and very high FHSZ. We performed the same analysis using WUI and housing density data. We cropped coastal sites accordingly to avoid incorporating a proportion of the ocean. With each site, and for each buffer size, we aggregated the proportion of underlying area surrounding the site-by-site cluster, identifying both proximate and distal patterns in FHSZ, WUI, and housing density. We hypothesized the Firewise sites would be predominantly located in very high FHSZs, Interface/Intermix WUIs, and medium-low housing density due to underlying hazard and state mitigation regulations [15].
To understand the influence of prior wildfires on Firewise site creation, we calculated the total number of acres burned in each documented fire in each year and calculated the distance of each fire to each Firewise site. Aggregating the number of fires and areas burned within 50 km of each site cluster allowed for the exploration of the association between proximate, large, wildfires and Firewise site generation (i.e., whether Firewise site generation can be considered reactive vs. proactive). The wildfire data from CAL FIRE was available only through 2020, so we could not examine any relationship between wildfires and site generation for the year 2021. We hypothesized that Firewise sites would be predominantly established rapidly in response to large regional wildfires [11][12][13].

American Community Survey
We extracted estimates from the U.S. Census Bureau at the block-group and zip-code tabulation area (ZCTA) geographic levels using the IPUMS database [43]. We used American Community Survey (ACS) estimates at both geographic levels. Table 1 articulates the variables used from each geographic level, as well as those derived from CalFire.
Block-group estimates were used to summarize the underlying demographics of sites to provide the most granular neighborhood estimate (figure 2). Specifically, we extracted race estimates (table B02001) (table B15003), and poverty estimates (table C17002) from the 2010-2014 and 2015-2019 ACS; earlier estimates for these variables were not available [43]. Given the longitudinal aspects of the data, we pulled variables from the block-group level for several iterations of the ACS to explore any demographic shifts over time, and found that the demographics of site clusters remained relatively stable over the study period (table S3).
Based on prior literature, we hypothesized that (1) predominantly white census block groups would have more Firewise sites than census blocks with predominantly minority populations [21], (2) census blocks with Firewise sites would have predominantly low poverty levels [13,[17][18][19], (3) census blocks with Firewise sites would have high proportions of elderly residents [13,17,18], and (4) census blocks with Firewise sites would have high proportions of residents with a bachelor's degree or higher [17,19].
For the purposes of spatio-temporal modeling of site proliferation, we abstracted five-year ZCTA estimates from 2015-2019 to all sites, regardless of when the site was established based on data availability and approximate timing of site formation. We chose the ZCTA-level geography for more reliable estimates and to have more sites nested within the same geographic unit for better model convergence. Specific ZCTA area attributes included total population size (

Derived variables
We used four derived variables in this analysis: site cluster, fire count, fire area, and fire hazard (%). Site cluster was created by using the k-means methods outlined in 2.1.1 to place Firewise sites into four distinct geographic regions in California (Sierras, Southern California, Northern California, Bay Area). Fire count was calculated by running a spatial intersection between wildfires over 5000 acres in size to the ZCTA polygon for each year of the analysis (2005-2021). We used the frequency of overlapping polygons to determine the frequency of large wildfires within a given ZCTA. We next computed Fire Area by calculating the area of overlap between a given ZCTA and a large wildfire perimeter. Finally, we calculated Fire Hazard using a spatial intersection of ZCTA polygons and the FHSZ dataset provided by CAL FIRE. We captured the percentage of area within each ZCTA polygon that was composed of moderate, high, and very high FHSZs.

Descriptive statistics
After integrating block-group demographics, we grouped findings of each demographic variable by site clusters, presented in figure 2. We performed the non-parametric 2-sample Wilcoxon test to compare each site clusters median demographics against those of the state of California writ large. Mean and standard deviations of these variables were calculated across at least two iterations of the ACS to assess stability of the estimates of time (table S3).
To assessing the buffer analysis, we selected buffer distances of 1-, 5-, and 25 km for each site. The proportion of area occupied by each classification of FHSZ, WUI, and housing density were calculated for each site and presented in figures (figures 3, S4 and S6). Generalized linear models were also performed to explore the interaction of underlying area across site clusters (tables S4-S6).

Moran's I
We next considered if Firewise site generation was a function of proximity to previously established sites. We hypothesized that Firewise sites were predominantly established near other Firewise sites [32]. Relying on the coordinate data and approval time for each Firewise site as provided by the NFPA, we generated spatial neighborhoods using the spdep package in R [44]. We then used K-nearest neighbors to generate a neighborhood list of each Firewise site's 20-nearest sites. We chose 20 neighbors, as they maintained the clustering of site groups on the Sierras, Northern California, Bay Area, and Southern California (i.e., very few neighbors existed across clusters), but similar results were also obtained with higher numbers of neighbors. We next calculated Global Moran's I to test for spatial autocorrelation of the approval date of Firewise sites. From there, we calculated Local Moran's I to classify low-low, low-high, high-low, and high-high clusters of spatial autocorrelation. We then generated P-values for each site's spatial autocorrelation and adjusted the false discovery rate using the Benjamini-Hochberg procedure [45]. Significance of the adjusted p-values was set to 0.05.

Spatio-temporal modelling
We modeled Firewise site proliferation by applying a longitudinal multi-level modeling approach in which the rate of Firewise Sites established per 1000 people in each ZCTA are nested within a given year. We used a spatio-temporal conditional autoregressive model (CAR) to account for spatio-temporal autocorrelation between ZCTAs over time. Applied using the CARBayesST package developed by Lee et al [46], our approach models the prevalence of Firewise sites per 1000 residents, while accounting for spatio-temporal random effects by employing a multivariate AR(1) temporal process and a spatially autocorrelated precision matrix [46,47]. The AR(1) temporal process helps adjust for temporal autocorrelation (values at one time point influence values in the next time point), and the spatially autocorrelation precision matrix takes into account the spatial neighborhood of a ZCTA, and the fact the neighboring ZCTAs have more similar traits than distant ZCTAs. The outline of the generalized linear model is given in equation (1) Y kt |µ kt ∼ f(y kt | µ kt , νˆ2) for k = 1, . . . , K, t = 1, . . . , N, where Y kt is the Poisson-distributed count of Firewise sites in a given ZCTA, k, at year t. O kt , represents the offset of ZCTA population size; this is assumed to be constant throughout the study period (2005-2021). x ⊤ ktβ is a vector of the model covariates for each ZCTA, k, at time t. β represents a vector of regression parameters with a multivariate Gaussian prior with mean µ β and diagonal variance matrix Σ β . ψ kt denotes the spatio-temporal autocorrelated random effects, which are expanded upon in equation (2) ψ kj = ϕ kt , where ϕ kt is a vector of random effects at time t for ZCTA k, and evolves over time as a consequence of the multivariate AR(1) process (ρ T ). τ 2 Q(W, ρ s ) −1 denotes the spatial autocorrelation directed at the variance.
Q (W, ρ s ), is a precision matrix first introduced by Leroux et al [48]. While the number of site-specific variables were limited enough for parsimonious model building, the sheer number of available variables from ACS at the ZCTA level made the prospect of a formal model-fitting process unwieldy. Therefore, we established models a-priori based on previous literature around Firewise and engagement in community wildfire protection. Specifically, we consider the percentage of the population with mortgages due to the requirement to have fire insurance for homeowners with mortgages, who are homeowners, who earn greater than 1.5x the poverty line, who are college-educated (bachelor's degree or higher), who are non-white, and who are over the age of 65. In addition to demographic variables, the proportion of area in a given ZCTA labeled at High or Very High FHSZ, the total count of fires over 5000 acres occurring in at least part of a given ZCTA in a given year (from the derived variable fire area), and the total acreage burned in each ZCTA in a given year were also considered.
The above model relied on Markov-chain Monte-Carlo (MCMC) methods for parameter estimation. Specifically, we used a Metropolis-adjusted Langevin algorithm (MALA) for the estimation process. The MALA uses Langevin diffusion to propose new steps in the state-space and converges well in a high-dimensional state-space [49]. We used parallel computing to simultaneously produce 11 MALA-derived MCMC chains to generate model estimates. To ensure convergence of these estimates, we ran each chain to generate 700 000 total samples, with a burn-in rate of 500 000 samples, and a thinning interval 300 samples generated 666 viable estimates per chain, for a total of 7326 estimates, in which all Gelman diagnostics were below 1.1 (table S7).
Finally, to check the robustness of the spatio-temporal model, we performed a sensitivity analysis. We ran a generalized linear mixed model with random slope/intercept, with an AR(1) temporal random effects process to appropriately account for temporal dependence.
All analyses and summaries were generated using R statistical software, version 4.1.2 [50].

Limitations
This research study focuses exclusively on Firewise sites in California, which represent 500 of the more than 1800 sites around the country. Demographic and participation trends among Firewise sites in California may differ from Firewise sites in other states, particularly those that have not experienced such severe or frequent recent wildfires. In addition, we consider Firewise sites within four regions identified by k-clustering rather than examining each Firewise site as an individual site. Furthermore, the Firewise data from NFPA are self-reported information from local Firewise leaders and have not been reviewed or verified. The geospatial data provided by NFPA were single coordinate points selected by local Firewise community leaders rather than parcels that trace the outline of the site's participating residences. Therefore, all of our subsequent analyses examining trends in socioeconomic and demographic characteristics, location within FHSZ levels, prior wildfire history, and proximity to other Firewise sites are based on single coordinate points that represent each Firewise site.
Furthermore, we recognize that aligning data from the Census Bureau to coordinates of Firewise sites is potentially problematic as this exposes our research to a potential ecological fallacy in which the demographics of the sites may not reflect the true demographic conditions of the Census geographies. We acknowledge that Firewise communities do not align perfectly with census geographies and therefore likely elicit some mismatch. However, given the data collected by NFPA, we pursued this approach to place sites within a socio-economic and demographic context. In addition, we also obtained Census estimates at different geographies (ZCTA and block-group) and at different years to provide the most transparent and robust assessment of our analysis. Based on spatial clustering, we identified four regions of Firewise clusters in California: Southern California, Bay Area, Sierras, and Northern California ( figure 1(b) [51].

Firewise site characteristics
Based on a questionnaire of annual Firewise activities filled out by 487 sites in 2021, sites emphasized community-wide education rather than active mitigation (table S2). The most common activities included coordinating a community-wide awareness/educational activity to increase risk reduction and preparedness (308 sites), placing a wildfire-related article in the community newsletter (285 sites), and hosting a presentation detailing the need for and importance of individual wildfire preparedness (280 sites). The least common activities were fire-resistant plant species workshops (26 sites) and holding an insurance policy clinic (27 sites).

Census block-group demographics
We examined the proportion of non-white and non-Hispanic residents in census block-groups with Firewise sites. Drawing on prior literature, we hypothesized that predominantly white census block-groups have more Firewise sites than census block-groups with predominantly non-white populations [21]. As expected, we found that Firewise sites are in predominantly white and non-Hispanic census block-groups across all regions and years in which Firewise sites are established, and at comparatively higher levels than the median estimates for California ( figure 2(a) and (b)).
We next hypothesized that census blocks with Firewise sites would have predominantly low poverty levels (median income <1.5x poverty level) [13,[17][18][19]. We found that Firewise sites are in predominantly high-income census block-groups, although sites in Northern California are contained within areas with consistently lower income levels than those in other regions (figures 2(c) and (d)). Sites in the Bay Area had the highest income and lowest poverty levels, reflecting the socioeconomic characteristics of the region.
We also hypothesized that areas with Firewise sites have higher proportions of residents with a bachelor's degree or higher [17,19]. We found differences in education between regions (figure 2(e)); Bay Area sites are generally in well-educated census block-groups, while sites in Southern and Northern California had the lowest levels of educational attainment and are comparable to the median estimates for California.
Finally, we hypothesized that census block-groups with Firewise sites would have higher proportions of elderly residents [13,17,18], based on the proportion of residents over 65 years old. Census block-groups of Firewise sites have consistent proportions of residents over 65 years (approximately 20%), substantially higher than the median estimate for California (approximately 10%) (figure 2(f)).

Fire hazard, history, and WUI
CAL FIRE divides California into moderate, high, and very high FHSZs that reflect levels of underlying fire hazard based on local characteristics such as fire history, fuels, typical fire weather, and terrain. We hypothesized that Firewise sites are predominantly located in very high FHSZs due to underlying hazard and state mitigation regulations [15]. Using buffer analysis, we found that Firewise sites in the Sierras, Northern California, and Southern California tend to be in areas with high proportions of very high FHSZs (figure 3). By comparison, sites in the Bay Area tend to be in areas with a comparable mix of moderate, high, and very high FHSZs. Therefore, communities in the Sierras, Northern California, and Southern California establish sites predominantly in very high FHSZs, whereas communities in the Bay Area form Firewise sites despite their mixed underlying fire hazard.
In addition, we hypothesized that Firewise sites are predominantly established quickly in response to large regional wildfires [11][12][13]. Calculating the cumulative area of wildfires within 50 km of sites over time, we identified two distinct patterns of Firewise site formation following wildfires (figures 4 and S6). Communities in the Bay Area and Sierras rapidly establish many sites in response to the number of acres burned. By contrast, communities in Northern and Southern California establish few Firewise sites despite repeated large wildfires.
In examining Firewise sites in relation to WUI classifications, we found that sites in the Sierras were in predominantly Intermix zones (74%). Bay Area sites were located mainly in Interface zones (57.6%) (tables S1 and S5; figure S3). Further analyses of housing density data reveal that most site-clusters were in Moderate density areas; while the Bay Area had substantially higher amounts of high-density housing (table S6). Buffer analysis confirms these general trends from 1 to 25 km ( figure S5).

Proliferation of Firewise sites
We next explored the possibility of a 'domino' effect for sites. We hypothesized that the establishment of sites across California was neither explicitly random nor purely a function of fire hazard or history, but rather that proximity to other sites may spur proliferation. We used Moran's I statistic to reveal statistically significant  spatial autocorrelation in the approval date of Firewise sites. Figure 5(a) depicts the Moran scatterplot with test statistic and p-value, indicating that sites near one another have similar approval dates. Figure 5(b) shows clustering of sites with statistically significant Local Moran's I statistics, with sites in Northern California designated as low-low clusters, meaning that sites near one another were established early in the study period. Meanwhile, the Bay Area exhibits a large high-high cluster; sites in the Bay Area were therefore established later in the study period with spatial autocorrelation with nearby sites, indicating a domino effect occurring later in the study period.

Spatio-temporal analysis
Finally, we aimed to understand the relationship between site creation and demographic information and fire hazard. Spatio-temporal modeling (table 2) demonstrates that the prevalence of Firewise sites per capita within ZCTAs is strongly associated with underlying demographics. We find that fire hazard, mortgage percentage, homeowner percentage, non-white population percentage, and percent of the population over 65 were statistically significant, accounting for substantial variability on the rate of Firewise site creation. Fire Hazard, Homeowner percentage, and population over 65% were positively associated with Firewise site creation, meaning that higher values of these variables were predictive of a higher rate of site creation. The percentage of homes with mortgages and non-white populations were negatively associated with Firewise site creation. The proportion of ZCTA area designated as High or Very High FHSZ was statistically significant. No other wildfire-related variables were significant. These results further underscore the role that demographic and socio-economic characteristics play in site proliferation, diminish the role of large/proximate fires, and reinforce our findings in section 3.3 in which FHSZ seem to correspond with the placement of Firewise sites.

Discussion
The rise in participation in the Firewise USA program reveals growing engagement in community-wide wildfire preparedness and mitigation, with the biggest increase in Firewise sites throughout the state occurring after repeated record-breaking large and destructive wildfires between 2017 and 2020 [35,36]. In California, regional participation trends reflect underlying fire hazard (Northern California, Southern California, Sierras), prior fire history (Bay Area, Sierras), and proximity to other Firewise sites (Northern California, Bay Area). Across all regions, sites tend to be in census block-groups that have low proportions of non-white and non-Hispanic populations, and with approximately 20% of the population over 65 years old.
Other research into Firewise sites in California has similar demographic findings, though our work extends this effort by incorporating regional differences [52]. Bay Area sites tend to be in more well-educated and wealthier census block-groups than other regions, while Northern California sites are in less well-educated and wealthy census block-groups than other regions. These findings reveal distinct regional factors influencing grassroots wildfire preparedness; Firewise sites, like wildfire risk, are not homogenous within California.
Our results therefore further highlight the importance of local context in climate adaptation [53]. Societal norms, values, and cultures influence the adaptive capacity of communities to respond to climatic stressors like wildfires [54]. The Firewise USA program offers neighbors a framework to advance their own grassroots community mitigation activities, but each Firewise site is self-organized and acts individually. Though we consider these individual sites within larger regions, each site is distinct and reflects local risks, concerns, and cultures. Our spatio-temporal and demographic analyses indicate that Firewise sites often meet similar demographic profiles (e.g., whiter, older, and more well-educated than the median population of California), potentially leaving out many communities that do not meet this demographic profile but face severe risks from wildfires [20]. More research on factors influencing Firewise USA participation at the individual site scale and regional trends across the United States could better inform trends in community wildfire mitigation. In addition, future research should identify what factors-such as lack of awareness of Firewise, low concern of wildfire risk, or limited resources-may influence why communities choose not to participate in the Firewise USA program.
This research builds on prior studies exploring why individuals and communities are motivated to pursue wildfire mitigation. For example, homeowners who believe that mitigation activities are effective at reducing their wildfire risk will be more motivated to pursue these activities, and previous mitigation actions are often an indication of future interest and intention [55]. Homeowners are also more likely to pursue wildfire mitigation following a major wildfire, especially in the case of a near-miss event, which can prompt community organization, outreach, and education [56]. In addition, other research has discussed four WUI community 'archetypes' for wildfire planning and mitigation ranging from formalized suburban to working landscape communities in the Western United States [57,58]; though we do not group individual Firewise sites into these archetypes, we similarly emphasize regional differences and how they can inform increased interest in wildfire mitigation.
Our findings also reveal new opportunities for improved community wildfire protection. For example, Northern and Southern California have few Firewise sites despite high prevalence of hazardous areas and recent destructive wildfires, indicating minimal influence of prior wildfire history on site formation. Unlike communities in other regions which form Firewise sites in areas with very high FHSZs, communities in the Bay Area form Firewise sites despite a mix of moderate, high, and very high FHSZs, potentially indicating greater wildfire concern across areas of lower hazard in the Bay Area. Results from our Moran's I analysis show that many sites may form at similar times, such as in Northern California and the Bay Area. Local fire departments or Fire Safe Councils in both regions could share more information on community wildfire mitigation programs to help propagate additional sites and clusters of additional sites. In addition, Firewise sites tend to emphasize education (e.g., distributing information or hosting community meetings on wildfire preparedness) rather than mitigation (e.g., volunteering to remove fuels for an elderly neighbor, organizing evacuation drills). Education likely improves community awareness of wildfire risk and hazard reduction activities [59,60] but falls short of actual mitigation [61]. Future iterations of Firewise USA could establish more stringent requirements to reduce wildfire risk beyond education and outreach [23,62] or incorporate professional training to ensure adequate protection [63].
Greater emphasis on active mitigation (home hardening and defensible space) could also help address the ongoing insurance crisis in which homeowners face limited or very expensive fire insurance options. In 2017 and 2018, insured losses from wildfires in California exceeded $23 billion [64,65], prompting private insurance companies to issue non-renewals and the California Department of Insurance to issue moratoria on non-renewed or canceled insurance policies to protect nearly 20% of policyholders in the California insurance market; homeowners with mortgages are required to have fire insurance [27]. Fourteen insurers in California covering 40% of the state's housing market offer discounts based on home and neighborhood mitigation, with six insurers specifically providing discounts to Firewise sites [66]. More insurance companies might offer discounts to participating homes if they had greater assurance of annual risk reduction activities by homeowners, similar to the 'Safer from Wildfires' framework issued in California to identify mitigative actions that insurance companies should consider in their policy coverage [67].
Current investments by Firewise households are likely far short of the hundreds to thousands of dollars needed to install and maintain structure protection measures [10]. Mitigation costs may quickly become prohibitive, even for motivated homeowners highly engaged in wildfire mitigation. Homeowners may look to government grants or loans to support structure protection funding [19], but current investments are inadequate [68]. Although California committed nearly $1 billion USD in wildfire resilience in the 2021-2022 budget plus an additional $200 million USD over the next six years, that funding was primarily dedicated to wildfire suppression and forest health rather than structure protection [69]. Greater emphasis on community protection, similar to the Firewise USA program, combined with a significant increase in resources dedicated to widespread structure protection, is critical to reducing neighborhood risk and structure losses from wildfires.

Data availability statement
Firewise communities: Source data on Firewise communities are available upon request from the National Fire Protection Association.
Demographic and socioeconomic information: Source data on demographic and socioeconomic information are available from the U.S. Census Bureau's American Community Survey. Block-group and ZCTA shape files are available from the TIGER database.
Fire perimeters and risk: Source data on fire perimeters and fire hazard severity zones are available from the California Department of Forestry and Fire Protection's Fire and Resource Assessment Program.
The data that support the findings of this study are openly available at the following URL/DOI: https:// github.com/kampfsca/Firewise. risk GIS files. We thank An-Min Wu and Caroline Ferguson for their comments on draft versions of this manuscript.

Ethics declarations
R K M had a professional services agreement with the National Fire Protection Association to produce a report on the Firewise Sites of Excellence pilot project. No Sites of Excellence are located in California.

Authors' contributions
A R K and R K M conceived and designed the research. A R K designed and conducted the statistical analysis. A R K and R K M wrote the paper.