Factors associated with single-family home survival in the 2018 Camp Fire, California

The 2018 Camp Fire, which destroyed 18,804 structures in northern California, including most of the town of Paradise, provided an opportunity to investigate vegetation and housing factors associated with home loss and determine whether California's 2008 adoption of exterior building codes for homes in the wildland-urban-interface (WUI) improved survival. We randomly sampled single-family homes constructed: before 1997, 1997 to 2007, and 2008 to 2018, the latter two time periods being before and after changes to the building code. We then quantied the nearby overstory canopy cover and the distance to the nearest destroyed home and structure from aerial imagery. Using post-re photographs, we also assessed re damage and assigned a cause for damaged but not destroyed homes. Our analysis of post-re outcomes in the town of Paradise suggested that both the proximity to other burning structures and nearby wildland fuels factored in the probability of home survival, with several measures of distance and density of destroyed structures and nearby overstory canopy cover emerging as signicant explanatory variables. The relative importance of nearby burning home variables versus surrounding vegetation in explaining outcomes has varied among studies, with Gibbons et al. (2012) reporting canopy cover within 40m of the home to be the strongest predictor. Number of buildings within 40m was also a signicant variable in their analysis. Even though both nearby burning homes and vegetation variables were included in the same models in our study, interpretations about relative strength of these two sets of factors are tempered by limitations of the vegetation data, with overstory canopy cover an imperfect measure of wildland fuel hazard.

The increase in destructive wild re events has been linked to changes in re frequency, development patterns, and climate. Loss of indigenous burning and active re suppression over the past 150 or more years following Euro-American expansion into California reduced the incidence of re in many areas. As a result, surface and vegetative fuels have increased, especially in forest and woodland areas that historically burned most frequently, leading to more severe re when it does burn (Steel et al. 2015). Such res are also often more intense because re suppression has effectively eliminated much of the lower-intensity burning under more benign weather conditions. When landscapes now experience re, most often it is when wild re escapes initial attack under worst-case scenario weather conditions (Calkin et al. 2014). In addition, over the last several decades, warmer temperatures and longer re seasons (Westerling et al. 2006) have increased fuel volatility and the probability of ignitions coinciding with extreme weather conditions. In other areas with longer historical re return intervals and stand replacing re the norm, such as chaparral ecosystems in southern California, the re regime has not changed as dramatically (Keeley and Fotheringham 2001).
Further complicating the wild re challenges, human populations have increased nearly ten-fold over the last 150 years, with a substantial proportion of houses built within or among wildland vegetation (Radeloff et al. 2018). Partly due to the effectiveness of re suppression, most of these homes were not built or maintained with the goal of being able to withstand wild re in the absence of re suppression resources. In addition, home design or construction codes and standards to enhance a building's exterior resistance to wild re are relatively recent (International Code Council 2003), with substantial development having occurred prior.
Post-wild re analyses provide an opportunity to investigate why some houses survive and learn how to better co-exist with wild re in re-prone environments. During wild re, buildings can be subjected to three different wild re exposures -wind-blown embers, radiant heat, and direct ame contact . Embers are produced when vegetation ignites and burns (Koo et al. 2010). In large, fast-moving wild res burning under extreme conditions, embers can be transported several kilometers or more (Koo et al. 2010) and ignite buildings directly or indirectly . A direct ember ignition includes embers igniting decking or siding by accumulating on or next to the material, or penetrating vents or open windows and entering the building (Quarles et al. 2010;Hakes et al. 2017). In contrast, indirect ignitions occur when embers ignite combustible materials such as vegetation, bark mulch, leaf litter, neighboring buildings, or near-home objects such as stored materials, decks, or wood fences (Quarles et al. 2010;Hakes et al. 2017). Indirect ignition scenarios ultimately result in radiant heat and/or ame contact to the home or building. Direct ame contact and extended radiant heat exposures can ignite siding and other exterior-use construction materials or break glass in windows. Radiant heat exposure often occurs when a neighboring structure ignites. Evaluations of wild re home losses have frequently found large-scale destruction to be the result of direct or indirect ignition by embers rather than high-intensity re in wildland fuels directly impacting the home, with burning homes then leading to house-to-house re spread (Murphy et al. 2007;Cohen and Stratton 2008). However, the potential in uence of housing density on structure losses in wildl res has varied, with some studies nding a greater probability of loss at higher housing densities (Price and Bradstock 2013;Penman et al. 2019), while other studies have reported a greater risk at lower housing densities (Syphard et al. 2012;Syphard et al. 2014). Amount of nearhome combustible vegetation has also been linked to the probability of home loss in wild res (Price and Bradstock 2013;Syphard et al. 2014;Penman et al. 2019).
Even within these record-setting res, California leads the United States in having a building code with the objective of limiting the impact of wild res on the built environment. In the 1960s, the state began requiring homeowners to implement defensible space fuel modi cations within the rst 9 m (30 ft) of a building, a distance which was expanded to 30 m (100 ft) in high re hazard zones in 2005 (California State Board of Forestry and Fire Protection 2006). Work on standardized test methods to evaluate exterior-use construction materials for re performance began in the late 1990's and later incorporated into Chap. 7A, an addition to the California Building Code which was adopted in 2008. As a part of Chap. 7A, an approved list of exterior construction materials for roof coverings, vents, exterior walls, and decks, was created. Chapter 7A applies to new construction as well as remodels of existing residential and commercial buildings in some jurisdictions. The 2018 Camp Fire, which destroyed much of Paradise, California, provided an opportunity to evaluate the performance of buildings constructed after the adoption of Chap. 7A and explore factors associated with home survival.
The Camp Fire started on the morning of 8 November 2018, with the failure of an electrical transmission line and spread rapidly through wildland fuels comprised of mixed conifer forest, brush, grass, and dead and down surface fuels (Maranghides et al. 2021). Surface fuels were unusually dry due to persistently low relative humidity throughout the summer and fall and the late onset of fall rains (Brewer and Clements 2019). Driven by strong NE winds, the fastmoving re quickly reached the towns of Concow, Paradise, and Magalia, and became the most destructive wild re in California history. At least 85 people were killed and 18,804 structures were destroyed. A total of 62,053 ha (153,336 acres) were burned before the re was contained. A high proportion of the home and business losses occurred in Paradise -the largest town within the re footprint. The re passed from one side of Paradise to the other during one burn period over less than 12 hours (Maranghides et al. 2021). With the focus on saving people's lives, very few homes were subject to re-ghting efforts, and survival was therefore largely a function of characteristics of the home and surrounding environment. Previous similar analyses have typically combined data across multiple res and years, with variable degrees of re ghter intervention. The massive home loss in a single burn period with the Camp Fire presents an opportunity to investigate factors with fewer confounding variables.
The objective of this research was to answer three questions: 1) did proximity to nearby burning structures factor into the probability of home survival; 2) did fuels associated with nearby vegetation factor into the probability of home survival; and 3) was the full adoption in 2008 of Chap. 7A into the California Building Code associated with improved odds of home survival?

Methods
The Butte County Assessor's database, dated June 1, 2018, was used to extract 11,515 parcels within the Paradise city limits (Fig. 1). Parcels were sorted by use code and 7,949 single-family dwellings were selected, after discarding 89 without a listed build year. Mobile homes, businesses, and other non-single-family structures were excluded. We then linked Damage Inspection (DINS) data, obtained from CAL FIRE, to parcel number to ascertain damage sustained in the Camp Fire and whether the building was destroyed, partially damaged, or had no impact from the Camp Fire. We lumped homes classi ed as "damaged" into the "survived" category, because in most instances, the damage, based on photos included with the DINS data, was minor -e.g. cracked windows, bubbled exterior paint, or melted gutters, with the structure itself intact.

Sample population
For our analyses, we randomly selected 400 single-family dwellings in Paradise, strati ed by three time periods (Fig. 1): Time 1 = homes built before 1997, while Time 2 (homes built from 1997 to 2007) and Time 3 (homes built from 2008 to 2018) represented the two eleven year periods on either side of the 2008 adoption of Chap. 7A in the California Building Code. If the changes to the building code improved home survival, survival percentage in time period 3 should be signi cantly higher than survival in time period 2, especially after adjusting for any potentially confounding variables. The strati cation was done to ensure a large enough sample size in Time Period 3. Two hundred homes (out of 7288) were randomly selected in Time 1, one hundred homes (out of 519) were selected in Time 2, and 100 homes (out of 142) were selected in Time 3 (Fig. 1). More homes were selected during Time 1 because such a low percentage (13%) of older (pre-1997) homes survived. Of the population of homes that were randomly selected by the construction period, 24 of the surviving homes were noted as damaged in the DINS report, the rest undamaged. Damage was listed as "affected (1-9%)" for 23 of the damaged homes, and "minor (10-25%)" for one.

Variables
For each randomly selected home, we used Google Earth to measure the distance from the edge of the home (as de ned by edge of the roof, using pre-re images when destroyed) to the closest edge of the nearest home and nearest structure, as well as the nearest home and nearest structure that burned. "Nearest structure" was in most cases another single-family home, but also included mobile homes, businesses, detached garages, or outbuildings such as larger sheds. Small sheds -those < 120 ft 2 , where a building permit isn't required -were excluded. Such smaller sheds may have posed a threat to the home as well but were more challenging to consistently quantify, especially if under a tree canopy. We determined the density of structures in the surrounding area by counting the number of single-family homes, partially-built homes, mobile homes, and businesses (excluding small sheds) with midpoints (based on a visual estimate) included within a 100 m radius centered on the target home. We then counted how many of those structures were destroyed. We visually estimated the percentage cover of overstory vegetation from Google Earth images taken prior to the re in 2018 and/or 2017 within a 30 m radius circle centered on the selected home and between 30 m and 100 m from the selected home. Cover of the understory of grass and/or shrubs or landscape plantings was not estimated, as overstory canopy cover was relatively high, and this often obscured the understory. Some larger mid-story shrubs might have been included with the overstory due to the di culty in distinguishing them from trees. Lot size was provided in the Butte County Assessor's data. Whether the house was located in the Wildland Urban Interface (de ned as developed areas that have sparse or no wildland vegetation but are near a large patch of wildland) or the Wildland Urban Intermix (de ned as areas where houses and wildlands intermingle) was determined by overlaying a University of Wisconsin data layer on the city of Paradise (Radeloff et al. 2005). We used Radeloff et al. (2005) to de ne the interface as census blocks with at least 6.17 housing units km − 2 that contained < 50% wildland vegetation but were within 2.4 km of a heavily vegetated area (> 75% wildland vegetation) larger than 5 km 2 . Intermix was de ned as an area with more than 6.17 housing units km − 2 but dominated by wildland vegetation. Percent slope was calculated as the rise between the lowest and highest point along a 100 m radius circle centered on the home.

Analysis approach
Outcome and possible explanatory variables (S1 Table) were rst analyzed individually using a generalized linear model in SAS PROC GENMOD and assuming a normal distribution to evaluate whether they differed by time period or by outcome (survived, destroyed). To account for the sampling scheme, in this and all subsequent analyses, each observation was weighted by the inverse of its probability of selection -i.e., homes from Time Period 1 had a weight of 7288/200, homes from Time Period 2 had a weight of 519/100, and homes from Time Period 3 had a weight of 142/100. Comparisons among main effects (outcome, time period) and interactions (outcome × time period) were determined using Tukey's HSD test for multiple comparisons, when signi cant.
To determine whether proximity to nearby burning structures or overstory canopy cover were associated with home survival, we used a generalized linear model t for binary response data, with a logit link function and weighting to account for the sampling scheme. Variables in the initial model were: 1. Y-variable: Outcome (Survived/Destroyed); X-variables: construction time period, year built, land use category (Wildland Urban Interface/Intermix), distance to nearest destroyed structure, total structures destroyed within 100 m, overstory canopy cover within 30 m, overstory canopy cover between 30 m and 100 m, slope, and the interaction of each with the construction time period.
When independent variables were highly correlated (R > 0.6), only the one most clearly mechanistically linked to outcome was included. For example, 'Distance to nearest structure' was highly correlated with 'Distance to the nearest destroyed structure' and 'Total structures − 100 m' was highly correlated to 'Total structures destroyed − 100 m' (Table 1), so only the latter were included. Lot size was not included as there was no clear mechanistic link with home survival and we hypothesized that elements contributing to re behavior would be captured by correlated variables. The land use category was included to quantify differences in vegetation arrangement at scales larger than 100 m. Non-signi cant interactions and non-signi cant main effects for variables that did not have a signi cant interaction with time were sequentially removed to produce the nal model. To determine whether homes constructed after the Chap. 7A building code update survived at a signi cantly higher rate after factoring in all other possible confounding variables, the same analysis was conducted except without interactions with the construction time period. Table 1 Signi cance of individual factors by time period, outcome (destroyed, survived), and outcome x time period for a subset of single-family homes in Paradise, CA. Means for time period, outcome, and outcome × time period (when interaction was signi cant) are provided below (standard error in parentheses). Levels within variables followed by different letters were signi cantly different (P < 0.05).

2008-2018
Surv. 44 --54.0 a (7.7) We then designed models to rst test the effect of variables that may have directly in uenced home survival during the re and second, to test the effect of just the variables available prior to the re. The latter variables were ones that might be mitigated preemptively through planning, retro tting, or vegetation management. For each of these models, we determined the effect size and performed a regression tree analysis. Variables included for each approach (accounting for the re, pre-re only): 1. Y-variable, accounting for the re: Outcome (Survived/Destroyed); X-variables: year built, distance to nearest destroyed structure, total structures destroyed within 100 m, canopy cover within 30 m, canopy cover between 30 m and 100 m, slope. 2. Y-variable, pre-re only: Outcome (Survived/Destroyed); X-variables: year built, distance to nearest structure, total structures within 100 m, canopy cover within 30 m, canopy cover between 30 m and 100 m, slope.
To quantify the relative strength of continuous variables for explaining home survival, each of the dependent (x) variables were centered and scaled to have a mean of zero and standard deviation of one. Logistic regression (McCullagh and Nelder 1989) was then used to calculate coe cients and compare effect sizes. The logistic regression models were t using the svyglm function from the survey package in R (Lumly 2020). A decision tree for predicting home survival was produced using the rpart function in the rpart package (Therneau and Atkinson 2019) in R, t for binary response data with a logit link function (Breiman 1998). This approach is similar to logistic regression, where the linear predictor is a decision tree model. To determine the number of splits in the decision trees, we performed cross-validation 10,000 times to compute the optimal pruning parameters. We then used the average of the 10,000 optimal pruning parameters as the pruning parameter in the nal decision tree. The latter group of statistical analyses were completed using R version 4.0.0 (R Core Team 2020). Figures were made in R using the ggplot2 package (Wickham 2016).

Visual evaluation of damaged homes
To learn more about vulnerabilities of the Paradise home sample and gain insight into potential points of re entry, we reviewed the CAL FIRE damage inspection (DINS) spreadsheet (obtained from CAL FIRE 12/18/2018) and downloaded photographs of all damaged homes (N = 310 homes with pictures) associated with the damage assessment at: https://www.arcgis.com/apps/MapSeries/index.html?appid=af7e5bb3960a48c096ed910c640a30b3) Photographs typically keyed in on the damage, and we reviewed each, along with notes about damage in the DINS summary. Observed home damage was assigned to radiant heat, direct ember ignition, or ame impingement categories (S2 Table). Homes where ame impingement was recorded were further split into three categories: 1), caused by fuel continuity with the broader landscape (which allowed re to reach the home), 2) indirect ember ignition (e.g., gutter contents, near-home fuels) with ames then impacting the home, or 3) unknown/undetermined. [The DINS assessment gathered similar information, but the full suite of data was not collected for over a quarter of homes and ember ignition was not separated into direct and indirect categories.] Where DINS data were collected, our evaluation was often in agreement, but there were a few instances where we differed. For example, if the DINS assessment noted "direct ame impingement" but the photo showed no charring or near home fuels consumed, we listed the damage cause as "radiant heat". Gutter res were variously categorized but we assigned all to the "indirect ember ignition" category. The DINS assessment also only lists a single cause of re damage when a considerable number of homes displayed multiple causes.

Results
Overall, most (86%) of the single-family homes in Paradise were built before 1990, and homes of this age fared poorly, with only 11.6% surviving the Camp Fire (Fig. 2). Survival increased to 20.6% for homes built between 1990 and 1996, 34.3% for homes built between 1997 and 2007, and 43.0% for homes built between 2008 and 2018. The 400 randomly selected homes in our sample had similar survival rates to the full population of single-family homes − 11.5% vs.
13.3%, respectively, for the < 1997 time period (Time = 1), 37.0% vs. 34.3%, respectively, for the 1997-2007 time period (Time = 2), and 44.0% vs. 43.0%, respectively, for the 2008 to 2018 time period (Time = 3). Many of the potential explanatory variables differed over the three time periods as well, and were therefore confounded with potential construction or building code differences (Table 1). Older homes (< 1997) were on average in areas with higher housing density and had more homes burn within 100 m than homes built from 1997-2007 (Table 1). Homes built prior to 1997 had a higher average overstory canopy cover in the rst 0-30 m from the home than homes built afterwards ( Table 1). The 'Distance to nearest destroyed structure' × Time interaction was signi cant, with surviving homes a greater distance from the nearest destroyed structure in time periods one and three. This difference was especially pronounced for the newest homes (Table 1). While average lot size increased numerically over time, the differences were not signi cant (Table 1). Overstory canopy cover 30-100 m from the home was signi cantly lower for surviving homes (37.0 %) than destroyed homes (50.4%) but did not differ between time periods (Table 1). With most houses situated on top of a plateau, the average percent slope was relatively low and did not differ signi cantly among outcomes or time periods (Table 1). None of the variables differed between time periods 2 and 3 -immediately pre-and post-Chap. 7A adoption.
Many of the continuous variables we analyzed were signi cantly correlated with each other, with distance to nearest structure and distance to nearest destroyed structure (r = 0.625) and total structures within 100m and total structures destroyed within 100m (r = 0.926) being the most strongly correlated (Table 2). Factors in uencing home survival Eliminating the two most highly correlated variables (distance to nearest structure and total structures per 100m) and analyzing the remaining variables together in the same model showed that both nearby destroyed structures and overstory canopy cover were signi cantly associated with home survival. The 'distance to nearest destroyed structure' × construction time period interaction was signi cant (Table 3), with a much higher survival probability when homes were a larger distance from a destroyed structure, especially for homes built 1997built -2007built and 2008built -2018. Total structures destroyed within 100 m also was strongly linked to home survival (Table 3), with a much higher survival probability when fewer surrounding homes burned (Fig. 3b). For the vegetation variables, the 'CanopyCover 0-30m' × construction time period interaction was signi cant (Table 3). Higher survival was noted with lower canopy cover for homes built since in 1997 and after, but was not related to survival in older (< 1997) homes (Fig. 3c).
CanopyCover 30-100m also was highly signi cant, with a higher survival probability at lower canopy cover percentages across times (Table 3, Fig. 3d). Land use category was signi cant, with a higher survival rate for homes in the wildland urban intermix (29.3%) than homes in the wildland urban interface (16.0%).
Year built [within construction time period] and slope were not signi cant and did not make it into the nal model (Table 3). For the next set of analyses, separate models (this time without specifying construction time period) were run on normalized data for 1) variables in play during the Camp Fire (including re-related variables), and 2) variables present prior to the Camp Fire (i.e., variables that might factor into pre-re planning). For the rst model, distance to the nearest destroyed structure had the largest effect size, suggesting that the greater the distance to a burning structure, the higher the probability of survival (Fig. 4a). Also signi cant were canopy cover within 30-100 m and the number of destroyed structures within 100 m. Both the latter two variables had a negative relationship with survival, with higher survival where canopy cover within a 30-100 distance was lower, and number of destroyed structures within 100 m was fewer (Fig. 4a). Year built, slope, and canopy cover within 0-30 m all had con dence intervals that overlapped with zero. When only pre-re variables were included, housing density had the largest effect size, with greater survival when the number of structures within 100 m was low (Fig. 4b). Canopy cover within 30-100 m had the second largest effect size, with greater survival at lower canopy cover levels (Fig. 4b). Distance to nearest structure, year built, slope, and canopy cover within 0-30 m all had con dence intervals that overlapped with zero (Fig. 4b).
Decision tree analysis using variables present during the re indicated a threshold of 18 m from nearest destroyed structure best predicted whether a home survived or not. Survival probability for homes < 18 m to the nearest destroyed structure was very low (0.058), compared with a 0.354 survival probability for homes ≥ 18 m from the nearest destroyed structure (Fig. 5a). Based on our sample, a majority (73.6%) of the homes in Paradise were < 18 m from a destroyed structure. For the 26.3% of homes ≥ 18 m from a destroyed structure, if the overstory canopy cover was also < 53% within 30-100 m, the survival probability improved to 0.481 (Fig. 5a). If the home was also built during or after 1973, the survival probability improved to 0.606 (Fig. 5a). The nal split, involving just 10.2% of the homes in Paradise, suggested that for homes meeting these criteria (i.e., ≥ 18 m from the nearest destroyed structure, < 53% canopy cover within 30-100 m, and built ≥ 1973), the survival probability improved to 0.733 if slope was less than 8.2%. For the decision tree including just pre-re variables, year built was the rst split, with a probability of survival of only 0.111 for homes built before 1996 (90.8% of homes in Paradise), compared with 0.396 for homes built during or after 1996 (9.2% of homes) (Fig. 5b). For homes in this latter category, survival probability improved to 0.766 if the overstory canopy cover within 30-100 m was < 33%. If canopy cover within 30-100 was ≥ 33%, the survival probability fell to 0.239.
Damaged homes -nature of damage and cause In our review of photographs of the 310 damaged homes in Paradise, 63% had radiant heat damage (Fig. 6a), mostly to windows and exterior walls (Fig. 6b). Window damage consisted of cracked or broken glass and damaged window framing, but frequently included both. Blistered paint or melted/sagging vinyl siding were the most common wall (siding) damages. In most cases, the source of the radiant heat was di cult to assess, as the photos focused on the damage. However, a closer investigation of 20% of randomly sampled of homes where radiant heat damage was identi ed demonstrated that all had at least one neighboring structure that was destroyed during the re. The average distance to the destroyed structure was 12.1 m. Flame impingement was the next most common cause of damage (44% of damaged homes) (Fig. 6a). In most ame impingement cases (28% of the total damaged homes), the damage was interpreted to be the result of indirect ember ignition. For only 10% of damaged homes was the continuity of fuels from the broader surroundings (often needle or leaf litter) identi ed as the likely reason for ame impingement. For another 10% of damaged homes, whether needle or leaf litter was continuous with the surroundings or just localized next to the home could not be determined from the photograph. [Note -these three ame impingement categories do not add to 44% because some houses showed evidence of multiple ame impingement causes.] For the cases of ame impingement via indirect ember ignition, embers ignited near home ammable objects (e.g., fences, patio furniture, stored lumber), near home leaf litter, near home vegetation (or litter under that vegetation), leaf litter in gutters, or wood bark mulch, in order of frequency from most to least (S2 Table). Direct ember ignition was identi ed as the likely cause of damage for fewer than 6% of homes (Fig. 6a). The most common locations for embers to ignite were attached wood stairs, decking, and window trim. Counting either direct ember ignition or ame impingement due to indirect ember ignition, embers were implicated as a cause in 33% of damaged homes.

Burning structures and wildland fuels both in uence home survival
Our analysis of post-re outcomes in the town of Paradise suggested that both the proximity to other burning structures and nearby wildland fuels factored in the probability of home survival, with several measures of distance and density of destroyed structures and nearby overstory canopy cover emerging as signi cant explanatory variables. The relative importance of nearby burning home variables versus surrounding vegetation in explaining outcomes has varied among studies, with Gibbons et al. (2012) reporting canopy cover within 40m of the home to be the strongest predictor. Number of buildings within 40m was also a signi cant variable in their analysis. Even though both nearby burning homes and vegetation variables were included in the same models in our study, interpretations about relative strength of these two sets of factors are tempered by limitations of the vegetation data, with overstory canopy cover an imperfect measure of wildland fuel hazard.
One possible clue to the relative importance of adjacent structures burning comes from the different outcomes for wildland urban intermix and interface homes. Houses built amongst wildland vegetation (intermix) survived at a higher rate (29%) than houses built in more of a subdivision arrangement with wildland fuels nearby (interface) (16%). The higher survival may then have been more a function of greater average distance to the nearest destroyed structure (24 m vs. 11 m in the intermix and interface, respectively) and lower average density (7.7 vs. 11.1 structures within 100 m in the intermix and interface, respectively). If proximity to wildland fuels had been the dominant driver, greater percentage losses in the wildland urban intermix would have been expected. Kramer et al. 2019, in an analysis of three-decade's worth of wild res in California, also reported higher survival of homes in the wildland-urban intermix compared to the wildland-urban interface, and together with our results provide some additional evidence of the importance of nearby burning structures to home loss, relative to variables associated with wildland fuels. However, in our study, other factors were likely in play as well, with intermix homes being somewhat newer. In Paradise, an increasing percentage of homes were located in the intermix vs. the interface over time: 66% in time period 1, 80% in time period 2, and 88% in time period 3.

Homes as fuel
Distance to nearest destroyed structure and the total number of destroyed structures within 100 m were consistently the strongest predictors in all of our analyses. This makes intuitive sense because burning structures produce a substantial amount of radiant heat, which can ignite adjacent homes or break glass in windows, allowing embers to enter the home. Nearby burning structures are also a source of embers, which can result indirect or indirect ember ignitions of nearby structures. Our visual analysis of 310 damaged homes corroborated the results of the statistical analyses, with more homes showing evidence of damage from radiant heat exposure (often from adjacent structures burning) than from ame impingement. Our ndings are consistent with other analyses of destructive wild res showing housing density to be strongly associated with home loss (Price and Bradstock 2013;Penman et al. 2019), but in contrast to Syphard et al. (2012Syphard et al. ( , 2014 and Syphard and Keeley (2020), who have reported reduced probability of home loss at higher housing densities. The difference between studies likely has to do with variation in density ranges evaluated, as well as variation in vegetation type and housing arrangement. Syphard et al. (2012) sampled large re-prone regions with shrub-dominated vegetation in southern California, ranging from outlying WUI areas to denser cities that did not burn to answer the question of housing arrangements most prone to loss in a wild re. Since the entire scope of our analysis was within the Camp Fire perimeter, our research question differs: when burned, what factors in uenced survival? In any case, the interpretation of Syphard et al. (2012), that higher density development reduced the likelihood of loss may better apply to denser development patterns than present in Paradise, where housing densities were intermediate to low and interspersed with native (and non-native) vegetation. Such lower density wildland urban intermix and interface development is prevalent in foothills and lower mountainous regions of central and northern California (Hammer et al. 2007).
At what distance an adjacent burning structure presents vulnerability is not well studied. Our analyses identi ed a threshold of 18 m from the nearest destroyed structure that best differentiated surviving and destroyed homes (Fig. 5a). Price and Bradstock (2013) found the presence of houses within 50 m to be predictive of loss. Radiant heat ux, which is inversely related to distance from the aming source, can be a factor up to 40 m from a burning structure (Cohen 2000). Cohen (2004) reported that models predicted ignition of wood walls when less than 28 m from a crown re in forested vegetation, with actual experimental crown res nding ignition at a 10 m distance, but not 20 m or 30 m. The radiant heat ux adjacent to burning structures is different and likely more sustained than a similar heat ux adjacent to crowning wildland vegetation.
Between home spacing has been evaluated in post-re assessments conducted after the Witch Fire in San Diego County, California (Insurance Institute for Business & Home Safety 2008), the Waldo Canyon Fire in Colorado Springs, Colorado (Quarles et al. 2013), and the Black Bear Cub Fire in Sevier County, Tennessee (Quarles and Konz 2016). During each of these res, home-to-home spread was observed with spacing less than 10 m. The IBHS Witch Fire report (Insurance Institute for Business & Home Safety 2008) referred to home-to-home spread as "cluster burning", which was not observed when homes were located more than 14 m apart. Our nding of an 18 m threshold is similar to the IBHS Witch Fire results. Regardless of the actual ideal home separation level, many homes in Mediterranean ecosystems are commonly on lot sizes with less than 18 meters of separation between buildings.

Wildland fuels and defensible space actions
Overstory canopy cover was a signi cant predictor of home survival in the statistical models, with the canopy cover 30-100 m away having a larger effect size than canopy cover in the immediate vicinity of the home (0-30 m) (Fig. 4a,  b). This result (and other evidence, below) suggests that overstory canopy cover may only be correlated to factors that contributed to re spread and increased the threat to homes, rather than a direct contributor. Wildland re spread is dependent on surface fuels -litter, duff, and dead and down woody material, which would be expected to be most abundant and continuous under or adjacent to overstory tree canopy. The link between overstory canopy cover and surface fuel abundance may have been weaker from 0-30 m than distances farther removed from the home because of the greater likelihood that such surface fuels were better managed near homes, perhaps as a result of defensible space activities. In addition, the continuity of vegetative fuels is more likely to be broken up by lawns, driveways, or irrigated landscaping near the home. While vegetation abundance within 30 m has been reported to be associated home loss in southern California res burning in shrubland vegetation types (Syphard et al. 2014), Alexandre et al. (2016) found vegetation near a building not to be a strong factor in models of loss for res in southern California and Colorado. They theorized that the connectivity of vegetation to the home was more critical than vegetative cover.
While burning trees and associated vegetation may generate substantial ame lengths and embers which can then threaten homes, the overstory tree canopies themselves did not appear to drive re intensity in most cases. With the Camp Fire, many overstory trees located away from burning homes survived Cohen and Strohmaier 2020) (Fig. 7). Rather than tree torching directly impacting nearby structures, the torching of trees and other vegetation appeared from photographs and personal observation to frequently be caused by heat from nearby burning structures. Additionally, a substantial proportion of the canopy of native tree vegetation in Paradise at the time of the re was comprised of California black oak (Quercus kelloggii Newb.), a native deciduous species that would have shed at least a portion of its leaves by the time of year when the Camp Fire burned through Paradise. Even when fully leafed out, the crowns of black oak trees are relatively open with low canopy bulk density. Deciduous oak litter breaks down faster than conifer litter, and the light fuel loads in pure black oak stands tend to promote low-intensity surface re rather than crown re (Skinner et al. 2006). Ponderosa pine (Pinus ponderosa Lawson & C. Lawson) was the other major native tree species. Leaf and needle litter can carry ames to the home or provide receptive fuels for ember ignitions, and would likely have been positively correlated to overstory tree canopy cover, especially in the fall. Embers can also ignite litter that has accumulated in gutters and roofs. High overstory canopy cover may also indicate areas where associated vegetation and surface fuels had developed to the greatest extent in the absence of re and active management, especially at a distance from homes. With the lands in the Paradise area having no record of re in modern recorded history (Maranghides et al. 2021), considerable vegetative ingrowth and accumulation of dead and down surface fuels was likely, especially relative to historical amounts. Ingrowth could have included brush and smaller conifers that acted as ladder fuels, leading to torching and ember generation.
Even though our data showed a stronger association between overstory tree cover and home survival for distances beyond which defensible space is typically mandated (100 ft, or 30 m), this does not mean that vegetation modi cation within 30 m is unimportant. For reasons described earlier, the fuel hazards contributing to outcome were likely not well captured by the overstory canopy cover variable, especially in this near-home zone. In addition, once structures become involved, defensible space vegetation modi cation to 30 m (100 ft) may be insu cient to mitigate ember and radiant heat exposures contributing to home loss. In an analysis of CAL FIRE DINS data over multiple res, including the Camp Fire, Syphard and Keeley (2019) reported that defensible space was a poor predictor of outcome, with structural variables (e.g., eave construction details, numbers of windowpanes (double vs. single), vent screen size) more highly correlated with home survival. The low predictive power of defensible space may be partially due to the coarseness with which defensible space is classi ed in the DINS data, with broad distance categories not fully capturing spacing, composition, or ammability of the vegetation. In addition, in many destructive wild res, a large portion of homes are lost through direct or indirect ember ignition and not ame impingement associated with the continuity with wildland fuels (Murphy et al. 2007;Cohen and Stratton 2008). With embers capable of igniting fuels over 1-2 km away, the protective effect of vegetation modi cation within 30 m of the house does not guarantee survival when re-ghting resources are not present. Modi cations in this region, however, do provide access and a safer means of protecting a home when re ghting resources are available.
Our analysis relied upon aerial photo interpretation, and we could not assess surface fuels under dense tree canopies.
As a result, and because of the likely indirect effect of leaf litter coming from the canopy, we caution against the interpretation of the canopy cover variables used in this study, and cover percentages in the decision trees, as guides to forest thinning targets. Furthermore, surface and near-ground live fuels are considered the priority for altering re behavior and in uencing re hazard (Agee and Skinner 2005). Higher canopy cover may be correlated to the rate of surface litter and woody fuel accumulation but does not necessarily directly translate to high re hazard if these surface fuels are managed and maintained at low levels. Patchiness and arrangement relative to prevailing winds can also reduce threat posed by near-home vegetation (Gibbons et al., 2018).
Did the adoption of Chap. 7A into the California Building Code in uence survival?
While the survival rate for homes built in the 11 years after the adoption of Chap. 7A to the California Building Code in 2008 was numerically slightly higher than the survival rate of homes built in the 11 years immediately before, the difference was not statistically signi cant. It is possible that signi cance might have been found with a larger sample size, but even so, any in uence of the building code update was likely swamped by other factors. This was not a surprise because of the many interacting variables that affect building performance, in addition to building products rated to resist exterior re exposures. The 2008 Chap. 7A building code update institutionalized several important and worthwhile changes to construction in high re hazard zones, including the use of ember and ame-resistant vents.
These changes may improve the probability of survival for some types of wild re (e.g., vegetation and wind-driven res); however, the changes were apparently not su cient to fully protect buildings from radiant heat exposures from nearby burning structures. Radiant heat can break window glass and allow embers to enter the building (Penman et al. 2019). Chapter 7A mandated the use of tempered glass in one pane of a double-paned window, but the magnitude of radiant heat exposure was likely still too much in many cases, or other vulnerabilities remain.

Variation in factors contributing to home loss across construction time periods
In models for predicting survival, the signi cant interaction of several of the potential explanatory variables with construction time period suggested that the factors most strongly in uencing home vulnerability differed for homes of different ages. Homes built in the most recent two eleven-year periods (1997-2007 and 2008-2018) survived at a signi cantly higher rate than homes built prior to 1997. Factors potentially contributing to this increase include trends towards a longer average distance to the nearest structure and nearest destroyed structure, and a larger average lot size. Newer homes had lower overstory canopy cover in the immediate vicinity (0-30m), whereas the older homes tended to be concentrated near the center of Paradise, where overstory tree cover was higher. The two most recent construction time periods also saw changes in building construction including roo ng materials having longer periods of robust performance (i.e., 30-50 years of service life), double-pane windows (as a result of changes to the energy code), and increased use of noncombustible ber-cement siding. Many of these improvements, which potentially make newer homes less vulnerable to wild re exposures, occurred well before the 2008 Chap. 7A update to the building code. Older homes may also have developed vulnerabilities resulting from overdue home maintenance. We speculate that with a higher proportion of newer homes surviving the ember onslaught, outcome then depended to a greater extent on degree of radiant heat exposure from nearby burned structures. This hypothesis is supported by the much stronger in uence of distance to nearest burned structure and the number of structures burned within 100 m for newer (1997 and after) than older < 1997) homes. A substantially lower proportion of older homes survived regardless of the distance to or density of nearby burned structures, suggesting other vulnerabilities (such as maintenance issues). Another factor that may have increased the survival probability of newer homes was simply less time for occupants to accumulate combustible items on their properties (e.g., sheds, stored objects, wood piles, play structures). The difference between distance to nearest home and distance to nearest structure was much greater for older than newer homes (data not shown), indicative of structures such as sheds, detached garages, or other outbuildings being added to properties over time. Our summary of damage location and cause for damaged homes as well as rst-hand accounts (Maranghides et al. (2021); N. Wallingford, personal communication) indicated such nonvegetative items were frequently ignited by embers and the reason for a ame impingement exposure.
Di culties in post-wild re interpretation A primary challenge in determining the potential causes for building survival after wild re can be the variation in re behavior experienced and/or re ghter response across the population of homes evaluated. However, in this study the home losses largely occurred during one burn period under relatively consistent burning conditions, and with the focus of rst responders on evacuation, very few homes likely saw any intervention by re ghters. The DINS assessment indicated that only seven of the 400 randomly selected homes (1.7%) experienced some defensive action by re ghters, with six of these homes surviving and one destroyed by the re. Therefore, the Camp Fire provided a more homogenous burn environment than in many other post-re evaluations of home survival, most of which combined data across multiple res and years (e.g., (Syphard et al. 2012;Alexandre et al. 2016;Penman et al. 2019;Syphard and Keeley 2019)). However, while similar factors may be pertinent in other wild res, it is still important to recognize that the variables identi ed here were speci c to the housing, vegetation, and topographic conditions found in Paradise, and may not apply elsewhere.
Determining pre-re structural characteristics post-re is challenging and availability of such data is generally limited . Details about near-home vegetation, especially within the rst 1.5 m of the structure, which has been shown to be an especially vulnerable location for ember ignition, were not available. We were also not able to quantify the presence and distance to small sheds and other storage structures, the age and condition of the roo ng, or individual residents' maintenance practices. The DINS data (e.g., extent of vegetation clearing for defensible space, siding type, type of window glass (single or multi-pane), deck construction, and presence of attached fencing) have value, but missing data and lack of information for structures not damaged or destroyed limit the utility for some analyses. We instead focused on variables that could be consistently evaluated on every home, such as overstory canopy cover and distance to the nearest destroyed structure. Our vegetation variables were, however, coarse, and likely missed factors that contributed to home survival.
Lastly, for the damaged home cause and area of damage summary, it is important to acknowledge that the vulnerabilities may differ for damaged and destroyed homes. With evidence for what contributed to loss no longer available for destroyed homes, damaged homes provide a picture of the different vulnerabilities, but the relative contribution of factors involved may not be the same.

Conclusions
The results of this study support the idea that both proximities to neighboring burning structures and surrounding vegetation in uence home survival with wild re. Denser developments, built to the highest standards, may protect subdivisions against direct ame impingement of a vegetation re, but density becomes a detriment once buildings ignite and burn. Recent examples of losses in areas of higher density housing include the wind-driven 2017 Tubbs Fire in northern California, where house-to-house spread resulted in the loss of over 1400 homes in the Coffey Park neighborhood , and the wind-driven 2020 Almeda Fire in southern Oregon, which destroyed nearly 2800 structures, many in denser areas in the towns of Talent and Phoenix (Cohen and Strohmaier 2020). Once re becomes an urban con agration, proximity to nearby burned structures becomes especially important because occupied structures contain signi cant quantities of fuel, produce substantial heat when burned, and are a source of additional embers. For density to be protective, home, and other structure ignitions would need to be rare. Fifty-six percent of homes in Paradise built during or after 2008 did not survive, illustrating that much improvement is needed in both current building codes and how we live in wild re prone WUI areas before proximity to nearby structures becomes a bene t rather than a vulnerability. The threat posed by nearby burning structures as well as our nding of an apparent strong in uence of vegetation 30-100 m from the home -a distance that in most cases encompasses multiple adjacent properties -demonstrates that neighbors need to work together to improve the overall ability of homes and communities to resist wild re exposures.
To maximize survivability, homes need to be designed and maintained to minimize the chance of a direct ame contact, resist ember ignition, and survive extended radiant heat exposure. Our analyses demonstrating the strong consists of one level, so is binary in nature in that a building either needs to comply, or it does not. Interaction between components, for example, siding, window, and the under-eave area on an exterior wall, is not considered. The Australian building code for construction in bush re prone areas, AS 3959 (Standards Australia 2018) incorporates six different construction classes based on anticipated radiant heat, ame, and ember exposure levels.
Our summary of damaged but not destroyed homes in Paradise was in line with other reports showing a high proportion of home ignitions indirectly resulting from embers (Mell et al. 2010). Embers frequently ignited near home combustibles such as woody mulch, fences, and receptive vegetation with ames and/or associated radiant heat then impacting the home itself, supporting awareness of the importance of combustibles within the rst 1.5 m (5 ft) of the building on home survival. A re-interpretation of defensible space fuel modi cations is needed to increase the building's resistance and exposure to embers and direct ame contact, especially in the area immediately around a building and under any attached deck or steps. This does not diminish the value of defensible space fuel modi cations 9 to 30 m (30 to 100 ft) away from the home, which not only reduces fuel continuity and the probability of direct ame contact to the home, but also provides re ghters a chance to intervene.
While our data show a relationship between home loss and vegetative fuels (high overstory canopy cover likely associated with a greater litter and woody fuel abundance, as well as other wildland understory vegetation) that can contribute to re intensity and ember generation, the WUI re loss issue has been described as home ignition problem more so than a wildland re problem (Cohen 2000;Calkin et al. 2014). The damaged home data were in line with this view, with few homes showing evidence of continuity with wildland fuels that would contribute to ame impingement, but numerous homes with near home fuels, both from manmade and natural sources, that led to direct or indirect ember ignitions.
California's Mediterranean climate will continue to challenge its residents with regular wild re exposure throughout the state. Whether through modifying the nearby surface and vegetative wildland fuels or the home itself, adapting to wild re will take time. The good news is that the trend in survival is improving with newer construction practices. However, with 56% of houses built after 2008 still succumbing to the Camp Fire, much room for improvement remains. Our data suggest it is possible to build (and maintain) buildings that have a high probability of surviving a worst-case scenario type of wild re, even in re-prone landscapes such as the Paradise area. Newer homes built after 1972, where the nearest burning structure was > 18 m away, and fuels associated with vegetation 30-100 m from the home kept at moderate and lower levels (< 53% canopy cover) had a 61% survival rate -an approximately 5-fold improvement over the Paradise housing population as a whole. Survival percentages substantially higher still are potentially possible if all components of risk, including ember generation in nearby wildland fuels, continuity of wildland and other fuels on the property, and home ignitability are su ciently mitigated.  Figure 1 Map showing the perimeter of Paradise, California, with the location of 400 randomly selected homes built during three time periods (pre-1997, 1997-2007, and 2008-2018).

Figure 2
Percentage of surviving single-family homes in Paradise by decade of construction.

Figure 3
Probability of home survival with a) distance (m) to nearest destroyed structure, b) the number of destroyed structures within a 100m radius, c) overstory canopy cover within 0-30m, and d) overstory canopy cover within 30-100m, for homes built during three time periods (before 1997, 1997-2007, and 2008-2018). A vertical dotted line in (a) shows the 18 m threshold between survival and destruction identi ed by the regression tree analysis (Figure 5a).

Figure 4
Effect sizes for two logistic regression models of home survival in the town of Paradise during the 2018 Camp Fire, including continuous variables (a) present during the re, and (b) only variables present pre-re. Regressions were based on a random sample of 400 homes.

Figure 5
Regression trees for predicting home survival in the town of Paradise in the 2018 Camp Fire, with models including continuous variables (a) present during the re, and (b) only variables present pre-re, both based on a random sample of 400 homes. Survival proportion is listed in bold under each branch, along with the percentage of homes in Paradise that each branch applied to (in parenthesis).

Figure 6
Percentage of damaged but not destroyed homes in Paradise by a) re damage cause category and b) re damage location. Fire damage cause was either radiant heat, direct ember ignition, or ame impingement. Flame impingement was further subdivided into ame impingement due to indirect ember ignition, fuel continuity with the broader landscape, or unknown. Numbers were based on visual assessment of photos taken by the CAL FIRE inspectors and information in the CAL FIRE DINS (damage inspection) data. Totals exceed 100% because some homes had multiple sources of re damage.

Figure 7
Aerial image showing a portion of Magalia just NW of Paradise, illustrating a gradient of re damage to overstory vegetation with distance from destroyed homes. At least in some areas, burning homes may have in uenced the effects to overstory vegetation more than burning canopy vegetation in uenced the outcome to homes. Photo: Owen Bettis, Deer Creek Resources.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download. CampFire400HomeSample.csv CampFireParadiseDamagedHomes.xlsx MetadataCampFireHomeSurvivalKnappetal.docx