The impact of land ownership, firefighting, and reserve status on fire probability in California

The extent of wildfires in the western United States is increasing, but how land ownership, firefighting, and reserve status influence fire probability is unclear. California serves as a unique natural experiment to estimate the impact of these factors, as ownership is split equally between federal and non-federal landowners; there is a relatively large proportion of reserved lands where extractive uses are prohibited and fire suppression is limited; and land ownership and firefighting responsibility are purposefully not always aligned. Panel Poisson regression techniques and pre-regression matching were used to model changes in annual fire probability from 1950–2015 on reserve and non-reserve lands on federal and non-federal ownerships across four vegetation types: forests, rangelands, shrublands, and forests without commercial species. Fire probability was found to have increased over time across all 32 categories. A marginal effects analysis showed that federal ownership and firefighting was associated with increased fire probability, and that the difference in fire probability on federal versus non-federal lands is increasing over time. Ownership, firefighting, and reserve status, played roughly equal roles in determining fire probability, and were found to have much greater influence than average maximum temperature (°C) during summer months (June, July, August), average annual precipitation (cm), and average annual topsoil moisture content by volume, demonstrating the critical role these factors play in western fire regimes and the importance of including them in future analysis focused on understanding and predicting wildfire in the Western United States.


Introduction
Globally, wildfires have increased in both number and severity (Pechony and Shindell 2010, Liu et al 2010, Flanigan et al 2013. This increase is even more pronounced in the western United States (Agee and Skinner 2005, Westerling et al 2006, Dennison et al 2014, Abatzoglou and Williams 2016. While fires are a natural part of most western ecosystems, current regimes vary greatly from historic baselines in terms of both frequency and severity (Hessburg et al 2015, Miller et al 2012b, Westerling 2016. Thousands (Collins and Roller 2013, Regan et al 2010, Zedler et al 1983, smoke has impacted public health , and lives have been lost.
Although climate change plays an undeniable role in this increase (Westerling et al 2006, Rhodes and Baker 2008, Preisler et al 2011, Miller and Safford 2012, Berner et al 2017, its exact contribution is unclear. Factors such as firefighting, vegetation management, and land ownership are known to impact fire in complex ways (Westerling and Bryant 2008, Miller et al 2012a, Barbero et al 2014, Parks et al 2016, though there is limited research on the role each plays in influencing fire probability. Teasing these drivers apart has proven to be challenging, yet understanding their impacts is crucial to management and policy. Methodologically, disaggregating the impacts of land ownership and firefighting is difficult. Federal ownerships-especially reserved lands-often have vastly different ecological characteristics than private lands: they are generally less productive, steeper, and more isolated than non-federal ownerships. Since these factors also influence wildfire occurrence, comparisons of federal and non-federal lands are compromised by different baseline fire regimes. Likewise, federal agencies typically only fight fires on federal lands, while state agencies focus on non-federal lands, preventing a clear comparison of the possible influence of federal and state firefighting. California provides a unique opportunity to analyze how land ownership, firefighting, and reserve status affect wildfire trends. First, half of wildlands in California are in non-federal ownership, providing ample areas to compare federal and non-federal lands. Second, there are considerable areas of reserved lands-designated areas where vegetation management activities and fire suppression actions are limited (USDOI and USGS 2016, Christensen et al 2016)-on both federal and non-federal ownerships.
California also serves as an ideal 'natural experiment' due to a policy referred to informally as the 'balance of acres' arrangement (aka the Cooperative Fire Protection Agreement; see Giambattista 2005). Under this agreement, state and federal agencies trade firefighting responsibility to maximize efficiency, particularly in areas where a legacy of 19th century railroad grants has created 'checkerboard' ownership patterns of federal and non-federal ownerships. As a result, some federal lands are protected by state fire protection and vice versa (table 1). This produces four combinations of ownership and firefighting: federal lands with federal firefighting, federal lands with state firefighting, nonfederal lands with state firefighting, and non-federal lands with federal firefighting. In sum, with comparable areas of federal and non-federal ownerships, large areas of reserve as well as non-reserve land across federal and non-federal ownerships, and areas with traded firefighting responsibilities on reserve and non-reserve lands, California is an ideal area to test for their effects on fire probability.
Using a combination of pre-regression matching and panel Poisson models, coupled with spatially explicit data on fire parameters, biophysical drivers of fire, vegetation, and human settlements, we estimated the relative effects of land ownership, firefighting, and reserve status on average fire probability in California from 1950-2015 in reserve and non-reserve areas across four different vegetation types (forest, shrublands, rangelands, and forest without commercial species). Specifically, for each vegetation type, ownership, firefighting responsibility, and reserve status we asked: 1. How are ownership, firefighting, and reserve status associated with annual fire probability from 1950-2015?
2. How is ownership related to fire probability over time?
3. How is firefighting responsibility related to fire probability over time?  non-profit owned lands. In total, there were 10 136 190 ha of federal land in our dataset and 10 264 144 ha of non-federal land (table S3).

Study area
To classify a piece of land as 'reserved' or not, we used the United States Geological Survey's Protected Areas Database under the United States Gap Analysis Program (GAP) dataset to divide land into four categories: (1) lands primarily managed for biodiversity, (2) lands managed mostly for biodiversity, (3) lands that allow multiple uses, and (4) lands with no known mandate for protection (USDOI and USGS 2016). GAP 1 and 2 lands were combined into a 'reserved' category to represent areas where limited management is allowed. GAP 3 and 4 lands were classified as 'non-reserved' to represent areas where more extensive management is permitted, but not necessarily practiced. For forest lands, we applied the USFS Forest Inventory and Analysis program definition of 'reserved forest land' as 'land permanently reserved from wood products utilization through statute or administrative designation' (Christensen et al 2016).
Actual firefighting responsibility was determined using the Direct Protection Area designation (figure 2, table S4), which categorizes firefighting as state or federal. Our area of analysis included 10 251 477 ha of lands with federal firefighting and 10 148 875 ha with state firefighting (table S3). Direct Protection Area, State Responsibility Area, vegetation type, and reserve status were assumed to be unchanged throughout the study period. While there have likely been small changes in these variables over time, we were unable to find data documenting these changes. Therefore, we assume that any differences were not of a sufficient magnitude to impact our results.
Topographic, demographic, and climatic variables were used in our model to control for the influence that local geographic variation can exert on ecosystem susceptibility to wildfire (McKenzie et al 2004, Narayanaraj and Wimberly 2012, Westerling et al 2006. Variables included in the model were elevation (m), slope (degree), aspect, distance (m) to the closest major roadway, distance (m) to and population of the nearest city, average maximum temperature ( • C) during summer months (June, July, August), average annual precipitation (cm), and average annual topsoil moisture content by volume (mm) (table S4). Climatic variables were averaged over the following time periods : 1950-1966, 1967-1983, 1984-2000, and 2001-2015. These time periods also served as panels for the panel regression analysis. The use of four periods captured broad changes over time while avoiding the highly variable year-to-year rates of fire.
Fire history data were acquired from FRAP maps of historic fire perimeters in California (table S4). This dataset dates back to the 1880s, but due to concern about data accuracy in earlier periods, we started our analysis from 1950. Each fire recorded in the dataset included the year the fire burned and its perimeter. If an area burned more than once since 1950, it had multiple fire perimeters associated with it. In total the dataset contained over 13000 unique fires. The average fire size in the FRAP dataset was approximately 692 ha (1711 acres) (table S10 available at stacks.iop.org/ERL/13/034025/mmedia). Approximately 99% of the total area burned came from fires greater than 121.41 ha (300 acres) in size. Fires greater than 121.41 ha (300 acres) had an average size of 1265 ha (3310 acres) (table S10).
Estimating the influence of ownership, firefighting, and reserve status on fire probability For our analysis, we randomly drew 100 000 points each (using GIS software ArcMap) from forested and non-forested land types. The number of points was selected such that after the matching procedure there were sufficient observations in each strata of ownership, reserve status, and firefighting agency to provide robust estimation. We found that 100 000 pre-matched points proved sufficient for this purpose (table S10). For each point, we then assigned the spatial attributes described previously as well as how many times it had burned in each time period. This value-the number of times a point burned in each time period-became our dependent variable for the regression statistics. We chose random selection of the initial points in order to assure that our sample was a valid representation as a whole.
The panel structure of our data allowed us to address two issues which can conflate the relationship between cause and effect (Angrist and Pischke 2008). The first issue occurs if our data is subject to observable bias between data points in the different ownership, firefighting, and reserve status categories (Caliendo and Kopeinig 2008). We suspected that there may be large observable ecological differences between federal and non-federal land, as federal lands are often of lower site quality, at higher elevations, and further from developed areas than non-federal lands (Joppa and Pfaff 2009), potentially obscuring comparisons of fire probabilities on federal and non-federal lands (Andam et al 2008).
To control for differences in geographic characteristics by ownership, we used a matching algorithm (nearest neighbor matching without replacement) to find points under federal ownership that were geographically similar to points under non-federal ownership (Ho et al 2007). We decided to match on ownership, since firefighting and reserve status vary across ownership. We matched points individually for each vegetation type and matched observations on a number of characteristics including: elevation (m), slope (degree), aspect, distance (m) to the closest major roadway, distance (m) to and population of the nearest city, average maximum temperature ( • C) during summer months (June, July, August), average annual precipitation (cm), and average annual topsoil moisture content by volume (mm), and GAP status. The matching algorithm essentially created four new datasets-each comprised of an equal number of federal and non-federal data points-that were similar in the matched covariates. This matching procedure greatly reduced bias in our dataset by allowing us to control for differences in the characteristics of federal and non-federal land. Specifically, we show reduction in the mean differences between matched and unmatched observations (table S5, table S6, table S7, table S8). The matching process reduced the overall number of points used for analysis (table S10), but ensured that federal and non-federal lands selected for analysis were similar in their characteristics, allowing us to better isolate the effects of our variables of interest on fire probabilities.
Second, it is likely that unobservable factors in our dataset played a role in determining the number of times a point burned. For instance, although we observed many factors about a particular point in time, there may have been unobserved fine scale or historical factors that we were not able to control for in our sample selection. Therefore, we used a random effects panel model, which can help control for time invariant factors that are not directly observable (Wooldridge 2012).
All told, using the datasets comprised of the observations that are matched between federal and non-federal lands, we estimated the following panel regression: where NB was the number of times a pixel burned per time period, Ownership represented federal or nonfederal ownership, Firefighting designated either federal or state fire protection, Reserve Status indicated reserve or non-reserve lands, and the Control variable, which included the list of covariates we expected to influence the probability: elevation (m), slope (degrees), aspect, distance (m) to the closest major roadway, distance (m) to and population of the nearest city, average maximum temperature ( • C) during summer months (June, July, August), average annual precipitation (cm), and average annual topsoil moisture content by volume (mm). We included all possible interactions between Firefighting, Time Ownership, and Reserve Status, in order to allow the effect of firefighting and ownership to vary across time and reserve status. The one exception to this is the model for non-commercial species where a lack of reserved lands makes this full saturation impossible. In this case we interact Firefighting, Time and Ownership. u i is the observation-specific (i.e. point) random effect, and e it is the error component for each point in each time period. Our dependent variable was the number of times a given point burned in each panel, described as a count (i.e. 0, 1, 2,…). Therefore, we used a Poisson panel model specification to estimate the model. We choose the Poisson model over alternatives such as the negative binomial model because robust standard errors can be more easily calculated in Poisson models, and can typically correct for bias in standard errors even when the model is overdispersed (Wooldridge 2012).

Validation, accuracy assessment, robustness check
We performed a number of robustness checks on our results. First, we were concerned that the results may be sensitive to our choice of four time periods. We ran two alternative models: one with a breakpoint in 1983 and another with a breakpoint in 2000. Results did not change quantitatively. In addition, we suspected that areas that had burned during the previous time period would be less likely to burn in the next period. We therefore included a lag term indicating if a pixel had burned in the past time period. The inclusion of this lagged variable did not change our results and the coefficient was not statistically different from zero, so we report regression results without it.

Effects of land ownership, firefighting, and reserve status on annual fire probability
Using pre-regression matching to control for differences in geographic characteristics by ownership and  1967-1983 1984-1999 2000-2015  the regression analysis to control for local variables that influence fire probability, we found that average annual fire probability was nearly always higher for points with federal ownership, federal fire protection, and reserve status (table 3). This trend held across all vegetation types (figure 1, table S1). The increase in fire probability associated with the marginal effect of federal ownership and federal firefighting were very similar in magnitude for each vegetation type except rangelands, where the effect of ownership was nearly double the effect of firefighting (table 3). Reserve status had the largest effect on forest lands and the least effect on shrublands (table 3). Reserve status had a larger influence on fire probability than ownership or fire protection for forest lands and a lesser impact on shrublands. Non-federal, non-reserved forests with state fire protection had the lowest fire probability on average, with the exception of shrublands, where non-reserve lands with non-federal ownership and federal firefighting consistently had the lowest fire probability on average (figure 1, table 2, table S1).

Effects of federal firefighting on fire probability over time
The increase in fire probability associated with federal firefighting became stronger over time for forests, rangelands, and shrublands (table 4). For forests, rangelands, and shrublands, the magnitude of the effect increased from nearly zero in the second time period to 0.32, 0.27, and 0.39 percentage points, respectively, by the fourth. In forests without commercial species, the effect of federal firefighting was not as straight forward. While it increased from 0.01 in the second time period to 0.16 percentage points in the fourth, it peaked at 0.31 in the third period (table 4).

Effect of climate variables on fire probability
In order to determine potential climate effects on annual fire probability over the study period , we included average maximum temperature ( • C) during summer months (June, July, August), average annual precipitation (cm), and average annual topsoil moisture content by volume (mm) for each time period in our marginal effects analysis (table 3). These variables were found to have very little effect for each vegetation type compared to federal ownership, firefighting, and reserve status, with the exception of shrublands, where the effect of reserved status was very low (0.07) and therefore similar to the effect of temperature (0.05) and soil moisture (−0.06).

Discussion
There is considerable discussion about what is driving increases in wildfires in the Western US, with scientists pointing to a host of potential drivers from increased human ignitions to fuel buildup to climate change (Dennison et al 2014, Schoennagel et al 2017, Westerling 2016, Littell et al 2016. Our results clearly demonstrate that non-federal lands have seen far less of an increase in fire probability than have federal lands. Our results also showed a consistent pattern of lands with federal firefighting-regardless of ownership or reserve status-nearly always having a higher fire probability than lands with state firefighting (figure 1,  table 2, table S1). The effect of federal ownership may be linked to changes vegetation management associated with ownership. From the 1950s to the 1970s, vegetation management practices on federal and non-federal lands were broadly similar (Stewart et al 2016). In nonreserve roaded areas, vegetation management activities such timber harvesting, non-commercial thinning of trees and shrubs, grazing by domesticated animals, and prescribed burning were commonly practiced if they produced net revenue for private owners or were supported by federal agencies (Dana andFairfax 1980, Williams 1989). Beginning in the 1970s, parallel state and federal laws addressing environmental quality, clean water, clean air, and endangered species led to a greater analysis of the potential environmental and ecological impacts of typical practices and to increased emphasis on avoiding them (Burnett andRoberts 2015, Koontz 2007). This reduced the removal of timber and forage (both of which are also flammable fuels) on both federal and non-federal lands, with much larger declines in the removal of timber and forage on federal lands, leading to greater fuel accumulation on federal lands. (Stewart et al 2016, Spiegal et al 2016 Substantial management changes on federal lands in California occurred again in the 1990s under the Northwest Forest Plan (USDA and USDOI 1994) and the California Spotted Owl guidelines (Verner et al 1992). These policies, designed to protect desired habitats for certain species and to reduce potential risks to other resources, led to substantial decreases in timber harvesting on federal lands (Stewart et al 2016, McIver et al 2015. Grazing on federal lands also declined approximately 30% over this time period (Wiles andWarren 2016, USFS 2016). This led to a divergence over time in the probability of active vegetation management between federal and non-federal ownerships.
Firefighting practices in California also diverged over time. California's state firefighting agency, CAL FIRE, has consistently stated that their 'fire mission is to protect natural resources on state responsibility areas [SRA] from damage by wildfire' (CAL FIRE 1990). Beginning in the 1960s, the National Park Service began allowing lightning-ignited fires to burn (van Wagtendonk 2007, Stephens andRuth 2005). The federal shift away from a more aggressive suppression strategy was formalized in a major review of federal wildland fire management in 1995 that began with the 'protection of human life is reaffirmed as the first priority,' and highlighted that 'wildland fire, as a critical natural process, must be reintroduced into the ecosystem' (USDOI and USDA 1995). While state and federal fire agencies share common training, terminology, and incident command structures (National Wildfire Coordinating Group n.d.), the operational differences can be characterized by a relatively greater focus on aggressive direct attack by CAL FIRE (Haight and Fried 2007) compared with more indirect attack by federal agencies (Katuwal et al 2016, Stonesifer et al 2016. A low fire probability may be more desirable where primary management objectives pertain to economic or carbon benefits (Berner et al 2017, Hicke et al 2013, but for lands where management goals are primarily ecologically driven (Stephens et al 2013, Hessburg et al 2015, Keeley and Safford 2016, increased fire frequency is often a goal for establishing more resilient ecosystems. That is, returning to pre-European settlement fire regimes by allowing fires to burn would necessitate an increase in area burned and an increase in fire probability. Stephens et al (2007) estimated that pre-European-settlement, 1 800 000 ha burned per year in California; our data showed approximately 240 000 ha burned per year on average in the most recent time period (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015). The considerably higher fire probability on reserve lands across all vegetation types (figure 1, table S1) may suggest that federal and state land management agencies are successfully pursuing strategies to reintroduce fire into those ecosystems, or that a lack of vegetation management has increased the likelihood of fire. More information is needed to fully assess this.
As with any study, there are certain levels of complexity we were unable to control for. In our case, we have no information on the landscape structure and composition of historical vegetation types. While forest structure is generally similar between landownership types in California (Stewart et al 2016), we are unable to control for landscape scale differences which may impact fire frequency. Likewise this study did not examine the effects of fire severity. Without such an analysis, it is impossible to determine the full ecological implications of the increase in fire probability over time. At the time of this study, large-scale fire severity data was only available through the Monitoring Trends in Burn Severity (MTBS) program and did not include data for fires that burned solely on non-federal lands. Regional studies using MTBS fire severity data on federal lands have often found fire severity to be increasing Safford 2012, Lutz et al 2009), but tree-mortality based classes and MTBS based classes had poor agreement with severity metrics in a study of National Forests in Oregon and Washington (Whittier and Gray 2016). Finally, we did not examine if ignition sources play a role in the differences described here. This is an area of potential future research, as examination of ignition sources in other Mediterranean regions has shown that ignition source can influence the timing, spatial distribution, and size of fires (Curt et al 2016).
A surprise in our result was the relatively minor effect of climate variables (average maximum temperature ( • C) during summer months (June, July, August), average annual precipitation (cm), and average annual topsoil moisture content by volume (mm)) compared to ownership, firefighting, and reserve status (table 4). Previous studies have correlated increases in the length of fire season, summer temperature, fuel aridity, and droughts have played a significant role in increased fire probability in the western United States (Westerling et al 2006, Spracklen et al 2009, Barbero et al 2015, Abatzoglou and Williams 2016, USFS 2015, Westerling 2016. However, if we compare the effect of one degree Celsius increase in temperature (0.03) to the effect of federal ownership for forest lands (0.14), the corresponding increase in fire probability associated with switching all lands to federal ownership is equivalent to a 4.5 • C increase in temperature (calculated by dividing the effect of federal ownership by the effect of a 1 • C increase in temperature). Considering the wealth of scientific study showing substantial pre-climate change anthropogenic influence on fire probability (Skinner and Chang 1996, Graham et al 1999, Agee and Skinner 2005, Westerling and Bryant 2008, Miller et al 2012a, Collins 2014, Stephens et al 2007, Parsons and DeBenedetti 1979, Pyne 1982, Steel et al 2015, it is logical that anthropogenic influence in the form of choices with respect to vegetation management and firefighting continue to play a major role in fire probability even as our climate changes, though the relative role of climate will likely increase.
Our modeling techniques introduce novel ways of controlling for generic differences between federal and non-federal lands by using matching techniques that have been applied for impact evaluation in many other fields (Brandt et al 2015, Butsic et al 2011. Likewise, the use of a panel data structure allows for control of time invariant unobserved variables that may likewise obscure results. As with all models, there are some caveats. The structure of our data, where many points do not burn, suggest that a model that accounts for a preponderance of zeros such as a hurdle model or a zero inflated Poisson model may be a good fit. However, it is difficult to incorporate such models into a panel setting, and abandoning the panel structure of our data would force us to rely on the assumption that the error term associated with a given point is not correlated over time, a very unlikely case. Therefore, when faced with an imperfect model choice, we believe a panel Poisson model, coupled with pre-regression matching offered us the best choice.

Conclusion
Our data showed continually increasing wildfire probability across all ownerships, firefighting agencies, reserve statuses, and vegetation types. While our analysis did not show a large effect from three climate variables (average maximum temperature ( • C) during summer months (June, July, August), average annual precipitation (cm), and average annual topsoil moisture content by volume (mm)), its relative role may increase over time as our climate changes. However, our analysis suggests that ownership, firefighting decisions, and reserve status play a large role in wildfire probability in the study region.