Insights into Forest Composition Effects on Wildland–Urban Interface Wildfire Suppression Expenditures in British Columbia

Abstract

Burned area, fire severity, and suppression expenditures have increased in British Columbia in recent decades with climate change. Approximately 80% of suppression expenditures are attributable to wildfires near the Wildland–Urban Interface (WUI). Evaluating the potential for fuel management to reduce suppression expenditures is essential to mitigating demands on fire response resources and reducing impacts on communities. One management approach is to increase the proportion of deciduous tree species, which have a lower propensity for crown fire. Using fire suppression expenditure data from 1981 to 2014, we applied the machine learning method causal forests (CFs) to estimate the effect of the proportion of conifer forest cover on suppression expenditures for WUI fires and how these effects varied with other influential factors (i.e., heterogenous treatment effects). Across all fires, the effect of conifer cover on suppression expenditures was stronger on private land compared to public land, under high fire danger measured by daily severity ratings (DSRs), which reflect wind speed and fuel moisture, and for fires igniting earlier in the calendar year, based on Julian day. These findings provide insights into prioritizing wildland fuel treatment when budgets are limited. The CFs approach demonstrates potential for broader applications in fire risk mitigation and analysis beyond the scope of the current data. CFs may also be valuable in other areas of forest research where heterogenous treatment effects are common.

1. Introduction

Wildfire size, severity, and burned area have increased in the Canadian province of British Columbia (BC) in recent decades [1,2], consistent with patterns across western North America [3]. These trends are largely driven by anthropogenic climate change, warmer temperatures, drier fuels, and longer fire seasons and are expected to intensify through the mid-to-late 21st century [4,5,6,7,8,9,10,11]. Correspondingly, fire suppression expenditures in BC have risen on both a total annual expenditure and a per-hectare basis [12].

Rising expenditures are a proxy for increasing demands on fire management resources, firefighters, and managers. Fire management in BC increasingly prioritizes protecting values at risk in the Wildland–Urban Interface (WUI), where wildfires may trigger evacuations and damage or destroy homes and community assets. The potential for such losses motivates aggressive, and therefore costly, suppression efforts near the WUI [13,14,15]. As a result, the suppression of fires near the WUI is a major contributor to provincial firefighting expenditures. Fuels are the one driver of fire behavior that can be manipulated. Thus, understanding how fuels contribute to suppression expenditures is central to evaluating the cost-effectiveness of risk mitigation strategies.

Forest fuel treatments such as reducing ladder fuels and crown density can lessen expected suppression expenditure by lowering crown fire risk [16], limiting fire size, frequency, and severity [17]. A simulation study predicted that combining mechanical thinning and prescribed fire reduced severe fire proportions and suppression expenditures in Arizona, with costs declining by 6.43% for every 1% decrease in severe fire and by 4.91% for each 1% decrease in mixed-severity fire [14]. A case study of the Carleton fire complex in Washington State also found that thinning and prescribed underburns reduced fire spread and severity [18]. Examination of a thinning and prescribed fire experiment burned in a California wildfire suggested a reduction in fire severity that persisted up to 20 years after treatment [19]. However, treatments are expected to be less effective during extreme fire weather [18,20].

Fuel type alone has consistently been identified as key driver of suppression expenditures [21]. Regression analysis shows that fires in standing timber in US National Forests cost 61%–62% more to suppress than those in heavy brush, while grass fires cost 44%–45% less [22]. In Alberta, deciduous and mixedwood forests have lower burn rates compared to pure conifer stands [23]; analysis of experimental burn data confirms difficulty in torching and sustaining crown fires in aspen compared to conifer stands, including pine and spruce species and Douglas-fir [24]. The overstory of aspen and other boreal broadleaf species (birch, poplars) does not readily support crown fire, in part because of the very high (>140%) moisture content of the foliage, and insufficient surface fire intensity to initiate crowning after leaf flush in the understory in summer [25,26]. Most of the fire spread in aspen are surface fires in the spring before leaf-out and before understory vegetation has flushed [27].

This study builds on earlier analysis [12] to focus on the effects of forest fuels on suppression expenditures in BC. Because costs of managing forest fuel are highly variable and context-dependent [28], analytical approaches capable of determining how the treatment effects of forest fuels on suppression expenditures vary with other factors such as fire danger or land tenure (i.e., heterogeneous effects, as opposed to homogenous treatment effects where a treatment has the same outcome everywhere) are needed. Treatment effect here refers to the associated impact of a covariate on the suppression expenditure. Technically, in a model allowing for heterogeneous effects, the treatment effect itself becomes a function of other variables or groups, rather than a single fixed number. Although traditional regression models can include interaction terms to account for heterogeneity, they may not effectively capture intricate or nonlinear interactions. Meanwhile, ad hoc analyses of treatment effects within subgroups of covariates risk false discovery by manipulating the analysis until statistically significant results are found [29]. Recognizing this heterogeneity allows policymakers to understand not just the average effect of a factor, but how that effect changes in different contexts or locations. This could lead to more accurate risk assessment, improved resource allocation and policy guidance for land use and development.

Causal forests (CFs) [29,30] offer a nonparametric machine learning framework for systematically estimating heterogeneous treatment effects. While machine learning has been applied in wildfire science [31], its use in suppression expenditure modelling remains limited. Notable exceptions include Random Forests (RFs) and gradient boosting applied to drivers of expenditures [12] and Double Machine Learning applied to the effects of reporting delay [32]. A recent review found that CFs have mainly been used in health and economics research, with relatively few applications in environmental studies [33,34,35] and only one in forest policy research [36]. We are not aware of any CFs applications in wildfire research. In the following paragraphs we briefly describe the causal forest methodology and contrast it with the more familiar Random Forests.

RFs algorithm is a widely used machine learning method for prediction and classification. Each decision tree in a RFs is trained on a bootstrap sample of the training data, comprising covariates or features and dependent variables or outcomes. At each node a random subset of covariates is considered as candidate split variables, and the data are split by covariate to minimize prediction error. Predictions from many trees are aggregated by averaging in regression or majority vote in classification. RFs also provide measures of variable importance, indicating the relative contribution of each covariate to predictive accuracy.

CFs [29,30] extend RFs by enabling statistical inference, where one covariate is explicitly designated as the “treatment variable,” and the objective function for each decision tree is not focused solely on predictive accuracy, but instead optimizes a combination of (a) the treatment variable’s explanatory power within each branch and (b) the ability of the partition to estimate the corresponding treatment effect. CFs are a flexible and data-driven method for estimating how treatment effects vary across different subgroups of covariates [37,38]. Unlike in a regression model, where varying an input variable is expected to change the outcome at the rate determined by the corresponding regression coefficient, a causal forest can predict this rate of change for a specific set of input covariates. CFs are especially useful in observational studies because they adjust for observed confounding and flexibly model complex relationships. Through residualization (or orthogonalization), they separate out the parts of the treatment and outcome that are explained by other covariates, helping it uncover treatment effects more accurately, even in high-dimensional data. Additionally, CFs use sample splitting where one part of the data is used to build the trees, and another part is used to estimate treatment effects. This separation prevents the model from “cheating” by using the same data to both decide where to split and estimate the effect, which helps reduce overfitting and improves the reliability of the results. Additional details are provided in the methodology section.

The novel contribution of this analysis lies in applying the CFs methodology to characterize the relationship between suppression expenditures and wildland fuels and determining how these effects vary with other influential covariates; it demonstrates an application of CFs in forest research. CFs enable the identification of subgroups with covariates with particularly strong or weak effects, providing nuanced insights that are valuable for targeting interventions. Specifically, we evaluate the impact of the percentage of conifer cover as a proxy for wildland fuel influence on suppression expenditures, offering new evidence for managers seeking to develop fire risk mitigation strategies.

2. Data and Materials

2.1. Study Area

British Columbia, the westernmost province of Canada, has a land area of 95 million hectares. Coniferous stands dominate 60 million hectares of forest cover in the province, with lodgepole pine and spruce being the most common [39]. Roughly two million hectares of land has been converted for human use and development, supporting a population of about five million people. The population distribution is spatially heterogeneous, with human settlements concentrated in the south-west, coast, and in lower elevation valleys in the south and central interior [40].

The British Columbia Wildfire Service (BCWS) is responsible for wildfire management over most of the province, and its primary focus is to ensure human safety and protect valuable assets when determining the appropriate response level to wildland fires. Modified response may be considered for fires deemed to be of lower risk—this consists mainly of observing the fire, with only limited action to protect isolated values [34]. Following wildfire discovery, small three or four person crews, helicopters, or air tankers may be used in initial attack. While 94 percent of all new wildfires are controlled by initial attack, larger crews, incident management teams, and other resources such as heavy machinery, water tenders, and other equipment may be used in an extended or sustained attack until the fire is contained. Additional firefighting resources including structure protection units, contract firefighting personnel, and mutual aid personnel from other jurisdictions and from the armed forces may be utilized in particularly severe wildfire seasons or situations [41].

The use of such resources during sustained attack attracts additional expenditures in the form of firefighter overtime, food and shelter, aircraft hourly rates, fuel and fire retardant, and equipment contract or lease costs. In general, an aggressive suppression strategy that uses more resources on a fire will incur more expenditures and as such may be reserved for fires with more values at risk. Other factors that may increase fire-specific expenditures include fire intensity and spread, as well as the potential difficulties encountered by crews due to complex terrain or remote location.

2.2. Fire Suppression Expenditures

British Columbia has experienced several severe wildfire seasons in recent years, with the 2023 fire season setting a record for the largest area burned. Wildfire suppression expenditure alone totaled $1094.8 million CAD—exceeding 1% of the province’s annual budget [42,43]. The 2.84 million hectares burned represent only a fraction of the broader impact on human settlements and natural resources. The damage to communities, economies, infrastructure, ecological values, and carbon emissions cost is significant but remains poorly quantified.

The specific factors influencing individual fire suppression expenditures are studied using data from the BC Wildfire Service incident database, for the years 1981–2014. Only sustained attack fires of at least 4 ha in size are considered, as in [12,44], since such large fires require deployment of more firefighting resources and thus constitute a different suppression expenditure regime than the initial attack fires. Reported suppression expenditures are deflated to 2002 Canadian dollars using the Consumer Price Index for British Columbia [45]. Decile values for total suppression expenditures and suppression expenditure per area variables are presented in

Table 1. Decile values of the expenditure and expenditure per hectare for wildfire suppression, both deflated to 2002 dollars.

Decile	Expenditure	Expenditure Per Hectare
min	7	0
10	1965	106
20	6259	319
30	15,159	685
40	29,403	1228
50	50,204	2092
60	80,424	3334
70	122,448	5256
80	217,098	8231
90	493,243	14,276
max	25,193,978	224,529

.

Larger-than-average wildfires burning over 361 hectares of land but costing less than $16 per hectare to suppress are assumed to be a modified response fire and are thus removed from subsequent analysis. The chosen expenditure threshold is a natural boundary in the population for larger-than-average fires (Figure A1) and was found to be able to remove the large, low expenditure per area fires primarily in the northern and high-elevation areas that were likely to be modified response fires.

In order to study how expenditures vary for fires near human settlements, the analysis was restricted to fires that occurred within 30 km of a WUI [46]; these fires accounted for 80 percent of the total suppression expenditures (Figure 1). The majority of such fires were suppressed for less than $60,000 per fire; however, the top five percent of fires in terms of expenditures cost over a million dollars each to suppress and in aggregate accounted for nearly 60 percent of the total provincial suppression expenditures incurred between 1981 and 2014. Such variation in suppression expenditures motivates a fire-specific approach to understanding how fire covariates may influence firefighting costs at the fire level, as is done in this study. A total of 4224 extended attack wildfires is included in this study, after removing 58 modified response wildfires and 1213 fires that occurred at a distance greater than 30 km from the nearest Wildland–Urban Interface.

2.3. Variable Selection and Data Compilation

The variables included in the present analysis are similar to those included in [12] to allow for comparison (

Table 2. Variables used in the suppression expenditure causal forest model.

Variable	Variable Definition	Source
Fire Characteristics
ln (Fire size)	Natural log of final fire size in hectares	BC Wildfire Service
ln (Fire perimeter)	Natural log of final fire perimeter in kilometers.	BC Wildfire Service
Duration of fire	The duration from ignition to control	BC Wildfire Service
Julian day of the year	Julian day of fire ignition	BC Wildfire Service
Fire year	Year of fire	BC Wildfire Service
Cause of fire	Dummy variables for human caused and lightning caused fires	BC Wildfire Service
Fire Environment
ln (DSR)—Daily severity rating	An index of fire danger based on wind speed and fuel moisture	NRCan
ln (Slope)	Natural log of the percentage slope of the area	NRCan
In (Elevation)	Natural log of the elevation of the fire ignition point in meters	NRCan
Aspect	The sin and cosine of the aspect at the fire ignition point in radians	NRCan
ln (Topographic roughness)	An index of topographical roughness Index (TRI)	NRCan
Fuel type	Dummy variables or Forested Area and Grassland	NFI
Percent coniferous	The percent of trees that are coniferous near the ignition point.	NFI
Eco-province dummy variables	10 dummy variables for the eco-provincial regions of BC	NRCan
Values at Risk
ln (Population within 30 km)	Population of within 30 km of fire	Stats Canada
ln (Distance WUI density 3+)	Distance to the nearest level three or higher density WUI	BC Wildfire Service
Land tenure dummies	Dummy variables for private land, crownland, parks, and other land.	BC Wildfire Service
Fire Response
ln (Detection time delay)	Natural log of hours between fire ignition and discovery	BC Wildfire Service
ln (Discovery size)	Natural log of final fire size when discovered	BC Wildfire Service
ln (Fire load anomaly)	Difference between the number of fires burning and the average amount.	BC Wildfire Service

). Quadratic variables explicitly supplied in [12] are not included in the present analysis since quadratic terms, if present, are expected to appear in the conditional average partial derivative obtained from the causal forests. Large fire interaction terms are not used in the present analysis as different models are trained for different fire size regimes, making the use of an explicit large fire interaction term redundant.

Fire characteristic variables obtained from the BC Wildfire Service consist of the natural logarithm of the final fire size (in hectares), the natural logarithm of the final fire perimeter (in kilometres), the duration of the fire (in days), the Julian day of the year of ignition, the fire year, and whether the fire was human or lightning caused. Larger fires and fires that require prolonged suppression efforts are in particular expected to attract higher expenditures.

The natural logarithm of the daily severity rating (DSR) is obtained from a fire duration model dataset [47]. The DSR, being a transformation of the Canadian Fire Weather Index (FWI) with $DSR=0.0278\times FW{I}^{1.78}$, is an indicator of expected fire intensity given the environmental conditions present, i.e., temperature, relative humidity, precipitation, and wind [48]. The DSR is averaged over the duration of the fire to represent the fire weather present while the fire was being fought and is expected to be positively correlated to suppression expenditures because a fire-favorable environment with dry fuel or strong winds will likely encourage fire spread.

The natural logarithm of the elevation, of the slope, and of the topographic roughness index (TRI), as well as the sine and cosine aspects of the ignition point, are obtained from the Canadian 50 m Digital Elevation Model [49] and provide insight into how topography near the ignition point might influence suppression expenditures. Fuels near the ignition point are determined using the National Forest Inventory (NFI) 250 m maps [50], where the open fuel types are grassland. Percent conifer cover (in percentages) adds another variable to further differentiate the possible fuel types present near the point of ignition. Fuel conditions were derived from a single available year, which may not accurately reflect the actual forest conditions during specific wildfire events or capture variability across years. An eco-province dummy variable provides additional location information to quantify how local conditions influence suppression costs.

Human settlement values-at-risk are quantified using the natural logarithm of the nearby census population [51,52], with any community within 30 km of the fire being considered affected. The natural logarithm of the distance to the nearest class 3 density WUI is obtained from [46] and is utilized as another predictor. Land tenure, another indicator of human settlement values that are threatened by ignition, is supplied as a dummy variable.

The natural logarithm of the time between ignition and detection (in hours) and the natural logarithm of the fire size at detection (in hectares) are indicative of the fire conditions present at the beginning of suppression, whereas the natural logarithm of the difference between the average and actual number of concurrent fires is indicative of resource availability for suppression efforts. All such fire response variables are obtained from the BC Wildfire Service.

3. Methods

CFs are designed to estimate treatment effects on an outcome, conditional on observed covariates. Like RFs, each causal tree is trained on a random sample of the data, comprising a specified treatment variable and covariates. CFs use an “honesty” principle, in which the data are divided into two parts. One subsample is used to determine the structure of the tree. At each node, a random subset of covariates is considered as candidates for splitting. The algorithm evaluates possible ways of dividing the data for each covariate in the subset and selects the covariate and split point that yield subgroups with the greatest difference in estimated treatment effects. A separate subsample is used to estimate treatment effects within the resulting leaves. A leaf is the terminal node of a decision tree: a subgroup of observations that have similar covariate values after recursive splitting. When the treatment is binary, the treatment effect in a leaf is estimated by comparing average outcomes of treated and control observations. When the treatment is continuous, the treatment effect corresponds to the local slope of the outcome with respect to the treatment variable within the leaf, conditional on covariate and after adjusting for confounding. To adjust for confounding, nuisance models, such as outcome regressions and (generalized propensity) scores, are used to residualize outcomes and treatments before tree building. For example, terrain, elevation, climate, fuel load, and land ownership are all potential confounders affecting both forest composition and how costly fires are to suppress, creating bias if not properly adjusted for. Unlike traditional regression, which relies on predefined functional forms and may miss complex interactions, causal forests use a flexible residualization approach to remove the spurious associations between treatment and outcome that arise from these observed covariates even when relationships are non-linear or vary across subgroups. Finally, treatment effect estimates from many causal trees are aggregated to produce estimates of the Conditional Average Treatment Effect (CATE), while also revealing heterogeneity in causal effects across covariates.

In this study, we estimated CFs using the percentage of conifers as the continuous treatment variable, conditional on the rich set of covariates in Table 2. We estimated the causal forests in R using the grf package which supports both binary and continuous treatments. For binary treatment variables, grf provides estimates of the Conditional Average Treatment Effect (CATE), which is the treatment effect specific to individuals based on their characteristics, and the Average Treatment Effect (ATE), which is the overall treatment effect across the training population. CATE is given by E[Y(1) − Y(0)|X = x] and ATE is given by E[Y(1) − Y(0)]; Y(1) and Y(0) are the two possible potential outcomes, X is a covariate vector, and x is a specific vector of characteristics. For continuous treatment variables, the package supplies estimates of the Conditional Average Partial Effect (CAPE) and the Average Partial Effect (APE). CAPE is given by Cov[Y, W|X = x]/Var[W|X = x], and APE is given by E[Cov[W, Y|X]/Var[W|X], where W denotes the continuous treatment variable—the percentage of conifers—while Y represents individual wildfire suppression expenditures, and X comprises the covariates listed in Section 2.3 and Table 2. The estimates reported in this study should thus be interpreted as partial effects for continuous variables. However, for simplicity, we will refer to partial effects as treatment effects for the study. The treatment effect refers to the estimated impact of a specific variable or intervention on an outcome of interest, while controlling for other influencing factors. In our case, it measures how changes in the proportion of conifers affect suppression expenditures, independent of other variables such as terrain, fire weather, and proximity to the WUI.

To control for unobserved heterogeneity in wildfire conditions, we estimated three separate causal forests for different wildfire size regimes [32] and an additional causal forest for all wildfires. Prior to determining the key covariates that significantly influence treatment effect heterogeneity, we evaluated the overlap assumption and goodness-of-fit to examine how well the causal forest is detecting heterogeneity. The overlap assumption means that for every combination of observed covariates, there is a nonzero probability of receiving either the treatment or the control. To assess whether the overlap assumption was upheld, we plotted propensity score histograms. The causal forest function (grf::causal_forest) estimates generalised propensity scores [53] for continuous treatments. For binary treatments, the overlap assumption bounds the propensity scores between zero and one. A concentration of propensity scores around 0 and 1 would thus indicate deterministic treatment based on our observed covariates and prevent us from establishing inference effects. In the case of a continuous treatment and generalised propensity scores, the assumption bounds the values away from zero [54]. Therefore, generalised propensity scores bunching near 0 would also suggest the presence of deterministic treatment. To assess the goodness of fit, we followed [30] in using both the grouping by CATE method and the calibration function (grf::test_calibration). The grouping by CATE method groups observations based on whether they have high or low CATE relative to the median effect, estimates the ATE for each group, then checks whether the two estimates are statistically different. The calibration function evaluates fit by estimating the true CATE as a linear function of the out-of-bag predicted CATE and returning two coefficients: the mean forest prediction and the differential forest prediction. A statistically significant mean forest prediction of 1 suggests that the average prediction is correct. If the differential forest prediction, which measures how the predicted CATE covaries with true CATE, is 1 (or at least positive) and statistically significant, then heterogeneity is present and correctly being detected by the causal forest.

To identify the most important covariates driving treatment effect heterogeneity, we followed the Best Linear Projection (BLP) approach to estimate a linear approximation of the CATE outlined in [55]. In essence, BLP linearly regresses the CATEs against a pre-selected set of covariates using a doubly robust estimator, allowing researchers to conveniently infer significant areas of heterogeneity. We selected covariates based on variable importance, i.e., the weighted number of times a variable was split on during the learning process. Using the variable importance function (grf::variable_importance), we used different criteria, including choosing the five features that received the most splits or those that had more than the average number of splits, to subset variables. We then input the resulting subset into the best linear projection function (grf::best_linear_projection) to examine how those variables affect CATE.

4. Results

Causal forest models were trained for three different wildfire size categories as well as all fires. We used the grf causal forest function, which estimates the nuisance parameters with separate regression forests. Using the default settings of the package, our model includes 25 of 31 covariates at each split, 50% of the total sample for each tree, grew to 2000 trees, and had a minimum terminal node size of five observations. The average treatment effect of percentage of conifers on expenditures is reported (

Table 3. Average treatment effects for the causal forest model of extended attack WUI fire suppression expenditures in BC (1981–2014).

	MacMillan et al. [12] (n = 5459)		Causal Forest
			4–40 ha (n = 2997)		40–200 ha (n = 810)		>200 ha (n = 417)		All Fires (n = 4224)
Percent conifer	0.0020	***	0.0031	***	0.0065	***	0.0050	**	0.0037	***
	(0.0005)		(0.0007)		(0.0014)		(0.0018)		(0.0006)

Note: The standard errors are in ( ). Significance is indicated with *** if p < 0.001, ** if p < 0.01.

), which is representative of the expected influence of the variable on the logarithm of the suppression expenditure. The percentage of conifers was found to be significant for all the fire size regimes, with the computed average treatment effects consistent with the regression coefficients reported by [12].

To ensure the causal forest exhibits good fit for heterogeneous effects analysis, we did several tests on the treatment effect of percentage conifer. First, we plotted the propensity score histogram for small, medium, and large wildfires (Figure 2). They are all concentrated away from 0, suggesting that the overlap assumption holds; treatment is not deterministically decided by the covariates. We further grouped the CATEs into high and low estimates and calculated the ATE for each group, finding that the two groups are statistically different at the 95% confidence level for all fire size regimes. Lastly, the mean forest prediction and differential forest prediction estimates returned by the test calibration function (

Table 4. Causal Forest Calibration Diagnostics for all fire sizes.

	4–40 ha		40–200 ha		>200 ha		All Fires
Grouping by CATE Estimate	0.0030	***	0.0064	***	0.0042	**	0.0037	***
	(0.0007)		(0.0014)		(0.0017)		(0.0006)
Mean forest prediction	0.9226	***	0.9946	***	1.0338	**	0.9063	***
	(0.2551)		(0.2117)		(0.3832)		(0.1651)
Differential forest prediction	1.5247	***	1.2585	*	−0.9101		1.8208	***
	(0.3862)		(0.7467)		(1.4997)		(0.3597)

Note: The standard errors are in ( ). Significance is indicated with *** if p < 0.001, ** if p < 0.01, and * if p < 0.05.

) are close to one and statistically significant for the most cases suggesting that the causal forest is well-calibrated and capturing heterogeneity. The exception is that the differential forest prediction is negative and insignificant for large fires, indicating the absence of heterogeneity. Hence, the overall average treatment effect for large fires in Table 3 may need to be interpreted with a degree of caution despite the estimate being statistically significant at the 95% confidence level.

In

Table 5. Best linear projection for all fire size regimes based on five most important variables.

	4–40 ha (n = 2997)		40–200 ha (n = 810)		>200 ha (n = 417)		All Fires (n = 4224)
ln(DSR)	0.0025	**			−0.0020		0.0019	*
	(0.0009)				(0.0031)		(0.0008)
ln (Topographic roughness)	−0.0006						0.0003
	(0.0005)						(0.0004)
ln (Distance WUI density 3+)	−0.0007		0.0017				0.0001
	(0.0005)		(0.0011)				(0.0005)
ln (Deviation count)	−0.0003				−0.0011
	(0.0005)				(0.0016)
Private land	0.0053	**					0.0056	***
	(0.0016)						(0.0014)
ln (Detection time delay)			0.0088	**
			(0.0033)
Julian day of the year			−0.0001	*	0.0000		0.0000	*
			(0.0000)		(0.0001)		(0.0000)
In (Slope)			0.0000
			(0.0008)
Duration			−0.0004
			(0.0002)
ln (Discovery size)					0.0014
					(0.0010)
ln (Fire perimeter)					−0.0013
					(0.0019)

Note: The standard errors are in ( ). Significance is indicated with *** if p < 0.001, ** if p < 0.01, and * if p < 0.05.

, we employed the best linear projection on the five most important variables to investigate the heterogeneous effect of percentage of conifers on expenditures. We found that the Daily Severity Rating (DSR) and whether land was private were significant sources of heterogeneity for small fire size. The estimates are positive, suggesting that the effect of percentage of conifers on expenditures is stronger for fires with a higher DSR or on private land. For medium fires, the detection-time delay (time from ignition to detection) and Julian day of the year are significant sources of heterogeneity. The estimate for detection-time delay is positive, suggesting that the effect of percentage of conifers on suppression expenditures is augmented for wildfires that are detected later, while the estimate for Julian day of the year is negative, suggesting that the effect is weaker later in the wildfire season. However, it should be noted that the coefficient for Julian day of the year in the best linear projection is very small. For large fires, the best linear projection approach did not reveal heterogeneity. The relatively low sample size of large fires with 417 observations as opposed to 810 medium fires and 2997 small fires could have contributed to the lack of heterogeneity for large fires. Alternatively, there is simply no heterogeneity among large wildfires. For all fires, we found that the DSR, Julian day of the year, and private land indicator are significant sources of heterogeneity. The estimates for the DSR and private land indicator are positive, suggesting that the effect of conifers on expenditures is stronger for fires with a higher DSR or on private land. The estimate for Julian day of the year is negative and suggests that the effect of conifers on expenditures is weaker further into the fire season, although the coefficient for Julian day of the year is again small. The causal forest did not find fire size to be a source of heterogeneity.

For robustness, we fitted two additional best linear projections, one using variables that had more than the mean number of splits and the other using variables with more than the median number of splits. We found that the results are consistent as above for all fires, except that topographic roughness also emerges as a source of significant heterogeneity for medium fires (results available upon request); the effect of conifers on expenditures appears stronger for fires burning on rough terrain.

5. Discussion

5.1. Treatment Effects

Both average treatment effects and heterogenous effects of the percentage of conifer cover to the natural logarithm of wildfire suppression expenditures were obtained by use of the machine learning approach causal forests in this study. The present study also separated the dataset by fire size to determine if the variable association with wildfire suppression expenditures change with fire size.

Overall, we found that suppression expenditures grew with the proportion of coniferous forest cover. As shown in Table 2, a one percentage point increase in the percentage of conifers increases costs by 0.37% when considering fires of all sizes. Practically speaking, increasing the percentage of coniferous trees from 50% to 100% increases suppression expenditures by approximately 18.5%. This finding corroborates [12] which found using a negative binomial model that increasing the percentage of conifers from 50% to 100% is associated with an 11% rise in expenditures. This increase is because fires in conifer forests can move into the canopy and spread more quickly at higher intensity. Consequently, it can be expected that by replacing coniferous with deciduous species, or increasing the conifer composition of mixedwood forests, suppression expenditures for sustained attack wildfires could be reduced by between 0.1 and 0.9 percent per percent point of conifers, representing potential suppression expenditure savings for the province. Given the BC’s 10-year average wildfire suppression cost at $316.9 million, fuel treatment reducing provincial conifer cover by 1% could reduce provincial suppression expenditure by $0.3–$2.8 million dollars per year. Our findings highlight the potential benefits of reducing conifer dominance and promoting forest diversification. Deciduous stands, though capable of carrying fire in dry spring conditions, generally burn less intensely after leaf-out [26,56], which helps reduce suppression expenditures and potential damages. Species like aspen and birch, with higher moisture content and lower flammability, can act as natural fire breaks especially near communities and infrastructure. While this strategy shows promise, replacing conifers comes with trade-offs. BC’s forestry sector is heavily conifer-based, and deciduous species differ in growth rates, wood properties, and market value. Site conditions also influence species suitability. A full cost–benefit analysis is needed, but integrating deciduous species more intentionally into fire-prone landscapes could be a valuable adaptation strategy. By exploring the heterogeneous effects of percentage of conifers on suppression expenditures using the best linear projection, we found that when considering all fires regardless of size, higher DSR and private land strengthened the effect of conifers on expenditure, and higher Julian day of the year weakened the effect. In other words, reducing the proportion of conifers is more effective for fires with higher DSR, on private land, or in locations that burn earlier in the calendar year. The DSR during a fire and whether it will burn earlier in the year are uncertain, making targeted fuel treatment according to these variables less than desirable. Still, we could turn to historical data and future forecasts, tailoring fuel treatment to areas that exhibit consistently high DSRs or early burning times.

We found that the positive effect of the proportion of conifers on suppression expenditures is stronger on private land. This is likely because private forests are frequently located near homes and infrastructure, requiring more aggressive and costly protection. These results suggest that minimising suppression expenditures through forest thinning and diversification would be more beneficial on private land. Land ownership is largely constant, making it a reliable characteristic to base fuel treatment on.

5.2. Data Limitations

There are several limitations to the data used in this case study. While the dataset tracks individual wildfires from 1980 to 2014, fuel conditions such as the percentage of coniferous forest cover had to be calculated separately using a single available year of data from the National Forest Inventory [50]. Consequently, the fuel variables do not necessarily reflect the actual forest conditions for wildfire in a particular year. Our study would have benefited from richer data collection at the outset, as retroactively estimating the forest conditions that were present during a wildfire is challenging. The assumption that all conifers have similar effects on fire behavior and suppression is also a simplification, since species–specific differences such as fuel moisture, crown structure, and flammability can significantly influence fire dynamics. Furthermore, the dataset tracks conditions near the ignition point of the fire and does not account for spatial and temporal heterogeneity as a wildfire progresses. In light of this limitation, the causal forest we estimated for small fires is possibly more reliable than our other estimates since smaller fires are more likely to have conditions consistent with the ignition point. Dependence on the ignition point was a common weakness in most suppression expenditure models until recently [57]. Our study reinforces that the use of advanced analytical methods in fire management is limited by data availability in many cases. Fortunately, intricate geospatial data on wildfires are becoming increasingly available. Thus, future research can improve the accuracy of this causal forest approach by rounding out the data inputs to incorporate spatial and temporal heterogeneity. In addition, future research could specifically investigate the effect of private land on suppression expenditures, given its current disaccord in the literature [12,58]. Furthermore, the research does not account for the subjective nature of decision-making, which can significantly influence suppression expenditures. Suppression decisions are often made under pressure, with limited information, and are shaped by human judgment, risk perception, and institutional priorities. Decision-makers may choose aggressive tactics not solely based on fire behavior. Variability in experience, training, and local knowledge among decision-makers can also lead to inconsistent suppression strategies across similar fire events. While fire behavior and fuel types set the context, it is often the human element driven by urgency and uncertainty that determines the final expenditure. To assess whether expenditures are justified or excessive, the USDA uses the Stratified Cost Index (SCI), a regression-based tool comparing actual costs to predicted values based on fire size, fuel type, intensity, and proximity to communities [59,60]. By controlling operational and environmental factors, SCI identifies outliers and supports benchmarking. It can be enhanced with metrics like resource intensity, WUI proximity, and fuel treatment history to inform more efficient, risk-based suppression strategies. Future research could build on this approach to better account for human and institutional drivers of suppression expenditures. Finally, although CFs addresse confounding through a revisualization (orthogonalization) procedure, which is a key strength of the method, unobserved confoundedness due to missing variables such as local fire management strategies or historical land use could still bias the estimates. It also requires enough variation in treatment across different covariate profiles to produce stable results.

5.3. Causal Forest Approach

Despite the limitations of the current data, the causal forest approach demonstrates potential for broader applications for fire risk mitigation and analysis beyond the scope of the current data. The fire suppression expenditures associated with a particular fire can be influenced by different predictors in unique ways, depending on how large a fire might become or other factors that are present at ignition. As such, effectiveness of a given fuel treatment can be expected to differ between various scenarios and circumstances. Causal forests could therefore be used as a tool to aid decisions involving the mitigation of fire risk or damage potential, since the effectiveness of different fire mitigation strategies can be predicted for specific scenarios and not by a regression fit to a general set of data.

In addition, the causal forest model’s ability to predict treatment effects for any given set of covariates allows the model to generate probability distributions for the predicted suppression expenditure savings based on known covariates of a proposed treatment (e.g., the location and fuel type, proposed reduction in conifer species). A Monte Carlo-based approach could then generate the low, expected, and high estimates of suppression expenditure savings due to a proposed treatment, allowing for a situation-specific cost–benefit analysis. Such computed distributions may be useful when quantifying the probability of treatment success for a given set of conditions (e.g., location or fuel), since a one-tailed p test could be performed on a tailored distribution that represents the predicted treatment effects for the conditions of interest.

Forests are inherently variable in their species composition, structure, and site conditions. Many key processes (regeneration, growth, and mortality) as well as disturbances (wildfire, insects, and pathogens) are shaped by a complex array of environmental factors. This variability creates the potential for heterogeneity in treatment effects, which is well recognized in studies of fuel treatments [61], silviculture [62], growth and yield [63], and other areas of forest research. Indeed, classical experimental design techniques (blocking, factorial design, randomization) were developed to reduce heterogeneity [64] which is considered an extraneous factor, or to account for it statistically (analysis of covariance). However, understanding heterogeneity in treatment effects is important in real-world applications. Causal forests offer a promising approach to reveal these nuances, particularly in observational studies where experimental controls may not be feasible.

6. Conclusions

We trained causal forest models to assess the treatment effect of percentage of conifer on suppression expenditures in the Wildland–Urban Interface in British Columbia using fire suppression expenditure data from 1981 to 2014. The treatment effect refers to the estimated impact of the percentage of conifer on suppression expenditures, while holding other factors constant. It helps isolate the influence of conifer-dominated fuels from other variables such as terrain, proximity to the WUI, and fire weather. Our findings show that conifer cover significantly increases suppression expenditures. This effect is especially pronounced on private land, under higher Daily Severity Rating (DSR) conditions, and earlier in the fire season. These results suggest that conifer fuels are a key driver of suppression expenditures, and their impact varies depending on land ownership, fire severity, and timing.

The results offer insight into how provincial fire suppression expenditures could be influenced by forest fuels management, in particular by the reduction in conifer fuel or by the addition of deciduous fire breaks. The findings support the use of landscape fuels management as a meaningful tool to potentially reduce wildfire suppression expenditures incurred by fire managers. Strategic placement of broadleaf fire breaks and the promotion of mixed coniferous–deciduous forest cover near the WUI has the potential to reduce firefighting expenditures, as well as bring a multitude of benefits associated with maintaining biodiverse forests, for example in the form of increased avian diversity [65] or in the form of human well-being attained from peri-urban forests that are perceived to be biodiverse [66]. The results also suggest that we should prioritize wildland fuel treatment on private land.

For future study, the specific expenditures savings offered by coniferous–deciduous species replacement could also be quantified with use of wildfire simulations such as Burn-P3. Suppression expenditures could, for example, be modelled for a baseline fire scenario with untreated, coniferous stand, and compared against an alternative with a fraction of the coniferous trees replaced with deciduous fuels. Cost–benefit analysis of the fuel treatment could be explored with different percentage points of conifers replaced, or by altering the geometry of fuel treatment, for example by comparing how belts and blocks of deciduous trees might mitigate wildfire spread and consequently reduce suppression expenditures, possibly in a manner similar to [20].