The Impact of Wildfires and Wildfire-Induced Air Pollution on House Prices in the United States

Property Data

Our research is based on ZTRAX (version 5, April 2021, Zillow Group 2021), which contains the nationwide housing transaction and assessment records. We create a repeat-sale dataset from 2010 to 2018 by merging the housing transaction and assessment datasets. We chose the years 2010–2018 because the nationwide housing market experienced a significant change due to the shock of the financial crises between 2007 and 2008, and the wildfire dataset we use provides data through 2018. Our focus is specifically on properties with repeat sales, which allows us to address certain data limitations. By using parcel fixed effects, we can effectively control for time-invariant house characteristics, thereby reducing bias associated with omitted variables, such as unobservable house attribute variables.

We discard transactions with missing sale prices and geographic location information and keep only deed transfers of single-family residential properties. We further eliminate transactions that are labeled as non–arm’s length sales and intrafamily transfers, as well as residences with frequent transactions, to reduce the bias created by transactions that do not reflect sales under typical conditions. Furthermore, because we use repeat-sale data and control for the parcel fixed effects, we only consider residences that have not had any remodeling, new building, or major rehabilitation between the first and last transaction dates of 2010 and 2018. Any observations reporting a construction year that follows the sale year are also excluded. To minimize data entry errors, we trim the top and bottom 1% of the data using total rooms and lot size. We eliminate the outliers with sale prices less than $10,000 and more than $10,000,000, as well as properties with more than 20 rooms and 100,000 square feet. The sale prices are adjusted using the monthly housing consumer price index from the U.S. Bureau of Labor and Statistics (n.d.). The detailed data processing procedures and the figure of the sample distribution across the United States are presented in Appendix A.

Wildfire and Air Pollution Data

The data on wildfires used in this study were sourced from the FPA FOD (Short 2021), which documents 2,166,753 wildfire events that occurred in the United States from 1992 to 2018. By using the wildfire centroid coordinates in the database, which offer a higher level of location specificity than county-level data, we were able to accurately pinpoint the locations of the wildfires and correlate them with the coordinates of local residences and wind pattern maps to generate measurements of the wildfires. Excluding observations from Puerto Rico, Alaska, and Hawaii, we obtained a total of 2,120,520 observations. Because most of the dates on which the wildfires were declared contained or controlled are unavailable, we relied on the wildfire discovery date for our analysis. We assume that if there was no record of a wildfire in a particular county on a given date, then no wildfire occurred. Figure 2 shows the distributions of wildfires of all sizes and those that burned at least 300 acres at the county level across the contiguous United States. This figure illustrates that wildfire incidence rates vary across the country, with the western and southern states experiencing more frequent and intense flames, a trend consistent with the occurrence of extreme heat and drought events in these regions. Further details on the measurements of the wildfires are discussed below.

• Download figure
• Open in new tab
• Download powerpoint

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Distribution of Wildfires (2010–2018): top, Total Occurrences for All Wildfires Size Classes; bottom, Total Occurrences for Wildfires with at Least 300 Burned Acres

Source: Fire Program Analysis Fire-Occurrence Database (Short 2021).

We use PM_2.5 to measure wildfire-induced air pollution in this article. The first reason is that PM, particularly PM_2.5, is the primary pollutant of concern from wildfire smoke; PM_2.5 accounts for over 90% of the total wildfire smoke particle mass released by wildfires (Stone et al., 2019). Second, according to the EPA’s “Integrated Science Assessment (ISA) for Particulate Matter” report (U.S. EPA 2019), causal relationships between health effects and PM_2.5 are relatively more likely to exist among the various size fractions of PM, which means that the level of PM_2.5 is more likely to affect health status and thus influence household purchase decisions. The monthly total concentration estimates of ground-level PM_2.5 data in the United States originate from the Atmospheric Composition Analysis Group (ACAG) (van Donkelaar et al. 2019; Hammer et al. 2020).7 We spatially join each property’s coordinate to the shapefiles of monthly PM_2.5 and then extract the values of the local PM_2.5 level. The ACAG dataset is gridded at the finest solution (0.01 × 0.01), allowing us to obtain the monthly PM_2.5 level for the property’s surrounding area (about 1.11 km × 1.11 km). We use the mean level of PM_2.5 at the census tract level for properties for which we cannot extract a value from the shapefile directly. If the zonal mean at the census tract level is still unavailable, we use the county-level zonal mean of PM_2.5.8 Figure 3 shows the distributions of mean PM_2.5 at the county level across the contiguous United States. Generally, the distribution is consistent with the ISA report (U.S. EPA 2019, 2020) that the eastern areas of the country suffer higher but a more uniform level of PM_2.5 than western areas, whereas California has a significantly higher level of PM_2.5 than the surrounding states.

Figure 3

Average Ground-Level Fine Particulate Matter (2010–2018)

Source: Atmospheric Composition Analysis Group.

Other Data

We control for weather, demographics, and surrounding environmental factors. The monthly data of precipitation, pressure, humidity, and temperature can be obtained from the National Aeronautics and Space Administration Phase 2 of the North American Land Data Assimilation System.9 The monthly local precipitation, pressure, humidity, and temperature are extracted from the shapefiles using a similar processing method to PM_2.5, and if the values are missing,10 we use the zonal means at the census tract or county level.

The neighborhood features that we consider include vegetation coverage, population density, home density, the ratio of white people, and the distance between properties and the WUI. While families living in or near the WUI enjoy the benefits of living close to forests, their homes are more susceptible to the risks of wildfires. This trade-off may explain why more households are choosing to live in or near the WUI. To investigate this trade-off, we use two variables to control for potential confounding factors. First, we consider the proportion of vegetation at the census block level, which is a proxy for the benefits of living near the woods. Second, we measure the distance between each property and the nearest intermix or interface WUI, which is a proxy for wildfire risks. This approach helps us evaluate the benefits and dangers of living near the woods and mitigates the endogeneity problem of wildfires by controlling for other factors that may affect the relationship between wildfires and house prices.

To collect the data, we use the 1990–2010 WUI of the contiguous United States geospatial data, which provides details on housing and population density in 2000 and 2010, the proportion of vegetation coverage in 2001 and 2011, and WUI areas in 2000 and 2010.11 Using ArcGIS, we join the coordinates of properties with the shapefiles and extract values for housing and population density and the proportion of vegetation and calculate distances between residences and the nearest WUI. Race data at the census block group level are obtained from IPUMS National Historical Geographic Information System (Manson et al. 2021), which originates from the U.S. Census Bureau 2000 and 2010 census data. To expand our observations beyond the available years, we apply extrapolation methods to obtain data between 2011 and 2018 for housing density, population density, race, and distances between WUI and houses. For the proportion of vegetation coverage, we use interpolation to obtain the data for 2010 and extrapolation to get observations between 2012 and 2018.

To create the wildfire measurements, we also calculate the distances between wildfire centroids and properties and extract wind pattern data at wildfire centroids. Distances are calculated with ArcGIS. The wind data originates from the NLDAS-2 (Mocko and NASA/GSFC/HSL 2012; Xia et al. 2012). The wind direction at the wildfire centroid is determined by the monthly zonal and meridional wind speeds. Similarly, wind speeds are extracted directly from the shapefiles.

Finally, we obtained a repeat-sale dataset between 2010 and 2018, which covers 48 contiguous states, Washington, DC, and 1,834 counties. There are 3,945,340 transaction records of 1,886,684 houses.12 About 41.36%, 28.64%, 17.19%, and 12.81% of the sample is distributed in the South, West, Midwest, and Northeast regions, respectively. Most of our observations are in the areas with more frequent wildfire occurrences. In the next section, we introduce the variables in more detail. Table 1 presents the definitions of variables used in the analysis. Table 2 presents the summary statistics of the main variables used in the empirical analysis.

Direct and Indirect Impact of Local Wildfires

We present the empirical methods used to estimate the impact of wildfires and wildfire-induced air pollution on house prices based on the hedonic price model. Note that we only keep those properties that sold at least twice; thus, we use a repeat sales framework here. We first estimate the total impact of wildfires, as presented in equation [8]: 8 ln(P_hym) is the natural log of the adjusted sale price of house h that sold in month m year y in county c. As shown in Figure 1, when creating the wildfire measurements, we consider the following factors: wildfire frequency, wildfire size, wildfire causes, wind pattern, the timing of wildfire, and the distance between the house and the wildfire. Thus, Fire_hym is a vector of local wildfire measurements, including the weighted sum of wildfire occurrences, Local_hym (we also divide the local wildfires into two categories: upwind wildfires, Local_Up_hym, and downwind wildfires, Local_Down_hym) and two additional considerations about extreme wildfire exposures, the number of days since the most recent wildfire that occurred within 80 km of the house since 1992, Days_hym, and the distance between the property and the nearest wildfire that happened over five years before the transaction month m, Distance_hym.13 Figure 4 provides examples to illustrate the upwind and downwind wildfires and the upwind and downwind areas. Most wildfires, unlike other natural disasters such as hurricanes, are sparked by human activity; therefore, they may not be wholly exogenous. As a result, we include the ratio of wildfires caused by natural causes as an additional control variable.

To assess local effects, in our main analysis, we only consider wildfires that have burned at least 300 acres14 within 30 km of the property h over 12 months before the transaction month m year y, denoted by J_hym. The weighted sum of wildfire occurrences, Local_hym, is defined in equation [9]. For each wildfire event (j ∈ J), its impact is discounted by the distance (km) between the house and the wildfire, d_hj. The weighted upwind and downwind wildfire occurrences are defined in equations [10] and [11], respectively. If the angle between the wind vector at the wildfire centroid and the vector from the wildfire centroid to the property (θ_hj) is less than 90°, we consider wildfires to be upwind, and the property is in a downwind direction. Wildfires, on the other hand, burn in the downwind direction of the property. Because homebuyers typically make their housing purchase decision before the transaction date, changes in local environmental amenity levels should influence their decision-making process before the transaction date; we consider the wildfire events that occurred during the 12-month period before the transaction month m for the short-term effect. Figure 5 shows examples of constructing wildfire measurements.

9

10

11

W_hym denotes a vector of time- and location-variant factors including the ratio of natural-caused wildfires,15 average temperature, precipitation, humidity, and pressure over the study period of the surrounding area (about 13.88 km × 13.88 km) of the property, housing density, population density, and ratio of vegetation by census block level in year y, ratio of white people by census block group level in year y, and distance between house and intermix/interface WUI in year y. τ_c × σ_y denotes the county-by-year fixed effects, which control for the unobserved constant factors in each year of each county. η_m denotes the month-of-year fixed effects, which control for the monthly variations over the annual cycle. ∂_h denotes the property fixed effects. The standard errors are clustered at the property level.

Figure 4

Examples of Upwind and Downwind Directions

The next step is to test whether local wildfires have a significant impact on local air quality, as shown in equation [12]. If wildfires can significantly influence local air pollution levels, we consider whether air pollution, specifically PM_2.5, can be regarded as the channel through which wildfires influence house prices. PM_hym denotes the average PM_2.5 level of the house’s surrounding area (about 1.11 km × 1.11 km) for the previous 12 months before transaction month m year y. We use the weighted sum of wildfire occurrences, Local_hym (or Local_Up_hym and/or Local_Down_hym), to measure wildfires. The ratio of nature-caused wildfires is controlled as well. The covariates W_hym are the same as those in equation [8]. County-by-year fixed effects, τ_c × σ_y, month-of-year fixed effects, η_m, and property fixed effects, ∂_h, are also included. The standard errors are clustered at the property level. 12 Last, we test whether wildfires influence house prices through PM_2.5. We add PM_2.5 as an additional variable in equation [8], as shown in equation [13]. If PM_2.5 has a significant effect on house prices (β_p) as well as wildfire exposure (α_f in equation [8] and β_f in equation [13]), then we think there is clear evidence that wildfires indirectly influence house prices through PM_2.5. 13

Figure 5

Wildfire Measurements

Addressing Endogeneity of PM_2.5

When evaluating the impact of PM_2.5 as per equation [13], one concern is that PM_2.5 may be linked with unobserved factors in the error term (i.e., violate the second assumption of sequential ignorability), although we control for several factors and fixed effects. For example, we cannot fully control for the sectoral makeup of the economy, crime rate, and school district boundaries. The variation in these neighborhood attributes may influence consumer choices and home prices. As suggested by Imai et al. (2011), we apply an instrumental variable method to address the endogeneity issue of the mediator. Therefore, we need to find at least one excluded exogenous variable as our instrumental variable (Wooldridge 2010). Previous studies have proposed various instrumental variables for air pollution, including policy interventions like the U.S. Clean Air Act’s nonattainment status designation and the Chinese River Huai policy, which were used in studies by Chay and Greenstone (2005) and Huang and Lanz (2017), respectively, to create exogenous variability in air pollution. Another instrumental variable is the externality associated with transboundary air pollution, which has been used in studies by Bayer, Keohane, and Timmins (2009), Luechinger (2010), Khawand (2015), Zheng et al. (2014), Barwick et al. (2018), Yang and Zhang (2018), Williams and Phaneuf (2019), Zheng et al. (2019), and Chen et al. (2021). These studies constructed distant air pollution as an instrumental variable for local air pollution because air pollutants can travel a long distance by wind, and distant air pollution emissions are unlikely to correlate with local economic activities. In a study from Indonesia, Tan-Soo (2018) considered the influence of forest fires on local air quality and used a similar source-receptor logic to build wind- and distance-based fire hotspots as an instrument for local PM_2.5. Building on these previous studies, we use distant large-scale natural-caused wildfires as an instrumental variable to estimate the impact of wildfire-caused PM_2.5 on house prices.

Following these studies, we create a variable, DistPM_cym, to measure distant air pollution and study the spillover effects of air pollution. DistPM_cym is defined as equation [14] and the example of the construction of the distant all-source PM_2.5 can be found in Appendix Figure B1. DistPM_cym denotes the average monthly imported PM_2.5 from counties (k ∈ K(d_ck ≥ 100km)) that are at least 100 km from county c where the property h is located over the previous 12 months from the transaction month m year y.16 Through regression analysis, we find a significant positive relationship between distant PM_2.5 and local PM_2.5 (results are available on request). Appendix Figure B2 depicts the spatial distribution of the distant PM_2.5 from 2010 to 2018, which is the average of DistPM_cym for house h located in state s. This figure highlights the states that suffered more PM_2.5 from distant counties over the 2010–2018 period. Influenced by the wind direction, the distribution of distant PM_2.5 shows some differences with the local PM_2.5, but in general, the eastern and inland areas suffered more imported all-source air pollution.

Furthermore, the exclusion restriction assumption of the instrumental variable requires that the instrumental variable should not affect house prices through any channels other than PM_2.5. Hence, given that wildfires are among the major sources of PM_2.5, and there is considerable evidence of transboundary PM_2.5 spillover effects, using wildfires as the instrumental variable can help test the indirect or mediation effect of wildfires. Because local wildfires can have a direct impact on house prices, we instead use distant wildfires as the instrumental variable to tackle the endogeneity problem and investigate whether PM_2.5 can be considered as a pathway through which distant wildfires affect house prices. 14 Based on Tan-Soo (2018), we create distant large-scale natural-caused upwind wildfires, DistFire_Up_hym, as an instrumental variable. A difference is that we only consider the natural-caused wildfires to further increase confidence in the exogeneity of the instrumental variable. This instrumental variable is obtained using equation [15]. We also create distant large-scale natural-caused downwind wildfires (defined in equation [16]) and all-cause upwind wildfires (same construction method as in equation [15] but including all-cause wildfires) to examine robustness. Only natural-caused wildfires that burned at least 1,000 acres and occurred at least 100 km from the property during the 12-month period before the transaction month m year y are considered, which are denoted by S. Similarly, the distance and wind weights are taken into consideration. 15 16 Appendix Figure B3 depicts the spatial distribution of the predicted value of PM_2.5 attributed to distant natural-caused upwind wildfires from 2010 to 2018 for house h in state s. This predicted value can be obtained by equation [17]. The distribution presents a different picture to that of the local PM_2.5 and wildfire occurrence. First, because the western and southern regions experienced significantly more frequent and intense wildfires than the eastern and northern regions and we apply the distance discount when we create distant natural-caused upwind wildfires, it is reasonable to find that the western and southern regions suffer more from the PM_2.5 attributed to distant wildfires. Second, perhaps because the sea wind clears the air pollutants in coastal states as compared to inland states, the coastal areas in the west and south suffer less PM_2.5 attributed to distant large-scale natural-caused upwind wildfires relative to inland places. 17 We then can apply two-step regression to get the instrumental variable estimate, , as presented in equations [18] and [19], which is the local average treatment effect of PM_2.5 on house prices. In the first stage, we regress the endogenous variable, PM_hym, on instrumental variables (DistFire_Up_hym) and all the other exogenous variables from the previous models as well as county-by-year fixed effects, month-of-year fixed effects, and house fixed effects to obtain the predicted local PM_2.5. In the second stage, we substitute PM_2.5 with the predicted value obtained in the first stage. The exogenous variation of PM_2.5 in the second stage is attributed to the exogenous variation of distant upwind natural-caused wildfires. As a result, the instrumental variable estimate measures how PM_2.5 that is attributed to distant wildfires affects local house prices, that is, the indirect impact of distant wildfires. If the components of PM_2.5 do not have significant changes, we can assume that the indirect effects of local wildfires and distant wildfires are the same. That is, a one-unit increase in PM_2.5 attributed to wildfires (distant or local) leads to a percentage change in house prices. If the overall exposure to local upwind (downwind) wildfire increases by one unit, then the house prices will increase by a () percentage point. In addition, if the overall exposure to distant upwind wildfire increases by one unit, the house prices will increase by a percentage point. 18 19 Next, we discuss the validity of instrumental variables. The valid instrumental variable should satisfy two conditions: relevance and exogeneity. First, the air quality in other counties should have an impact on local air quality. Previous research has found evidence of a significant transboundary spillover effect of air pollution, which is mostly driven by wind and is related to distance (Bayer, Keohane, and Timmins 2009; Banzhaf and Chupp 2010; Luechinger 2010; Khawand 2014; Zheng et al. 2014, 2019; Barwick et al. 2018; Yang and Zhang 2018; Chen and Ye 2019; Williams and Phaneuf 2019; Chen et al. 2021). Second, distant wildfires can produce a large amount of PM_2.5, which can be carried by the wind and affect local air quality as well. The t-statistic of the instrumental variable obtained in the first-stage regression (presented in Section 5) also shows that the instrumental variable is a strong predictor of local PM_2.5.

Second, our instrumental variable should not directly influence house prices or influence house prices through channels other than PM_2.5. To reduce the possibility that distant wildfires are correlated with variables affecting property prices in the focal county, we create a 100 km buffer zone and only focus on counties outside the buffer zones. Other factors used to construct instrumental variables include exogenous wind direction, geographic distance, and natural-caused wildfires, further ensuring the exogeneity of the instrumental variables. To improve confidence in the instrumental variable, robustness examinations using distant natural-caused downwind wildfires and distant all-cause upwind wildfires were performed. Although DistPM_cym cannot be used to get the indirect impact of wildfires, we use DistFire_Up_hym and DistPM_cym as instrumental variables to conduct the overidentification test. The robustness examinations and overidentification test results suggest that our results are robust, and we cannot reject the hypothesis that our instrumental factors are valid.

There may also be a concern that other air pollutants emitted by wildfires can be carried by wind to the local area and thus affect house prices. Air pollutants may influence people’s decisions because they can lead to limited visibility and adverse health outcomes. As we discussed earlier, wildfire smoke has complex components, and PM is the major component of wildfire smoke. These particles tend to be very small, and about 90% of total particle masses consist of PM_2.5 (Stone et al. 2019). The impact of other components on visibility is comparatively smaller than that of PM_2.5 and thus are less likely to affect purchase decisions. In terms of health concerns, three pollutants (PM, ozone, and carbon monoxide) may pose health threats during wildfire events (Stone et al. 2019). First, PM_10-2.5 (PM₁₀ is composed of PM_2.5 and PM_10-2.5) is not a major concern, because about 90% of total particle masses consist of PM_2.5 (Stone et al. 2019). Second, carbon monoxide dilutes rapidly, so it is rarely a concern unless people are in very close proximity to wildfires (Stone et al. 2019). As a result, it is improbable that carbon monoxide travels to the focal county and influences local health. Third, ozone is not directly emitted from a wildfire but forms in the plume as wildfire smoke moves downwind (Stone et al. 2019), so ozone can be another channel through which distant wildfires influence local health. Ozone is not the major component of wildfire smoke and is invisible. Although ozone can affect health, it is less likely to be recognized by people. Thus, ozone can be another potential channel, but for the reasons outlined above we assume that the impact of this channel is minor. We also conduct an empirical study to verify this assumption in the future.

Long-Term Effects and Spillover Effect of Wildfires

In the main analysis, we focus on the impact of wildfires that occurred in the 12 months leading up to the transaction month. However, as McCoy and Walsh (2018) found that household perception of wildfire risk tends to return to baseline levels after two to three years, we investigate longer-run impacts in this section by examining the weighted sum of wildfires over the past two and three years. To account for the declining influence of past wildfires on current perceptions of risk, we apply a discount of one-half for wildfires that occurred two to one year before the transaction and a discount of one-third for wildfires that occurred three to two years prior to the transaction. This approach enables us to capture the effects of wildfires over a longer time horizon.

Our instrumental variable strategy enables us to analyze the spillover effects of distant wildfires. If we assume that the marginal effect of PM_2.5 attributed to distant wildfires and that attributed to local wildfires on house prices are the same, then we can compare the indirect impact of local wildfires and spillover effects of distant wildfires by comparing the magnitudes of effects of distant and local wildfires on PM_2.5, as discussed above.

Robustness Check

We performed extensive robustness examinations, including (1) testing the impact of wildfires that burned at least 100 acres or at least 1,000 acres, (2) constructing instrumental variables with buffer zone radius sizes of 80 and 120 km, (3) assessing the impact of natural-caused wildfires, (4) estimating heterogeneous impacts across different regions, (5) excluding states with nondisclosure requirements for house sale prices, and (6) conducting a placebo test using randomly assigned wind pattern weights for wildfire occurrences. For further details, refer to Appendix C.