Wildfires exacerbate inequalities in indoor pollution exposure

Wildfires lead to dramatic increases in fine particulate matter pollution concentrations. Based on the premise that higher-income households purchase more defensive investments to reduce the degree to which outdoor pollution infiltrates indoors, in this study, we investigate how income contributes to outdoor–indoor pollution infiltration rates during wildfire events. Using crowd-sourced data from the PurpleAir Real-Time Air Quality Monitoring Network and econometric models that explore variations in monitor readings over time, we find increases in outdoor pollution lead to significant increases in indoor pollution, but disproportionately so in lower-income areas. The results highlight a new inequality in pollution exposure: not only are outdoor pollution levels higher for lower-income individuals, but indoor pollution levels are higher even for similar outdoor pollution levels.


Introduction
Wildfires pose a significant threat to society.They lead to dramatic increases in fine particulate matter (PM 2.5 ) pollution, a pollutant that travels deep into our bodies to cause considerable harm to health [1][2][3].The higher pollution exposure occurs not only outdoors but also indoors given how rapidly and effectively PM 2.5 penetrates indoors [4].This is particularly concerning because people spend the vast majority of their time indoors [5,6], thereby limiting the apparent value of seeking refuge inside one's home during a wildfire.
People can, however, make defensive investments to improve indoor pollution levels resulting from wildfires: installing appropriate filters into existing heating, ventilation, and air conditioning (HVAC) systems or using stand-alone air purifiers can effectively remove particles from the air [7][8][9].Given the costs involved in purchasing and running these devices, higher-income households are more likely to own and use them [10,11].Based on this pattern, we hypothesize that indoor pollution levels increase more rapidly during wildfires for lowerincome families.Such a finding would highlight a new inequality in pollution exposure: not only are outdoor pollution levels higher for lower-income individuals [12], but indoor pollution levels are higher conditional on experiencing the same outdoor pollution level.
Here, we investigate how income contributes to outdoor-indoor pollution penetration rates during wildfire events.We use crowd-sourced data on outdoor and indoor PM 2.5 levels in California in 2020, a time period that includes the August Complex fire, which involved a series of wildfire events that led to elevated PM 2.5 levels for over two months.After merging census tract-level information on wildfire smoke exposure and income levels, we analyze the moderating role of income on the relationship between outdoor and indoor PM 2.5 levels.We use graphical evidence and panel data regression models that include census tract fixed effects (FEs).By including these FEs, we explore how daily changes in smoke-induced outdoor PM 2.5 relate to daily changes in indoor PM 2.5 within a census tract, thereby removing the influence of time-invariant characteristics of a census tract.We interact income with outdoor pollution to explore if the correlation in daily changes in outdoor and indoor PM 2.5 differs by income levels in a census tract.Consistent with previous evidence, our analysis reveals that increases in outdoor pollution lead to significant increases in indoor pollution.However, the impact is considerably larger in lower-income areas, consistent with our hypothesis that defensive expenditures that improve air quality are less common in these areas.

Methods
Our primary source of data is outdoor and indoor PM 2.5 levels in California in 2020, obtained from crowd-sourced data from the PurpleAir (PA) Real-Time Air Quality Monitoring Network (www2.purpleair.com/).PA produces a retail, residential air pollution monitor that measures airborne particulate matter at 120 s intervals.PA sensors use laser particle counters to measure particle number concentrations (number of particles per deciliter, dl) and an algorithm based on assumptions about the particulate mix to report measures of PM 2.5 in µg m −3 .Although not approved by the Environmental Protection Agency for regulatory purposes, evaluations indicate high reliability of these monitors [13].
The individual monitor data include its geocoded location, which we use to build our dataset for analysis.To attribute an outdoor PM 2.5 measure to each indoor monitor, we take an inverse distance-weighted average of all outdoor monitors within 10 km of each indoor monitor.We then calculate the daily average indoor and outdoor concentration in each census tract.Limiting our sample to indoor monitors that operated from 2019 through 2021 with at least 100 d of observations in 2020, our final dataset, aggregated to the daily level, includes 110 563 observations from 450 indoor monitors located in 345 census tracts.
We merge several external sources of data at the census tract level.For wildfire smoke PM 2.5 concentration, we rely on the data from Childs et al [14], which provides daily average estimates of census tract-level PM 2.5 originating from wildfires.We assign the median household income at the census tract, obtained from the 2019 Census of Population and Housing.We use daily weather data from NOAA's Integrated Surface Database, based on hourly readings from monitors with less than 25% missing observations in 2020.To interpolate missing values, we calculate inverse distance-weighted averages from all monitors within 50 km.We attribute census tractlevel measures by taking the average of all stations within the tract or the values of the closest station within 50 km if no monitor is present within the tract [15].
To explore the relationship between outdoor and indoor pollution, we conduct multivariable regressions to adjust for potential confounding factors.The regressions include controls for: (1) weather (second-order polynomials for air temperature, dew point temperature, wind speed, and precipitation, with separate terms for positive and negative air and dew point temperatures) given its established impact on pollution levels; (2) indicator variables for each month to adjust for seasonality; (3) indicator variables for each day of the week to account for daily patterns in lifestyle, such as cooking behavior, that may affect indoor pollution levels; and (4) census tract fixed effects to account for all time-invariant characteristics within a census tract.We specify both indoor and outdoor PM 2.5 in logs to account for the skewed nature of the data as shown in figure 1, a point further discussed in the supplementary information.Given the log-log regression specification, we interpret the coefficient as an elasticity, or the % change in indoor pollution from a 1% change in outdoor pollution.All regressions are clustered on the census tract using Liang-Zeger standard errors to allow for arbitrary serial correlation within a tract.
To explore the moderating role of income, we include an interaction term between outdoor pollution and high-income in the regression equation, where we define high-income as the median household income of the census tract being above the median for all census tracts observed in California.We omit the below-median income group as the reference, which enables us to interpret the non-interacted outdoor PM 2.5 as the elasticity for below-median income and the interacted term as the differential elasticity for the high-income group relative to lowincome.We also estimate models using log income as a continuous variable, with the interaction term interpreted as the differential elasticity in indoor PM 2.5 from a 10% increase in income.Results using quartiles of income are in table S1 of the supplementary information.

Results
Figure 1 displays kernel densities and the means of indoor and outdoor PM 2.5 levels by wildfire status: days with no smoke, days with below-median smoke (low-intensity wildfires), and days with abovemedian smoke (high-intensity wildfires).Average indoor PM 2.5 is considerably lower than outdoor within wildfire status.Both outdoor and indoor PM 2.5 , however, increase dramatically during wildfires.In fact, indoor PM 2.5 during high-intensity wildfires (mean of 18.1 µg m −3 ) surpassed outdoor PM 2.5 during periods of either no smoke (8.6 µg m −3 ) or low-intensity wildfires (12.4 µg m −3 ), suggesting considerable infiltration of outdoor PM 2.5 into the indoor environment.
Figure 2 provides a glimpse into the effects of wildfires on PM 2.5 levels by income.The top panel plots the mean daily indoor PM 2.5 level (and 95% confidence interval) by decile of daily wildfire smoke PM 2.5 , separately for census tracts above and below the median household income.The bottom panel produces the same plot for outdoor PM 2.5 levels.Three patterns emerge.First, as smoke intensity rises, outdoor and indoor pollution levels increase for all census tracts.Second, the correlation between smoke intensity and indoor PM 2.5 levels rises at a faster rate for monitors in low-income census tracts.At low levels of smoke, indoor PM 2.5 levels are comparable by income.At higher levels of smoke, indoor PM 2.5 is higher in low-income tracts, with the differences statistically significant for the 7 highest smoke intensity deciles.At the top smoke decile, the difference in indoor PM 2.5 is 10.1 µg m −3 .Third, in stark contrast to indoor levels, outdoor PM 2.5 levels rise comparably for high-and low-income census tracts, with mostly non-significant and small differences at all smoke intensity deciles.This pattern suggests differential defensive investments by income that underlies the differences in indoor PM 2.5 levels.
In table 1, we show indoor and outdoor PM 2.5 levels according to wildfire status as defined in figure 1, separately for high-and low-income census tracts, with the last column providing results of a t-test for differences between the two.On days without wildfire smoke, outdoor PM 2.5 levels are roughly 3 µg m −3 higher in low-income tracts, whereas the difference in indoor levels of 0.4 µg m −3 is quite small in magnitude despite being statistically significant.Moving to low-intensity wildfires, both indoor and outdoor PM 2.5 levels rise, but the differences by income are still small (1.1 for indoor and 1.8 for outdoor).Once wildfire smoke reaches elevated levels, indoor and outdoor PM 2.5 levels further increase.Consistent with figure 2, the difference of 1.7 µg m −3 in outdoor pollution is statistically insignificant, whereas the difference of 7 in indoor pollution is statistically significant and large in magnitude.
In our first set of regression results, we focus on outdoor PM 2.5 as the independent variable.Although outdoor PM 2.5 can be affected by various sources, almost all of the high levels of PM 2.5 in our sample are attributable to wildfires.As shown in table 2, column (1), a 10% increase in daily outdoor PM 2.5 leads to a statistically significant 8.63% increase in indoor Figure 2. Mean indoor and outdoor PM2.5 levels by income and deciles of wildfire smoke.The dots plot the mean daily indoor PM2.5 (and 95% confidence intervals) for days without smoke and for deciles of daily wildfire smoke separately for census tracts below (red) and above (dark blue) the median household income (top panel, left axis).The crosses plot the mean daily outdoor PM2.5 levels (and 95% confidence intervals) for days without smoke and for deciles of daily wildfire smoke PM2.5, separately for census tracts below (red) and above (dark blue) the median household income (top panel, left axis).Estimates from high-income tracts that do not overlap with confidence intervals from low-income tracts represent statistically significant differences.

Table 1.
Means and standard deviations (in parentheses) of indoor and outdoor PM2.5 levels by wildfire status-no wildfire smoke (top), below-median wildfire smoke (center), above-median wildfire smoke (bottom) -separately for high-(column 1) and low-income (column 2) census tracts.Column 3 provides the results of a t-test for differences between column 2 and column 1 with the standard errors in brackets.PM 2.5 on the same day for low-income census tracts.

Above-median
The statistically significant coefficient on the interaction term between outdoor PM 2.5 and income implies indoor penetration is 1.79 percentage points lower for high-income census tracts.Based on the means from the third panel in table 1, these estimates translate to a marginal impact of a 0.30 µg m −3 increase in indoor PM 2.5 in low-income tracts and 0.17 µg m −3 in high-income tracts from increasing outdoor PM 2.5 by 1 µg m −3 .Extrapolating to outdoor PM 2.5 levels of 60 µg m −3 , a common value when wildfires occur, implies an increase in indoor PM 2.5 of 26 in lowincome and 15 in high-income tracts.Column (2) suggests a similar pattern of results when we specify income continuously, with a 10% increase in income resulting in an 11% decrease in indoor PM 2.5 .
In the second set of regression results in table 2, we focus on the smoke treatment intensity as defined in table 1 as the independent variable, with 'days with no smoke' as the reference group.The results from column (3) indicate that low-intensity wildfire smoke days increase indoor PM 2.5 by 21.4% and high intensity increases by 148% in low-income census tracts.The coefficients on the interaction term with the income variables are negative for both treatment groups, indicating the degree indoor penetration is smaller for high-income tracts.For example, the coefficient of −0.313 for above-median wildfire suggests indoor PM 2.5 increases by 117% for high-income tracts (as opposed to 148% for low-income).Similar patterns are found using income as a continuous variable, though the interaction terms are not statistically significant (column (4)).

Discussion
In this study, we find increases in outdoor pollution resulting from wildfires lead to sharp increases in indoor pollution, but disproportionately so in lowerincome census tracts.Indoor PM 2.5 levels increase by as much as 10 µg m −3 more in lower-income tracts.This suggests the defensive investments undertaken by families in higher-income tracts help to buffer the penetration of outdoor pollution into the indoor environment.Evidence indicates that stand-alone air purifiers and filters used in central air conditioning systems can reduce PM 2.5 levels by as much as 80-90%.Since these devices are disproportionately owned by higher-income households, this is a likely explanation behind the improved indoor air quality in higher-income areas.
Our results are concerning because wildfires are increasing in intensity, with burned acreage in the United States due to wildfires considerably increasing over the past 40 years [16], and Canada having experienced its largest wildfire season in history this summer [17].Higher temperatures and increased drought due to climate change are leading explanations behind these trends [18,19].Therefore, our results suggest a new implication from these increased wildfires and thus climate change: inequality in indoor PM 2.5 exposure.
In contrast to anthropogenic sources of PM 2.5 , where there are strong differences in outdoor PM 2.5 by income [12], we find minimal differences in wildfire-induced outdoor PM 2.5 by income.(Consistent with [12], we find larger income differences in PM 2.5 on days without wildfires.)This is likely due to the fact that anthropogenic sources are more likely to be located in lower-income neighborhoods [20], whereas wildfires occur sporadically and often in distant locations, thereby limiting the effectiveness of residential sorting that might give rise to differential outdoor pollution levels.Yet the decisions of households to protect themselves from this harm leads to vast differences in indoor PM 2.5 levels.
Given the duration of the fires and the wellestablished health effects that arise at the levels of PM 2.5 described here, these differences in indoor exposure likely contribute to health inequities.Although we do not directly measure health outcomes, studies indicate that indoor air filtration leads to improvements in cardiovascular and respiratory health [7].We find as much as a 10 µg m −3 differential in indoor PM 2.5 exposure by income, with the increased PM 2.5 levels from these wildfires lasting for several months.This difference in indoor PM 2.5 concentrations is comparable to the pollution changes in air purifier intervention studies and, thus, potentially large enough to induce health impacts.
In a closely related paper, Burke et al explore how exposures and behaviors in response to wildfires vary by income levels [21].They find that residents in higher-income areas are more likely to stay home, search for information on protective investments, and own indoor pollution monitors; these patterns are consistent with wildfires having larger impacts on lower-income individuals.In contrast to our study, however, they find that PM 2.5 infiltration only slightly varies with income levels.The most likely explanation for this difference is their use of a linear instead of a log-log model that we estimate.As discussed in the supplementary information, we provide several tests that support the log-log model as the more appropriate specification.Combined with the visual evidence in figure 2 of the differences in indoor PM 2.5 by income, we contend that income leads to important differences in indoor PM 2.5 levels.
There are several important limitations in our analysis.By relying on data from PurpleAir, our measures of indoor and outdoor pollution come from a select sample, both in terms of which census tracts and which households within a census tract are represented.Higher-income tracts are more likely to be included in our sample, thereby limiting the generalizability of our findings to lower-income tracts within the state.Table 1 supports this pattern: we observe PA monitors in 289 census tracts above the median income and only 55 below; if they were uniformly distributed, we would expect an equal amount above and below the median.Higher-income census tracts also contain a higher volume of monitors (384 indoor monitors in higher-income census tracts vs. 64 in lower-income census tracts), which increases the precision of pollution readings within a tract.
Knowing which households within a tract own monitors can introduce measurement error in our analysis since the income data is measured at the tract level.If, for example, the same pattern of monitor adoption across tracts holds within tracts-higher-income households within a tract are more likely to purchase monitors-this would result in non-classical measurement error that leads to an ambiguous bias in our estimates.Absent characteristics of those who purchased monitors, we are unable to ascertain the extent of this issue.Future studies that better characterize the owners of monitors would help address these limitations.
Another limitation is that we consider income as a proxy for defensive investments.Ideally, we would have direct measures of these investments, such as air filtration ownership and usage, and test their specific contribution.Absent direct measures, we instead interpret income as reflecting the ability to own them, contending that these devices are 'normal goods' more likely to be purchased by those with higher incomes.Consistent with this, higher income households are more likely to have central air conditioning [10] or any air conditioning [22], to purchase standalone air filters [11] (though we note this study is based in China), and to search for information on protective investments [21].The above evidence that higher-income families spend more time at home during large wildfires [21] is also consistent with the possibility that they are aware of the improvements in indoor PM 2.5 due to their defensive investments.
A concern that stems from both the lack of direct measurement of defensive investments and lack of representativeness of monitors is that households that own monitors may be those most likely to engage in defensive behaviors, which may also bias our estimates.Incorporating direct measurement of defensive investments, in addition to improved representativeness, is an important step for improving our understanding of the potential mechanism behind the patterns found in this analysis.
Despite these limitations, we believe our results point to an interesting finding that wildfires exacerbate inequalities in exposure to pollution, but through changes in indoor as opposed to outdoor pollution.Given the expected increase in wildfires due to climate change, this suggests another channel by which climate change will contribute to health inequality.Efforts to further quantify these effects, both by estimating impacts in other settings and by calculating the potential health impacts from differential exposure, including their monetized value, represent a fruitful area for future research.
If these results extend to high pollution events more generally, then this also broadens our understanding of factors that contribute to differences in pollution exposure and the resulting health impacts, highlighting the role of the indoor environment.Since California wildfires are particularly salient events that occur with increasing regularity and receive considerable media attention, however, avoidance behavior may be particularly large in this context.An important direction for future research is to better understand the impact of other high-pollution events (or wildfires in other locations) on disparities in indoor pollution levels.

Figure 1 .
Figure 1.Kernel densities of indoor and outdoor PM2.5 levels by wildfire status: no wildfire smoke (top left), below-median wildfire smoke (top right), above-median wildfire smoke (bottom).The y-axes show the densities.The x-axes show the level of PM2.5 in µg m −3 .Solid lines correspond to observations from indoor monitors, and dashed lines to observations from outdoor monitors.Only observations of PM2.5 below 100 µg m −3 are included.

Table 2 .
Coefficients and standard errors (in parentheses) of OLS regressions of log indoor PM2.5 on shown variables.All regressions include control variables for weather (second-order polynomials for air temperature, dew point temperature, wind speed, and precipitation, with separate terms for positive and negative air and dew point temperatures) and indicator variables for each month, day of the week, and census tract.Standard errors are clustered on census tracts.