Outside in: the relationship between indoor and outdoor particulate air quality during wildfire smoke events in western US cities

For this work, we used low-cost (<$300) monitors available in the public PurpleAir network. Sensor lists were downloaded from the PurpleAir API (https://api.purpleair.com/), and archived data was downloaded from the ThingSpeak API (thingspeak.com). The network consists of two main types of monitors: PA-II (or PA-II-SD, a version enabled to store data without a WiFi connection on an SD card), designed for outdoor or indoor use, and PA-I-Indoor, designed for indoor use only. The PA-II (and PA-II-SD) monitor contains two Plantower PMS5003 sensors, while the PA-I-Indoor monitor contains one Plantower PMS1003 sensor. The two Plantower sensor models are nearly identical but have slightly different laser wavelengths (650 ± 10 nm for PMS1003, and 680 ± 10 nm for PMS5003) and air flow pathways within the sensor (Kelly et al 2017, Sayahi et al 2019). Plantower sensors estimate particle mass concentrations by converting observations of particle number via light attenuation to particle mass concentrations. Plantower reports two mass concentration estimates using different conversion factors 'cf_1' and 'cf_atm' recommended for the factory environment and outdoor use, respectively, according to the Plantower manuals. Each PurpleAir monitor also measures temperature and relative humidity. These data are available to download for public monitors in the PurpleAir network. Monitors are labeled as located indoors or outdoors by the user during the set-up process. As of 13 April 2021, over 6000 indoor and 18 000 outdoor PurpleAir monitors with publicly available data had been deployed globally, according to the monitor list from the PurpleAir API (although not all of the monitors are currently active).

From the PurpleAir monitors available, we identified pairs of indoor and outdoor monitors by first locating the nearest outdoor monitor to every indoor monitor. If the nearest outdoor monitor was greater than 1000 m away, the indoor monitor was removed from our analysis. With this criteria, we identified 5051 monitor pairs (5051 unique indoor monitors and 3345 unique outdoor monitors), globally. A single outdoor monitor could be paired to multiple indoor monitors (i.e. it is the closest outdoor monitor to multiple indoor monitors). We downloaded all available 10 min average observations from 2017 to 2020 for these monitor pairs from the PurpleAir network. These monitors came online, and some went offline, at different points throughout this time period, thus most do not provide complete observations from 2017 to 2020. We focused our analysis in the contiguous western US (west of 100° W, 4466 monitor pairs) in 2020, due to greater monitor data availability and a high number of intense smoke events in several major western US cities. Figure S1 shows the count of monitors used in our study by data completeness for 2020. We note indoor environments with PurpleAir monitors may not be a representative sample of all indoor environments and may be biased towards homes of some demographic groups over others. We also note that the COVID-19 pandemic likely led to unique building occupancy and behavioral patterns in 2020 that may influence both indoor and outdoor PM_2.5 concentrations (e.g. Mousavi and Wu 2021). We discuss potential implications of these factors on our analysis in the results section.

We followed suggested sensor performance guidelines from the Plantower manuals to clean data from the PurpleAir sensors. First, we removed monitors with a reported 'cf_1' PM_2.5 concentration outside 0–500 µg m⁻³, the sensor effective range reported by the Plantower manual, which removed 0.07% of the original 154 277 421 10 min observations (indoor and outdoor observation total) from indoor-outdoor monitor pairs in the western US in 2020. We also removed data points reporting temperatures outside the range 14–140 °F and relative humidity outside the range 0%–99%, the reported sensor working ranges in the Plantower manual. This removed 0.36% and 0.6% of the 10 min observations, respectively. For the PA-II and PA-II-SD monitors, which contain two Plantower PMS5003 sensors (labeled 'A' and 'B'), we filtered observations for agreement between the two sensors following the reasonable agreement reported by the Plantower manual: for 'cf_1' PM_2.5 observations <100 µg m⁻³, we removed data where the absolute difference between A and B sensor-reported 'cf_1' PM_2.5 is >10 µg m⁻³. For 'cf_1' PM_2.5 observations >100 µg m⁻³, we removed data where the percent difference between A and B sensor-reported 'cf_1' PM_2.5 is >10%. This data cleaning step reduced the total 10 min PM_2.5 observations from the co-located monitor pairs by 2.66% of the original total. Overall, 3.67% of the original observations were removed by the data cleaning, leaving 148 621 084 10 min observations in total (but 121 483 358 unique observations, as outdoor monitors can be paired to multiple indoor monitors) in the western US in 2020.

Many previous studies have shown that PurpleAir monitors, and Plantower sensors in general, have high precision but low accuracy compared to federal reference method monitors, and thus should have a correction factor applied for analysis. Because the PurpleAir monitors rely on light-scattering, which is sensitive to particle size, composition, and hygroscopicity, PurpleAir performance and correction factors vary for different particle sources (Singer and Delp 2018, Tryner et al 2020b) and ambient conditions (Sayahi et al 2019, Malings et al 2020, Tryner et al 2020a, Barkjohn et al 2021). Several correction factors are available on the PurpleAir website. At the time of this writing, these include: 'US EPA' (developed for outdoor monitors across the US; Barkjohn et al 2021), 'AQandU' (winter in Salt Lake City Utah; https://aqandu.org/), 'LRAPA' (woodsmoke in Oregon; www.lrapa.org/), and 'WOODSMOKE' (woodsmoke in Australia; Robinson 2020). While there are several woodsmoke-specific correction factors (Delp and Singer 2020, Holder et al 2020, Mehadi et al 2020, Robinson 2020), we chose to adjust both indoor and outdoor data using the correction factor developed by Barkjohn et al (2021).

The Barkjohn et al (2021) correction model was developed by comparison of outdoor PA-II monitors to PM_2.5 monitors in the Environmental Protection Agency's (EPA) air quality system, which follow a federal reference method. The application of this model to indoor PurpleAir monitors is an extrapolation as the model has not yet been tested for PA-I-Indoor monitors, designed for indoor use, which contain a different Plantower sensor. We discuss this issue further in the limitations section. The Barkjohn et al (2021) correction factor is:

$$\begin{equation*}{\text{P}}{{\text{M}}_{2.5}} = {\text{ }}0.524{\mkern 1mu} \times {\mkern 1mu} {\text{P}}{{\text{A}}_{{\text{cf}}\_{\text{1}}}}{\mkern 1mu} - {\mkern 1mu} {\mkern 1mu} 0.0862{\mkern 1mu} \times {\mkern 1mu} {\text{RH }} + {\text{ }}5.75,\end{equation*}$$

where PA_{cf_1} is the average of the A and B sensor-reported PM_2.5 with the 'cf_1' conversion factor, and RH is relative humidity in percent. The correction leads to PM_2.5 estimates similar to those produced by smoke-specific correction models at high concentrations (Holder et al 2020, Barkjohn et al 2021). Finally, we calculated daily averages of the corrected PM_2.5 observations, removing times when less than 50% of the day (72 10 min averages) reported PM_2.5 measurements. This criteria removed 5.6% of the daily, co-located PM_2.5 observations. In our analysis, we did not attempt to remove indoor-generated PM_2.5 peaks from sources such as cooking as these sources contribute to the overall PM_2.5 exposure indoors. In this work, we are focused on quantifying PM_2.5 exposures indoors and outdoors during smoke-free and smoke-impacted periods, rather than attempting to estimate the fraction of outdoor-generated PM_2.5 that infiltrates indoors on smoke-impacted and smoke-free days.

We identified smoke-impacted time periods for each monitor by combining satellite-based smoke-plume estimates with the PurpleAir surface PM_2.5 observations. The National Oceanic and Atmospheric Administration (NOAA) Hazard Mapping System (HMS) produces a smoke plume product that identifies locations where there is smoke somewhere in the atmospheric column based on the inspection of visible satellite imagery by trained satellite analysts (Ruminski et al 2006, Rolph et al 2009). HMS smoke plumes may miss dilute smoke, cannot determine the vertical distribution of smoke, and are limited only to daylight hours (Rolph et al 2009, Brey et al 2018a). Because HMS smoke plumes cannot distinguish the vertical extent of smoke plumes, co-located PurpleAir indoor and outdoor monitor observations were labeled as smoke-impacted on days when both of the following criteria were met: (a) an HMS smoke plume is identified at the monitor location, and (b) the daily-mean outdoor PM_2.5 concentration is >1.5 × standard deviation + the monitor's 2020 annualmean outdoor PM_2.5 concentration on days when there is no HMS plume at the monitor location. We repeated our analysis using only HMS plumes to identify smoke-impacted observations and found that removing the PM_2.5 > 1.5 × standard deviation requirement does not qualitatively impact our main conclusions. Smoke-free observations were defined as all observations at monitors on days without an HMS smoke plume aloft. Because there are non-smoke related seasonal patterns in PM_2.5, we tested the sensitivity of our analysis to the inclusion of all 2020 smoke-free days versus only smoke-free days within the smoke season (June–November) in figures S2–S4. The inclusion of smoke-free days outside of the smoke season does not qualitatively impact conclusions, thus we decided to use all smoke-free observations in 2020 in our main analysis. With this method, we identified 33 965 smoke-impacted paired monitor-days and 234 877 smoke-free paired monitor-days across year 2020 observations within the western US. These observations come from 3175 monitor pairs, 1264 of which report observations for at least 10 smoke-impacted and 10 smoke-free days in 2020. In figure S1 we show the count of monitors by number of observed smoke-impacted and smoke-free days. The 2020 annual median smoke-impacted and smoke-free indoor and outdoor PM_2.5 concentrations are calculated for all monitor pairs with at least 10 observations of both smoke-impacted and smoke-free days.

We focus the remainder of our analysis on regions surrounding six western US cities with a sufficient number of indoor and outdoor monitor pairs and which experienced smoke-impacted air quality events in 2020. With these criteria, we selected areas around the following cities: San Francisco, CA; Los Angeles, CA; Salt Lake City, UT; Denver, CO; Seattle, WA; and Portland, OR. The areas were defined by selecting counties surrounding each city, listed in table S1, and monitors were included in the region if they were located within one of the surrounding counties determined by cross-referencing monitor latitudes and longitudes in the PurpleAir data with county shapefiles form the 2018 social vulnerability index (SVI) created by the Center for Disease Control (CDC)/Agency for Toxic Substances and Disease Registry (ATSDR, CDC/ATSDR 2018). Seattle and Portland were combined in our analysis due to their proximity and low monitor count. The areas, San Francisco, Los Angeles, Salt Lake City, Denver, and Seattle and Portland, contained 2322, 181, 59, 39, and 84 (Seattle and Portland combined) indoor and outdoor monitor pairs reporting in 2020, respectively. The count of monitors by data completeness for each region is shown in figure S1. Within each region, we identified the distribution of socioeconomic status represented in these co-located monitor pairs. However, we do not know whether the indoor monitors within each census tract are located in residential or commercial buildings. We counted the number of monitor pairs by census tract (cross-referencing monitor latitudes and longitudes provided in the PurpleAir data with the 2018 SVI census tract shapefiles) and identified the social vulnerability of each census tract with the 2018 SVI from the CDC/ATSDR (Flanagan et al 2011, 2018, CDC/ATSDR 2018). The CDC/ATSDR's SVI is a national percentile ranking of each census tract by social vulnerability across multiple economic and demographic indicators; higher values indicate higher vulnerability. Distributions of monitors by SVI were compared to the region's population distribution by SVI.

We investigated the relationship between indoor and outdoor PM_2.5 concentrations on smoke-impacted and smoke-free days across each region and at individual monitor pairs within each region. To evaluate the relationship across each region, we grouped all daily indoor and outdoor observation pairs within each region by smoke-impacted and smoke-free days. For the outdoor PM_2.5 distributions, we weighted outdoor observations by the number of indoor monitors assigned as its pair. We also grouped indoor PM_2.5 by outdoor PM_2.5 AQI. The PM_2.5 AQI bins, as currently defined by the EPA, are as follows, 'good': PM_2.5 < 12 µg m⁻³, 'moderate': 12 µg m⁻³ < PM_2.5 < 35 µg m⁻³, 'unhealthy for sensitive groups': 35 µg m⁻³ < PM_2.5 < 50 µg m⁻³, 'unhealthy': 50 µg m⁻³ < PM_2.5 < 150 µg m⁻³, 'very unhealthy': 150 µg m⁻³ < PM_2.5 < 250 µg m⁻³, and 'hazardous': 250 µg m⁻³ < PM_2.5. Grouping all monitors in each region allowed us to investigate potential regional differences in relative indoor and outdoor PM_2.5 concentrations across the different indoor environments within the region. We also evaluated the indoor and outdoor relationship at each individual monitor pair in the region by calculating the ratio of daily-average indoor PM_2.5 concentrations to daily-average outdoor PM_2.5 concentrations (indoor/outdoor ratio).

The 2020 annualmedian PM_2.5 concentrations on smoke-impacted and smoke-free days for co-located indoor and outdoor PurpleAir monitors in the western US are shown in figure 1. Following the data cleaning and smoke filtering described in sections 2.1 and 2.2 and excluding monitors with fewer than 10 smoke-impacted or 10 smoke-free days in 2020, there is a total of 1264 indoor-outdoor monitor pairs in the western US at which smoke-impacted and smoke-free annual medians are calculated. A majority of the co-located monitor pairs with sufficient observations are located in California with additional clusters of monitors in and around Denver, Salt Lake City, Seattle, and Portland. By definition of the smoke-impacted days (HMS smoke plume aloft and outdoor PM_2.5 greater than 1.5 standard deviations above the non-smoke impacted annual mean), the outdoor PM_2.5 on smoke-free days is lower than outdoor PM_2.5 on smoke-impacted days, on average (figures 1(a) and (b)). We also find the annualmedian indoor PM_2.5 is, on average, higher for smoke-impacted days compared to smoke-free days (figures 1(c) and (d)). Across all monitors, median indoor PM_2.5 is, on average, 82% (interquartile range, IQR: 43%–135%) or 4.3 µg m⁻³ (IQR: 2.0–7.2 µg m⁻³) higher on smoke-impacted days. Thus, smoke events degrade indoor air quality across a large number of indoor environments in the western US.

Zoom In Zoom Out Reset image size

Of the 1264 monitor pairs for which annual medians are calculated, annual median indoor PM_2.5 concentrations on smoke-impacted days (figure 1(d)) are lower than annual median outdoor PM_2.5 on smoke-impacted days (figure 1(b)) at 1235 monitor pairs. However, the annual medians of smoke-free days were lower for indoor monitors at only 789 monitor pairs of the 1264 monitor pairs (figures 1(a) and (c)). Thus, at 38% of the monitor pairs, median indoor PM_2.5 is greater than median outdoor PM_2.5 on smoke-free days, but fewer than 3% of the monitor pairs have a smoke-impacted annual median indoor PM_2.5 concentration that is greater than the smoke-impacted annual median outdoor PM_2.5 concentration. For smoke-impacted days, median indoor PM_2.5 across all monitor pairs is 58% (IQR: 40%–71%) lower than median outdoor PM_2.5 across the monitors. For smoke-free days, annual median indoor PM_2.5 across all monitors is only 7.9% (IQR: −8.6%–23%) lower than annual median outdoor PM_2.5. Finally, we find a slightly higher correlation between daily-average indoor and outdoor PM_2.5 on smoke-impacted days (Spearman's r of 0.66) compared to smoke-free days (0.63), see figure S5. However, at nearly half (45%) of the monitors, there is a higher correlation between indoor and outdoor PM_2.5 for smoke-free days than for smoke-impacted days.

In figure 2, we provide a map of co-located PurpleAir monitors in each region and the representation of the co-located monitor locations in terms of social vulnerability by census tract. Figures 2(a)–(d) show the monitors selected to represent each region of interest. A large number of monitor pairs shown in figure 2 were installed directly after smoke events occurred in the region, and hence not every point in figure 2 provides data during smoky time periods. A large increase in installed PurpleAir monitors after severe smoke events in California has been previously noted (Delp and Singer 2020, Krebs et al 2021, Liang et al 2021). Figure 2(f) shows the cumulative distribution of the population and number of co-located PurpleAir monitor pairs for each region and the full US by SVI (higher values indicate higher vulnerability). Across all regions, and the US at large, there is a higher number of co-located monitors in census tracts of lower social vulnerability compared to the population in those census tracts. For example, in San Francisco, about 40% of the population is located in census tracts with an SVI above 0.5 while only about 15% of the monitors are located in these census tracts. Thus, those of higher social vulnerability are underrepresented by these co-located PurpleAir monitors.

Zoom In Zoom Out Reset image size

Figures 3(a) and (b) show the indoor and outdoor distribution of daily-mean PM_2.5 concentrations for smoke-free observations and smoke-impacted observations, respectively, for each region. Indoor and outdoor concentrations in smoke-free and smoke-impacted conditions occasionally exceed 100 µg m⁻³, which may be due to indoor pollution events, highly localized outdoor sources, dense smoke plumes, or erroneous observation values caused by monitor malfunction that were not removed during data cleaning. The median and IQR for indoor PM_2.5 during smoke-free conditions across the regions are remarkably similar, where the median indoor PM_2.5 concentrations for San Francisco, Los Angeles, Seattle and Portland, Salt Lake City, and Denver are 4.83 µg m⁻³, 5.29 µg m⁻³, 4.52 µg m⁻³, 4.65 µg m⁻³, and 5.51 µg m⁻³, respectively. Outdoor concentrations show slightly more variability. Across the five regions, Los Angeles has the highest median daily-average outdoor PM_2.5 concentration across all area monitors (7.09 µg m⁻³), while the Seattle and Portland area has the lowest median concentration (4.07 µg m⁻³). Across all regions, daily indoor PM_2.5 is similar to outdoor PM_2.5 on smoke-free days. There is a notable difference for the Los Angeles area monitors, which show the largest shift between the distributions of indoor and outdoor PM_2.5 on smoke-free days (a 25% difference in the median PM_2.5), where the indoor PM_2.5 is lower. In contrast, there is a smaller shift in the distribution of indoor and outdoor PM_2.5 on smoke-free days in the Seattle and Portland region (11%) and the Denver region (7.4%), where the indoor PM_2.5 is slightly higher.

Zoom In Zoom Out Reset image size

On smoke-impacted days, there is a clear difference between the distributions of indoor and outdoor PM_2.5 in each region, where the outdoor PM_2.5 distribution is shifted towards higher concentrations than the indoor PM_2.5 concentrations. Median smoke-impacted indoor PM_2.5 concentrations in the five regions are 40%–71% lower than median smoke-impacted outdoor PM_2.5 concentrations. The distributions of smoke-impacted daily indoor PM_2.5 are again similar across the regions. There are regional differences in daily outdoor PM_2.5 on smoke-impacted days. Median daily outdoor PM_2.5 is highest in Los Angeles (24 µg m⁻³), followed by the Seattle and Portland region (23 µg m⁻³) and San Francisco (21 µg m⁻³). Although the Seattle and Portland region has one of the highest median smoke-impacted outdoor PM_2.5 concentrations, it has the lowest median smoke-impacted indoor PM_2.5 concentration amongst the regions. However, of the regions in figure 3, Seattle and Portland have the fewest PM_2.5 observations. In all five regions, indoor PM_2.5 is generally lower than outdoor PM_2.5 on smoke-impacted days.

Figures 3(a) and (b) show that both indoor and outdoor PM_2.5 concentrations are elevated on smoke-impacted days, compared to smoke free days in each region (although, outdoor PM_2.5 is required to increase on smoke-impacted days by our definition). Median indoor PM_2.5 concentrations are 50%–93% higher on smoke-impacted days than smoke-free days across the regions. The difference between the smoke-free and smoke-impacted distributions of dailymean indoor PM_2.5 concentrations are statistically different in all regions according to a two sample Kolmogorov–Smirnov test at a 95% confidence interval. However, we note that there may be some temporal autocorrelation in daily-average indoor PM_2.5 concentrations in our dataset. We do not test whether the outdoor smoke-impacted and smoke-free distributions are statistically different because outdoor PM_2.5 is required to increase on smoke-impacted days by definition and there is likely a high degree of spatial autocorrelation for areas like San Francisco with a high density of outdoor monitors sampling similar air masses.

Relative indoor and outdoor PM_2.5 concentrations during smoke events can vary significantly across different indoor environments (Henderson et al 2005, Kirk et al 2018, Shrestha et al 2019). We show the median ratio of daily-average indoor to outdoor PM_2.5 for smoke-impacted and smoke-free observations at each monitor in the regions in figure 4. Black dashed lines across the figure indicate where the smoke-impacted and smoke-free ratios are equal to 1. Smoke-impacted ratios for nearly all monitor pairs (1055 of the1068 monitor pairs across the regions for which individual indoor-outdoor ratios are calculated) in the five areas lie below the 1 line, indicating that for these indoor environments, indoor PM_2.5 is, in general, lower than outdoor PM_2.5 on smoke-impacted days. Monitor pairs are more evenly distributed about the 1 line for smoke-free observations, where 61% of monitor pairs have a smoke-free indoor/outdoor ratio <1. In figures S6–S10, we show two-dimensional histograms of hourly indoor PM_2.5 versus hourly outdoor PM_2.5 for smoke-impacted and smoke-free observations in each city. The figures, in agreement with figures 3 and 4, show that on smoke-impacted days, indoor PM_2.5 concentrations are predominantly lower than outdoor PM_2.5, while on smoke-free days, the indoor and outdoor PM_2.5 concentrations are often of a similar magnitude.

Zoom In Zoom Out Reset image size

The indoor/outdoor ratios on smoke-free days show some differences between regions. On smoke-free days, a majority of monitor pairs in San Francisco, Los Angeles, and Salt Lake City (60%, 81%, and 68%, respectively) are below the 1 line. In contrast, there are more monitor pairs above the 1 line on smoke-free days in the Seattle and Portland region (87%) and an even distribution in the Denver region (50%). Although there are only a few monitor pairs in Seattle and Portland, they all report relatively low indoor to outdoor ratios on smoke-impacted days (<0.51). Regional differences may be driven by differences in outdoor, rather than indoor PM_2.5 concentrations in these areas, as indicated by the higher intra-regional variability in outdoor PM_2.5 compared to indoor PM_2.5 in figure 3.

The regional differences in indoor/outdoor PM_2.5 ratios suggest there may be different PM_2.5 infiltration rates by region, possibly driven by regional differences in climate, window opening, home age, building type, and/or air conditioning (AC) in addition to differences in outdoor PM_2.5. In particular, the presence of AC and building age have been shown to impact PM_2.5 infiltration on smoke-impacted days in residential buildings with PurpleAir monitors in San Francisco (Liang et al 2021). The 2019 American Housing Survey (AHS) provides some insight into regional differences in these housing unit characteristics. The AHS is a housing characteristics survey of a weighted, representative sample of approximately 3000 housing units in different metropolitan areas around the US (US Census Bureau 2019). We report the fraction of AC use and fraction of units built before 2000 estimated by the 2019 AHS for the metropolitan areas included in our study alongside our regional indoor/outdoor PM_2.5 ratios in table S2. We note AHS geographical definitions of these areas differ from the definitions used here and Salt Lake City was not surveyed in the 2019 AHS. Los Angeles, the area with the highest fraction of units with AC (mean of Los Angeles and Riverside metropolitan area AC use), has the lowest indoor/outdoor PM_2.5 ratio on smoke-free days compared to the areas with a lower fraction of units with AC. On smoke-impacted days in our study, we find regional ratios somewhat follow the opposite, unexpected pattern where regions that have a higher fraction of units with AC have higher indoor/outdoor PM_2.5 ratios. Our regional indoor/outdoor ratios on smoke-free or smoke-impacted days do not follow regional home age patterns (see table S2). We note that although the AHS is a representative sample of these metropolitan areas, paired indoor and outdoor monitors in the PurpleAir network are not a representative sample, as was shown in figure 2. Thus, without knowledge of occupant behavior and details on the individual indoor environments in which these monitors are located, both of which impact indoor PM_2.5 concentrations (e.g. Liang et al 2021), it is challenging to distinguish regional differences. Distinguishing differences is especially challenging in locations with few monitor pairs, like Seattle, Portland, and Denver. Although it is difficult to establish regional differences in indoor/outdoor ratios, a large majority of indoor environments in all these regions have an average indoor/outdoor ratio <1 on smoke-impacted days.

In the two cities with the largest numbers of co-located PurpleAir monitors, San Francisco and Los Angeles, we evaluated indoor and outdoor PM_2.5, as well as indoor/outdoor PM_2.5 ratios, as a function of SVI on smoke-impacted and smoke-free days in figures S11–14. In San Francisco, we find generally higher outdoor PM_2.5 levels on smoke-free days for the few monitors in high SVI census tracts (figure S12) in agreement with previous works on outdoor particulate air quality and socioeconomic status (Hajat et al 2015). However, small sample sizes in the higher vulnerability census tracts make distinguishing any additional patterns difficult, especially for indoor PM_2.5 and indoor/outdoor PM_2.5 when the type of building (i.e. residential vs. commercial) is unclear. According to the 2019 National AHS, lower-income households live in smaller, older homes (US Census Bureau 2019), which are typically more leaky (Chan et al 2005), suggesting indoor exposure to outdoor air pollution may be higher in more socioeconomically vulnerable communities. However, there are additional factors that may also follow SVI patterns and impact the relationship between indoor and outdoor air quality including building type (i.e. apartment vs. stand-alone structure), AC prevalence, number of occupants, etc. It is not currently possible to establish a relationship between SVI and indoor/outdoor PM_2.5 during smoke-free or smoke-impacted periods with the PurpleAir monitor network because of the lack of monitoring in high SVI areas, which is an important limitation of the present work and the PurpleAir monitor network in general.

We explore the relationship between indoor PM_2.5 concentrations and indoor/outdoor ratios as a function of outdoor PM_2.5 for summer and fall smoke-impacted observations in figure 5. Figure 5 panels (a)–(e) show daily-average indoor PM_2.5 concentrations binned and colored according to the corresponding outdoor PM_2.5 levels associated with the EPA's AQI levels. The EPA's AQI is widely used across the US for public communication on air quality and mitigation strategies during pollution episodes, including recommendations to remain indoors. PurpleAir monitors have been found to show a linear response to PM_2.5 up to around 200 µg m⁻³, and thus are likely reliable for PM_2.5 AQI levels up to the 150–250 µg m⁻³ bin (Holder et al 2020, Mehadi et al 2020). In San Francisco, there were 225 smoke-impacted monitor days with daily-average outdoor PM_2.5 concentrations associated with the 'very unhealthy' (150 µg m⁻³ < PM_2.5 < 250 µg m⁻³) AQI. These occurred at 218 unique monitor pairs on 10–11 September and 1–2 October 2020. In the Seattle and Portland area, there were 13 monitor days with daily-average PM_2.5 concentrations associated with the 'very unhealthy' AQI, occurring 11–17 September 2020 at 7 unique monitor pairs. Across most regions, the majority of the smoke-impacted outdoor PM_2.5 concentrations are in the 12–35 µg m⁻³ range.

Zoom In Zoom Out Reset image size

The daily mean of total indoor PM_2.5 concentrations, plotted on the y-axis in figure 5 panels (a)–(e), generally increase as the outdoor PM_2.5 concentrations increase across each AQI bin in each region. The only exception is for the <12, 12–35, and 35–55 µg m⁻³ outdoor PM_2.5 bins in Seattle and Portland (figure 5(e)), where there is no distinct change in the indoor PM_2.5 distributions. On average, median indoor PM_2.5 increases by 48% per AQI bin across the regions. Because AQI bins are nonlinear, we also calculate the linear relationship using an ordinary least squares regression between indoor and outdoor PM_2.5 concentrations for each indoor-outdoor monitor pair in the western US with at least 10 smoke-impacted days and 10 smoke-free days. We found indoor PM_2.5 concentrations increase by a median of 1.5 µg m⁻³ (IQR: 0.6–3.4 µg m⁻³) for every 10 µg m⁻³ increase in outdoor PM_2.5. Of the 1264 indoor-outdoor monitor pairs for which we calculated smoke-impacted indoor versus outdoor PM_2.5 slopes, 76% are statistically significant at a 95% confidence interval using a 2-sided Wald test with a null hypothesis of a slope equal to zero. The statistically significant slopes are slightly higher on average with a median of a 2.1 µg m⁻³ (IQR: 1.0–4.3 µg m⁻³) increase in indoor PM_2.5 for every 10 µg m⁻³ increase in outdoor PM_2.5. We find 77 indoor environments (6% of indoor environments for which slopes were calculated) had negative slopes, indicating indoor PM_2.5 decreased as outdoor PM_2.5 increased. However, only three of these slopes were statistically significant.

The daily-average indoor PM_2.5 concentrations rarely reach the same AQI level as the outdoor PM_2.5, especially for outdoor PM_2.5 AQI levels above the 12–35 µg m⁻³ bin. Above this 'moderate' PM_2.5 level, only 7% of indoor PM_2.5 concentrations are at or above the outdoor PM_2.5 AQI level. However, the daily-average indoor PM_2.5 concentrations are frequently elevated above the 'healthy' (<12 µg m⁻³) AQI level on smoke-impacted days. In figure S15, we show the cumulative distributions of indoor PM_2.5 concentrations at paired monitors on all smoke-impacted days and heavily smoke-impacted days (outdoor PM_2.5 > 55 µg m⁻³). The figure shows that 32% of observed daily-average indoor concentrations on all smoke-impacted days exceed the 'healthy' AQI level. Further, 3% of observed daily-average indoor concentrations on all smoke days and 19% of observed daily-average indoor concentrations on heavily smoke-impacted days exceed the EPA's current 24 h outdoor PM_2.5 standard of 35 µg m⁻³.

Figure 5 panels (f)–(j) show the ratio of daily-average indoor to outdoor PM_2.5 on smoke-impacted days, binned by outdoor PM_2.5 AQI. Although the indoor PM_2.5 increases across each AQI bin, the ratio of indoor to outdoor PM_2.5 generally decreases. On average, the median indoor/outdoor ratio across the regions decreases by 29% per outdoor AQI bin. As shown in figures 5(f)–(j), the absolute decrease in the indoor/outdoor ratio is larger for the lower AQI bins, compared to the higher AQI bins in Los Angeles and San Francisco. In the Seattle and Portland (up to the 150–250 µg m⁻³ bin), Salt Lake City, and Denver regions the decrease across AQI bins appears more linear on the log scale. The distribution of indoor/outdoor ratios in 150–250 µg m⁻³ outdoor PM_2.5 bin in Seattle and Portland deviates from this trend and is much wider than the other distributions. However, we note it is one of the smallest bins plotted in figure 5, representing only 13 observations. It is thus unclear if this is a true deviation from the observed trend or simply an artifact of the small sample size. For all AQI bins above the <12 µg m⁻³ bin, the indoor/outdoor ratio is <1 for 97% of the observations. Thus, although the indoor PM_2.5 increases as outdoor PM_2.5 increases on smoke-impacted days, the absolute increase in the indoor PM_2.5 is smaller than the increase in outdoor PM_2.5. Further, the relatively smaller increase in indoor PM_2.5 does not typically result in indoor PM_2.5 concentrations at the same AQI level as the outdoor concentrations but can result in indoor PM_2.5 concentrations above the 'healthy' AQI level.

Our analysis on the influence of smoke on indoor air quality is limited by a lack of detail on indoor environment characteristics where monitors are placed and occupant behavior. There are multiple potential confounding variables that we are unable to properly control for when using a large set of monitors from the PurpleAir network. These confounding variables include characteristics of individual indoor environments such as building age, sources of heating and cooling, stove type, level of air filtration, location of monitor, among other factors, which have been shown to impact indoor air quality (Allen et al 2003, Shrestha et al 2019). In addition, we do not have a record of occupant behavior such as cooking times, opening windows, etc, which are also known to impact indoor air quality (e.g. Farmer et al 2019, Patel et al 2020). Several recent studies using PurpleAir monitors have been able to identify some of these building characteristics including distinguishing residential and commercial buildings, AC presence, and home age through surveys or by matching monitor locations to residences on Zillow (Liang et al 2021, May et al 2021). As mentioned previously, some of these factors, like AC presence, may follow regional and socioeconomic patterns. Due to the large number of monitors in our analysis and the limitations of accessible datasets on additional factors like home value and AC presence, we choose to not attempt to control for these variables in our study.

In addition, due to the COVID-19 pandemic, indoor behavior in 2020 may differ from a typical year. The stay-at-home orders and increased remote work likely led to higher-than-normal occupancy of residential spaces and lower-than-normal occupancy of non-residential spaces. There were observed 17%–24% increases in indoor PM_2.5 during the 2020 COVID-19 lockdowns at PurpleAir monitors in California, compared to 2019 levels (Mousavi and Wu 2021). However, these values were found to return to normal levels post lockdown (Mousavi and Wu 2021). Because we are unaware of the type of indoor environment (e.g. office space, residential building, etc) and the abnormal work and life circumstances created by the COVID-19 pandemic, it is challenging to control for these confounding variables through other indicators such as weekend/weekday variables. Despite this lack of detail on the indoor environment and unique circumstances created by the COVID-19 pandemic, we find a consistent impact of smoke on indoor air quality.

There are likely biases present in the indoor environments monitored in the PurpleAir network, which may impact this work. Individuals who purchase PurpleAir monitors for their personal residences are likely invested in the air quality in their home and may be more likely to take extra steps to improve their indoor air quality compared to the general public. Community groups, government agencies, and scientists also purchase and deploy PurpleAir monitors. Placement of these monitors may also add a bias depending on project goals. In addition, as shown in figure 2, census-tracts of higher social vulnerability are not well represented in the PurpleAir indoor and outdoor monitor pairs used here. Similarly, Liang et al (2021) show PurpleAir monitors in California represent residences with an estimated property value 21% greater than the local average. Previous work has shown communities of lower socioeconomic status are often subject to higher levels of outdoor air pollution (Hajat et al 2015). However, the relationship between socioeconomic status and indoor air quality, particularly smoke-impacted air quality, remains unclear.

There are additional challenges with measuring PM_2.5 concentrations with the PurpleAir network and low-cost sensors. PurpleAir monitors are often purchased and installed by citizens at their homes. This may impact data quality as the monitors are not calibrated in the local environment and may not be located in ideal sampling locations. Data may not completely span the day as the PA-II and PA-I Indoor monitors transmit data through a wireless internet connection, which may be unstable. Because PurpleAir monitors rely on light-scattering, reported PM_2.5 mass concentrations are sensitive to particle size, composition, and hygroscopicity. Therefore, PurpleAir performance can change by aerosol source type (Singer and Delp 2018, Tryner et al 2020b) and the ambient environment (Sayahi et al 2019, Tryner et al 2020a), which can differ indoors and outdoors. Mehadi et al (2020) found PurpleAir performance is also sensitive to woodsmoke composition. The Barkjohn et al (2021) correction factor used here was developed for all-source outdoor PM_2.5 observations across the US and agrees with woodsmoke correction factors at high concentrations (Holder et al 2020). There has been less validation of sensors in an indoor environment, but Delp and Singer (2020) found a factor of ∼2 overprediction for an indoor sensor during smoke events, consistent with the typical offset found for outdoor monitors. We repeated our analysis using the indoor-developed scaling from Delp and Singer (2020) for indoor monitors and with the Lane Regional Air Protection Agency (LRAPA) correction factor (www.lrapa.org/), developed in an area impacted by woodsmoke from home heating in winter, and found lower indoor concentrations by AQI bin and a smaller decrease in the indoor/outdoor ratios with increasing outdoor AQI bins. These correction factors also resulted in lower indoor/outdoor ratios on smoke-free days. However, our main conclusions do not change. At present, these limitations are inherent with any use of the indoor PM_2.5 observations from PurpleAir.