Using low-cost air quality sensors to estimate wildfire smoke infiltration into childcare facilities in British Columbia, Canada

The health risks associated with wildfires are expected to increase due to climate change. Children are susceptible to wildfire smoke, but little is known about indoor smoke exposure at childcare facilities. The objective of this analysis was to estimate the effects of outdoor PM2.5 and wildfire smoke episodes on indoor PM2.5 at childcare facilities across British Columbia, Canada. We installed low-cost air-quality sensors inside and outside 45 childcare facilities and focused our analysis on operational hours (Monday–Friday, 08:00–18:00) during the 2022 wildfire season (01 August–31 October). Using random-slope random-intercept linear mixed effects regression, we estimated the overall and facility-specific effects of outdoor PM2.5 on indoor PM2.5, while accounting for covariates. We examined how wildfire smoke affected this relationship by separately analyzing days with and without wildfire smoke. Average indoor PM2.5 increased by 235% on wildfire days across facilities. There was a positive relationship between outdoor and indoor PM2.5 that was not strongly influenced by linear adjustment for meteorological and area-based socio-economic factors. A 1.0 μg m−3 increase in outdoor PM2.5 was associated with a 0.55 μg m−3 [95% CI: 0.47, 0.63] increase indoors on non-wildfire smoke days and 0.51 μg m−3 [95% CI: 0.44, 0.58] on wildfire-smoke days. Facility-specific regression coefficients of the effect of outdoor PM2.5 on indoor PM2.5 was variable between facilities on wildfire (0.18–0.79 μg m−3) and non-wildfire days (0.11–1.03 μg m−3). Indoor PM2.5 responded almost immediately to increased outdoor PM2.5 concentrations. Across facilities, 89% and 93% of the total PM2.5 infiltration over 60 min occurred within the first 10 min following an increase in outdoor PM2.5 on non-wildfire and wildfire days, respectively. We found that indoor PM2.5 in childcare facilities increased with outdoor PM2.5. This effect varied between facilities and between wildfire-smoke and non-wildfire smoke days. These findings highlight the importance of air quality monitoring at childcare facilities for informed decision-making.


Introduction
1.1.Health effects of wildfire smoke PM 2. 5 The health risks associated with extreme events like naturally occurring and human-caused wildfires are expected to increase because of climate change [1][2][3].Exposure to wildfire smoke is known to have important impacts on human health and is predicted to increase in coming decades as climate change drives conditions towards more frequent and extreme wildfires.Wildfire smoke is a complex mixture of gases and airborne particulate matter [4].Particles less than 2.5 micrometers in aerodynamic diameter (PM 2.5 ) are considered the most health-relevant constituents of this mixture for two key reasons.First, when inhaled, PM 2.5 can penetrate deep into the lungs and cause a variety of adverse systemic health effects.Short-term exposure to wildfire smoke has been associated with a wide range of acute outcomes, from exacerbation of respiratory symptoms to increases in all-cause mortality [5,6].There is also growing evidence that the longer-term effects of wildfire smoke exposure are consistent with those of PM 2.5 from other sources including increased population mortality and respiratory morbidity [7].Second, PM 2.5 can be transported long distances from wildfires.As a result, exposure is often so widespread that millions of people can be affected [4,8].
Children are particularly susceptible to experiencing adverse health effects from exposure to wildfire smoke.Exposure among children is associated with increased rates of acute respiratory outcomes, impaired lung development, and other potential long-term health consequences [9][10][11][12].Increased susceptibility in children may be driven by their higher baseline respiratory rate, their higher metabolic rate and oxygen demand, their higher activity levels, and their developing respiratory system [11].Young children are also vulnerable because they lack the ability to take autonomous actions to lower their exposure to air pollution.Instead, caregivers are in the position to initiate health protective measures including keeping children indoors, closing windows, and using air cleaners.
1.2.Indoor exposure to wildfire smoke PM 2. 5 Most of the epidemiologic evidence describing the health effects of wildfire PM 2.5 is based on exposures measured outdoors [6,8,12,13].However, data from North America indicate that people spend the majority (85%-90%) of their time inside [13][14][15].A growing number of studies have found that outdoor PM 2.5 can infiltrate easily into some indoor environments [16][17][18], but there is limited research describing how people are exposed to PM 2.5 in different indoor environments during wildfire smoke events [19,20].Despite this, many local, national, and international public health agencies recommend that people stay indoors during wildfire smoke events [21,22], highlighting the need to better characterize indoor exposure during wildfire smoke events.
Estimates of wildfire smoke PM 2.5 infiltration into the indoor environment indicate that there is large between-building heterogeneity [23,24].This wide variability in infiltration estimates is explained by differences in multiple factors related to environmental conditions, human behaviors, and building characteristics.For instance, the amount of wildfire smoke that penetrates indoors is determined by levels of outdoor PM 2.5 , indoor and outdoor temperatures, occupant actions such as opening or closing windows, and the tightness of the building envelope [16,19,24].For example, May et al [17] found that infiltration of wildfire smoke PM 2.5 was higher for schools than for residential buildings, with more variability between residential locations.This previous work underscores that wildfire smoke infiltration needs to be measured for individual buildings to develop adequate interventions to prevent exposure because of high between-building infiltration variability.
Recent increases in the availability of crowdsourced data from widely deployed low-cost air-quality sensors has enabled researchers to begin measuring wildfire smoke infiltration into thousands of locations by pairing indoor and outdoor sensors located near one-another.However, this approach comes with inherent uncertainty in quantifying the association between indoor and outdoor PM 2.5 for a given location, because sensors are paired based on spatial proximity and are typically not physically located at the same building.For example, although O'Dell et al [25] was able to estimate infiltration into more than 1200 locations in the western USA using crowdsourced data from PurpleAir, paired sensors could be located up to 1 km apart.Instead, co-location of sensors inside and outside of the same building would provide a more reliable and accurate approach for estimating infiltration of outdoor PM 2.5 into individual buildings.
It is particularly important to understand the indoor air quality impacts of wildfire smoke at childcare facilities because many children spend a significant amount of time in such locations.Data from the Survey on Early Learning and Child Care Arrangements in Canada reported that 52% of children younger than 6 years were enrolled in childcare in 2022, with 31% of children in childcare facilities [26].In the US, a 2019 survey found that children in care programs spent 31-47 h per week in those facilities [27].Moreover, children may spend most of their time indoors at childcare facilities.For example, a study using personal GPS monitoring among preschool children (n = 46) attending child daycare facilities in Seattle, WA showed that children spent from 63% to 88% of their time indoors [28].

Objectives
In this project, we used a community science approach to enable childcare facilities across British Columbia (BC), Canada to install pairs of indoor and outdoor low-cost air-quality sensors.Our project aims were three-fold: (1) to provide facilities with air quality sensors that they could use indefinitely to monitor the air quality in and around their site, (2) to estimate the relationship between indoor and outdoor PM 2.5 as a proxy of PM 2.5 infiltration, and (3) to assess how the relationship between indoor and outdoor PM 2.5 varied between wildfire and non-wildfire days and between facilities.We explore how our findings can help childcare facility managers make informed decisions to protect the health of their children and staff.Our project also builds from our previous work that estimated PM 2.5 infiltration at a health care facility [23], by now measuring smoke infiltration into multiple childcare facilities.Furthermore, we enhance the accuracy and reliability of our infiltration estimates compared to the wider literature by physically locating air quality sensors inside and outside of the same building.

Study context
The province of BC is located on the west coast of Canada and spans almost one million km 2 .Approximately two-thirds of BC is forested [29] and there are wildfires every summer with high between-year variability.Climate projections for BC indicate drier, hotter, windier, and longer summers in future.Combined, these factors will lead to an increasing risk of lengthier and more intense wildfire seasons [30] and subsequent smoke exposures.
In this project, data were analyzed from 01 August to 31 October 2022, which was the peak of the 2022 wildfire season in BC.In 2022 there were 1801 wildfires in the province which burned a total of 135 235 ha [31].While this is lower than the 10 year average (1483 wildfires; 407 373 ha), the 2022 wildfire season and associated smoke impacts were unique because they started later than usual and extended further than normal into the autumn season due to a prolonged drought [32].Wildfires and associated smoke-impacts were widespread in BC during the study period (figure 1).
The air intake openings on the protective housing of the project EGGs were enlarged by the manufacturer at the request of project collaborators (Health Canada).The enlargements were intended to reduce particle buildup on the sensors by decreasing air velocity, which had been identified by collaborators as a concern (personal communications).We tested each indoor-outdoor sensor pair prior to field deployment to ensure that sensors measured PM 2.5 accurately and reliably.Details on the sensor housing modification, testing protocols, and testing results are given in the supplemental material, section 1.

Sensor correction and indoor peak removal
Prior to analyses, we corrected all EGG PM 2.5 data (1 h averages) using regression equations trained on the association between outdoor PurpleAir sensors co-located at 25 federal equivalent method PM 2.5 monitoring stations [34].This equation was developed to correct sensors on days with conditions ranging from typical ambient air pollution to heavily impacted by wildfire smoke.The PurpleAir correction equation was deemed appropriate because PurpleAir sensors and EGGs use the same Plantower PMS5003 sensors for measurement of PM 2.5 .Following corrections, all values less than 1.0 were set at 1.0 µg m −3 and values greater than 500 µg m −3 were excluded because they are outside the effective range of the sensor [35].
After correction, we removed indoor peaks in PM 2.5 due to non-ambient indoor sources (e.g.cooking, humidifiers, barbeques) using a previously published method [23].We assumed that indoor peaks were comprised of sudden large increases in indoor PM 2.5 concentrations, followed by a decline to pre-peak levels once the indoor source diminishes.To do this, we identified the start of indoor peaks when (1) indoor PM 2.5 was at least twice as high as outdoor PM 2.5 (i.e.indoor/outdoor I/O PM 2.5 ratio ⩾ 2) and (2) concentrations continued to increase in the subsequent minutes.Peaks were considered finished when the I/O PM 2.5 ratio returned to parity (i.e.I/O PM 2.5 ratio ⩽ 1).This definition of indoor peaks avoids removing data when indoor PM 2.5 concentrations from wildfire smoke infiltration remain high even while outdoor PM 2.5 levels drop (e.g.due to wind and rain), which may contribute to I/O PM 2.5 ratios exceeding 1.0.All EGG PM 2.5 data reported and analyzed in this manuscript refer to corrected values with indoor peaks removed, unless otherwise specified.

Recruitment and sensor deployment
Working with regional partners, we contacted all registered child daycare facilities (henceforth referred to as childcare facilities) in Interior Health Authority, Northern Health Authority, and Vancouver Coastal Health Authority (figure 1).In all cases a letter was sent to the facilities describing the project and inviting participation.We communicated with interested facilities to assess whether their sites met the technological requirements of the EGGs, such as compatible Wi-Fi and access to an outdoor power outlet.We sent each participating facility one indoor and one outdoor EGG along with an information packet describing how and where to set up the sensors and with tips for interpreting EGG data.Facilities were instructed to set up their outdoor EGGs in a location protected from precipitation and to place both the indoor and outdoor EGGs away from local sources of air pollution and direct sunlight.The project team helped facilities install EGGs when the site managers requested aid.This was done via email, phone, and in person as needed.
Uncorrected sensor data were available to facilities on the LED screen built into each EGG.Outdoor corrected data were made available on the University of Northern British Columbia air quality map [36].Indoor and outdoor corrected data were made available on a private Rshiny [37] dashboard developed for the facilities as a part of the project.

Data analysis 2.5.1. Wildfire days
All days in the study period were classified as wildfire days or non-wildfire days at each facility using methods adapted from O'Dell et al [25].We defined wildfire days using plume coverage data from the US National Oceanic and Atmospheric Administration's Hazard Mapping System (HMS) and PM 2.5 data from the outdoor EGGs.First, we linked HMS data spatially to the location of each childcare facility to assign each day as impacted by a smoke plume or not.Next, we calculated the facility-specific mean 24 h outdoor PM 2.5 concentration and standard deviation (SD) on all days not impacted by smoke plumes during the study period.Finally, we defined the facility-specific wildfire PM 2.5 threshold as the mean value from the previous step plus 1.5 * SD.Days were then defined as wildfire days at each facility if the 24 h outdoor PM 2.5 concentration at the facility was greater than the wildfire PM 2.5 threshold.Days were additionally defined as wildfire days if they were within 2 d before or after another wildfire day (based on criteria above) and if the 24 h outdoor PM 2.5 concentration was within 15% of the wildfire PM 2.5 threshold.We added this second criterion to account for days between wildfire days when the outdoor PM 2.5 concentration was still elevated but just below the wildfire day threshold.

Daily operational hour averages and exclusion criteria
Data analysis was performed on average PM 2.5 concentrations during childcare operational hours, from 08:00 to 18:00 on Monday-Friday.To calculate daily means, we computed the average of 1 h EGG data for PM 2.5 (corrected), CO 2 , temperature, humidity, and pressure.Daily means were excluded if more than 40% of the 1-minute data were missing for that day during operational hours [38].Facilities were then excluded if more than 40% of study days had missing data or if they experienced no wildfire days.We focused our main analysis on operating hours to estimate indoor PM 2.5 when children were present onsite and to mitigate the influence of varying occupancy patterns or changes in ventilation during non-operating hours [24].16) situational vulnerability (quintiles 1-5).Facilities were classified as residential or commercial by examining images of the facility on Google Street View (Google, California).Residential facilities included single and multi-family homes and apartments, while commercial facilities included schools and other non-residential properties.When the distinction was not clear or there were no pictures available, we used the zoning designation for the address.The four dimensions of the CIMD are indicators of area-level socio-economic status [39].Each dimension is calculated based on different variables in the 2016 Canadian Census (e.g.proportion of the population that is low income, proportion of the population participating in the labor force) and is reported as quintiles where 1 is the least deprived and 5 is the most deprived.CIMD data were linked to each facility by the dissemination area.Dissemination areas are the smallest standard geographic unit for which all census data are disseminated and had an average population of 400-700 persons across Canada in 2016.

Indoor and outdoor PM 2.5
The aim of this project was to estimate the relationship between indoor and outdoor PM 2.5 at each facility as a proxy for daily PM 2.5 infiltration, while accounting for potential confounders.To do this, we used linear mixed effects regression (LMER) with indoor PM 2.5 as the outcome and outdoor PM 2.5 as the explanatory variable.To allow the relationship between indoor and outdoor PM 2.5 to vary by facility, we included a random slope and intercept for outdoor PM 2.5 within each facility (equation ( 1)): PM 2.5 indoors = PM 2.5 outdoors + Random slope and intercept (PM 2.5 outdoors |Facility) . ( We then added each potential confounder to the model and compared the unadjusted (equation ( 1)) and adjusted (equation ( 2)) effect estimate of outdoor PM 2.5 to evaluate whether the potential confounder changed the effect: PM 2.5indoors = PM 2.5 outdoors + Confounder k + Random slope and intercept (PM 2.5 outdoors |Facility) .(2) Variables were considered confounders and were retained in the final LMER if they changed the effect estimate of outdoor PM 2.5 by ⩾10% and were independently associated with indoor PM 2.5 (likelihood ratio test p-value ⩽ 0.1) in a bivariable LMER containing only that variable and a random slope for facility.
To assess the effect of wildfire days on the relationship between indoor and outdoor PM 2.5 , we refit the LMER with the data stratified by non-wildfire and wildfire days.All analyses were conducted in R (version 4.2.0) and RStudio (version 2022.2.3.492)[40,41].

Follow-up analyses: timing of infiltration
We performed follow-up analyses to explore the timing of PM 2.5 infiltration.First, we assessed whether infiltration was different between operational and non-operational hours by rerunning the models specified above, but for daily non-operational hours (averaged between 00:00-07:59 and 18:01-23:59 on the same calendar date as the operational hours average).Second, we fit LMER models (random intercepts for facility) to the uncorrected minute-level PM 2.5 data to estimate the indoor PM 2.5 concentration as a function of the cumulative lagged effect of outdoor PM 2.5 , using methods defined by Nguyen et al [23].This approach models the indoor PM 2.5 concentration at the current minute (t) as the proportion of outdoor PM 2.5 that infiltrated indoors and remained suspended for a specified number of minutes (equation (3)): PM 2.5 indoors t = β 1 PM 2.5 outdoors t−1 + β 2 PM 2.5 outdoors t−2 + . . .+ β 120 PM 2.5 outdoors t−120 + intercept (3) where each β coefficient represents the proportion of outdoor PM 2.5 that infiltrated indoors for the listed minute.We ran separate lag models by increments of 10 min intervals up to 120 min (e.g. 10 min, 20 min, …, 120 min).For each model, we calculated total infiltration over the specified interval as the sum of the lagged coefficients (equation ( 4)): Finally, we compared the lagged infiltration coefficients across the models to describe how the lagged relationship of outdoor PM 2.5 on indoor PM 2.5 varied across the different periods of time.
Indoor and outdoor PM 2.5 concentrations across all facilities increased significantly (Mood Median Test MMT p-value < 0.001) during wildfire days compared with non-wildfire days (table 1; figure 3).The median (IQR) indoor PM 2.5 concentration increased by 235% from 4.3 µg m −3 (2.9-6.7)during non-wildfire days to 14.4 µg m −3 (9.8-20.6)during wildfire days.The median outdoor PM 2.5 concentration increased by 229% from 6.3 µg m −3 (3.7-10.5) to 20.7 µg m −3 (16.8-31.2).Average indoor PM 2.5 was lower than outdoor PM 2.5 , but the I/O PM 2.5 ratio was lower during wildfire days than during non-wildfire days (MMT p-value < 0.001).The median (IQR) I/O PM 2.5 ratio was 0.76 (0.56-0.97) on non-wildfire days and was 0.64 (0.49-0.83) on wildfire days.Additionally, on both wildfire and non-wildfire days the I/O PM 2.5 ratio gradually decreased as the outdoor PM 2.5 concentration increased (figure S5.1).However, there was substantial variation in the I/O PM 2.5 ratio both within and across facilities and 32 facilities experienced at least one day with the I/O ratio ⩾1.0 (figure S5.2).

LMER modeling of indoor and outdoor PM 2.5 3.2.1. All study days (unstratified)
In LMER modeling, no covariates confounded the relationship between indoor and outdoor PM 2.5 , such that the addition of each covariate had minimal impact on the effect estimate for outdoor PM 2.5 .The median (range) change in the effect estimate of outdoor PM 2.5 was −0.02% (−1.3%-0.34%).Further, no covariates substantially impacted the model fit.The conditional R 2 (variance explained by both the fixed and random effects) for the unadjusted model (equation ( 1)) was 0.91 and ranged from 0.90 to 0.91 with the addition of each covariate (equation ( 2)).
The final unstratified model for all study days contained only outdoor PM 2.5 with a random slope and intercept for each childcare facility (equation ( 1)).In this model, indoor PM 2.5 increased by an average of 0.57 µg m −3 [95% CI: 0.51-0.63]per 1 µg m −3 increase in outdoor PM 2.5 .The adjusted intraclass correlation coefficient (ICC) was 0.61, indicating that 61% of the total variance of indoor PM 2.5 was explained by the random slope and intercept for outdoor PM 2.5 in each facility.The relationship between indoor and outdoor PM 2.5 varied widely between facilities, with the effect estimate for each 1 µg m −3 increase in outdoor PM 2.5 ranging from 0.15 to 0.91 µg m −3 indoors between facilities (figure 4).

Stratification by wildfire day
Like the unstratified model, the final model for both non-wildfire and wildfire days contained only outdoor PM 2.5 with a random slope and intercept for each childcare facility.No covariates substantially changed the effect estimate of outdoor PM 2.5 in either model (figure 5) and none substantially impacted the model fit.The conditional R 2 for the unadjusted final model on non-wildfire and wildfire days was 0.85 and 0.86, respectively, and both were unaffected by the addition of each covariate (equation ( 2)).In the final stratified LMERs, indoor PM 2.5 increased linearly with outdoor PM 2.5 on both non-wildfire and wildfire days.However, the magnitude of this relationship was slightly higher during non-wildfire days than wildfire days (figure 5).On non-wildfire days, indoor PM 2.5 increased by an average of 0.55 µg m −3 [95% CI: 0.47-0.63]per 1 µg m −3 increase in outdoor PM 2.5 .On wildfire days, indoor PM 2.5 increased by an average of 0.51 µg m −3 [95% CI: 0.44-0.58]per 1 µg m −3 increase in outdoor PM 2.5 .This suggests that about 55% and 51% of outdoor PM 2.5 got inside these facilities, on average, during non-wildfire and wildfire days, respectively.Despite these overall relationships, there was substantial between-facility variation in the association between indoor and outdoor PM 2.5 during non-wildfire and wildfire days.The random slope and intercept explained 62% (ICC = 0.62) of the total variance in indoor PM 2.5 on both non-wildfire and wildfire days.This high between-facility variability meant that the effect estimates for outdoor PM 2.5 ranged widely between facilities.On non-wildfire days, each 1 µg m −3 increase in outdoor PM 2.5 ranged from an increase of 0.11 to 1.03 µg m −3 indoors across facilities, and it ranged from 0.18 to 0.79 µg m −3 on wildfire days (figure 6).In some facilities, the relationship between indoor and outdoor PM 2.5 during wildfire days was influenced by several high-leverage points (Vancouver Coastal 1 and 7) because of a limited number of wildfire days.The change in association between indoor and outdoor PM 2.5 during wildfire days compared to non-wildfire days was also facility specific (figure 6).At the different facilities, the effect estimates for outdoor PM 2.5 increased, decreased, or remained unchanged on wildfire days compared with non-wildfire days.The effect estimates for outdoor PM 2.5 changed by less than 10% (range: −5%-9%) on wildfire days compared with non-wildfire days in 6 facilities, while they increased by more than 10% (range: 11%-205%) in 15 facilities and decreased by more than 10% (range: −12% to −59%) in 14 facilities.There was some evidence that this change was associated with the facility type.Of the 13 residential facilities, 9 (69%) decreased by more than 10% during wildfires, while 14 of 22 (64%) commercial facilities increased by more than 10% (figure S6.1).This contrast between facility types was driven by differences in infiltration during non-wildfire days.On non-wildfire days, the median (range) effect estimate of outdoor PM 2.5 was 0.65 (0.34-1.03) µg m −3 across residential facilities and 0.44 (0.11-0.94) µg m −3 across commercial facilities.On wildfire days, the effect estimates were more similar, with the median (range) 0.50 (0.27-0.78) across residential facilities and 0.54 (0.18-0.79) across commercial facilities.

analysis: timing of infiltration
Rerunning the final model on non-operational hours did not result in substantial changes to the effect estimates for outdoor PM 2.5.Overall, the magnitude of the effect estimate decreased slightly during non-operational hours, but there was no clear trend in the direction of change across facilities.For example, during wildfire days, the median (IQR) changes in the effect estimate for outdoor PM 2.5 during non-operational hours compared to operational hours was 0.1% (−14%-7%).
When analyzing the cumulative lagged effect of outdoor PM 2.5 on indoor PM 2.5 , we found that most PM 2.5 infiltrated indoors in the first 10 min.The data suggests that during non-wildfire and wildfire days, respectively, 89% and 93% of the total infiltration that occurred over the previous 60 min occurred in the first 10 min (figure S7.1).

Summary
In this project, we deployed pairs of low-cost air quality sensors inside and outside of 45 childcare facilities across BC, Canada.While average PM 2.5 concentrations were lower indoors than outdoors, indoor PM 2.5 concentrations more than tripled during wildfire days compared with non-wildfire days.Overall, we found that indoor PM 2.5 concentrations tracked outdoor PM 2.5 regardless of other variables, including indoor CO 2 , local meteorological conditions, and area-level socio-economic factors.Across all facilities, our results indicate that slightly more than half of the outdoor PM 2.5 penetrated indoors on average, with slightly less infiltrating during wildfire days.Additionally, most of the infiltration that did occur appeared to happen within the first 10 min following an increase in outdoor PM 2.5 .Despite these findings, the effect of outdoor PM 2.5 on indoor PM 2.5 was heterogeneous across facilities, but was relatively stable within facilities during non-wildfire and wildfire days.Notably, the influence of wildfire days on the relationship between indoor and outdoor PM 2.5 was different for each facility.

Indoor PM 2.5 concentrations were lower than outdoors, but increased substantially during wildfire days
Indoor PM 2.5 tended to be lower than outdoor PM 2.5, including during wildfire-impacted days.In general, this provides some evidence that indoor air in these childcare facilities was better quality than outdoor air during wildfire smoke events, and that recommendations for children to spend more time indoors during wildfire smoke events are warranted to impart some protection [42].However, the fact that indoor PM 2.5 at these facilities increased by 235% on wildfire days suggests that a significant portion of wildfire smoke PM 2.5 penetrated indoors, despite any actions that may have been taken to reduce smoke at each facility.Further, the wide range of I/O ratios within facilities indicates that the amount of outdoor PM 2.5 penetrating indoors fluctuated.In addition, most facilities experienced at least one day with an I/O ratio greater than or equal to 1.0 indicating that in some buildings, indoor air offered little protection.
There are many well documented actions that facilities can take to reduce smoke indoors, including closing windows, limiting building entry to one door, using portable air cleaners or building do-it-yourself air filters [43], and upgrading, maintaining, and using heating, ventilation, and air conditioning (HVAC) systems [44,45].A more comprehensive investigation of the effectiveness of different actions taken by facilities is beyond the scope of this community science project because we do not have facility-level information on actions taken to reduce smoke.As a result, it is unclear whether indoor PM 2.5 increased on wildfire days because actions were not taken to reduce indoor air pollution or because ineffective actions were taken.Further, these are data from operating childcare facilities such that there may be real-world barriers to the implementation or effectiveness of smoke reduction behaviors and there may be local sources of outdoor PM 2.5 that we were unable to account for.

Indoor PM 2.5 increased as outdoor PM 2.5 increased
Increasing outdoor PM 2.5 was strongly associated with increasing indoor PM 2.5 and indoor concentrations were significantly higher on wildfire days.This aligns with previous research which found that indoor PM 2.5 increased considerably during wildfire smoke events when outdoor PM 2.5 concentrations were high [19,25,46].In our models, outdoor PM 2.5 explained 85% to 91% of the variation in indoor PM 2.5 , suggesting that within each facility indoor PM 2.5 varied in a predictable way with outdoor PM 2.5 .Further, this relationship was primarily driven by the outdoor PM 2.5 concentration over the previous 10 min, indicating that indoor air quality was affected nearly immediately by wildfire smoke.
That indoor PM 2.5 closely tracks outdoor PM 2.5 , suggests that readily available PM 2.5 monitoring data [36] is useful for estimating wildfire-smoke related PM 2.5 exposure indoors where people may spend the majority of their time.Future work should develop models of indoor PM 2.5 to evaluate whether outdoor PM 2.5 can reliably predict indoor exposure across different locations, using widely available data from low-cost sensors.Such models could be used to broaden public health surveillance and recommendations to include exposures that people experience indoors as well as to improve the accuracy of estimating PM 2.5 exposure in epidemiological studies, which often rely on outdoor concentrations only.

The relationship between indoor and outdoor PM 2.5 was highly variable across facilities
We observed substantial between-facility variation in the relationship between indoor and outdoor PM 2.5 such that unmeasured factors that varied between facilities explained approximately 60% of the total variation in indoor PM 2.5 .There were three important features of this between-facility variability.First, there were differences in the relationship between indoor and outdoor PM 2.5 .At some facilities, indoor air quality closely tracked outdoor air quality, while at others, this relationship was attenuated, perhaps due to less PM 2.5 getting inside or more effective indoor air filtration.Second, there was facility-level variation in the effect of wildfire days on the relationship between indoor and outdoor PM 2.5 .In some facilities, the relationship between indoor and outdoor PM 2.5 was reduced during wildfire days, while in others it remained unchanged or increased.Third, the relationship between indoor and outdoor PM 2.5 within each facility was relatively constant on non-wildfire and wildfire days.
Overall, the high degree of between-facility variability is likely the result of many site-specific factors that influence infiltration, and thus the relationship between indoor and outdoor PM 2.5 .For example, after measuring indoor and outdoor PM 2.5 in 18 different buildings for 6-12 months, Stamp et al [24] concluded that real-world measures of PM 2.5 infiltration were a metric of both building characteristics and occupant behaviors.They found that the I/O PM 2.5 ratio was strongly influenced by ventilation rates, window use, and occupancy patterns.Other studies have similarly observed high levels of between-location variability in PM 2.5 infiltration [17,19,25,47].Burke et al [48] found that indoor PM 2.5 concentrations varied by more than 20 times among neighboring households during the same wildfire smoke events.This suggests that the relationship between indoor and outdoor PM 2.5 is a location specific parameter largely driven by different combinations of building envelope, mechanical ventilation, indoor filtration, and behavioral factors, rather than local environmental (e.g.meteorological) conditions.
Although wildfire days affected the relationship between indoor and outdoor PM 2.5 differently across facilities, this relationship was relatively constant during non-wildfire and wildfire days within each facility.This suggests that there were a set of unique factors at each facility that influenced infiltration relatively constantly, but that changed or had different impacts during wildfire days compared with non-wildfire days in some facilities.For facilities with a consistently low I/O PM 2.5 ratio, it is possible that the buildings were tightly sealed and running HVAC/air filtration systems during the entire study period, thereby always effectively reducing indoor PM 2.5 [16,17].Alternatively, for facilities with a consistently high I/O PM 2.5 ratio, windows may have been left open continually, leading to constant rates of natural ventilation and infiltration across the study period.For facilities where the I/O PM 2.5 ratio was attenuated during wildfires, managers may have closed windows, turned on HVAC systems, or initiated other air filtration whenever it was smoky [16].For facilities where the I/O PM 2.5 ratio was magnified during wildfire events, indoor PM 2.5 levels may have increased due to reduced ventilation rates from closing windows and subsequent accumulation of indoor PM 2.5 .For example, Liu et al [49] determined that the deposition rate of indoor PM 2.5 varied with the ventilation rate, such that closing windows caused PM 2.5 to remain suspended for a longer period of time, leading to indoor accumulation.Finally, regardless of the direction of the change during wildfire days, the relationship between indoor and outdoor PM 2.5 was not substantially different during non-operational hours.This indicates that the factors determining infiltration were consistent through the day and night (e.g.windows remained closed, HVAC remained turned on).PM 2.5 I/O ratios were higher in residential facilities than commercial facilities during non-wildfire days, but they became more similar during wildfire days.As such, the I/O ratio was more likely to decrease in residential facilities and to increase in commercial facilities during wildfire days.This may be due to differences in air filtration and ventilation systems in residential versus commercial buildings.For example, residential buildings in North America often rely on natural ventilation (i.e.open windows) and they have less standardized HVAC regulations than commercial settings [50].The residential facilities in this study may have been more likely to have open windows during non-wildfire periods resulting in the observed higher levels of infiltration.However, the interaction between facility type and infiltration was difficult to interpret for two reasons.First, the facility-type variable was based on a visual assessment of images of each facility from the outside, such that we do not have verifiable information on specific differences between these facility types.Second, the number of wildfire days was limited, with 9 facilities experiencing less than 10 wildfire days.As a result, there was a relatively sparse amount of data available to make robust comparisons between facility types during wildfire and non-wildfire days.

The case for installing air quality sensors
Despite the many potential explanations for between-facility variability (see section 4.4), we were unable to measure many facility-level factors that might influence the relationship between indoor and outdoor PM 2.5 .Regardless, these factors are likely to have different impacts in different facilities, interact with other factors, and change over time.Characterizing each of these factors and their interactions with site-specific conditions and occupant behaviors becomes intractable at a large-scale and generalizability would be limited.As such, it may be more important to help facilities install and use their own low-cost air quality sensors to monitor conditions in real-time rather than seeking to develop and rely on a comprehensive list of infiltration determinants that facilities may not have the capacity to change (e.g.purchasing HVAC).
On-site sensors can potentially help facility managers make real-time decisions to take actions such as closing windows, hosting recess indoors, turning on air filtration devices, increasing ventilation, and maintaining or upgrading HVAC systems.It is important to make these decisions at the facility-level with real-time local air quality data because they intersect with the susceptibility of the occupants, which cannot be generalized to a broad scale.For example, facilities supporting children with pre-existing respiratory conditions might make different decisions regarding air quality than other facilities.There are also distinct advantages associated with deploying sensors inside and outside of a given facility.In many cases, decisions are made based on the absolute values of outdoor air pollution, such as in response to air quality alerts [22].With sensors located indoors and outdoors a given building, facility managers can act based on the relative air quality indoors versus outdoors.Sensors could be linked to alert staff if indoor PM 2.5 becomes much worse than outdoor PM 2.5 , thus informing decisions to open windows, to take children outdoors, or to attempt to reduce indoor sources of PM 2.5 .Conversely, if PM 2.5 becomes worse outdoors, facilities may decide to keep the children indoors and take measures to limit infiltration.
Despite these possibilities, more work is needed to understand how facility managers use this air quality information to mitigate indoor PM 2.5 during wildfire events.If facilities find it challenging to take real-life actions based on sensor data, it will be crucial to first understand those barriers and then address them when implementing air quality sensor networks, such as enabling text alerts when indoor PM 2.5 becomes worse than outdoor PM 2.5 .

The relationship between indoor and outdoor PM 2.5 was not dependent on meteorological conditions and socio-economic factors
The relationship between indoor and outdoor PM 2.5 at each facility did not change with linear adjustment for several meteorological parameters including temperature and humidity, as well as indoor CO 2 and differences in neighborhood socio-economic factors.Some of these factors have been found to effect PM 2.5 infiltration in previous studies [51].For example, Liang et al [52] found that indoor and outdoor temperature were important predictors of PM 2.5 infiltration into 60 homes in Georgia, USA.Our results may have been different because the impacts of these factors may be important in some locations and not others.Additionally, our study period was not long enough to detect the impacts of larger seasonal changes in meteorological factors and our facility sample size may not have been large enough to detect effects of socio-economic factors.These previous studies also used different modeling methods that may have impacted which covariates were important.For example, Liang et al [52] modeled the I/O PM 2.5 ratio and Hystad et al [51] used non-mixed linear regression.Despite this, our findings that outdoor PM 2.5 at each facility explained 85% to 91% of the variation in indoor PM 2.5 suggest that changes in indoor PM 2.5 were largely a function of outdoor PM 2.5 .Differences in the magnitude of the relationship between indoor and outdoor PM 2.5 and the total indoor PM 2.5 concentration are due to facility-level factors we did not measure, including the baseline levels of indoor PM 2.5 , the behaviors of occupants, ventilation rates, and filtration efficiency.The combined importance of these factors is indicated by the substantial impact of the random intercepts and slopes in our models.

Limitations
The community-science nature of this project was associated with several limitations.First, there was a large amount of missing data in some facilities because the EGGs sometimes stopped reporting data.For example, at some sites the sensors frequently lost Wi-Fi connectivity because of changes to the Wi-Fi network password or because the Wi-Fi went offline.Monitoring the data feed from these facilities has required ongoing efforts by the study team and when sensors stopped reporting it sometimes took a significant amount of effort to contact facility managers and to help them reinstall the EGGs.Second, we did not have facility-level information on sources of indoor PM 2.5 .Although we attempted to remove indoor peaks due to indoor sources, our algorithm was focused on removing peaks when indoor PM 2.5 concentrations were double the outdoor concentrations.As a result, we were unable to account for sources that contributed to more moderate increases in the indoor concentration.We also were not able to account for local, non-wildfire sources of outdoor PM 2.5 .Third, our data-driven definition of wildfire days may not have accounted for the public's perception of poor air quality, and thus may not have captured the threshold at which people take action to prevent smoke from infiltrating indoors.Indeed, consistent with O'Dell et al [25], we see that as ambient conditions become progressively smokier (figure S5.1), the I/O PM 2.5 ratio becomes increasingly smaller, indicating that at higher smoke concentrations facilities may have taken more actions to mitigate indoor smoke.Since this relationship appears to be gradual it is unlikely that a dichotomous definition of wildfire day could adequately capture when people decide to implement smoke mitigation measures.
Finally, smoke impacts during the 2022 wildfire season were moderate compared to some previous years, such that we were not able to capture changes in the relationship between indoor and outdoor PM 2.5 that may occur at higher concentrations.For example, in 2017 the maximum 24 h average PM 2.5 concentration measured at 62 regulatory air quality monitoring stations across BC was 883 µg m −3 [53].In comparison, in this project, the maximum observed 12 h and 1 h concentrations during operational hours measured with facility EGGs were only 145 µg m −3 and 292.7 µg m −3 , respectively (table 1).

Conclusion
We found that indoor PM 2.5 concentrations in BC childcare facilities can increase substantially during wildfire events, and that indoor PM 2.5 is strongly and almost immediately influenced by outdoor PM 2.5 .We also found that there was substantial variability in the relationship between indoor and outdoor PM 2.5 across facilities and that the effect of wildfire days on this relationship was not the same in all facilities.Taken together, this suggests that PM 2.5 infiltration is highly dependent on facility-level characteristics and occupant behaviors.This underscores the need to monitor local, building specific air quality in order to accurately estimate and mitigate exposures to wildfire smoke PM 2.5 indoors.Public health practitioners should consider working with facilities that serve susceptible populations, including childcare facilities, to help them develop and implement wildfire smoke monitoring and decision-making plans, including the deployment and interpretation of site-specific air quality sensors.

Figure 1 .
Figure 1.Childcare facility locations and timeseries of outdoor PM2.5 concentrations.(A) Counts of participating childcare facilities by regional Health Authority (Interior, Vancouver Coastal, Northern).Red points show the locations of wildfires greater than 200 ha in size [33] that occurred during the study period.(B) Timeseries of daily mean operational hour PM2.5 concentrations across all outdoor EGGs.The red ribbon shows the daily minimum and maximum of PM2.5 concentrations across all EGGs.The dotted horizontal line (gray) shows the median wildfire day threshold in this project (16.4 µg m −3 ).

Figure 2 .
Figure 2. Diagram of factors potentially influencing the concentration of indoor PM2.5 in study sites.

Figure 3 .
Figure 3. Boxplots showing the distribution of daily mean PM2.5 concentrations during operational hours across childcare facilities for non-wildfire (blue) and wildfire (red) days.Boxplots show the distribution of PM2.5 concentrations (A) inside and (B) outside of each childcare facility.The (C) distribution of the daily mean I/O PM2.5 ratios across facilities.For each distribution, the boxplot edges indicate the interquartile range; the black horizontal line shows the median; whiskers extend to the min and max.

Figure 4 .
Figure 4. Linear relationship between the indoor and outdoor PM2.5 concentration at each childcare facility across the entire study period.The plot demonstrates the between-childcare facility variability in the relationship between indoor and outdoor PM2.5.Lines show the predicted random linear relationship between indoor and outdoor PM2.5 at each childcare facility while points show the observed relationship between indoor and outdoor PM2.5.Predicted values are derived from a linear mixed effected regression model a with indoor PM2.5 as the outcome and outdoor PM2.5 as the predictor, with random slopes and intercepts for PM2.5 at each childcare facility.The dashed (gray) line indicates a 1:1 relationship between indoor and outdoor PM2.5. a PM 2.5 indoors = PM 2.5 outdoors + Random slope and intercept (PM 2.5 outdoors |Facility) .

Figure 5 .
Figure 5. Influence of each covariate on the linear relationship between indoor and outdoor PM2.5 across childcare facilities.Shown are the regression coefficient estimates and 95% confidence intervals for outdoor PM2.5 unadjusted a (red) and adjusted b (black) individually for each variable shown.Global effect estimates are derived from a linear mixed effects regression model with indoor PM2.5 as the outcome, outdoor PM2.5 as the predictor, and with random slopes and intercepts for outdoor PM2.5 in each childcare facility.Results are stratified by wildfire day.Deviation from the vertical (red) dashed line indicates that inclusion of the variable changed the effect of outdoor PM2.5 on indoor PM2.5.Changing the coefficient estimate of outdoor PM2.5 by at least 10% was a key criterion for whether each variable was considered a confounder of the relationship between indoor and outdoor PM2.5. a PM 2.5 indoors = PM 2.5 outdoors + Random slope and intercept (PM 2.5 outdoors |Facility) b PM 2.5 indoors = PM 2.5 outdoors + Confounder k + Random slope and intercept (PM 2.5 outdoors |Facility) .

Figure 6 .
Figure 6.Linear relationship between the indoor and outdoor PM2.5 concentration at each childcare facility stratified by non-wildfire (blue) and wildfire (red) days.Plots show the between-childcare facility variability in the effect of wildfire days on the relationship between indoor and outdoor PM2.5.Lines show the predicted linear relationship between indoor and outdoor PM2.5 at each childcare facility while points show the observed relationship between indoor and outdoor PM2.5.Predicted values are derived from a linear mixed effected regression model a with indoor PM2.5 as the outcome and outdoor PM2.5 as the predictor, with random slopes and intercepts for PM2.5 at each childcare facility, stratified by wildfire day.a PM 2.5 indoors = PM 2.5 outdoors + Random slope and intercept (PM 2.5 outdoors |Facility) .

Table 1 .
Air quality EGG data summary.The EGG data are averaged daily during childcare operating hours, Monday-Friday, 0800-1800 h.