The MultiWiSE metrics were developed and applied in BC, which is the westernmost province in Canada. The province is heavily forested with complex topography, defined by the Coast Mountains on the west coast, the Rocky Mountains along the eastern border with Alberta, and the dry interior plateau between them. Like other regions on the west coast of North America, BC has a long history of severe wildfire seasons with a marked increase in wildfire activity over the past 20 years [32]. It provides an ideal setting to test the MultiWiSE methods due to the significant spatiotemporal variation in WFS and generally low background PM2.5 concentrations.
Analyses were conducted at the level of the national census subdivision, of which there were 652 in the 2016 census that were populated and had PM2.5 data available. Census subdivisions are relatively stable geographic areas with highly variable population density depending on their urbanicity. The analyses covered the 14-year period from 2010 to 2023, or 730 consecutive weeks, with above-average wildfire seasons in 2010, 2014, and 2015, and severe wildfire seasons in 2017, 2018, 2021, and 2023 [32]. The wildfire season is defined here as May 1 through October 31, which captures the vast majority of wildfire activity in BC while largely excluding controlled burning and residential woodsmoke in the shoulder seasons [33].
PM2.5 data
Daily estimates of total (i.e., all-source) PM2.5 were taken from the Canadian Optimized Statistical Smoke Exposure Model (CanOSSEM), which is described in detail elsewhere [34]. Briefly, CanOSSEM combines PM2.5 measurements with satellite observations of wildfires, aerosol, and meteorological parameters to generate daily estimates of average PM2.5 at a 5 km x 5 km resolution for all populated areas of Canada. The random forest model performs well, with a root mean squared error (RMSE) of 2.96 μg/m3 and 96% of estimates within 5 μg/m3 of the observed values for the original 2010-2019 training period [34]. With the inclusion of data from 2020 to 2023, the RMSE increased to 3.73 μg/m3, and 96% of estimates still remained within 5 μg/m3 of the observed values (unpublished data). The increased RMSE was largely due to the unprecedented 2023 wildfire season across Canada [32].
Data from CanOSSEM have been used for multiple epidemiologic studies [35,36,37] and will provide the foundation for a national program of research on chronic health effects associated with WFS PM2.5. For these analyses, we first aggregated the daily CanOSSEM PM2.5 estimates to the census dissemination area-level, of which there were 7,368 in the 2016 census that were populated and overlapped with the available PM2.5 estimates. Dissemination areas have a population of approximately 400-700 persons and are the smallest geographic unit at which census results are reported. Each dissemination area maps to a single census subdivision, which represents municipalities or municipal equivalents in Canada. Next, we calculated daily population-weighted PM2.5 estimates for the 652 census subdivisions. Weekly sums and averages of daily PM2.5 were then calculated for each census subdivision by epidemiologic week, where the first epidemiologic week of each year starts on the first Sunday of the year [38]. Only complete epidemiologic weeks between 2010 and 2023 were used for analysis, leading to 730 consecutive weeks between January 3, 2010, and December 30, 2023. All 652 census subdivisions had PM2.5 data available for 99.5% or more of the 730 weeks.
Establishing the location-specific counterfactual and separating WFS from non-WFS PM2.5
The methods described here were designed to require only a multiyear time series of daily total PM2.5 estimates or measurements, so they can be easily applied in any location without the need for other complex datasets. Similar to O’Dell et al. [29], the approach is predicated on determining the normal range of PM2.5 concentrations in the absence of WFS during the multiyear period of interest. Once established, we use that baseline to define a counterfactual PM2.5 concentration for WFS-impacted weeks (i.e., the PM2.5 concentration expected in the absence of WFS), then separate total PM2.5 into its non-WFS PM2.5 and WFS PM2.5 components. The following section describes the process in detail (also see Fig. S1).
First, daily total PM2.5 is used to calculate the sum and average of PM2.5 for each epidemiological week in the multiyear period. Next, the distribution of the weekly sum of total PM2.5 is used to identify all values with modified z-scores ranging from −2 to +2. The modified z-score is a robust alternative to the z-score when working with skewed data [39]. The median of these values is set as the counterfactual value, meaning the weekly sum of PM2.5 expected in the absence of WFS. Next, weeks during the wildfire season (May 1 to October 31 used here) with a modified z-score greater than +2 are identified as WFS-impacted, indicating that one or more days during that week were affected by WFS. For WFS-impacted weeks, the non-WFS PM2.5 is set to the counterfactual value, and the WFS PM2.5 is set to the difference between total PM2.5 and non-WFS PM2.5. For non-WFS-impacted weeks, the total PM2.5 is retained as the non-WFS PM2.5 and WFS PM2.5 is set to zero. The total PM2.5, non-WFS PM2.5, and WFS PM2.5 are then used to create cumulative exposure trajectories using weekly sums and exposure time series using weekly averages, where the weekly sums are divided by 7 (Fig. 2 shows examples at two locations).
The panels display the A cumulative and B average weekly total, WFS, and non-WFS fine particulate matter (PM2.5) and C distribution of the weekly sum of total PM2.5 for each census subdivision. The distribution plots highlight the range and median of normal weekly sums (modified Z-score between −2 and +2), which are used to establish the counterfactual value and extract the WFS and non-WFS contributions from total PM2.5. The locations of the two census subdivisions were chosen to indicate low (Kitimat-Stikine C) and high (Vernon) WFS impacts, and are shown in Fig. S2. The Multiyear Wildfire Smoke Exposure (MultiWiSE) metrics for the two locations can be found in Table S1.
Multiyear wildfire smoke exposure (MultiWiSE) metrics
Once the location-specific counterfactual has been established and the weekly averages and sums of WFS PM2.5 and non-WFS PM2.5 have been calculated, the estimates of WFS PM2.5 can be used to generate the twelve MultiWiSE metrics to comprehensively characterize exposure to WFS (Fig. 1). The MultiWiSE metrics are calculated using weekly PM2.5 to balance the need for a time scale suited to studying the health effects of multiyear exposure with the ability to detect short-term peaks in PM2.5. The metrics fall into four groups described as cumulative exposure, weekly exposure, episode exposure, and recovery, described in more detail below. Here, the metrics are calculated for 2010–2023, but this exposure window can be flexibly defined to capture any time period of interest (e.g., exposure in the 5 years prior to a health outcome). The MultiWiSE metrics can also be used to examine spatial and temporal trends in WFS exposure, which we demonstrate by comparing the metrics across BC census subdivisions for the entire study period and for the first and last 7 years. Once the twelve MultiWiSE metrics are calculated, Spearman correlation coefficients can be used to evaluate how well the metrics capture different features (frequency, intensity, and duration) of WFS PM2.5 exposure.
Cumulative exposure: These two metrics describe cumulative exposure to WFS over the entire exposure window, reflecting overall intensity of the exposure. Metric 1 is cumulative WFS PM2.5, which reflects the total sum of exposure over the multiyear period in mg/m3. It is calculated by adding together the weekly sums of WFS PM2.5 and represents the sum of daily average WFS PM2.5 across the entire period. Metric 2 is the WFS fraction of the cumulative total PM2.5 exposure (%), calculated by dividing the cumulative WFS PM2.5 by the cumulative total PM2.5.
Weekly exposure: These four metrics describe WFS-impacted weeks, defined as weeks during the wildfire season with a modified z-score greater than +2. Metric 3 is the average WFS PM2.5 concentration (μg/m3) across all WFS-impacted weeks. This metric is comparable to the annual or multiyear averages typically used in epidemiologic studies of long-term air pollution exposure. Metric 4 is any WFS, defined as the total count of WFS-impacted weeks (i.e., weeks with any WFS PM2.5) during the exposure window, and Metric 5 is the count of WFS-impacted weeks when the weekly average WFS PM2.5 is greater than 5 μg/m3, chosen to be consistent with the definition used by Casey et al. [30]. Metric 6 is the count of WFS-impacted weeks when the weekly average total PM2.5 is greater than 25 μg/m3, suggesting that on average, the 24 h PM2.5 concentrations were greater than 25 μg/m3 over the 7-day period, in part due to WFS. This value was chosen because it is the current 24 h air quality objective in BC, though any similar daily regulatory value could be substituted in its place (e.g., 35 μg/m3 is the National Ambient Air Quality Standard in the US).
Episode exposure: These five metrics describe WFS episodes, considering both all episodes and severe episodes. WFS episodes (i.e., all episodes) were defined to capture both short periods with extreme WFS as well as longer periods with more moderate WFS, and severe episodes were defined to isolate the more extreme WFS episodes. A WFS episode is defined as either (1) two or more WFS-impacted weeks separated by no more than three non-WFS-impacted weeks, or (2) a single WFS-impacted week where the sum of WFS PM2.5 > 250 μg/m3. A severe WFS episode is defined as any WFS episode that includes at least one week with a sum of WFS PM2.5 > 250 μg/m3, which could be a single week or a multi-week episode with one or more weeks over this threshold. The 250 μg/m3 value was chosen to capture periods with extreme air quality impacts, such as the highly publicized smoke in eastern North America during 2023 wildfire season [40]. Metric 7 is the total count of WFS episodes and Metric 8 is the count of severe episodes over the entire exposure window. Metric 9 is the longest episode (weeks), meaning the WFS episode with the longest duration between the first and last WFS-impacted week in the episode. Metric 10 is the worst episode (μg/m3), meaning the WFS episode with the highest average WFS PM2.5 concentration across all WFS-impacted weeks in the episode. Finally, Metric 11 is the fraction of cumulative WFS PM2.5 from severe episodes (%).
Recovery: This metric describes recovery, capturing the periods of time with little to no WFS exposure. Metric 12 is the average recovery period (weeks) between WFS episodes, calculated as average length of time between episodes, including periods before the first and after the last episodes in the exposure window.
Note that of the twelve MultiWiSE metrics, ten are calculated using only the estimates of weekly WFS PM2.5, while two metrics (2 and 6) also require data on total PM2.5. As such, most of the metrics can easily be calculated for any long time series of daily WFS-specific PM2.5 concentrations, such those generated by deterministic, statistical, or blended modeling approaches [29, 41, 42]. A key difference when using pre-existing estimates of WFS PM2.5 to calculate the MultiWiSE metrics is that WFS-impacted weeks need to be defined as weeks where the sum of WFS PM2.5 > 0 μg/m3, rather than defined using the modified Z-scores of total PM2.5.
Hypothetical locations
In addition to applying these methods and metrics to data from 2010-2023 across census subdivisions in BC, we applied them to 14 years of weekly PM2.5 data generated for the three hypothetical locations described in the introduction (Fig. 3). All three locations had the same cumulative WFS PM2.5 over the 14-year period, but the patterns of exposure were different. The first had low levels of WFS PM2.5 in most years, the second had high levels in a few years, and the third had moderate levels in approximately half of the years. These hypothetical locations serve to highlight differences between the twelve MultiWiSE metrics in areas with patterns of WFS impacts that are not temporally autocorrelated.
The panels display the A cumulative WFS fine particulate matter (PM2.5) and B weekly average total and WFS PM2.5 and WFS episodes for each location. The three locations have the same cumulative WFS PM2.5 with different frequencies, intensities, and durations of exposure. The Multiyear Wildfire Smoke Exposure (MultiWiSE) metrics for the three locations can be found in Table 2.