Pollution Model¶

Data¶

PM2.5 Data¶

The baseline projections are obtained from the ECCC using their GEM-MACH model, which gives us a prediction for background PM2.5 levels for the years 2026, 2031, and 2036. All other timepoints are calculated using linear interpolation:

Wildfire PM2.5 is based on a separate model, RAQDPS, which predicts PM2.5 from anthropogenic and non-wildfire natural sources, and RAQDPS-FW, which predicts PM2.5 from those same sources plus wildfires. The wildfire contribution for each month and census division is the difference between the two models’ predictions (RAQDPS-FW minus RAQDPS), averaged over 2018-2023 and floored at zero to remove small negative values caused by model noise. This historical monthly wildfire contribution is then scaled by a future climate scaling factor, derived from projected changes in wildfire-related PM2.5 under different IPCC Shared Socioeconomic Pathway (SSP) scenarios, to estimate wildfire PM2.5 at future timepoints.

The climate scaling factor for each SSP scenario is derived from Table S2 of [Xie, 2022], which reports, for three CMIP6-driven multiple linear regression models (CESM2, GFDL-ESM4.1, and CNRM-ESM2-1), the percent change in August-September mean PM2.5 over the western US by the late 21st century (2080-2100) relative to 1990-2010, under each SSP scenario. We take the average percent change across the three models as the total wildfire PM2.5 increase by 2100 for that scenario (SSP1_2.6 ~45%, SSP2_4.5 ~85%, SSP3_7.0 ~124%, SSP5_8.5 ~136%), and linearly interpolate the scaling factor between 1 (present day) and this value at 2100 for intermediate timepoints.

The data currently extends to December 2036, the last timepoint for which GEM-MACH background PM2.5 projections are available.

The resulting data is stored as a separate dataset (one .csv file per SSP scenario), each containing the columns below, including an SSP column identifying which scenario that dataset corresponds to:

wildfire_pm25_scaled = wildfire_pm25 * factor

total_pm25 = wildfire_pm25_scaled + background_pm25

Census Division Data¶

The pollution data is only available at the CDUID level, but agents in the model are only assigned a province. So, each agent is additionally assigned a census division within their province, sampled with probability proportional to that census division’s population in 2021 (see CensusDivision in census_division.py). For a given province with census divisions \(\{1, \dots, n\}\), an agent is assigned to census division \(c\) with probability:

where \(c'\) ranges over all census divisions in the same province, so the denominator is the province’s total population.

This is preferable to a uniform random assignment, since population is unevenly distributed across census divisions within a province.

Census division population data comes from Statistics Canada’s Table 98-10-0007-01, Population and dwelling counts: Canada and census divisions, saved as master_census_data_2021.csv: