The synergistic effects of PM_2.5 and high temperature on community mortality in British Columbia

Abstract

Exposure to fine particulate matter (PM_2.5), a primary component of wildfire smoke, and exposure to high temperatures both independently increase the risk of mortality, with evidence of synergistic effects. These environmental stressors often co-occur during wildfire season, and their synergistic effects are expected to worsen with climate change. However, the combined health risks of wildfire-related PM_2.5 and temperature remain poorly understood, limiting the effectiveness of public health interventions. This study investigated the joint effects of PM_2.5 and temperature on community all-cause mortality across 13 wildfire seasons (2010–2022) in southwest British Columbia, Canada. Daily estimates of ambient PM_2.5 and temperature exposure were assigned from a machine learning-based prediction model and satellite data, respectively. Using a case-crossover design and conditional logistic regression, we examined non-linear associations between co-exposures and mortality across deciles and absolute exposure ranges. We found significant, non-linear interactions, with the highest mortality risk observed on days with PM_2.5 levels of 12–14 µg/m³ and temperatures ≥ 26 °C (OR = 7.31, 95% CI: [5.34, 10.0]). Co-exposures exceeding the 90th percentile showed synergistic effects, contributing an excess mortality risk of 7.9% (95% CI: 6.1, 9.1). Moderate co-exposure levels posed substantial risks, underscoring the need for greater attention to their public health impacts. Our findings highlight the urgent need to understand and address the compounding effects of PM_2.5 and temperature to mitigate risks in a changing climate.

Introduction

Climate change is causing warmer and drier conditions across many regions, including North America, increasing the risk of more frequent and intense wildfires¹. Recent increases in wildfire occurrence have led to more episodic, short-term exposures to high concentrations of ambient fine particulate matter (PM_2.5), a primary component of wildfire smoke (WFS)². Given the large body of evidence on the adverse health effects of short-term PM_2.5 exposure, including during WFS events^3,4, it is imperative to better understand the risks associated with air pollution due to climate-related events, especially when exposure co-occurs with other climate-related hazards.

Concurrent with increasing exposures to WFS, climate change is leading to hotter weather and increasing the frequency of extreme heat events (EHEs). The detrimental health effects of heat exposure are well-documented in the scientific literature, particularly concerning all-cause mortality and cardio-respiratory outcomes^5,6. Recent research in the western region of North America has demonstrated the substantial overlap between WFS and EHEs^7,8,9. Epidemiological evidence further shows that elevated temperatures amplify the adverse health effects of PM_2.5 exposure, and vice versa^{9,10,11,12,13}.

Existing research on the health effects of PM_2.5 and temperature also suggests plausible non-linear relationships and interactions that pose challenges for offering clear public health guidance when the exposures are co-occurring. Specifically, evidence suggests that the impact of short-term PM_2.5 on mortality exhibits a log-linear relationship^{4,14,15,16,17}, indicating that the health risks of PM_2.5 on days with moderate WFS do not differ significantly from those on days with severe WFS (Fig. 1). In contrast, the relationship between temperature and mortality is exponential, with a marked increase in health risks from moderate to extreme heat days^{18,19,20,21,22}(Fig. 1). Differences in how PM_2.5 and temperature individually relate to mortality suggests a potentially complex interplay when the exposures co-occur, which may result in different magnitudes of co-exposure interactions along the range of each exposure. Most public health advice to date has focused on potential interaction at high levels of each exposure²³.

Fig. 1: Hypothetical non-linear exposure-responses for short-term PM_2.5 and temperature on mortality risk during the warm season.

Most research and public health guidance has focused on the co-occurrence of high exposures.

Previous methods of studying the combined effects of PM_2.5 and temperature on mortality have often simplified this interaction by categorizing levels of exposure into quantiles^10,11,12. This approach has demonstrated how higher quantiles of temperature can amplify the mortality risks within higher quantiles of PM_2.5, and vice versa. However, such an approach might not fully capture the nuanced, non-linear dynamics of their potential interaction. Therefore, a more detailed understanding of the precise nature of this interaction is needed to identify specific combinations of short-term ambient PM_2.5 and temperature exposure levels that result in joint effects on population health. Such insights would benefit the development of public health policies and communication strategies by providing a clearer picture of when and how these co-exposures to environmental stressors pose human health risks²³.

Our study aimed to elucidate these complex interactions by estimating the combined effects of daily ambient PM_2.5 and temperature on community all-cause mortality over 13 wildfire seasons in the southwest region of mainland British Columbia (BC), Canada (Fig. 2a). By employing a time-stratified case-crossover design, we investigated the joint impacts of co-exposure across multiple categories of ambient PM_2.5 and temperature exposure. Rather than limiting the analysis of co-exposures to quantile categories, we also defined both the temperature and PM_2.5 exposure categories using specific ranges. Concentration-based categories align with existing air quality communication tools, such as the Canadian Air Quality Health Index (AQHI)²⁴ and the American Air Quality Index (AQI)²⁵, which are familiar to both the public and policymakers. Though quite nuanced regionally, temperature-based measures also inform when heat alerts are issued, such as Canada’s public weather alerts²⁶ and the Heat Warning system in the US²⁷. Fixed ranges thus potentially offer a direct and understandable way to assess co-exposure to PM_2.5 and temperature, facilitating clearer public health advice and enabling individuals and communities to make informed decisions about their exposure and protective measures. This approach should advance the translation of complex air pollution epidemiological findings into practical, actionable guidance, enhancing the effectiveness of public health messages regarding the risks of PM_2.5 exposure, particularly during periods of elevated temperatures.

Fig. 2: Context of study area and period.

a Map of the study area which includes residential postal codes (purple dots) within the Local Health Authority (LHA) boundaries (polygons) that intersect with the mainland southwestern region of the BC Heat Alert and Response System^58,60. The map lines delineate study areas and do not necessarily depict accepted national boundaries. b Time series plots of average calendar week PM_2.5 concentrations for the study area stratified by year, with the overall daily mean PM_2.5 and average WFS day threshold indicated with the horizontal blue and red lines, respectively. c Time series plots of average weekly temperature for the study area stratified by year, with the overall daily mean temperature and average EHE day threshold indicated with the horizontal green and dark red lines, respectively.

Results

Exposure and outcome summary

The study included 21,220 community deaths. Of these decedents, 57% were 65 years or older and 63.5% were male. The average age at death was lower for males (63 years) than females (71 years). On average, each death had 3.4 control days.

Less than half (44%) of all case and control days had a low PM_2.5 exposure ( < 6 µg/m³), and approximately one third (34%) were between 6 µg/m³ and 8 µg/m³ (Table 1). PM_2.5 exposure as a continuous measure differed significantly (p < 0.001) between case and control days. While the percentage of days impacted by WFS did not differ significantly between case (9%) and control days (8%), categories of PM_2.5 did (Table 1). There were also significant differences in temperature, temperature categories, and EHE days between the case and control days. The algorithm used in this study to define a WFS day confirmed the occurrence of such events in each month of our study period.

Table 1 Summary statistics for PM_2.5 and temperature exposures among all-cause deaths in greater Vancouver, British Columbia, Canada between 2010 and 2022

There was considerable overlap in the PM_2.5 concentrations and temperatures on days impacted by WFS and EHEs (Fig. 3). On days classified as having WFS impacts, the median [interquartile range, IQR] PM_2.5 concentration was 13.6 µg/m³ [11.1, 19.6] and the median temperature was 19.7 ^oC [16.5, 22.3]. Days without WFS impacts had a median PM_2.5 concentration of 6.1 µg/m³ [5.0, 7.2] and a lower median temperature of 14.2 ^oC [11.2, 17.2]. EHE days had a median PM_2.5 concentration of 12.1 µg/m³ [8.9, 15.8], with more than half (52%) of EHE days also having evidence of WFS impacts. Days without an EHE had a PM_2.5 concentration of 6.2 µg/m³ [5.0, 7.5], with only 6.0% of days without an EHE also exhibiting WFS impacts.

Fig. 3: Density plots of 24 h mean PM_2.5 and temperatures based on different day classifications (e.g., WFS and EHE days).

WFS refers to days impacted by smoke from wildfires and EHE refers to days impacted by extreme heat events.

There was notable overlap between days with moderate-to-high PM_2.5 and warmer temperatures (Fig. S1, Supplemental Materials). Specifically, on days when PM_2.5 was 16 µg/m³ or higher, the majority ( ~ 53%) had daily mean temperatures ranging between 20 °C and 26 °C (Fig. 2B). However, the most extreme temperature and PM_2.5 categories rarely co-occurred (e.g., 1.2% for ≥26 °C and ≥16 µg/m³ co-occurrence). Notably, 18% of days in the 12–14 µg/m³ PM_2.5 category had temperatures ≥26 °C. The lowest PM_2.5 days predominantly occurred during cooler periods (Fig. S1, Supplemental Materials).

Associations between continuous PM_2.5, temperature, and mortality

In crude models for community all-cause mortality, a 1-unit increase in PM_2.5 (µg/m³) was associated an odds ratio (OR) of 1.004 [1.001, 1.006] and a 1-unit increase in temperature (°C) was associated with an OR of 1.027 [1.023, 1.032]. After mutual adjustment, only temperature remained significantly associated with community all-cause mortality, and the OR for PM_2.5 was attenuated to the null. There was no evidence of a significant interaction between PM_2.5 and temperature on a continuous scale (Table 2).

Table 2 Crude and adjusted odds ratios (OR) with 95% confidence intervals for PM_2.5 and temperature (T) on a continuous scale and their interaction effect on all-cause mortality

Associations between WFS, EHEs, and mortality

In models unadjusted for EHE days (Table 3), WFS days had a crude OR of 1.05 [0.98, 1.11] and EHE days had a crude OR of 1.50 [1.40, 1.60] when not adjusted for WFS days. In the mutually adjusted model, EHE days had an OR of 1.56 [1.45, 1.67], whereas the OR for WFS days was attenuated to 0.91 [0.85, 0.97]. Notably, there was a significant negative interaction between WFS and EHE days (Table 4).

Table 3 Crude and adjusted odds ratios (OR) with 95% confidence intervals for binary indicator of wildfire smoke (WFS) and extreme heat event (EHE) days and their interaction on all-cause mortality

Table 4 Multiplicative interaction between wildfire smoke (WFS) and extreme heat events (EHE) with all-cause mortality

Associations between PM_2.5 and temperature deciles and mortality

In adjusted models using deciles of PM_2.5, the highest decile of PM_2.5 exposure, ranging from 10.1–121.0 µg/m³, was associated with an OR of 1.15 [1.08, 1.22], compared with PM_2.5 below the 50^th percentile. Odds of mortality was significantly elevated for all temperature categories over the 60^th percentile category, with the highest ORs for the highest percentiles (Table 5).

Table 5 Crude and adjusted Odds Ratios (OR) for percentile-based indicators of PM2.5 and temperature exposures

There were significant interactions between the > 90^th decile category of the PM_2.5 and the three highest temperature decile categories, with the highest interaction effects observed for the > 90^th decile of PM_2.5 when the temperature was in the > 90^th decile (OR = 1.66 [1.37, 2.00]) (Fig. 4a and Table S1, Supplemental Materials). Notably, the > 90^th decile for PM_2.5 exhibited its largest overlap with the > 90^th decile for temperature (4% of days) compared with all other PM_2.5 decile groups (Fig. 4b), suggesting that the interaction effects are concentrated within the > 90^th decile exposure combination. The stratified models were consistent with the interaction modeling results, with the highest ORs observed in the > 90^th decile for PM_2.5 when temperatures were also in the > 90^th decile (Fig. 5a, Table S2, Supplemental Materials).

Fig. 4: Joint effects of decile-based categories on mortality.

a Odds Ratios (ORs) and 95% CIs for the interaction effect of PM_2.5 and temperature deciles on all-cause mortality. Effect estimates are displayed in descending order. b Percentage representation in the dataset for the different PM_2.5 and temperature co-exposure decile groups.

Fig. 5: Forest plots of adjusted odds ratios (ORs) for decile-based categorical exposures from stratified models.

a PM_2.5 deciles estimated from models stratified by temperature deciles and (b) temperature deciles estimated from models stratified by PM_2.5 deciles. The reference groups in each model are the values ≤ 50^th percentile.

Associations between absolute PM_2.5 and temperature categories and mortality

Relative to PM_2.5 < 6 µg/m³, only the 10–12 µg/m³ and 12–14 µg/m³ exposure categories were significantly associated with increased mortality, with crude ORs of 1.19 (1.10, 1.29) and 1.62 (1.48, 1.78), respectively (Table 6). After adjusting for temperature categories, only the 12–14 µg/m³ PM_2.5 exposure category remained significantly associated with mortality, and the effect was attenuated. In contrast, temperature exposure categories were significantly associated with all-cause mortality in both crude models and models adjusted for PM_2.5 categories (Table 6). In contrast to the PM_2.5 exposure-response, the temperature exposure-response suggested an exponential relationship with mortality, with a very large upward shift in ORs when temperature was ≥26 ^oC.

Table 6 Crude and adjusted Odds Ratios (OR) for absolute PM_2.5 and temperature exposure categories estimated from conditional logistic regression models

Examination of joint effects between the absolute PM_2.5 and temperature categories also indicated non-linear effect modification. Specifically, the multiplicative interaction model (Fig. 6a) indicated the largest interaction occurred between the 12–14 µg/m³ PM_2.5 and ≥ 26 ^oC temperature categories. However, the wide 95% confidence intervals indicate substantial uncertainty in this interaction effect estimate, suggesting caution in its interpretation

Fig. 6: Joint effects of absolute value categories on mortality.

a Results from a multiplicative interaction model between absolute PM_2.5 and temperature exposure categories. b Percentage representation in the dataset for the different PM_2.5 and temperature co-exposure categories.

In the stratified models (Fig. 7a), the ORs for the PM_2.5 categories from 10 to 16 µg/m³ became substantially larger when stratified by the highest temperature category ( ≥ 26 ^oC, Fig. 7a). The exposure-response relationship across temperature categories also varied by the PM_2.5 categories (Fig. 7b), and these joint effects exhibited a non-linear pattern. Notably, the largest ORs were estimated for PM_2.5 exposure categories that were only moderately elevated.

Fig. 7: Forest plots of adjusted odds ratios (ORs) from stratified models based on absolute value exposure categories.

a Absolute PM_2.5 categories, estimated from models stratified by absolute temperature categories, and (b) absolute temperature categories, estimated from models stratified by absolute PM_2.5 categories. Because of limited overlap between certain categories, OR estimates were excluded for co-exposures that had fewer than 10 days within any cell of a stratified model because these represent unstable estimates.

Supplemental analysis

Days with extreme PM_2.5 and extreme temperature ( > 90th percentile) were associated with the highest odds of mortality compared with days without extreme PM_2.5 or temperature ( < 50^th percentile) (Table 7). The REOI and AP values above 0 indicate significant additive interaction, meaning the combined effect of extreme PM_2.5 and extreme temperature exceeds the sum of their individual effects. The synergy index (S) is further suggestive of a significant synergistic effect. Additionally, there is a large increase in excess risk (ER) of community mortality, changing from 0.11% [−1.1, 1.4] and 1.3% [0.14, 2.3] for extreme PM_2.5 only and extreme temperature only, respectively, up to 7.9% [6.1, 9.1] for the extreme co-exposure category. Such a pattern of a significant increase in the exposure-response due to high co-exposure is generally consistent with the above co-exposure models (Figs. 4a, 5, 6a, and 7).

Table 7 Summary of all-cause mortality odds ratios (OR), excess risk (ER), and additive interaction indicators for co-exposure to days extreme PM_2.5 and/or extreme temperature ( > 90^th percentile)

Sensitivity analyses focused exclusively on the EHE and WFS definitions to explore whether varying the respective criteria influenced our inferential conclusions for these binary exposure variables. Varying the EHE definition based on different averaging days (2- versus 3-days) and percentile thresholds (95^th versus 98^th) did not change any of the overall observation that EHE exposure is associated with higher odds of mortality (Table S7, Supplemental Materials). Importantly, the EHE definition used for inference in our study (3-day average using 98^th percentile threshold) resulted in the largest exposure-response relationship. For the WFS definition, using the HMS plume only to indicate a WFS day resulted in a much more attenuated exposure-response relationship.

Discussion

The findings from this time-stratified case-crossover study highlight the significant and complex relationships between PM_2.5, temperature, and community all-cause mortality during wildfire seasons in the greater Vancouver area over a 13-year period. The results emphasize the combined effects of these environmental stressors, which have profound public health implications, particularly in the context of increasing frequency and severity with climate change. Most notably, the community mortality effects of PM_2.5—which were temperature-dependent—are concentrated within a narrow range of moderately elevated concentrations, suggesting that such acute exposure may be under-appreciated in terms of our understanding of PM_2.5 associated mortality risks.

Our study demonstrates a clear and strong association between hotter temperatures and higher community mortality during the greater Vancouver area’s warmest months. As expected, the exposure-response relationship between temperature and mortality appeared to be exponential, with a pronounced increase in the odds of death as mean daily temperatures exceeded 26 °C. When using EHE days to define exposure, effects on mortality were also significantly elevated. These findings align with the broader literature on the health effects of elevated daily temperature and extreme heat, which consistently shows up across multiple cities and countries as a critical risk factor for daily mortality^22,28,29.

The relationships between PM_2.5 exposure and community mortality were more nuanced compared with temperature, and therefore deserving of greater discussion. Specifically, same-day PM_2.5 exposure did not consistently demonstrate a strong association with mortality across all PM_2.5 exposure categories, particularly when adjusting for temperature. Another recent study from China also found that the exposure response relationship between PM_2.5 categories on all-cause mortality resulted in stronger effect estimates at lower PM_2.5 exposure categories compared with higher PM_2.5 exposure categories. Specifically, PM_2.5 exposure categories of 0–35 μg/m³, 35–75 μg/m³, and 115–150 μg/m³ were associated with a mortality increase of 1.07 (1.01, 1.13), 1.03 (1.00, 1.05), and 1.05 (1.01, 1.08), respectively³⁰. A much larger study conducted across 652 cities in 24 countries found that the exposure-response relationship between acute PM_2.5 exposure and mortality was steepest in cities with lower levels of PM_2.5, lending further evidence of possible non-linear effects on mortality with effects concentrated in the lower spectrum of exposure³¹.

A pattern of PM_2.5 exposure-response attenuation was observed with all other PM_2.5 exposure metrics (continuous, WFS day, deciles, absolute ranges) after controlling for temperature or EHE exposure metrics. While other research has found attenuation in the health effect of PM_2.5 after adjustment for heat exposure³², a possible explanation of these findings are that adjusting for hotter temperatures, to isolate the independent effect of increasing PM_2.5 exposure, may constitute over-adjustment. Because we observed evidence of additive interaction between extreme PM_2.5 and temperature exposures (Table 3), controlling for temperatures as a confounding factor might obscure the true combined effect of these co-exposures on mortality. This could contribute to biased estimation of the association of PM_2.5 with mortality.

Our results suggested a non-linear relationship between PM_2.5 and mortality. Only one of the exposure categories (12–14 µg/m³) was significantly associated with higher mortality after adjustment for temperature (Table 7), with the ORs diminishing or becoming null at the highest PM_2.5 categories. Although less pronounced, there was also evidence of non-linearity with the decile-based exposure metrics. This paradoxical finding may have multiple explanations, such as biological saturation, where the cellular response to PM_2.5 reaches a limit (59), or behavioral changes during high PM_2.5 events, such as staying indoors and keeping windows and doors closed on high WFS days, which would reduce ambient PM_2.5 exposure and likely contributes to exposure misclassification in our study³³. Additionally, it is important to note that extremely high temperatures ( ≥ 26 °C) rarely coincided with days where PM_2.5 concentrations exceeded 16 µg/m³. If we assume that the effect of PM_2.5 is amplified by more extreme high temperatures, the lack of such heat occurrences during extreme high PM_2.5 events could lead to a muted overall effect of PM_2.5 on mortality. This suggests that the full impact of PM_2.5 on mortality might be underestimated in situations where extreme heat is absent or infrequent, and that it would be valuable to repeat similar analyses in locations where both PM_2.5 and temperature conditions are more extreme.

An important implication of the apparent non-linearity in the effects of PM_2.5 on mortality is that current public health guidelines, such as the 24 h PM_2.5 BC Air Quality Objective (AQO) of 25 µg/m³, may not adequately protect public health. In our study, most mortality impacts appear to have occurred at levels well below this threshold, a finding also reported by others^15,34,35,36. The data from this study, along with findings from other research^37,38, suggest significant WFS impacts beginning at PM_2.5 levels much lower than the AQO. Taken together, moderately smoky conditions—rather than strictly extreme smoky conditions—tend to be the driver of mortality impacts. This indicates that the current AQO may need to be re-evaluated and potentially lowered to better account for the non-linear relationship between PM_2.5 exposure and mortality risk in regions like greater Vancouver, especially during wildfire season when elevated PM_2.5 concentrations are more frequent (Fig. 1).

The key finding from this study is the significant joint effects between PM_2.5 and temperature on community mortality. Combined exposure to high levels of PM_2.5 and temperature significantly increased the odds of mortality, but these joint effects were complex. The combined effects were most evident when PM_2.5 exposure exceeded typical (i.e., median) background PM_2.5 levels, where the odds of mortality increased substantially with increasing temperatures. This joint effect was markedly higher when mean daily temperatures exceeded 26 °C. Analyses of percentile-based exposure categories were generally consistent with this finding. Moreover, the influence of increasing PM_2.5 levels on the magnitude of temperature effects also appeared to be non-linear, highlighting the complexities of this interaction.

Our results also suggest that a binary exposure classification (e.g., WFS or EHE) may be too crude when characterizing joint exposure-response relationships for WFS-derived PM_2.5 and extreme heat. For instance, an apparent negative interaction was observed between WFS and EHEs on community mortality. This paradoxical result may be partially explained by the observed non-linear interaction associated with joint PM_2.5 and temperature exposure. Moreover, our use of a binary exposure classification—especially for WFS—is subject to misclassification. Consistent with other studies from North America^32,39,40,41, our WFS definition relied heavily on HMS satellite data from NOAA, an imperfect measure of actual surface-level smoky conditions. The sensitivity analyses showed that the exposure-response was even more attenuated when relying on the simpler metric of HMS plume only.

The uncertainty of HMS plume data as an indicator of WFS exposure stems from inherent limitations in spatial and temporal measurements of the satellite product. First, they do not delineate the vertical location of smoke in the atmospheric column, adding uncertainty as to whether the smoke is aloft at high altitudes or at the surface, especially for regions far afield from wildfire activity. There is added uncertainty spatially for the smoke plume edges, driven by the coarse spatial resolution of the satellite imagery and imagery displacement effects that are spatially dependent. Finally, HMS plume data are a snapshot in time, thus incapable of capturing the dynamic nature of smoke plume dispersion over hours or days. Even within an HMS polygon, there can be variations in the density (low, moderate, high) of smoke levels, adding to uncertainty in assigning PM_2.5 to wildfires and defining WFS days⁴².

Our post hoc analysis suggested a greater-than-additive effect of same-day extreme PM_2.5 and temperature exposure. This finding aligns with recent case-crossover studies employing time-stratified analyses, which also estimated significant greater-than-additive effects of extreme PM_2.5 and temperature co-exposures^9,13. Rahman et al. ⁹ reported a 5.4% (95% CI: 2.3, 7.8) excess mortality risk attributable to same-day co-exposure to > 90th percentile average PM_2.5 and maximum temperature. Similarly, Xu et al. ¹³ estimated a 2.8% increase in the excess mortality risk from myocardial infarction was attributable to co-exposure to extreme temperature (2-day > 95th percentile) and PM_2.5 (> 95th percentile) in Jiangsu province, China. Here, we estimated an excess community all-cause mortality risk of 7.9% (95% CI: 6.1, 9.1) attributable to same-day co-exposure to extreme ( > 90th percentile) average daily PM_2.5 and temperature.

While the synergistic effects on excess mortality in our study are notably larger than those observed in the two case-crossover studies described above, this discrepancy may be explained by our focus on deaths that occurred outside of healthcare settings. People in the community, as opposed to those under care, may be more vulnerable due to several key factors. Firstly, although individuals in care settings such as long-term care facilities or hospitals may be biologically more susceptible to adverse health effects from environmental stressors, this susceptibility is often mitigated by factors like central air conditioning systems⁴³, which can significantly reduce exposure to extreme PM_2.5 and temperatures. For instance, one study in a Vancouver in-patient care facility during wildfire smoke episodes found indoor PM_2.5 levels to be substantially lower than outdoor levels (29.6 ug/m³ vs. 72.0 ug/m³ respectively)⁴⁴. Similarly, during the 2021 heat dome, the same facility reported indoor temperatures that were considerably cooler than those outdoors (23.6 °C vs. 27.2 °C), demonstrating the protective effect of such environments⁴⁵.

In contrast, individuals in the Vancouver community typically lack access to central AC⁴⁶ and reside in older buildings that exacerbate their exposure to environmental extremes. Recent data indicate that AC ownership in the Vancouver metropolitan area is significantly lower ( ~ 20%) compared to the national average in Canada (61%)⁴⁶. This disparity is particularly problematic at night, when sustained warmer temperatures limit the body’s ability to cool down.

Moreover, while individuals in care settings benefit from proximal access to health monitoring and immediate medical interventions, those in the community must often travel to emergency rooms for medical assistance, which can be critical during emergencies. The lack of constant health monitoring and ready access to transportation exacerbates their vulnerability, especially during extreme events like heatwaves when emergency medical services are strained⁴⁷. The 2021 BC coroner’s report highlights that most fatalities during BC’s 2021 EHE were among older adults with chronic diseases, living alone without adequate cooling or ventilation and resided in socially vulnerable communities⁴⁸.

There are significant public health implications regarding the synergistic effects of co-exposure to PM_2.5 and temperature on community mortality risk. However, fully understanding these implications requires a more refined and comprehensive analysis of when these synergistic effects are likely to occur in communities across differing levels of PM_2.5 and temperature exposures. To date, only a few studies, including two in California^9,49, one in Jiangsu province, China¹³, and this one have employed a time-stratified case-crossover design to investigate the joint effects of PM_2.5 and temperature on individual-level mortality, yielding largely consistent results. These individual-level studies align with a broader body of ecological research that also demonstrates synergistic effects and additive excess risk of all-cause and cause-specific mortality from environmental co-exposures^{12,50,51,52,53}. The consistent finding of an additive interaction between PM_2.5 and temperatures over the 90th percentile suggests a possible temperature threshold where the combined effects begin to significantly exceed the independent effects^9,50.

We observed a marked increase in the impact of PM_2.5 on community mortality when temperatures reached and exceeded 26 °C. While this specific temperature threshold may not be generalizable to populations beyond greater Vancouver, it indicates that using absolute temperature categories, rather than only quantiles, can elucidate the daily temperatures at which PM_2.5-related mortality risks are most severe. As noted earlier, the largest mortality effects for PM_2.5 occurred far below the provincial 24 h AQO in BC and within a limited exposure range. This range also had extensive overlap with elevated temperature days. We note that one limitation of our PM_2.5 and temperature categorization approach is that it may be less objective compared with quantile-based parameterizations. However, our analyses using quantile-based groupings showed a similarly complex interaction between PM_2.5 and temperature.

As climate change continues to drive more frequent and severe wildfires and heatwaves, population exposure to these environmental hazards is expected to increase and complicate conventional public health approaches¹. Our findings underscore the need for integrated public health strategies that address the combined risks of air pollution and extreme heat. Such strategies could include the development and implementation of early warning systems for combined PM_2.5 and temperature, public education campaigns, and targeted interventions for susceptible populations to mitigate the adverse health effects when these environmental stressors co-occur²³.

While this study provides valuable insights into the health effects of extreme heat and wildfire-related PM_2.5 co-exposure, there are several limitations. First, exposure misclassification is a potential concern, as we relied on ML-based PM_2.5 exposure estimation and satellite-based temperature measurements rather than measuring individual-level exposures. However, any misclassification is likely to be non-differential, which would bias results toward the null rather than create spurious associations, particularly when population exposure is common⁵⁴. Second, although the time-stratified case-crossover design inherently controls for time-invariant confounders at the individual level, unmeasured time-varying confounders cannot be ruled out. We tested for potential confounding by holiday status and found no meaningful or significant impact on effect estimates. Additionally, while the COVID-19 pandemic occurred during our study period, its influence on exposure to wildfire-related PM_2.5 and extreme heat is unlikely to introduce bias, as there is no strong evidence linking the pandemic to these environmental stressors. Lastly, given the relatively wide confidence intervals for some interaction effect estimates, caution is warranted when interpreting findings for specific exposure categories with lower precision. Despite these limitations, our study design and analytical approach help minimize bias, and our findings contribute to the understanding of the health risks associated with heat and wildfire smoke co-exposure.

This study provides timely insights into the all-cause community mortality risks associated with PM_2.5 and temperature co-exposures in the greater Vancouver area. Our findings indicate significant, non-linear, and potentially synergistic interactions between these combined environmental stressors, underscoring the need for a comprehensive public health approach that addresses their combined effects under climate change. The evidence also suggests that current air quality standards may not fully protect public health, because significant mortality impacts occurred at PM_2.5 levels well below existing thresholds. Additionally, the interaction between PM_2.5 and heat appear to be more pronounced at specific temperature thresholds, emphasizing the importance of tailoring public health interventions to account for local contexts and susceptibilities, including the integration of early warning systems and targeted interventions for at-risk populations.

Methods

This study was conducted in the greater Vancouver region of British Columbia (BC), Canada between 2010 and 2022, during the warmer months (May-October) when wildfire activity is typical in BC (“wildfire season”) (Fig. 2b)⁵⁵. This study area includes BC’s largest metropolitan regions, with a 2016 population (mid-point of study period) of approximately 2.46 million people, or approximately 53% of the provincial population⁵⁶. The summer climate in the study area is typically temperate with excellent air quality. Over the study period, the mean temperature was 14.7 ^oC and the mean PM_2.5 concentration was 7.4 µg/m³. Even so, this region experienced catastrophic EHEs in 2009 and 2021⁵⁷, and had prolonged periods of WFS exposure in 2010, 2014, 2015, 2017, 2018, 2020, 2021, and 2022 (Fig. 2b)⁵⁵.

We selected this sub-region of BC because we aimed to study an area with relatively uniform exposures to hot temperatures and WFS impacts over a geographically dense urban population. This approach is crucial, given the regional differences in how populations adapt to hot temperatures, which can influence sensitivity for various health outcomes, including mortality.

Analyses were restricted to all 6-digit postal codes within Local Health Authorities (LHAs) in the southwestern mainland region of the BC Provincial Heat Alert and Response System (BC HARS). The BC HARS comprises five geographic regions each with its own temperature threshold criteria for issuing heat alerts to protect public health⁵⁸. Postal codes were used to define the study area because they represent the smallest geographic unit recorded in the BC vital statistics records, which are described in more detail below. Six-digit postal codes are expressed as geographic points and linked to Canadian census geographies using the national Postal Code Conversion File. These are called single link indicators (SLIs) and they represent the geographic area with the majority of dwellings assigned to a particular postal code (analogous to population centroid)⁵⁹.

Mortality and exposure data

Death records between 2010 and 2022 were extracted from the BC Vital Statistics Agency database, consistent with our previous work on the relationship between temperature and mortality in BC^58,60. In this dataset, each death record includes the date of death, primary cause of death classified according to the International Classification of Diseases, 10^th revision (ICD-10), age, sex, setting of death (e.g., acute care, private home, etc.), and the 6-digit postal code of the location where the death occurred⁶⁰. We included all recorded community deaths of individuals with a BC residential postal code in the study area. Community deaths are defined as those that occurred outside of a care-providing facility, such as a hospital or long-term care home. We focused on community deaths based on our prior internal work showing they are more strongly associated with PM_2.5 and temperature than mortality in all locations⁶¹. The weaker association with all deaths is likely due in part to people in care-providing facilities having protection from exposures due to air filtration and air conditioning requirements, as well as receiving medical or supportive care.

To provide a comprehensive assessment of the relationships between PM_2.5, temperature, and all-cause community mortality we generated multiple exposure metrics (Table 8). These ranged from conventional measures such as continuous and binary exposures, to metrics with multiple exposure categories. The specific metrics are described in more detail below, and all exposure metrics were linked with the mortality data using the 6-digit residential postal code of each decedent.

Table 8 Summary of daily mean PM_2.5 concentration, daily mean temperature (T), wildfire smoke (WFS), and extreme heat event (EHE) exposure metrics and their use in epidemiological analyses

Daily PM_2.5 concentrations were assigned to all 6-digit postal codes in the study area by intersecting the geographic points of postal codes with predicted daily grid cell PM_2.5 concentrations from the Canadian Optimized Statistical Smoke Exposure Model (CanOSSEM). CanOSSEM is a comprehensive machine learning model that combines satellite remote sensing data to optimize the estimation of PM_2.5 concentrations from WFS and residential wood burning. CanOSSEM predictions of PM_2.5 were generated at a spatial resolution of 5 km x 5 km grid cells, covering all populated areas of Canada, with high reported accuracy⁶².

To estimate the effect of PM_2.5 pollution on all-cause community mortality and its interactive effects with different temperatures, we analyzed PM_2.5 as a continuous exposure and alternatively classified PM_2.5 into distinct exposure categories (Table 8). Specifically, we categorized the daily PM_2.5 concentrations values into deciles and used values at or below the 50^th percentile as the reference group to compare against each of the decile-based exposure group above the 50^th percentile. As an alternative to using deciles, we also grouped PM_2.5 concentrations as <6 µg/m³, ≥6–8 µg/m³, ≥8–10 µg/m³, ≥10–12 µg/m³, ≥12–14 µg/m³, ≥14–16 µg/m³, ≥16 µg/m³. The reference category of <6 µg/m³ was selected because it approximates the overall PM_2.5 median during the study period (median = 6.1 µg/m³) and reflects periods where ambient air quality is not impacted by WFS in the study area (Fig. 2B). Establishing categories in 2 µg/m³ bins allows flexibility to explore how specific co-exposures to PM_2.5 and temperature categories are associated with mortality in a potentially non-linear relationship.

We also derived a binary indicator of WFS exposure by identifying days impacted and not impacted by WFS, using the approach described by O’Dell and colleagues (28). First, we identified grid cells in CanOSSEM that intersected with a smoke plume, indicated by the daily Hazard Mapping System (HMS) smoke plume polygons provided by the US National Oceanic and Atmospheric Administration (NOAA)⁶³. Based on assessment of satellite remote sensing data, HMS data provide an indication whether a plume was present in the atmospheric column (i.e., overhead) for an observed area. For each grid cell, we then calculated the long-term mean (mean_{no_plume}) and standard deviation (SD_{no_plume}) of CanOSSEM PM_2.5 concentrations on days that did not intersect with an HMS polygon. Next, we added the mean_{no_plume} and 1.5*SD_{no_plume,} which represents a grid cell specific threshold for WFS impacts. For a grid cell to be defined as impacted by WFS on any given day, two criteria had to be met: (1) the grid cell intersected with an HMS plume and (2) the PM_2.5 concentration exceeded the mean_{no_plume} + 1.5* SD_{no_plume} threshold. If these criteria were not met, the grid cell and intersecting postal code point(s) were classified as having no WFS impact on that day.

For temperature exposure assessment we used the average daily (24 h) value for each CanOSSEM grid cell and the overlapping postal code point(s), which were derived from the Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2)⁶⁴. The MERRA-2 has a spatial resolution of 0.5° × 0.625°.

Temperature exposure metrics were then derived using continuous exposure and the same type of decile and absolute categorical exposure classifications used for PM_2.5. The absolute temperature categories were: <14 °C, ≥ 14–17 °C, ≥17–20 °C, ≥20–23 °C, ≥23–26 °C, ≥26 °C. We chose < 14 °C as the reference category because it approximates the median temperature during the study period (median temperature = 14.5 °C).

Deciles of temperature exposure were calculated using the daily temperature values across the defined study area. Days with a mean temperature below the regional median ( < 50^th percentile) of mean daily temperature were assigned as a reference group, with increasing increments of deciles assigned for all other days (50^th–60^th, 60^th–70^th, 70^th–80^th, 80^th–90^th, >90^th).

Because EHEs typically involve consecutive days of exposure, we derived an EHE metric that accounts for this “heat wave” effect^65,66. Specifically, for each day, we calculated the consecutive 3-day average temperatures (2-days prior, 1-day prior, and same day) for each CanOSSEM grid cell. Next, we calculated the month-specific 98^th percentile of 3-day average temperatures by grid cell for the entire study period^{66,67,68,69,70}. This 98^th percentile was then used as a threshold to derive an indicator variable for extreme heat event (EHE) days within each postal code.

To facilitate comparison of our results with existing literature, we defined a co-exposure metric using mutually exclusive extreme exposure groups, similar to Rahman et al. ⁹ and Xu et al. ¹³. These two studies also used a case-crossover design to investigate the effects of co-exposure to PM_2.5 and temperature extremes on mortality. Like these studies, we classified days as having extreme PM_2.5 (>90th region-wide percentile) and/or extreme temperature ( > 90th location-specific percentile) and fit a model, using days without extreme PM_2.5 or temperature as the reference group, to estimate the effect of the following mutually exclusive exposure groups: extreme PM_2.5 days only, extreme temperature days only, and days with both extreme PM_2.5 and extreme temperature.

Study design

The study employed a time-stratified case-crossover design, in which every case acts as its own control. The case-crossover design is useful for examining specific individual-level acute health outcomes (i.e., all-cause mortality) associated with short-term exposures (i.e., PM_2.5 or temperature). By comparing exposure levels on the day of the event (case days) with those on similar days when the event did not occur (control days), the effect of the exposure on the event can be assessed. This approach helps adjust for unmeasured long-term and seasonal variable trends. Furthermore, it automatically adjusts for all individual-level time invariant (or slowly varying) characteristics (e.g., age, sex, lifestyle, comorbidities) as the comparisons are conducted within subject. We applied the time-stratified method, which selects control days from the same calendar month and weekday as the case day, minimizing biases related to time trends, seasonal exposures, and health event variations⁷¹.

Statistical analyses

Multivariable conditional logistic regression models were used to estimate the independent and joint associations between the exposure metrics and all-cause mortality. We fit several different models to provide a comprehensive view of the complex associations. All models were fit with decedents as their own stratum. For the stratified modeling, we subset the data to individuals who had a case or control day that fit the respective criteria. For example, a death was selected for a stratified analysis if any of its case or control days fell into the stratum of interest (e.g., ≥26 °C). Multiplicative interaction terms using most of the PM_2.5 and temperature exposure metrics were fit to test for statistical interactions on mortality. We conducted the analyses using the clogit function from the survival package (version 3.4-0) in R (version 4.2.2). The full modeling results are provided in the Supplemental Materials (Tables S1–S6).

We performed post hoc analyses evaluating additive interaction effects of extreme PM_2.5 and temperature by using the mutually exclusive extreme exposure groups^9,13. To test additive interaction, we estimated relative excess odds due to interaction (REOI), proportion attributable to interaction (AP), and synergy index (S)^13,72. We calculated confidence intervals (CIs) for these measures using the delta method, which approximates standard errors based on the regression estimates and their variances, following the approach recommended by Hosmer and Lemeshow⁷³. We also adapted the approach in Rahman et al. ⁹ to estimate all-cause excess risk (ER) of mortality due to co-exposure. Here, a bootstrapping procedure with 1,000 iterations was applied, resampling from the data with replacement to generate an empirical distribution of ER estimates. The 95% CIs for ER were obtain using the percentile method, taking the 2.5^th and 97.5^th percentiles of the bootstrap distribution.

Sensitivity analyses were conducted to address certain assumptions in our analyses. Specifically, other analyses explored the sensitivity of mortality models to varying the EHE definition based on the number of averaging days (2-days, 3-days) and the percentile used for deriving the EHE temperature thresholds (95^th percentile, 98^th percentile). Finally, we also tested the sensitivity of mortality models using an alternate WFS day binary definition by defining a day impacted by WFS where there was an HMS smoke plume detected without any PM_2.5 concentration threshold criterion.