The potential impact of wildfire smoke on COVID-19 cumulative deaths in the San Diego-Tijuana border region

2020 broke records for the most active fire year on the West Coast, resulting in the worst air quality observed in decades. Concurrently, the public health threat of COVID-19 caused over 1 million deaths in the United States (US) and Mexico in 2020 and 2021. Due to the effect of air pollution on respiratory diseases, wildfire-specific particulate matter is a hypothesized driver of COVID-19 severity and death. Capitalizing on wildfire smoke that hit the San Diego-Tijuana border region in September 2020, we applied synthetic control methods to explore its potential differential role in affecting COVID-19 mortality on both sides of the border. Daily data on COVID-19 cumulative deaths for US counties were obtained from the CDC COVID tracker and data for Mexican municipalities was obtained from the Mexican Secretary of Health. Counties and municipalities with wildfire smoke exposure were identified using the National Oceanic and Atmospheric Administration hazard mapping smoke product (HMS); a day where 90% of the area was covered by smoke was considered exposed for the main analyses. Unexposed counties/municipalities were considered as potential controls. The San Diego-Tijuana border region was covered by dense smoke by the 7th of September; 707 COVID-19 deaths had occurred in San Diego and 1367 in Tijuana. While a slight increase in cumulative mortality was observed in San Diego, no change was found in Tijuana; neither estimate indicated a strong precise effect of wildfire smoke on COVID-19 mortality. We hope this study will serve as an illustration of how border contexts can be used to investigate differential vulnerability to wildfire smoke for infectious diseases. Examining the interactive effect of COVID-19 and smoke can help in recognizing the implications of these dual health risks which will be increasingly important as wildfires become more frequent and severe in the context of climate change.


Introduction
Over 1 million deaths from COVID-19 occurred in the United States (US) and Mexico in 2020 and 2021 only [1]. Although biomedical understanding of the disease has drastically improved since the virus emerged in early 2020, many unknowns remain about what individual and environmental factors increase vulnerability to severe disease [2,3]. COVID-19 infections can range from being asymptomatic to causing premature death, and the case fatality rate varies with a range from approximately 1% in the general population to up to 37% for those admitted to intensive care units [4]. Understanding the role of environmental health risks in driving severe infection and disease can provide insight on the biological risk factors and help to recognize and understand the implications of these potentially compounded health risks.
The threat of the COVID-19 pandemic was concurrent to raging wildfires and record high smoke exposure in the summer and fall months of 2020 in the US West Coast [5]. Wildfire smoke includes high concentrations of inhalable ambient particulate matter, such as PM 2.5 , particles under 2.5 µm in aerodynamic diameter composed of metals, ions, organic compounds, or other materials that are harmful to human health [6]. The inhalation of particulate matter can produce oxidative stress and inflammation due to the generation of reactive oxygen species which triggers cellular damage and may increase the risk of cardiopulmonary disease [6]. Furthermore, the smallest particles can cross the alveolar membrane to the bloodstream and cause systemic effects on the human biological system [7]. It has been shown that pulmonary response to these stressors can affect airway function and decrease resilience to viral and bacterial infections [8]. As wildfire smoke and COVID-19 affect the same biological systems and organs, wildfire-specific particulate matter has been hypothesized and shown to increase the risk of severe infection due to its widespread effect on the respiratory and cardiovascular system [9]. There is strong evidence of the adverse health effects of wildfire smoke [10][11][12] and emerging evidence of its effects on COVID-19 transmission and severity [5,[13][14][15][16][17][18][19][20]. However, the majority of epidemiological studies on wildfire smoke have been conducted in the US, Canada, and Australia, with little evidence from other contexts [10].
Generating evidence from regions with differing population demographics and social conditions is important as effects of wildfire smoke differ based on the study population and their specific vulnerability. Socio-economic factors increase susceptibility to air pollution impacts by increasing risk of other conditions such as asthma and/or affecting development and resistance to other disease threats [21]. Also, neighborhood-level risk factors faced by those living in low-income neighborhoods may increase their susceptibility to environmental hazards [22]. Stressors, such as crime, noise, and traffic can lead to acute and chronic changes in the functioning of body systems and increase the effect of exposures such as wildfire smoke on the biological system [23]. Understanding the role of social factors in driving vulnerability to wildfire smoke is critical to informing and prioritizing efforts to reduce the burden and protect populations that are most at-risk [24].
Border regions, having similar climatic and environmental conditions but differing cultural and social contexts, offer a unique opportunity to study the role of social vulnerability in driving health impacts of environmental exposures. The San Diego-Tijuana border is one of the most densely populated international crossing in the world, and the largest bi-national conurbation shared between the US and Mexico [25]. Although both cities are juxtaposed, the socio-economic context in San Diego and Tijuana are severely different. The United Nations Development Program ranks the US 21st on the Human Development Index rankings, Mexico is ranked 86th [26]. Furthermore, Tijuana is ranked to have the 4th worst quality of life index out of all Mexican cities based on geographical, environmental, and social factors [27]. In contrast, San Diego is ranked the 5th richest city in the US [28]. Such differing social conditions can be used as a natural experiment to better understand how population socio-demographics modify the health risk of environmental stressors.
While international borders are a physical barrier to human mobility that leads to differing social contexts, environmental exposures such as wildfires igniting on either side of the border can drive harmful smoke exposure to the entire region. The San Diego-Tijuana border region is particularly affected by extreme weather events. Climatic trends resulting from global warming have increased the likelihood of fire weather, or periods that have a high probability of fire due to high temperature, low humidity, low rainfall, and/or high winds [29]. The amount of forest area burned by wildfires was found to be more than ten times greater from 2003-2012 than it was in 1973-1982 in the Western US [30]. Furthermore, models have shown that precipitation is expected to decrease by 50% in Southwest California and decrease by up to 75% in Baja California [31], increasing the risk of wildfires in future years. Predictions indicate that wildfire emissions will increase by 19%-101% by 2100 in California [32]. Furthermore, 2020 broke records for the most active fire year in California [33], and the smoke emitted by wildfires extended across the border to cities like Tijuana. Changes in economic and transport activity from the COVID-19 pandemic and associated lockdown also played a role in reducing global ambient PM 2.5 concentrations by 31% in the first months of the pandemic [34], therefore wildfire smoke may be a greater contributor to harmful air pollution exposure in this pandemic context.
Capitalizing on the timing of the 2020 wildfire smoke event that hit the San Diego-Tijuana border region and using other counties in the US and municipalities in Mexico that did not have smoke during this period, we applied synthetic control methods to understand the role of wildfire smoke in affecting COVID-19 mortality. This methodology has been previously applied to study the impact of wildfire smoke on COVID-19 case fatality ratios in the San Francisco Bay Area [13] and here, it is applied to investigate its role in cumulative COVID-19 deaths in the San Diego-Tijuana border region hypothesizing differential impacts on both sides of the border. Understanding the role of wildfire smoke as a risk factor for severe COVID-19 and differing vulnerability factors can be used to inform measures to protect populations most at risk in the context of recurring wildfires and the ongoing pandemic.

Overview of methodology
A synthetic control methodology was applied to study the impact of wildfire smoke on daily COVID-19 cumulative mortality in each San Diego County and the Municipality of Tijuana [35]. This approach capitalized on the timing of wildfire smoke as a natural experiment to estimate the short-term changes in COVID-19 cumulative mortality. Using data from areas not affected by the fire (control units), the methodology identifies a synthetic control that best estimates the COVID-19 cumulative mortality before the wildfire smoke starts which then represents the counterfactual of what the trend would have been if the smoke exposure had not occurred. Each county in the US and municipality in Mexico that did not have wildfire smoke exposure was considered to build synthetic controls for San Diego and Tijuana separately. The synthetic controls were identified separately so that San Diego County and Municipality of Tijuana are only compared to other counties/municipalities within the same country since US counties and Mexican municipalities may not be comparable. A counterfactual synthetic control was constructed by weighting control units to most closely match the level and trend of the treated unit before exposure. If an appropriate match is found (i.e. the synthetic control groups and exposed groups have the same trend before the wildfire event), the difference in the counterfactual synthetic control units and the treated unit post-wildfire exposure can be interpreted as the effect of interest. Generalized synthetic control methods are an extension of the synthetic control approach [36] which has more flexibility, particularly in controlling for time-varying covariates; it is considered more efficient than the traditional synthetic control [36,37]. Lastly, it uses bootstrapping to estimate uncertainty (such as 95% confidence intervals) that are useful in the interpretation of results. A traditional synthetic control was also applied as a sensitivity analysis and presented in supplementary materials.

Data sources
Daily COVID-19 cases and mortality were obtained from the CDC COVID-19 tracker for counties in the US [38] and from the Secretary of Health Epidemiological Surveillance System for Viral Respiratory Diseases for municipalities in Mexico [39] which are counts of COVID-19 deaths based on the patient's residence county or municipality. Daily mobility patterns for each state for both the US and Mexico were extracted from the Google Community Mobility Reports [40]. Differences in movement trends from residential locations were considered a proxy for population dynamics and accounting for potential differences in COVID-19 measures within each country [41]. Smoke imagery was obtained from the National Oceanic and Atmospheric Administration (NOAA) hazard mapping smoke product (www.ospo.noaa.gov/Products/land/hms.html) for the US and Mexico, which identifies smoke plumes from wildfires applying algorithms from visible imagery from various satellites [42]. These estimates were revised and modified by trained analysts to improve accuracy. Aerosol Optical Depth information was collected from the Geostationary Operational Environmental Satellite Aerosol and Smoke Product and smoke plumes were categorized by NOAA's Office of Satellite and Product Operations as light, medium, or heavy categories, corresponding to PM 2.5 concentrations of 0-10, 10-21, and 22+ µg m 3 , respectively [43]. Satellite imagery for the border region was obtained from the National Aeronautics and Space Administration Worldview visualization tool [44].

Analysis
The percentage of a county or municipality covered by heavy smoke was estimated for each day from July through September 2020 using NOAA's HMS product. Smoke exposure was considered to start the first day that over 90% of a county or municipality was exposed to heavy smoke based on this product; sensitivity analyses were conducted using 70% and 100% smoke coverage. The package 'gsynth' in R v4.1.0 [45] was utilized to apply a generalized synthetic control, which can be viewed as an extension of the canonical synthetic control approach. A traditional synthetic control was also applied using the 'synth' package Stata 16 SE [46] as a sensitivity analysis. These approaches identify counties (US) and municipalities (Mexico) that had a similar trend to each unit before the smoke exposure occurred. Any county or municipality with smoke exposure during the entire study period based on the exposure definition was excluded as a potential control so only areas with no wildfire smoke during the study period were eligible for selection for the synthetic control. A weighting procedure was applied to identify the most suitable synthetic control trends in the pre-wildfire smoke period and considered to represent the counterfactual trend to the treated unit for the post-wildfire smoke period. For the generalized synthetic control method, controls are decomposed with calendar time, lagged outcomes (in the pre-exposure period) and time-varying covariates, and an interactive fixed effects model is applied to identify a suitable control based on this time series in the pre-treatment period [37]. This informs a reweighting approach in which weights for control units are selected based on these decomposed estimates and are used to impute hypothetical trends for the treated unit if they had not been treated by predicting the outcome in the treated unit during the post-treatment period [36,37]. Generalized synthetic control has more flexibility in controlling for time-varying covariates, and daily mobility, COVID-19 cases same day, COVID-19 cases previous day, and the standard deviation of COVID-19 cases the four previous weeks were included as covariates to account for potential differences in the COVID-19 infection rates that vary day to day and could be affected by smoke. The difference in the trend after the wildfire smoke can be considered as the effect of this exposure on the COVID-19 cumulative mortality. The advantage of this methodology is that the analysis itself accounts for any difference between counties/municipalities by finding the best possible fit in trends before the wildfire smoke occurs. By doing so, any difference in socio-demographics, COVID-19 burden, or political action affecting the outcome of interest that is not time-varying is accounted for by design. Other variables included in the model were to account for potential differences in trends related to the progression of the pandemic and related public measures between regions and identify weighted controls that have similar COVID-19 transmission and burden. Missing case and mobility data were interpolated using the 'mipolate' command in STATA, which uses linear interpolation using known values before and after the missing values [47]. Placebo tests were conducted considering the timing of the wildfire smoke exposure to have started on 20th September, four days after wildfire smoke stopped affecting the region.

Results
The study period analyzed is from 1 July to 30 September 2020. Smoke from wildfires burning throughout California started to cover the San Diego-Tijuana border region in early September 2020. By 7 September the border region was covered by dense wildfire smoke which continued for ten days following the start of the smoke ( figure 1, table S1). From the start of 2020 to the first day wildfire smoke covered the border region, a total of 707 COVID-19 deaths had occurred in San Diego and 1376 in Tijuana. By 30 September 2020, at the end of the study period, 783 COVID-19 deaths had occurred in San Diego and 1491 in Tijuana (table 1). However, the number of positive diagnosed COVID-19 cases varied drastically between both sides of the border, with over 55 000 cases in San Diego and less than 8000 cases in Tijuana during the study period (table 1). This is partially driven by differences in testing, and likely more unreported cases in Tijuana.
For both San Diego and Tijuana, a synthetic control was identified according to the trend of cumulative mortality in each (figure 2). Although the synthetic control prediction was not matched seamlessly for San Diego or Tijuana, it did correspond to the trend in cumulative mortality for the majority of the pre-treatment period. The difference in cumulative mortality for the pre-treatment period is described in supplementary table S2 and the difference between the trend in San Diego and Tijuana never exceeded six deaths, with most of the differences of less than 2 (table S2). The counties/municipalities identified with the traditional synthetic control analysis and their weights are included in the supplementary material (table S3).
For San Diego, a slight increase in cumulative mortality was observed in the County when comparing to the trend in the synthetic control in both the main analysis using generalized synthetic control and the sensitivity analyses (figures 2, S1, and S2). In Tijuana, the cumulative mortality follows approximately the same trend as its synthetic control, showing no change after the wildfire smoke started. Although results suggest a small effect of wildfire smoke on COVID-19 cumulative mortality in San Diego, no effect is observed in Tijuana (figure 2). However, neither estimate shows a strong and precise effect at the 95% confidence level, suggesting that we are not able to confirm a robust effect of wildfire smoke on cumulative COVID-19 mortality in this context. Results of placebo tests using 20 September as the wildfire smoke exposure start date (when wildfire smoke had cleared the region) showed no effect in either San Diego or Tijuana ( figure S3).

Discussion
Overall, we did not observe a strong effect of wildfire smoke on COVID-19 cumulative mortality in the San Diego-Tijuana border region. In San Diego, we observed a slight increase in cumulative mortality (figure 2). In Tijuana, no effect was observed as the synthetic control followed a very similar trend to what was observed in the Municipality of Tijuana after the wildfire smoke exposure started affecting the region. Tijuana is socially disadvantaged when compared to San Diego; according to the 2010 census in both countries, the poverty rate was three times higher in Tijuana than San Diego while the percentage of the population with a high school education was three times higher in San Diego [48,49] (table S4). It was hypothesized that this  could lead to a differential effect of wildfire smoke due to the increased social vulnerability of Tijuana.
Although we cannot infer that wildfire smoke drove a change in COVID-19 mortality and that social vulnerability played a role in this context, we hope that these results can highlight the need to continue to study and understand the compounded impact of these joint exposures and associated vulnerability factors.
Examining the interactive effect of COVID-19 and extreme weather can help in recognizing the implications of these dual health risks and can be used to inform measures to protect those that are most vulnerable during wildfires. Several recent studies have identified links between particulate matter air pollution and COVID-19 severity [15,17,19]. A study considering the effect of PM 2.5 on COVID-19 in western US counties during the 2020 wildfires found that short term exposure to particulate matter drove higher COVID-19 cases and deaths [14]. However, considerable variability was observed with many counties showing no effect and some even indicating a negative association [14]. Similarly, previously published work by coauthors of this work found variability in the effect of wildfire smoke on COVID-19 case fatality ratios in the San Francisco Bay Area with some counties showing a precise association and others counties showing no effect [13]. Yu and Hsueh found the effect of wildfire smoke on COVID-19 mortality in California to be moderated by the availability of hospital and public housing resources at the county level and disproportionally affects counties with higher social vulnerability [20]. Our results in the context of the existing literature further reinforce the heterogeneity of the role of wildfire smoke on COVID-19 severity. Differences in patterns observed in these studies and in our work could be attributed to a range of factors including differences in COVID-19 response measures, population demographics, behavioral risks and protective factors. Previous research and this current work highlight the need to further disentangle the drivers behind the heterogeneous effects observed thus far.
Wildfire smoke and air pollution may play a role in the COVID-19 pandemic through other pathways. For instance, studies have found that wildfire smoke is not only a driver of disease severity, but can also increase incidence, test positivity rates and case rates [16][17][18]. A study in Reno found that a 10 µg m −3 increase in the seven-day average PM 2.5 concentration increases SARS-CoV-2 test positivity rate by 6.3% [50]. Individual-level studies in Sweden and the US have also shown that short-term exposure to particulate matter is associated with a higher risk of SARS-CoV-2 positive test results and COVID-19 mortality [51,52]. In the constantly evolving COVID-19 pandemic context, more evidence is needed to understand how air pollution or wildfire smoke may interact or modify the protective effects of vaccines, for example. There are many interesting avenues for future studies to explore the links between extreme weather such as wildfire smoke and various expressions of the COVID-19 pandemic.
We acknowledge our current work has limitations. First, we only consider one county in the US and one municipality in Mexico as case studies, i.e. we were not able to evaluate potential geographic differences between regions or within each country. Also, wildfire smoke is defined based on satellite imagery, therefore exposure misclassification is possible in this context since it may also capture smoke higher in the atmosphere. Nonetheless, validation of the HMS smoke product with ground-level monitors has been shown correlation with PM 2.5 concentrations [43]. Additionally, we only focused on cumulative mortality as the outcome of interest. There are many other measures of COVID-19 burden that would be worth exploring in future work such as case fatality rates, hospitalization rates and symptom severity; we chose to focus on cumulative mortality to retain consistency across datasets between both countries. Also COVID-19 deaths recorded at the county/municipality level based on patient's residence were used but there could be regional variations in how the data was recorded and particularly between San Diego and Tijuana. Furthermore, Tijuana experienced more COVID-19 deaths earlier in the pandemic which could decrease the pool of potentially vulnerable people and attenuate the impacts of wildfire smoke. Lastly, many residents of Baja California sought medical care in the US when hospitals exceeded capacity during waves of the COVID-19 pandemic [53]; this could play a role in results we observe. As the first study to consider this research question in a border region, we hope it will motivate further research to broaden this analysis to additional regions of the US-Mexico border region and other border regions around the globe.
In conclusion, we used the San Diego-Tijuana border as a unique setting to explore the potential differences in the effect of wildfire smoke on either side of the border, using it as a natural experiment to explore differential health risks related to socio-economic context. It is critical to continue to study what drives susceptibility of these compounded health risks to best protect populations from these dual public health risks.

Data availability statement
The data that support the findings of this study are openly available at the following URL/DOI: https:// github.com/benmarhnia-lab/border_smoke_covid [54].