Abstract
Background and objective Ecological studies indicate ambient particulate matter ≤2.5mm (PM2.5) air pollution is associated with poorer COVID-19 outcomes. However, these studies cannot account for individual heterogeneity and often have imprecise estimates of PM2.5 exposure. We review evidence from studies using individual-level data to determine whether PM2.5 increases risk of COVID-19 infection, severe disease, and death.
Methods Systematic review of case-control and cohort studies, searching Medline, Embase, and WHO COVID-19 up to 30 June 2022. Study quality was evaluated using the Newcastle-Ottawa Scale. Results were pooled with a random effects meta-analysis, with Egger’s regression, funnel plots, and leave-one-out and trim-and-fill analyses to adjust for publication bias.
Results N=18 studies met inclusion criteria. A 10μg/m3 increase in PM2.5 exposure was associated with 66% (95% CI: 1.31-2.11) greater odds of COVID-19 infection (N=7) and 127% (95% CI: 1.41-3.66) increase in severe illness (hospitalisation or worse) (N=6). Pooled mortality results (N=5) were positive but non-significant (OR 1.40; 0.94 to 2.10). Most studies were rated “good” quality (14/18 studies), though there were numerous methodological issues; few used individual-level data to adjust for confounders like socioeconomic status (4/18 studies), instead using area-based indicators (12/18 studies) or not adjusting for it (3/18 studies). Most severity (9/10 studies) and mortality studies (5/6 studies) were based on people already diagnosed COVID-19, potentially introducing collider bias.
Conclusion There is strong evidence that ambient PM2.5 increases the risk of COVID-19 infection, and weaker evidence of increases in severe disease and mortality.
Funding This review was completed as a Scholarly Intensive Placement project by NS, which received no funding.
Competing interests The authors declare no competing interests.
Registration This study was registered on PROSPERO on 8 July 2022 (CRD42022345129): https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42022345129
Introduction
A considerable body of ecological evidence suggests fine particulate matter air pollution ≤2.5μm (PM2.5) increases the risk of COVID-19 infection, severity, and death (1–3). However, this evidence relies on comparisons of geographic units, which do not account for individual-level differences and often misclassify exposures due to poor precision/resolution in PM2.5 estimates (3). Associations between PM2.5 and COVID-19 may therefore be spurious, confounded by socioeconomic differences that influence exposure to air pollution and COVID-19 risks (4).
Nevertheless, are several reasons to suspect that PM2.5 increases COVID-19 risks. PM2.5 increases expression of Angiotensin-Converting Enzyme 2 (ACE2), which the COVID-19 spike protein uses to bind to and enters host cells (3,5). Though there is limited evidence for ambient PM2.5, studies of cigarette smoking suggest it inhibits cell defence against infections (5). PM2.5 and COVID-19 may also operate in tandem, both independently worsening respiratory and cardiovascular health, leading the combination of exposures to increase the likelihood of severe disease and death (3,6).
This systematic review builds on previous reviews (1–3) by focusing on studies using individual-level data that can provide more precise exposure estimates and better account for confounders. We address the following questions:
Is ambient PM2.5 exposure predictive of COVID-19 infection, severe COVID-19 disease, and COVID-19 mortality?
Do discrete high PM2.5 events like wildfires predict COVID-19 infection, severe COVID-19 disease, and COVID-19 mortality?
Methods
This review is registered on PROSPERO (7) and is reported according to PRISMA 2020 guidelines (8). A completed PRISMA checklist is available on a public repository (9).
Inclusion/exclusion criteria
To be eligible for inclusion, studies had to analyse individual-level data on the association between PM2.5 and COVID-19 infection, severity, or mortality using either a case-control or cohort design. Studies needed to present original research in an English-language peer-reviewed journal no later than 30 June 2022.
Studies were ineligible if they used ecological, cross-sectional, case-series, animal, or in-vitro designs; studies with a mixture of methods that included either case-control or cohort design were considered eligible. Hypothesis, review, editorial, commentary, and opinion pieces were excluded, as were pre-prints and conference presentations. Studies not using PM2.5 or only examining indoor air pollution or tobacco smoke as the pollutant exposure were excluded.
Search strategy and screening
We searched Medline, Embase and the World Health Organization COVID-19 database using terms listed in the Appendix. In addition, we screened the reference lists of grey literature and previous systematic reviews on similar topics for studies meeting the inclusion criteria. Two study authors (NS & TL) independently screened abstracts and full-texts for eligibility. Disagreements were resolved between screening authors or, failing that, by a third author (MC).
Data extraction and quality assessment
Two authors (NS & TL) independently extracted data and assessed study quality, and a third author (MC) settled disagreements. Data extraction focused on characteristics of the study sample/population, operationalisation of PM2.5 measurement, and COVID-19 outcomes. Effect size and direction, coefficient type (e.g., Hazard Ratio, Odds Ratio), and confidence intervals were tabulated.
Quality was assessed using the Newcastle-Ottawa Scale (NOS) (10) and scores were converted to Agency for Health Research and Quality (AHRQ) standards using the rubric in Shamsrizi et al (11):
Good quality: 3 or 4 stars in Selection domain AND 1 or 2 stars in Comparability domain AND 2 or 3 stars in Outcome domain
Fair quality: 2 stars in Selection domain AND 1 or 2 stars in Comparability AND 2 or 3 stars in Outcome domain
Poor quality: 0 or 1 star in selection domain OR 0 stars in Comparability domain OR 0 or 1 stars in Outcome domain
Meta-analysis
Results were pooled using a random-effects meta-analysis with the metafor (12) and metaviz (13) packages in R (14). A Meta-analysis Of Observational Studies in Epidemiology (MOOSE) (15) checklist is available in our public repository, along with meta-analysis code and data (9). Studies were limited to those rated “good” or “fair”, with sensitivity analyses including all studies regardless of quality. Assuming inherent variance due to differences in populations and methods, we used random effects models and report the I2 statistic for heterogeneity. All outcomes were converted to Odds Ratios for synthesis. Egger’s regression and funnel plots tested for publication bias. While not specified in the original protocol on PROSPERO, we added trim-and-fill and leave-one-out sensitivity analyses to test the robustness of results.
Where studies reported multiple outcomes, we prioritised the following: lengthiest PM2.5 measurement; most comprehensive measure of outcomes (e.g., serology and self-reported symptoms rather than one or the other; hospitalisation+ rather than just hospitalisation or ICU admittance); complete rather than restricted samples/populations (e.g., analysis of the entire Ontario population rather than only test-takers in Sundaram et al (16)); models adjusting for socioeconomic factors; and the indicator of “least” severity (e.g., hospitalisation over ICU admittance (17)); continuous PM2.5 measures (only one study used a non-continuous measure). For the two studies by Mendy et al., we used only the more recent, larger study (18) since it included all participants from the earlier one (19). This approach to outcome selection was not specified in the protocol as outcome reporting preferences of studies were unforeseeable.
Results
Search results
Search strategy results are in Figure 1 below. The initial literature search of Medline, Embase and the WHO COVID-19 database yielded 1,442 studies, which was reduced to 18 after screening. One study was excluded even though it met the inclusion criteria because it reported only statistically significant results rather than all results regardless of significance (20). A full list of screened studies along with reasons for exclusion is available in our public repository (9).
Study characteristics
All 18 included studies used a cohort design and focused on background ambient PM2.5; none were case-control studies. No study investigated discrete, large-scale PM2.5 exposures, meaning we were unable to address our second research question.
Half the studies used North American data (N = 9), mostly from the US (N = 6), followed by Canada (N = 2) and Mexico (N = 1). The remainder mostly used European data (N = 8), primarily the UK (N = 4), followed by Italy (N = 2), Spain, and Poland (N = 1 each). The last study used Chinese data.
Study quality
Study quality is summarised in Table 1. Most (13 of 18 studies) were rated “good”. More detail is available in the Critical appraisal document on our public repository (9).
Several methodological limitations are worth nothing. Only four of the 18 studies included individual-level adjustments for socioeconomic factors (e.g., education, household income, insurance status). Of the remaining 14, three did not adjust for any socioeconomic factors. Three of the seven infection studies only included participants with a COVID-19 test (24,31,33), while the remainder either used entire cohorts regardless of whether there was a record of a COVID-19 test (32,34) or conducted analyses of the entire cohort as well as just those tested (16,28). Similarly, all but one study examining severity (28) and mortality (27) were limited to cohorts who were diagnosed with COVID-19, while three were restricted to patients hospitalised with COVID-19 (17,23,29). Restricted cohorts present a risk of collider bias, as PM2.5 exposure could influence both whether an individual sought testing for COVID-19 or was COVID-19 positive, resulting in distorted associations (35).
Other methodological issues were not captured by the NOS tool. Three studies (16,24,27) included multiple predictors of interest within a single model rather than build models around PM2.5 as an exposure. As these are not designed to account for how independent variables may interact (e.g., as mediators or colliders), the statistical associations are less reliable (36).
Some studies reported resolutions up to 100m2 (28,31), others used entire cities (29) or monitoring stations spaced tens of kilometres apart (32,37). Several did not specify PM2.5 resolution. The timeframe of PM2.5 measurement also varied considerably, from just the week prior to inclusion/recruitment (17) up to ten years (18,19) and nearly two decades (30).
All studies using UK data relied on the UK Biobank. While this sample is large at around 500,000 people, it is not considered representative of the UK population due to low participation rates and a skew towards older persons (38,39). Additionally, three used PM2.5 estimates from 2010 (24,27,31) – a decade old – and all used participant residences from 2006-2010. This does not account for change of address or the steady decline in PM2.5 in the interim (40).
Most studies used single-pollutant models, i.e., PM2.5 without any other air pollutants (N = 12). The remainder were evenly split between only multi-pollutant models (16,24,27) and both single and multi-pollutant models (29,34,37). Rzymski et al. (17) used a dichotomous indicator of PM2.5 based on whether the mean or maximum exceeded 20μg/m3 in the week before admission to hospital with COVID-19. All others used continuous mean PM2.5, while Mendy et al. (18) also used the maximum.
PM2.5 exposure and COVID-19 infection
Seven studies examined PM2.5 and COVID-19 infection, which are summarised in Supplementary Table 1. All were rated “good” quality and reported a significant and positive association. Pooled results indicated a 10μg/m3 increase in PM2.5 was associated with a 66% increase in the odds of COVID-19 infection (95% CI: 1.31 to 2.11), with 83% of the variance attributable to heterogeneity (p < 0.001). Egger’s regression suggested publication bias (p = 0.012). Trim-and-fill points could not be applied, though results from leave-one-out sensitivity analysis remained significant with estimates ranging from 1.48 to 1.78 (see Supplementary Figures 1 and 2).
Associations were not consistent across analyses, though no negative association was identified, i.e., all were positive or null. Kogevinas et al. (28) did not find an association with infection determined solely by serological tests within the subsample who agreed to take a test (n = 3,922). However, there was a significant association within both the serological test subsample and the full sample (n = 9,088) when infection was determined by combining serological tests and self-report indicators. The difference may be attributable to limited sensitivity of the serological tests, leading to false-negatives; only 70% of cases identified through self-reported indicators had detectable COVID-19 antibodies.
Sundaram et al. (16) categorised PM2.5 exposures into five ordinal categories, which exhibited J-shaped curve with COVID-19 infection; compared to the lowest exposure group (2-6μg/m3), COVID-19 infection risk was lower 6-7μg/m3, similar (7-8μg/m3), and then increasingly higher in the next two groups (8-9μg/m3 and ≥10μg/m3). This was the case whether the comparisons were between those testing positive for COVID-19 and not testing positive (i.e., testers and non-testers in the Ontario population; N = 14,695,579) or between those testing positive for COVID-19 and those testing negative (i.e., testers only; N = 758,791).
The single-pollutant model that accounted for socioeconomic factors in Veronesi et al. (34) found COVID-19 infections increased 3.6% (95% CI: 1.009-1.075) for every 1μg/m3 in PM2.5. The single-pollutant model that omitted socioeconomic factors was similar but with a slightly bigger effect (RR: 1.051; 95% CI: 1.027-1.075), which increased substantially when other air pollutants were added to the model (NO2: 1.347, NO: 1.105, O3: 1.107). This suggests multi-collinearity between air pollutants, which may bias the association between PM2.5 and COVID-19 infections. Otherwise, associations remained significantly positive in single and multi-pollutant models.
PM2.5 exposure and COVID-19 severity
Nine studies examined PM2.5 and COVID-19 severity, of which five were rated “good”, two “fair”, and two “poor”. These are summarised in Supplementary Table 2. Mendy et al. (18) was the only study not to find a significant association, though in a later study the authors found a significant association when the same participants were included in a substantially larger cohort (n = 1,128 versus n = 14,783) and PM2.5 estimates were updated by a year (19). Aside from Kogevinas et al. (28), all cohorts were restricted to those diagnosed or hospitalised with COVID-19. Severity was indicated in numerous ways including hospitalisation (N = 5), ICU admission (N = 3), requiring respiratory support (N = 3), clinical symptomatology (N = 1), oxygen saturation (N = 1), or multiple indicators (N = 1).
Pooled results from N = 6 studies indicate the odds of a severe outcome was 227% higher (95% CI: 1.41 to 3.66) for every 10μm/g3 increase in PM2.5. Nearly all the variance in effects was due to heterogeneity (I2: 97%; p < 0.001). There was no detectable publication bias (p = 0.132). Trim-and-fill points slightly attenuated the results (OR: 2.04; 95% CI: 1.29 to 3.21) and the association remained significant in all leave-one-out analyses (see Supplementary Figure 4).
Chen, Sidell et al. (37) found a consistent associations between PM2.5 and COVID-19 hospitalisation when using PM2.5 measured over the previous year. The associations were consistent but weaker when using PM2.5 from the previous month. This study also found that the association remained when controlling for another air pollutant, NO2. Similarly, Li et al. (29) found a positive association between PM2.5 and clinically-defined severe COVID-19 across four different lag periods (0-7 days to 0-28 days), which attenuated but remained mostly significant when adjusting for other air pollutants.
PM2.5 exposure and COVID-19 mortality
Five studies examined PM2.5 and COVID-19 mortality, which are summarised in Supplementary Table 3. Four of six studies found a significant positive association. The remainder were null. One study was rated “poor”; the rest were “good”.
Pooled results from n = 5 studies were positive but non-significant (OR: 1.40; 95% CI: 0.94 to 2.10), with heterogeneity explaining 75% of the variance (p = 0.010). There was no evidence of publication bias (p = 0.100). Trim-and-fill points could not be applied, though leave-one-out sensitivity analysis indicated the results remained positive but only became significant with the exclusion of Chen, Wang et al (26) (OR: 1.66; 95% CI: 1.06 to 2.59). These results are summarised in Supplementary Figures 5 and 6.
Elliot et al. (27) was the only study that did not restrict its sample to those diagnosed or hospitalised with COVID-19, avoiding associated issues of collider bias. It also had a null finding with a negative point estimate (OR: 0.94, 95% CI: 0.75-1.18). However, the model included multiple predictors rather than being built around a single exposure-outcome relationship, meaning associations were less reliable.
Of the remaining studies, all but Chen, Wang et al. (26) found a significant positive association between PM2.5 and COVID-19 mortality. Chen, Sidell et al. (37) found that PM2.5 was consistently associated with higher mortality rates across multiple models, regardless of whether PM2.5 was measured in the previous month or year and whether the model adjusted for NO2.
Discussion
We found strong evidence that PM2.5 exposure increases the risk of COVID-19 infection and weaker evidence that it increases severity and risk of death.
Studies of COVID-19 infection were generally of high quality and consistently demonstrated a significant association with PM2.5 across methodologies and populations. Most importantly, the effect was observed even when adjusting for individual-level socioeconomic indicators, probably the most important confounder (4). While there was evidence of publication bias in COVID-19 literature, leave-one-out sensitivity analysis was robust to exclusions.
The association with COVID-19 infection was observed in both single and multi-pollutant models across all studies, suggesting PM2.5 is not just an indicator of a generalised effect of poor air quality on risk of COVID-19 infection, but an independent, causal predictor. However, multicollinearity may be an issue, as indicated by increase in effect size when other air pollutants were added to PM2.5 models in Veronesi et al. (34).
The evidence on COVID-19 severity and mortality also indicates a positive association, though the quality of the research was weaker and pooled mortality results were non-significant. Nearly every study was limited to people already diagnosed or hospitalised with COVID-19, introducing potential collider bias, or more specifically endogenous selection bias (41). As the above results suggest that PM2.5 influences who gets COVID-19, it could also mean that the infected cohorts differ substantially based on their PM2.5 exposure. For instance, PM2.5 may expand infections into less-vulnerable populations, reducing baseline risk of severe infection and biasing the association with PM2.5 towards null.
Kogevinas et al. (28) was the lone COVID-19 severity study to include participants who were not already infected. It also designed statistical models to examine the effect of PM2.5 rather than including multiple predictors in a single model, and used high-resolution measures at 100m2, finding a positive association with COVID-19 severity. Elliot et al. (27) was the only mortality study to include participants not diagnosed or hospitalised with COVID-19. While it found no association with PM2.5 exposure, all predictors were included in a single model, making the results less reliable (36). However, its exclusion in leave-one-out sensitivity analysis did not meaningfully affect pooled results.
Despite the weakness of evidence for effects of PM2.5 on COVID-19 severity and mortality, there are still reasons to treat it as real. There is strong circumstantial evidence of a mechanism, including effects of PM2.5 on receptor expression, cell defence, and cardiovascular and pulmonary health (3,5,6) which may make infected persons more vulnerable to worse COVID-19 outcomes. Combined with the positive (if not always significant) associations identified in this review, PM2.5 air pollution should be treated as a risk factor for severe COVID-19 disease and death.
Evidence gaps
We identified two major evidence gaps. The first is a lack of cohort or case-control studies of COVID-19 severity and mortality that were not limited to those with COVID-19 and that built models specifically around PM2.5 exposure. The second gap is a lack of cohort or case-control studies on discrete, large-scale PM2.5 exposures such as smoke from wildfires. It remains unknown whether intensive PM2.5 exposure increases short and long-term risks of respiratory illnesses like COVID-19. There is some ecological evidence on an association, though this mainly focuses on concurrent PM2.5 exposure (42–44). In the months following the 2019-2020 Black Summer fires in New South Wales, Australia, areas with more burn coverage had higher rates of COVID-19. However, there was no detectable association with larger particulate matter, PM10, and the study did not investigate PM2.5 (45). We therefore have little idea whether and how long people may be at elevated risk of COVID-19 following major smoke exposures.
Strengths and limitations
Among this systematic review’s strengths are an inclusion criterion that limited evidence to studies using individual-level data, a quality assessment that indicated most were of good quality, and synthesis of data with a meta-analysis. This review covers studies published in the first 2.5 years of the pandemic, building on previous reviews with more up-to-date evidence.
There are some limitations. Operationalisation of PM2.5 exposure varied across studies, including when it was measured, precision, and time periods covered. No studies captured variations in individual exposure due to time spent outdoors or regular movement into areas like the workplace. The review was restricted to outdoor air pollution exposures, which, combined with most studies originating in developed countries, may not be applicable to lower-income countries where indoor air pollution from ‘dirty’ heating and cooking fuels are a greater threat. While all studies used individual-level data, many used aggregated indicators for important confounders like socioeconomic status. COVID-19 is still a relatively new illness, so this review can only be considered an early snapshot of the evidence.
Conclusion
There is strong evidence that PM2.5 increases COVID-19 infections. The evidence for effects on COVID-19 severity and mortality is weaker, but similarly suggests that PM2.5 exposure increases risk. When considered alongside evidence that PM2.5 worsens cardiovascular and pulmonary health, we see good reason to treat the association with severe illness and death from COVID-19 as real, if not yet fully established. No cohort or case-control studies focused on discrete, large-scale PM2.5 exposures such as smoke from wildfires, which will become increasingly important as climate change increases both the frequency and intensity of wildfires.
Data Availability
All data and materials are available online on a public repository: https://doi.org/10.26180/21529158.v3
Appendix search terms
The following terms were used to search for relevant studies in Medline, Embase, and the WHO COVID-19 database. In Medline and Embase, terms were searched as MeSH headings. The WHO COVID-19 database does not support MeSH headings, so terms were searched as key words.
Air pollution OR maximum allowable concentration OR threshold limit values OR petroleum pollution OR traffic-related pollution OR particulate matter OR particulate matter 2.5 OR coal ash OR dust OR smog OR smoke OR soot OR air pollutants OR gasoline OR vehicle emissions OR particle pollution
AND
COVID-19 OR SARS-CoV-2 OR coronavirus disease 2019 OR severe acute respiratory syndrome coronavirus 2
AND
Virulence OR patient acuity OR severity of illness index OR morbidity OR basic reproduction number OR incidence OR prevalence OR mortality OR fatal outcome OR survival rate OR death OR hospitalisation OR length of stay OR patient admission OR asymptomatic diseases OR asymptomatic infections OR critical illness