The protocol for this systematic review and meta-analysis was registered in the PROSPERO registry (https://www.crd.york.ac.uk/prospero/) under registration number CRD42023425327 prior to the preliminary search for studies. This systematic review and meta-analysis was reported according to the guidelines of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) standards [7], adapted to the type of information and procedures of observational epidemiological studies. Conformity of the manuscript with the PRISMA guidelines is provided in Supplementary Tables S17A, B, Supplementary File 1.

The research question was formulated according to the format of the Population, Exposure, Comparator, Outcomes and Study Design (PECOS) questions [9]. The content of the question was as follows:

In the general human population (P), what is the effect of long-term exposure to ambient particulate matter (PM₁₀, PM_2.5) (E) versus exposure to lower levels of air pollution (difference of 10 μg/m³) (C) on the risk of all-cause and cause-specific mortality (O), as reported in cohort studies (S)?

We searched for studies published in PubMed and Embase. The search timeline of the previous systematic review [4] included studies up to September/October 2018. Accordingly, our search started in September 2018 and ended in May 2023. The search strategy included a combination of free text and Medical Subject Headings (MeSH) terms related to exposure, outcome and study design. In addition, reference lists of selected studies and reviews were scanned for relevant studies not identified by the formal search strategy. The search lines for both databases can be found in Supplementary Table S18, Supplementary File 1.

We updated our search in January 2024 to assess the influence of more recent studies published after May 2023. The studies identified in this new search were not included in the meta-analysis, but were listed, their results summarised, and the effect estimates compared with our pooled estimates.

We included prospective and retrospective cohort study designs, but excluded case-control studies and related designs, panel and cross-sectional studies. Other reasons for excluding studies were occupational or indoor exposure only, qualitative designs, reviews and non-human studies (i.e., in vivo, in vitro). We considered long-term exposure (of the order of months to years) to ambient PM_2.5 and PM₁₀ from any source, expressed as a concentration unit (μg/m³). Indoor and occupational exposures to these pollutants were excluded from the analysis. This long-term exposure was compared with lower levels of PM_2.5 and PM₁₀ in the same or a control population. The systematic review by Chen and Hoek [4] included case-control designs, but only three of these studies were found, and they were not included in the meta-analysis. Because the current systematic review focuses on association values and the certainty of the evidence around them, we focused on cohort studies.

Outcomes were all-cause and cause-specific mortality, including all causes (A00-Z99), circulatory diseases (I00-I99), ischaemic heart disease (IHD) (I20-I25), cerebrovascular diseases (I60-I69), respiratory diseases (J00-J99), chronic obstructive pulmonary disease (COPD) (J40-J44, J47), acute lower respiratory infections (ALRI) (J12-J18, J20-J22) and lung cancer (C33-C34: malignant neoplasm of the trachea, bronchus or lung only). See Supplementary Table S19, Supplementary File 1, for a description of the inclusion of diseases in categories and subcategories according to the International Classification of Diseases (ICD-10), 10th edition codes. Although all-cause mortality includes accidental deaths, the proportion of deaths caused by accidents, etc., is typically small (<10%) in comparison with the other causes of death [3]. Therefore, as in the previous review, we considered natural causes to be equivalent to all-cause mortality. We also included post-neonatal mortality in children aged 28 days to 12 months, which was analysed separately in the previously mentioned meta-analysis of all-cause mortality. In terms of participants, the study included the general population, of all ages, in all locations, regardless of level of development, income or urbanisation. There were no geographical restrictions, but studies limited to patient cohorts were excluded from the analysis. It is worth noting that the systematic review by Chen and Hoek [4] included infant mortality and patient-only cohorts.

PO and ES independently screened the titles and abstracts of the retrieved studies. In the final step, the full text of potentially eligible studies was assessed by the same reviewers. For exclusions at this stage, the reason was recorded in the forms. In case of disagreement, the two reviewers discussed and reached a consensus on the inclusion of the studies. All studies flagged for inclusion at the full-text assessment stage were included in the systematic review, but for the meta-analysis some studies were excluded if there was total or partial overlap between cohorts, i.e., if two or more studies reported data on the same cohort, typically in large collaborative multicentre studies. To decide how much overlapping data was acceptable to consider the studies as independent, our criterion was that both studies were included if the number of additional years reported in a cohort-specific study was equal to or greater than the number of overlapping years between the two studies, or if the number of participants in the cohorts differed by more than 25%. When overlap was identified, one of the studies was selected and the other was excluded based on the same pair of criteria: 1) the study with the larger sample size, and 2) the study with the largest time period.

We extracted information on publication details (title, authors, year of publication), study characteristics (study design, location, study population), exposure to the air pollutants of interest and ambient pollution levels (mean and median concentrations), outcome (cause of death), relative risks (RR), hazard ratios (HR) and odds ratios (OR) and 95% confidence intervals (CI) with corresponding exposure increments derived from single-pollutant models. We also collected information on the shape of the concentration-response functions and on two-pollutant models. In some studies, authors reported more than one effect estimate for the same exposure-outcome pair. In this case, we selected the estimate for inclusion in the meta-analysis based on single-effect estimates derived from the model identified by the authors as “main” in the methods section. If we could not identify the main model, we selected the simpler model that included all or as many of the critical confounders as possible, i.e., age, sex, body mass index (BMI), and socioeconomic status (SES). For studies not included in the previous classification, a case-by-case assessment and other considerations, such as the size of the subcohort for the specific model or other relevant ad hoc criteria, were taken into account. In addition, some studies may have contributed more than one independent effect estimate for a given exposure-outcome pair, for example, when a study reported results from two or more different cohorts or sites. For this reason, the number of effect estimates may be larger than the number of studies included.

To assess the risk of bias in individual studies, we used a domain-based risk of bias tool specifically designed for air pollution epidemiology studies. This tool was developed for the 2021 WHO global air quality guidelines. A detailed description of this tool can be found on the WHO website [12] and was also explained in the previous review [4]. Briefly, this tool considers 13 items grouped into 6 domains: confounding, selection bias, exposure assessment, outcome measurement, missing data and selective reporting, and is based on the Risk Of Bias In Non-randomised Studies (ROBINS) instrument [13].

Effect estimates reported in individual studies could be expressed as RRs, HRs, ORs or percentage increase in the risk (Perc.-incr.), but pooled estimates were calculated and reported as RRs. HRs and ORs were considered to be numerically equivalent to RRs, in the latter based on the “rare disease assumption” [14]. Perc.-incr. were converted to RRs using the following equation: R R = P e r c . − i n c r . 100 + 1

The RRs were previously standardised to reflect a risk increase associated with a 10 μg/m³ increase in the pollutant of interest, as some studies reported estimates for this increase, while others reported values related to interquartile range or unit differences. Standardisation was performed using this equation: R R s t a n d a r d i s e d = e Ln ( R R o r i g i n a l ) × 10 I n c r e m e n t o r i g i n a l

We applied a random-effects model, because we assumed that some heterogeneity between studies was to be expected, given the observational nature of the included study designs. The DerSimonian-Laird estimator was chosen to estimate between-study variability [15], which was chosen in the previous review [4]. We also calculated the I² and 80% prediction intervals, a measure of the observed variance without the effect of sampling error and a measure of how much effects vary across populations, respectively [16]. Prediction intervals [17] were used as a sign of heterogeneity, with the following rule: if 1) the prediction interval includes unity, and 2) the prediction interval is wider than twice the 95% CI, then there may be concerns about heterogeneity. Irrespective of the degree of heterogeneity suspected, an attempt was made to identify possible sources of potential heterogeneity through subgroup analysis and meta-regression. We investigated potential differences between WHO regions by comparing the observed Q values with their expected values, assuming a Chi-Squared distribution, while the influence of air pollution levels on the pooled effect estimates was assessed by meta-regression.

Sensitivity analyses were performed to estimate the degree of influence of the risk of bias on the effect estimates. We excluded the studies with a high risk of bias in each of the six domains and recalculated the effect estimates for studies with a low and moderate risk of bias to see if there were any changes in the direction or significance of the effect estimates.

To assess the potential for publication bias, we calculated Egger’s test and generated funnel plots to look for asymmetries.

In addition to the effect estimates, we also retrieved information on the shape of the concentration-response functions reported in individual studies and developed a narrative description of these results. We also identified studies that analysed two-pollutant models and narratively described these results in comparison with the effects estimated from single-pollutant models.

All statistical analyses were performed using the “meta” package [18] of the R statistical software, version 4.2.2 (https://www.r-project.org/).

To assess the certainty of the evidence, we used a modified version of the Grading of Recommendations, Assessment, Development and Evaluation (GRADE) approach [19], developed by an expert group convened by WHO in the context of 2021 global air quality guidelines. A description of this tool can be found in Supplementary Table S22, Supplementary File 1.

In the case of studies with overlapping data, we kept the same criteria for selecting the study for inclusion in the meta-analysis as in the protocol. However, we refined the definition of what could be considered as overlap. We also developed a criterion for selecting the effect estimate to be included when the same study reported more than one effect estimate. To assess heterogeneity in a given exposure-outcome pair, we also considered the consistency in the direction of effect estimates between studies and the extent to which this potential variability was explained by the subgroup analyses. In addition, we developed a set of criteria to assess the potential presence of publication bias. All these procedures are described in detail in the relevant sections of this manuscript. Finally, according to the protocol, a likelihood-based random-effects model, restricted maximum likelihood (REML), should be used. However, we decided to use the more commonly used DerSimonian-Laird estimator, as this was the statistic used in the previous review.

The PubMed and Embase searches for this review identified 2,054 studies, plus two additional studies identified in the bibliography of relevant reviews. After duplicate records were excluded, 1,312 studies were assessed by title and abstract. A total of 211 full-text articles were assessed for eligibility, 89 were excluded as they did not meet the inclusion criteria, and 122 were finally included for the quantitative meta-analysis of PM_2.5, PM₁₀, NO₂ and O₃. For this study, 112 studies from this new search analysing exposure to PM_2.5 and PM₁₀ were initially included. Starting with these 112 studies and adding the 107 studies included in the previous review [4], we applied the process of excluding studies reporting overlapping data. After this process, we finally arrived at a set of 106 studies that were included in the meta-analysis of the effects of PM_2.5 and PM₁₀ on all-cause and cause-specific mortality, including studies selected for the previous review (46 studies) and for this update (60 studies). These 106 studies contributed 379 effect estimates for all exposure-outcome pairs considered, i.e., some studies reported more than one pollutant-outcome pair for analysis. The following descriptions and results in this and subsequent sections refer to both sets of studies, those from this update and those from the previous review combined. The flowchart for this selection process is shown in Figure 1, and the excluded studies, with reasons, are shown in the Supplementary File 2. Supplementary File 3 shows the process of replacing articles with overlapping data on a case-by-case basis.

The studies covered results from several countries and regions, including five WHO regions: The European Region (EUR) (31 studies), the Region of the Americas (AMR) (34 studies), the Western Pacific Region (WPR) (36 studies), the Eastern Mediterranean Region (EMR) (1 study), the South-East Asian Region (SEAR) (2 studies), and two studies covering more than one region. The total number of participants was 948,483,820 when all cohorts were included. The majority of studies focused on adult populations (96 studies), while some studies analysed all age groups (7 studies) and three studies analysed populations under 5 years of age. We have calculated the average (mean) value of the mean/median ambient PM concentration reported in each individual study, generally mean/median values per year or over the whole study period. There were 84 studies analysing exposure to PM_2.5 with an average of the mean/median concentration of 23.6 μg/m³, while 47 studies analysed PM₁₀ with an average of the mean/median concentration of 43.7 μg/m³. These average values for PM_2.5 are high, mainly due to the mean concentrations found in China. Information on the new studies included in this review and all studies used in further analyses can be found in Supplementary Files 4, 6.

Of the six domains analysed using the risk of bias tool, only one domain (missing data) was free of studies with a high risk of bias. All other domains had at least one study rated as “high risk of bias” for a specific exposure-outcome pair, although the proportions were very different. The domain with the highest proportion of studies with a high risk of bias was confounding (13%), followed by exposure assessment (2%). A larger proportion of studies were classified as having a moderate risk of bias in the domains of confounding (79%) or exposure assessment (26%). Among the potential confounders not taken into account, the most common was BMI, followed by ethnicity, smoking and diet. Other potential confounders not taken into account were physical activity, marital status, socioeconomic status and year of enrolment. All these results are presented in Supplementary Figure S1, Supplementary File 1. The detailed ratings and justifications for the risk of bias assessments are provided on a case-by-case basis in Supplementary File 5 for new studies and in Supplementary File 6 for all studies included in the meta-analysis. The overall results of the risk of bias assessment by domain are reported in Supplementary Tables S23, Supplementary File 1.

Comparisons between WHO regions are shown in Tables 2, 3, and the associated forest plots are shown in Supplementary Figures S31–S45, Supplementary File 1. For PM_2.5, the only statistical difference between regions was for lung cancer, with higher relative risks reported in the European Region and non-significant relative risks in the Region of the Americas. The same difference was observed for PM₁₀, with higher relative risks for the European Region, but non-significant relative risks for the Region of the Americas and the Western Pacific Region. Although the differences are not significant, when analysing exposure to PM_2.5, the relative risks appear to be consistently lower in the European Region for circulatory and IHD mortality, but higher for cerebrovascular, respiratory and COPD mortality. On the contrary, for PM₁₀ the higher relative risks come from the Western Pacific Region, with a value as high as 1.285 for COPD mortality, but generally with wide confidence intervals.

PM levels did not explain the heterogeneity in the associations between PM_2.5 and all-cause and cause-specific mortality. On the contrary, the associations with circulatory and cerebrovascular mortality were modified by the effect of ambient PM₁₀ levels, resulting in higher relative risks for increases in these concentrations, as shown in Supplementary Table S1, Supplementary File 1. We also compared smaller and larger studies, but found no significant differences, except for PM_2.5 and cerebrovascular mortality, with higher relative risks in smaller studies (Supplementary Tables S20, S21, Supplementary File 1).

Excluding studies with different types of bias led to changes in both directions (Tables 4, 5), i.e., higher and lower relative risks compared to the original value, but these changes were not important in terms of significance. However, there was one exception, the association between PM₁₀ and COPD, where the exclusion of two studies with high risk of confounding bias resulted in a non-significant but positive association.

Of the included studies, 21 reported the effect of a second pollutant. In general, the inclusion of gases (NO₂, NO_x, O₃, O_x, SO₂) or black carbon in the models did not change the associations between particulate matter and mortality [11, 29]. In a few cases the adjustment attenuated the effect, although still leading to significant associations [25, 26, 30, 41]. In contrast, a DUELS study found an increase in the association between PM_2.5 and non-accidental mortality after adjustment for NO₂ [42]. However, two studies found that the association with respiratory mortality was not significant after adjustment for NO₂ [27, 43]. The results from two pollutant models can be found in the Supplementary File 7 on a case-by-case basis.

A summary of the certainty of the evidence ratings for each exposure-outcome pair is given in Supplementary Table S16, Supplementary File 1. Both associations of all-cause mortality with PM_2.5 and PM₁₀ were rated as high certainty of the evidence. Of all the pairs analysed for cause-specific mortality, three combinations of PM_2.5 or PM₁₀ were rated as having moderate certainty of evidence, while all the other exposure-outcome pairs (12) were rated as “high”. A more detailed description of each domain is given in Supplementary Table S22, Supplementary File 1, while the full certainty of the evidence rating for each exposure-outcome pair is shown in Table 6 for PM_2.5 and all-cause mortality and in Supplementary Tables S2–S15, Supplementary File 1, for the other exposure-outcome pairs. It should be noted that the associations between PM₁₀ and ALRI and both associations with post-neonatal mortality were not assessed due to the small number of studies. The results of the certainty of evidence assessment by domain are presented in Supplementary Table S24, Supplementary File 1.

In January 2024, 201 new records were retrieved. Of these, 30 were included by title/abstract and finally five were included for a narrative description of the results and for comparison with our meta-analytic estimates. Some studies were excluded because they reported on cohorts already included in our meta-analysis, many studies considered particulate matter only to adjust for other risk factors in the models, and there were other reasons for exclusion, which can be seen in the Supplementary File 2. Of the five new studies identified, one study analysed data from the “Scottish Longitudinal Study (SLS),” which included people aged 17 and over in 2002 [44], and found estimates that were higher than our pooled values. For example, for PM_2.5 and all-cause or respiratory mortality, the hazard ratio was 1.035 (95% CI: 1.023–1.047) and 1.062 (95% CI: 1.028–1.096), respectively, per 1 μg/m³ increase in PM_2.5. Another study from northern Europe [45] also reported higher estimates for all-cause mortality with exposure to PM_2.5 in 1990, with a hazard ratio of 1.40 (95% CI: 1.04–1.87) per 5 μg/m³ increase. Three studies reported data from China with slightly different results. While one study found higher values for PM_2.5 and all-cause (HR: 1.25; 95% CI: 1.04–1.50) and cardiovascular (HR: 1.38; 95% CI: 1.02–1.86) mortality [46], another study based on data from the China-HEART cohort [47] reported lower values for PM_2.5 and all-cause (HR: 1.022; 95% CI: 1.014–1.030) and cause-specific endpoints, although the estimates were positive and significant in all cases. Specifically for lung cancer mortality, another study from southern China [48] reported estimates that were higher than our pooled estimates, for example, the hazard ratio between PM_2.5 and lung cancer was 1.042 (95% CI: 1.033–1.052) per 1 μg/m³ increase. Taken together, the new studies strengthen the evidence for an association between PM and mortality, as the information provided is in the same direction as our estimates.

When interpreting the results of the present study, several limitations must be taken into account, in addition to those of heterogeneity and publication bias already discussed. Statistical tests to evaluate heterogeneity and publication bias have been of limited value, so more qualitative methods were needed in this review to assess these issues, with their limitations related to subjective judgements. Another limitation is that although more recent results have been published from areas outside Canada, the United States and Europe, they still come from countries that are predominantly high- and middle-income economies.

In conclusion, this study updated the previous systematic review and meta-analysis that informed the 2021 WHO global air quality guidelines on the effects of long-term exposure to PM on mortality, by including a large number of new studies. The results provide up-to-date evidence on the influence of air pollution on mortality, a picture of the rapidly increasing number of studies being conducted on this issue, and an outlook on the challenges of interpreting and appraising the evidence.