Particulate Matter and Premature Mortality: A Bayesian Meta-Analysis

Background: We present a systematic review of studies assessing the association between ambient particulate matter (PM) and premature mortality and the results of a Bayesian hierarchical meta-analysis while accounting for population differences of the included studies. Methods: The review protocol was registered in the PROSPERO systematic review registry. Medline, CINAHL and Global Health databases were systematically searched. Bayesian hierarchical meta-analysis was conducted using a non-informative prior to assess whether the regression coefficients differed across observations due to the heterogeneity among studies. Results: We identified 3248 records for title and abstract review, of which 309 underwent full text screening. Thirty-six studies were included, based on the inclusion criteria. Most of the studies were from China (n = 14), India (n = 6) and the USA (n = 3). PM2.5 was the most frequently reported pollutant. PM was estimated using modelling techniques (22 studies), satellite-based measures (four studies) and direct measurements (ten studies). Mortality data were sourced from country-specific mortality statistics for 17 studies, Global Burden of Disease data for 16 studies, WHO data for two studies and life tables for one study. Sixteen studies were included in the Bayesian hierarchical meta-analysis. The meta-analysis revealed that the annual estimate of premature mortality attributed to PM2.5 was 253 per 1,000,000 population (95% CI: 90, 643) and 587 per 1,000,000 population (95% CI: 1, 39,746) for PM10. Conclusion: 253 premature deaths per million population are associated with exposure to ambient PM2.5. We observed an unstable estimate for PM10, most likely due to heterogeneity among the studies. Future research efforts should focus on the effects of ambient PM10 and premature mortality, as well as include populations outside Asia. Key messages: Ambient PM2.5 is associated with premature mortality. Given that rapid urbanization may increase this burden in the coming decades, our study highlights the urgency of implementing air pollution mitigation strategies to reduce the risk to population and planetary health.


Introduction
Environmental pollution is a global public health problem [1,2]. Despite various preventive strategies, air pollution continues to be a significant contributor to adverse

Methods
Medline, CINAHL and Global Health electronic databases were systematically searched (last accessed January 2020) using keywords and Boolean/phrase terms based on particulate matter and premature mortality (Table S1: Search strategy). The search was augmented from the reference lists of the included articles. The review was registered in the International Prospective Register of Systematic Reviews (PROSPERO), systematic review registry (CRD42019134760). The inclusion criteria of our systematic review were: 1.
Studies that measured PM 2.5 or PM 10 ; 2.
Studies based on any study design; 4.
From any population group (no ethnic groups were excluded); 5.
Published in English in a peer reviewed journal; 6.
Available in Medline, CINAHL and Global Health electronic databases from inception to January 2020.
The exclusion criteria were: 1. Studies which assessed pollutants other than PM 2.5 and PM 10, or measured these pollutants in combination with other pollutants; 2. Literature reviews; 3.
Conference papers, abstracts and editorials.
For the purpose of the Bayesian hierarchical meta-analysis, we also included three further selection criteria, namely, (i) Studies that showed log normality of data; (ii) Studies that did not derive PM based on satellite observations; and (iii) Studies providing point estimates with 95% confidence intervals. We excluded studies using solely satellite observations due to the uncertainties linked to satellite-based PM, namely, poor satellite coverage in specific regions, cloud contamination and year-to-year variability as such observations can substantially impact the estimates of premature mortality when compared to global models [14].
Two authors independently reviewed study titles and abstracts for detailed review of the full text (NTW and AC). All duplicates were removed after the initial search. Any disagreements were resolved by consulting with a third, senior author (MS). Studies were excluded after full-text review if they did not meet the inclusion criteria. Data extracted for analysis included the first author's name, publication year, country, exposure estimates and the method of exposure ascertainment, outcome definitions, the method of outcome ascertainment and key results.
The working definitions for exposures was PM. PM is the particle pollutant component in the atmosphere and is a mixture of solid particles and liquid droplets that can only be seen microscopically. Based on the size of the particles, PM is categorized as PM 10 and PM 2.5 , defined as follows: • PM 10 : inhalable particles, with diameters that are 10 micrometers and smaller; and • PM 2.5 : fine inhalable particles, with diameters that are 2.5 micrometers and smaller.
Our outcome, premature mortality, was defined as death that occurs before the average age of death in the specific population group. It is defined as potential years of life lost.
Quality of the included studies: We assessed study quality by using the Newcastle-Ottawa scale (NOS) for observational studies [15]. This scale is comprised of three elements: (i) Four stars are allocated to study group selection (the first element); (ii) Two stars are allocated to comparability of the groups (the second element); and (iii) Three stars are allocated to ascertainment of the exposure and outcome (the final element).
The NOS score ranges from 0-9 and a methodologically robust paper can achieve a total of nine stars; a perfect score. Based on the total number of stars achieved, a study was categorized as good (a total of seven or more stars), fair (five or six stars) or poor (four stars or less) quality.
Statistical analysis: We conducted a Bayesian hierarchical meta-analysis [16,17]. We conducted two analyses; a meta-analysis for PM 2.5 and a meta-analysis for PM 10 , in which we used as a non-informative prior an improper uniform distribution over the positive real number line, followed by a heterogeneity analysis. The basic steps followed were (i) Checking for log normality of the data; (ii) Removing studies that had not achieved log normality; (iii) Transformation of log normal to the normal distribution; (iv) Meta-analysis; and (v) Converting the estimates to their original scale.
Prior to conducting the Bayesian hierarchical meta-analysis, we adjusted for differences among the baseline population characteristics of the studies included for analysis. In the original studies, the numbers exposed to PM in each country varied. To avoid considerable disparity across studies, we calculated the premature mortality rate for each respective study year by dividing the country specific number of premature deaths by the population for the same year. Log normality of the mortality rates was assumed and checked using properties of the log normal distribution (Supplementary Material S1, S2, S3). Two studies did not satisfy the properties of the log-normal distribution and were excluded from the meta-analysis [18,19]. The transformations between the corresponding log normal and normal distributions were undertaken with the usual conversion equations in conjunction with exploiting properties of the log normal distribution in order to calculate the variances of the log normally distributed mortality rates [20,21], as detailed in Supplementary Material S4&S5 ( Figure S1). The meta-analysis estimates were then transformed back to their original scale.
The analysis was carried out in freeware R, version 2019 [22] using the bayesmeta package version 2019 [17] [https://cran.rproject.org/web/packages/bayesmeta/bayesmeta. pdf] (accessed on 30 January, 2020). The bayesmeta package derives the posterior distributions of the synthesized mean and heterogeneity parameter and their posterior joint distribution. The code used to carry out this analysis is available in Supplementary Material S5.
The forest plots that were generated from the Bayesian analysis demonstrate the log normal mortality rates with 95% credible and prediction intervals mapped to the normal distribution. Heterogeneity plots were generated to display the posterior joint density of the log normal mortality rate and heterogeneity (τ) parameters, with a darker shading area corresponding to a higher probability density.
We used a non-informative prior as opposed to an informative prior. In the absence of clear prior evidence for mortality rates, we chose this conservative option because non-informative priors have a minimal effect on the analysis [23]. Furthermore, we chose a random effect meta-analysis instead of fixed effect meta-analysis (a more conservative approach) which assumes the potential for the original study samples to arise from different populations. In Bayesian hierarchical meta-analysis, if the number of studies is less than 20, the random effect model is the analysis of choice [23].

Studies Not Included in the Bayesian Meta-Analyses
Twenty studies were not included in the meta-analyses. Figure 3 shows the number of publications and the area/country of origin of the studies that were not included in the Bayesian meta-analysis. The techniques used to assess exposures varied among studies. Most studies used spatial modelling, using different modelling techniques, and four studies used satellite-based measures [4,35,43,44] (Tables 1 and 2).

Studies Not Included in the Bayesian Meta-Analyses
Twenty studies were not included in the meta-analyses. Figure 3 shows the number of publications and the area/country of origin of the studies that were not included in the Bayesian meta-analysis.

Results of the Bayesian Meta-Analysis
The extracted premature mortality rate of the eligible studies was utilized for the meta-analysis (Table 2). Fifteen studies were included in the meta-analysis of PM2.5 and three in the meta-analysis of PM10, while two studies that investigated the outcome based on both exposures were included in the respective analyses.
The Bayesian hierarchical meta-analysis forest plots report on the stepwise analysis. This approach is hierarchical, which differs to conventional meta-analysis. The first level of the forest plot corresponds to the relevant results of the participants in the study and the second level is generated as the study participants are nested within a study and, here, we assume the sample derived is a randomly selected sample from the exposed population.
Studies included in the PM2.5 analysis were published after 2013 and represented a limited number of countries. Six studies were from China, three from USA, three from India with one study each from Canada, Korea and Yugoslavia. City specific information was available only in one study [28].
The analysis based on PM10 represented studies from France, China and two cities from India. Therefore, most evidence here came from the Asian continent.
Values in the PM2.5 forest plot indicate the log mortality rate mapped to their corresponding normal distribution values. After conversion back to the original scale, the annual estimate of premature mortality due to PM2.5 was 253 (95%CI: 90, 643) deaths per 1,000,000 population globally (Figure 4). The predicted value, overarching the sampling error of individual studies, is the expected mean value of a future study which is 269 (95%CI: 15, 3083) per 1,000,000 population.

Results of the Bayesian Meta-Analysis
The extracted premature mortality rate of the eligible studies was utilized for the meta-analysis (Table 2). Fifteen studies were included in the meta-analysis of PM 2.5 and three in the meta-analysis of PM 10 , while two studies that investigated the outcome based on both exposures were included in the respective analyses.
The Bayesian hierarchical meta-analysis forest plots report on the stepwise analysis. This approach is hierarchical, which differs to conventional meta-analysis. The first level of the forest plot corresponds to the relevant results of the participants in the study and the second level is generated as the study participants are nested within a study and, here, we assume the sample derived is a randomly selected sample from the exposed population.
Studies included in the PM 2.5 analysis were published after 2013 and represented a limited number of countries. Six studies were from China, three from USA, three from India with one study each from Canada, Korea and Yugoslavia. City specific information was available only in one study [28].
The analysis based on PM 10 represented studies from France, China and two cities from India. Therefore, most evidence here came from the Asian continent.
Values in the PM 2.5 forest plot indicate the log mortality rate mapped to their corresponding normal distribution values. After conversion back to the original scale, the annual estimate of premature mortality due to PM 2.5 was 253 (95%CI: 90, 643) deaths per 1,000,000 population globally (Figure 4). The predicted value, overarching the sampling error of individual studies, is the expected mean value of a future study which is 269 (95%CI: 15, 3083) per 1,000,000 population.
Similarly, the transformed results of the PM 10 forest plot ( Figure 5) indicate that the annual estimate of premature mortality due to exposure to PM 10 was 587 (95%CI: 1, 39,746) deaths per 1,000,000 population. However, when the sampling errors of individual studies were removed, the predicted mean result for a future study was 645 (95%CI: 0, 16,106) per 1,000,000 population. Similarly, the transformed results of the PM10 forest plot ( Figure 5) indicate that the annual estimate of premature mortality due to exposure to PM10 was 587 (95%CI: 1, 39,746) deaths per 1,000,000 population. However, when the sampling errors of individual studies were removed, the predicted mean result for a future study was 645 (95%CI: 0, 16,106) per 1,000,000 population.  Figures 6 and 7 illustrate the joint posterior density of heterogeneity τ and the effect µ (log mortality rate), for PM2.5 and PM10, respectively. The darker area on the plots indicates the area of higher probability density. Red lines represent the 50%, 90%, 95% and 99% credible intervals of the joint distribution. The blue solid line is the conditional posterior mean log mortality rate as a function of heterogeneity, with the blue dashed lines  Similarly, the transformed results of the PM10 forest plot ( Figure 5) indicate that the annual estimate of premature mortality due to exposure to PM10 was 587 (95%CI: 1, 39,746) deaths per 1,000,000 population. However, when the sampling errors of individual studies were removed, the predicted mean result for a future study was 645 (95%CI: 0, 16,106) per 1,000,000 population.  Figures 6 and 7 illustrate the joint posterior density of heterogeneity τ and the effect µ (log mortality rate), for PM2.5 and PM10, respectively. The darker area on the plots indicates the area of higher probability density. Red lines represent the 50%, 90%, 95% and 99% credible intervals of the joint distribution. The blue solid line is the conditional posterior mean log mortality rate as a function of heterogeneity, with the blue dashed lines  Figures 6 and 7 illustrate the joint posterior density of heterogeneity τ and the effect µ (log mortality rate), for PM 2.5 and PM 10 , respectively. The darker area on the plots indicates the area of higher probability density. Red lines represent the 50%, 90%, 95% and 99% credible intervals of the joint distribution. The blue solid line is the conditional posterior mean log mortality rate as a function of heterogeneity, with the blue dashed lines corresponding to the 95% credible interval. The green lines indicate the marginal posterior median and 95% credible intervals for both parameters.

Heterogeneity of the Studies
corresponding to the 95% credible interval. The green lines indicate the marginal posterior median and 95% credible intervals for both parameters.  The observed heterogeneity for the pooled studies for PM2.5 was 1.06 (95%CI: 0.23, 2.06) and for PM10 it is 1.9 (95%CI: 0.00, 10.50).
When the true heterogeneity is compared between the PM2.5 and PM10 meta-analyses, the between-study variance (true heterogeneity) was high among the studies that have assessed the outcome based on PM10.   The observed heterogeneity for the pooled studies for PM2.5 was 1.06 (95%CI: 0.23, 2.06) and for PM10 it is 1.9 (95%CI: 0.00, 10.50).
When the true heterogeneity is compared between the PM2.5 and PM10 meta-analyses, the between-study variance (true heterogeneity) was high among the studies that have assessed the outcome based on PM10. The observed heterogeneity for the pooled studies for PM 2.5 was 1.06 (95%CI: 0.23, 2.06) and for PM10 it is 1.9 (95%CI: 0.00, 10.50).
When the true heterogeneity is compared between the PM 2.5 and PM 10 meta-analyses, the between-study variance (true heterogeneity) was high among the studies that have assessed the outcome based on PM 10 .

Discussion
In this systematic review, we identified thirty-six studies of either good or fair quality assessing the association between ambient PM 2.5 and/or PM 10 and premature mortality. All studies reported a positive association. In the meta-analysis, in which we included sixteen studies, we observed that 253 premature deaths per million population are associated with exposure to ambient PM 2.5 . Prediction estimates indicated that the magnitude of the PM 2.5 -premature mortality relationship will increase in future studies. We obtained unstable estimates for PM 10 , most likely due to the high level of heterogeneity among studies included. This is the first systematic review and meta-analysis conducted to assess the association between ambient PM (both PM 2.5 and PM 10 ) and premature mortality. Our findings reflect a previous meta-analysis based on 53 studies that explored the association between ambient PM 2.5 and all-cause mortality, which found that a 1 µg/m 3 increase in PM 2.5 was associated with a significant 1.29% increase in all-age all-cause mortality [11]. Similarly, Hanigan and colleagues reported a positive association between anthropogenic PM 2.5 and premature mortality in Australia [59], albeit not a meta-analysis. A recent systematic review and a meta-analysis conducted by Jie et al. [60] also found that exposure to PM 2.5 and PM 10 increases mortality. In addition to mortality studies, studies which assessed Disability Adjusted Life Years (DALYs) and Health Adjusted Life Years (HALYs) as outcomes have also found that PM exposure increases health burden [61]. However, the DALYs does not include years of life lost and the HALYs calculation includes both morbidity and mortality data. In this study we did not include either of these measurements as outcomes thereby enabling us to understand the impact on years of life lost due to premature mortality, which is a long-term exposure to particulate matter pollution. Not including papers with DALYs and HALYs estimates did not bias our findings given only a limited number of papers were excluded.
It is important to highlight that our assessment has only focused on ambient PM. We considered household air pollution as a separate exposure, as in the Global Burden of Disease study. However, household air pollution and ambient air pollution are interlinked exposures as each one contributes to the other. Indeed, it has been found that emissions from the use of unclean fuels for domestic energy, when compared to other emissions such as industry and road traffic, have the largest impact on premature mortality globally [62].
The biological plausibility of the association observed cannot be underestimated. With increasing industrialization and urbanization in most regions, more PM is released into the environment, which has a negative impact on the cardiovascular, cerebrovascular and respiratory systems. This, in turn, increases the risk of mortality before the expected life expectancy. Moreover, the causal relationship we found is supported by many studies. Brook et al. [63] and Pope et al. [64] reported short term changes in PM 2.5 levels which lead to changes in daily mortality rates. Thurston et al. [65] also highlighted that PM 2.5 increases IHD (Ischemic Heart Diseases) and mortality and reported a dose response association. Many studies, including the Harvard Six Cities study, have also found that long term exposures to PM 2.5 (Dockery et al. and Pope et al.) increase mortality and that the overall reduction of PM 2.5 can reduce the mortality rates, confirming its causal association.
The advantage of our study is the statistical approach used. The Bayesian hierarchical meta-analysis, compared to the conventional meta-analysis, assesses the predicted credible intervals taking the weights of the reference population rather than the individual study results. When compared to a conventional meta-analysis, Bayesian hierarchical methods utilize a prior probability distribution in assessing this. Therefore, Bayesian hierarchical random-effect models can obtain accurate pooling effects, even with a limited number of studies in the meta-analysis. Furthermore, conventional meta-analysis cannot incorporate extreme values and small studies due to the systematic difference, limiting its application to our research question [66]. In contrast to the conventional meta-analysis, Bayesian hierarchical meta-analysis can address these issues [67].
While reading this review, an important point to note is that the strategies undertaken by individual countries to reduce the emission of PM are not uniform across the globe. Therefore, our pooled estimate of premature mortality may vary according to these varying mitigation strategies. The finding of our study is a concern pointing to the urgency of implementing strategies to mitigate this growing environmental risk factor for premature mortality; the impact of ambient PM 2.5 on premature mortality is remarkably high when considering the current global population and the predicted population growth in the coming decades. Although we observed an increased premature mortality for PM 10 , the confidence interval was extremely wide, indicating an unstable estimate. Results therefore should be interpreted cautiously. The considerable variation observed was most likely due to the heterogeneity among the studies, and future research efforts need to focus on the effects of PM 10 and premature mortality.
As with all studies, there are a number of limitations. First, heterogeneity among studies may have hidden the real burden of premature mortality due to PM exposure. The studies we analyzed do not represent the global burden of premature mortality due to PM, or the urban rural disparity, as we did not have data representing all countries of the world. Indeed, most of the studies included were conducted in China and India. Although these countries account for 36% of the world's population, they are also among the most polluted countries, so less polluted countries may be underrepresented in our study. Further, within an individual country, the available data only represents a sample of the population, which may not reflect the true impact. The majority of studies included in our study did not adjust for weather conditions or other associated conditions. The epidemiology-based exposure dose-response functions that were applied, how premature mortality was calculated, and other factors associated with life expectancy may also hide the true association. Second, we were unable to conduct a subgroup analysis, for example by region, due to the limited number of studies and lack of variation in the countries where research was conducted. None of the studies commented on causality rather than association. Third, meteorological effects on particulate matter pollution were not quantified in our analysis. Furthermore, we have excluded the satellite-based studies from our analysis.
Notwithstanding these limitations our approach, namely pooling of the available study results to obtain a summary measure and then to statistically model the reference populations of the included studies using Bayesian hierarchical meta-analysis, is meaningful for analyzing the impact of an environmental exposure(s) in contrast to analyzing a selected sample. This has enabled findings related to environmental exposures, such as PM, where the exposure cannot be confined to a sample population, and a key outcome such as premature mortality. We recommend that future systematic reviews consider this approach when collating evidence on environmental exposures and outcomes.

Conclusions
Existing evidence indicates a positive association between ambient PM 2.5 and premature mortality, even while accounting for heterogeneity between studies. Evidence for PM 10 remains inconsistent. This is one of few meta-analyses that has explored the causal association between PM and premature mortality, taking into account the heterogeneity found in the various reported studies. This study, therefore, strengthens our current knowledge of the important relationship between exposure to PM and health outcomes, highlighting the urgency to mitigate the growing exposure to air pollutants.

Supplementary Materials:
The following are available online at https://www.mdpi.com/article/ 10.3390/ijerph18147655/s1, S1: Addressing the differences among the baseline population of the included studies; S2: Requirements for the Bayesian Hierarchical meta-analysis; S3: The statistical analysis; S4: Testing the log normal assumption; S5: Calculating the variances of the log normal distributions; Figure S1: The results of the regression testing the assumption expressed in Equation (1).  Acknowledgments: We would also like to acknowledge Michael Brear and Robin Schofield from the Melbourne Energy Institute for their support.

Conflicts of Interest:
The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.