Concentration-Response Relationship between PM2.5 and Daily Respiratory Deaths in China: A Systematic Review and Metaregression Analysis of Time-Series Studies

The association between the particulate matters with aerodynamic diameter ≤ 2.5 μm (PM2.5) and daily respiratory deaths, particularly the concentration-response pattern, has not been fully examined and established in China. We conducted a systematic review of time-series studies to compile information on the associations between PM2.5 concentration and respiratory deaths and used metaregression to assess the concentration-response relationship. Out of 1,957 studies screened, eleven articles in English and two articles in Chinese met the eligibility criteria. For single-day lags, per 10 μg/m3 increase in PM2.5 concentration was associated with 0.30 [95% confidence interval (CI): 0.10, 0.50] percent increase in daily respiratory deaths; for multiday lags, the corresponding increase in respiratory deaths was 0.69 (95% CI: 0.55, 0.83) percent. Difference in the effects was observed between the northern cities and the south cities in China. No statistically significant concentration-response relationship between PM2.5 concentrations and their effects was found. With increasingly wider location coverage for PM2.5 data, it is crucial to further investigate the concentration-response pattern of PM2.5 effects on respiratory and other cause-specific mortality for the refinement and adaptation of global and national air quality guidelines and targets.


Introduction
Ambient air pollution (AAP) has become a major environmental and public health risk for human society globally. The World Health Organization's (WHO) Global Urban Ambient Air Pollution (AAP) Database indicated that 98% of cities in Low-and Middle-Income Countries (LMICs) are exposed to air pollution that far exceeds the WHO Air Quality Guidelines (AQGs) limits [1,2]. From the Global Burden of Disease (GBD) study, air pollution was ranked as the fourth leading risk factor accounting for more than 5.5 million premature deaths across the world each year and more than 50% of all AAP-attributable deaths occurred in China and India [2,3]. Ambient particulate matter (PM) pollution is specifically connected to 2.9 million deaths and 69.7 million disability-adjusted life-years (DALYs) in 2013 [2,3].

BioMed Research International
Fine PM or PM 2.5 , namely, particles with aerodynamic diameter of 2.5 micrometres or less, can lodge deeply into lung passageways and enter major organ systems [4,5]. Of all AAP, PM 2.5 poses the greatest health risks that are closely associated with a wide variety of acute and chronic illnesses and premature deaths, but predominantly from cardiovascular and respiratory outcomes [4,[6][7][8][9]. WHO set the AQGs of an annual mean of 10 g/m 3 and a 24-hour mean of 25 g/m 3 for PM 2.5 , representing the upper end of PM 2.5 concentration range below which short-term and long-term mortality risks are expected to be significantly reduced [10]. The global estimates of annual average PM 2.5 in 2013 showed that 87% of the world's population was exposed to PM 2.5 higher than 10 g/m 3 , with consistent increases between 1990 and 2013 in population-weighted mean concentrations particularly in Asia [11].
The geographic extent of PM 2.5 pollution in China is unprecedented, as only 0.4% of the Chinese population lives in areas that meet the WHO AQGs [1,11]. In 2013 alone, around 910,000 people in China died prematurely due to AAP [3] and 760,000 deaths were associated directly with PM 2.5 [12]. Real-time air quality data using PM 2.5 gauge in the 74 leading cities in China became officially available to public since 2012 and ambient PM 2.5 concentration is being monitored against the National Ambient Air Quality Standards [13][14][15][16].
The health effects of AAP especially on total mortality and cardiovascular diseases have been extensively studied across different countries and regions. Given that PM 2.5 also causes asthma, respiratory inflammation, jeopardizes lung functions, and even promotes cancers, its impact on human respiratory system should not be dismissed. However, epidemiological evidence for the PM 2.5 pollution on respiratory mortality has not been well synthesized, especially at high concentration areas such as in China. Respiratory deaths associated with PM 2.5 at relatively high concentrations (e.g., 24-hour mean exceeding 75 g/m 3 ) were sporadically studied and previous findings presented little evidence of damage threshold of concentration range [7,[17][18][19]. Most previous studies either were based on weighed exposure-response coefficients from epidemiological findings in Europe and North America or remain unexamined with a focus on shortterm effects based on projected or recently observed PM 2.5 level [15,20]. Plenty of research findings from countries with relatively low PM 2.5 concentration range have shown linear association between PM 2.5 exposure and premature deaths from respiratory outcomes [6][7][8]21], while some studies indicated lower relative risk (RR) with high PM 2.5 concentrations and the exposure-response curve turning flat at extremely high PM 2.5 levels [18,19,22,23]. However, these findings could not be readily applied to the AAP situations in LMICs because of the significant variations in meteorological conditions, PM sources and components, and population sensitivity to the PM 2.5 [17][18][19].
As systematic PM 2.5 data measurement in China became available and frequently used for research since 2012, recent studies looked into health effects of high PM 2.5 concentrations by cause-specific mortality indications [18,[24][25][26][27].
A most recent publication of a nationwide time-series study in China evaluated the short-term associations between PM 2.5 and daily mortality across 272 representative Chinese cities between January 2013 and December 2015. Comparing with similar multisite studies in Europe and North America, the findings from the study in China suggested weaker effects on daily mortality for each 10 g/m 3 increase in PM 2.5 concentrations [28]. A wider range of risk heterogeneity among different PM 2.5 sources and possible confounders including meteorological variables and exposure lag-day effects has been considered [29]. It is worthwhile to systematically review the recent studies and synthesize the epidemiological evidence on the health effects of PM 2.5 at a notably high concentration level that exceeds both WHO AQGs and interim targets. Research on PM 2.5 data specifically in China would be an important milestone adding value to future studies in populous LMICs in addition to the findings on the relatively lower PM 2.5 levels from Europe and North America.
Concentration-response functions drawn by metaanalysis are well accredited in supporting epidemiological evidence of the integrated information for health impact assessments [29]. A comprehensive overview of existing literature and the quantitative estimates of the correlations between PM 2.5 and respiratory deaths in China would provide valuable feedback to the current WHO AQGs Standards over mortality effects and potentially shed light on public health strategies in other developing contexts where AAP poses major health and development threats. With the increasing availability of epidemiological studies on respiratory health effects of PM 2.5 in China, we conducted a systematic review and meta-analysis of population-based studies in China to (1) compile and compare information from major literature databases on the fatal respiratory outcomes by PM 2.5 ; (2) retrieve evidence from identified studies regarding the association of PM 2.5 with respiratory deaths in China; (3) assess concentrationresponse relationships between PM 2.5 concentrations and their health effects.

Materials and Methods
The combinations of the following key terms for the literature search include (1) particulate matter, ambient particulate, PM 2.5 , ultrafine particulate, ultrafine particle, air pollution, air pollutants, and inhalation exposure; (2) respiratory tract diseases; (3) China, Chinese, names of major Chinese cities; (4) mortality and death. There were no language restrictions. The detailed log of search strategies with a complete list of key words and medical subject heading (MeSH) terms is shown in Boxes 1-5.

Inclusion Criteria and Identification of Articles.
Initially, we planned to include all the population-based prospective studies and time-series studies in our systematic review; thus articles eligible for meta-analysis met the following criteria: (1) they are original, population-based studies including prospective study, cohort study, nested case-control study, time-series study, and longitudinal study; (2) the main pollutants were ambient PM 2.5 or fine PM; (3) PM 2.5 concentration data in China were reported; (4) the endpoint of interest was mortality/deaths from respiratory outcomes; (5) the risk estimates and the associated 95% confidence intervals (CIs) were reported; (6) the risks were adjusted for potential confounders such as geographic regions, meteorological factors, or exposure lag in days.
Studies were excluded if they (1) were reviews, editorials, commentaries, letters, methodological papers, experimental, retrospective, or cross-sectional studies; (2) exclusively focused on high-risk groups such as smokers or patients with preexisting respiratory symptoms; (3) focused on nonrespiratory deaths and exposure to household or indoor air pollution, second-hand smoke, PM 10 , gaseous pollutants including carbon monoxide (CO), ozone (O 3 ), nitrogen dioxide (NO 2 ), and sulfur dioxide (SO 2 ).
Full-text articles were downloaded for all the abstracts that met the search and screening criteria and identified through inclusion criteria for further analysis. Specific literature identification steps with selection results were summarized in Figure 1.

Data Extraction.
For the selected studies, the information on the title, authors, year of publication, study location, geographic region, duration, PM 2.5 concentration, daily respiratory mortality, risk measurement, lag days, and adjustments was extracted and entered into a Microsoft Excel form. Two investigators (M. R. and X. Y.) independently conducted and checked the data extraction. Discrepancies in the extracted data were resolved by discussion, with involvement of the third reviewer (Y. C.) when necessary.
For PM 2.5 concentration ( g/m 3 ) data, the daily average was represented by median value and supplemented by the midpoint of lower and upper boundaries or daily mean when median value was not provided. For risk measurement, percent increase in respiratory mortality per respective unit increase in daily PM 2.5 concentration with 95% CI and standard error (SE) was extracted or calculated from relative risk (RR); that is, percent increase in respiratory mortality = (RR− 1) × 100%. Risk measurement with the regional divisions (north versus south) was recorded. Because pollution levels are often highly correlated and selecting a single best fitting lag might result in inconsistence across studies, it is important to consider the pattern of lag periods across the studies [42]. Therefore, we conducted subgroup analysis by categorizing the studies into single-day lags and multiday lags. Single-day lag means the mortality after 0, 1, or more days with exposure to the PM 2.5 concentration of the exposure day. Multiday lag means the mortality after 1 or more days with exposure to the moving average PM 2.5 concentration of 2 or more days. For studies having several single-day or multiday lag effects, the average effects were used in the synthesis. If the risk effects for a study were estimated in more than one model, only the overall value with adjustment for fewer confounders was used for pooled analysis to allow for higher homogeneity among the included studies. (2) exp air pollution/ (3) "pm2.5".ti,ab. (4) (pm adj2 "2.5").ti,ab. (5) ((air or atmospher * ) adj3 (pollut * or quality or ambient)).ti,ab. (6) (particulate matter or ambient particulate or ultrafine particulate * or ultrafine particle * or ufp or coarse particle * or dust or smog).ti,ab. (7) or/(1)-(6) (8) exp respiratory diseases/ (9) (respirat * or asthm * or pneumo * or lung * or pulmonar * or bronch * or pleur * or trache * ).ti,ab. (10) (8) or (9) (11) exp mortality/ (12) (mortal * or fatal * or death * or dead * ).ti,ab. (13) or/(11)- (12) (14) exp china/ (15) (china or chinese).ti,ab. (16) (guangzhou or shanghai or chongqing or beijing or peking or hangzhou or wuhan or chengdu or tianjin or xi'an or jinan or shenzhen or nanjing or shenyang or harbin or shantou or suzhou or dongguan or wenzhou or qingdao or quanzhou).ti,ab. (17) or/(14)- (16)   (18) (7) and (10) and (13)     The statistic 2 , a quantitative measure of inconsistency, was calculated to evaluate the statistical heterogeneity across studies [43]. 2 > 30% is considered moderate heterogeneity and 2 > 50% is considered substantial heterogeneity [44]. Both fixed-and random-effects meta-analysis were used when heterogeneity occurs across studies. Potential publication bias was assessed by Egger's test. Subgroup analysis for different lag-day structure (single-day lags and multiday lags) was conducted. Division by geographic regions (northern cities versus south cities) for the lag structures was tested for possible additional findings. Sensitivity analysis was performed, by omitting one study in each turn, to investigate the influence of a single study on the overall meta-analysis estimate. Random-effects metaregression was used to examine the linear trend of the percent increase in respiratory mortality across PM 2.5 concentrations. To maximize all the data for calculation of the pooled concentration-response, the restricted maximum likelihood (REML) approach proposed by Harbord, which provides improved estimation of the between-study variance, was used to estimate the regression coefficients [45]. Linear splines with knot at the 50th percentiles were used to assess potential nonlinear associations through metaregression analysis [46]. All analyses were performed in Stata   Texas, USA). A two-sided value < 0.05 was considered statistically significant, except where otherwise specified.

Search Findings and Study
Characteristics. The preliminary search yielded an initial total of 1,957 publications. Following the screening of titles and abstracts based on the inclusion criteria, 22 full-text articles were included for full eligibility review and one article was identified through reference hand searching. Finally, 13 studies [18,25,[31][32][33][34][35][36][37][38][39][40][41] met the search and screening criteria and were obtained for meta-analysis. Eleven studies were in English [18, 25, 31-37, 40, 41] and two were in Chinese [38,39]. All studies are timeseries studies and no population-based prospective studies or cohort studies met the inclusion criteria for further analyses in our searching period. The detailed article identification process adopting the PRISMA Flowchart model is shown in Figure 1 [30]. The identified studies were published between 2007 and 2016 and investigated major cities in China, namely, Beijing, Guangzhou, Shanghai, Shenyang, Xi'an, and Hong Kong ( Table 1). The study period ranged from 1998 to 2015, of which the PM 2.5 measurement records revealed a wide concentration range between 2 g/m 3 and 769 g/m 3 and the average PM 2.5 concentration of all studied cities was far beyond the WHO AQG limits of 10 g/m 3 for annual mean and 25 g/m 3 for 24-hour mean of PM 2.5 [10]. Only three studies were conducted after 2012 when China officially released PM 2.5 data. All the studies adopted the International Classification of Diseases revision 10 (ICD-10) for the coding of the death causes of which respiratory diseases (ICD-10 codes J00-J99) including subcategories such as chronic obstructive pulmonary disease and acute respiratory infection was classified in the outcome assessment. All English articles provided risk estimates by percent increase in respiratory mortality and two Chinese publications provided RR for deaths from respiratory diseases. Unified form of risk estimates was calculated and obtained as the percent increases of respiratory mortality per 10 g/m 3 increase in PM 2.5 concentration. All studies provided lag-day effect adjustment for risk measurement. In particular, Li et al. [18] provided risk estimates at different lag structures of both single-day lags and multiday lags up to eight days.

Publication Bias and Homogeneity.
There was observable publication bias among the included studies for single-day lags (Egger's = 0.033) but not for multiday lags (Egger's = 0.120). However, the asymmetric Egger funnel plot (Figures 2 and 3) indicated potential publication bias among the studies. In view of the asymmetric funnel plots shown in Figures 2 and 3, a nonparametric "trim-and-fill" method was used to account for the publication bias in the sensitivity analysis [47,48].    Statistically significant heterogeneity was found for all single-day lags by random-effects meta-analysis ( 2 = 84.1%, Figure 4). Significant heterogeneity was also found for the northern cities (Figure 4). For multiday lags, no statistically significant evidence of heterogeneity was found for either all cities ( 2 = 0%, Figure 5) or region-specific cities ( Figure 5).
The results from subgroup analysis show that the southern cities appear having higher percent increase. However, when we examined the region effect using metaregression technique, controlling for lag structures, no statistically significant difference was found between the southern cities and the northern cities. The regression coefficient for region is −0.25 (compared to the southern cities) and corresponding 95% CI is "−0.57, 0.06."

Sensitivity Analysis.
For single-day lags, when omitting one study in each turn, the combined estimates of percent increase of respiratory mortality changed little (percent increase ranging from 0.25 to 0.35, Table 3). However, when trim-and-fill method was used to consider for publication bias, the overall combined effect dropped from 0.30 to 0.06.
For multiday lags, the combined estimates were more or less constant ( Table 4). The percent increase ranged from 0.68 to 0.77 and the overall combined effects also changed little (from 0.69 and to 0.66). The sensitivity analysis indicated the robustness of the combined estimates for multiday lags.

Linear Relationship between PM 2.5 Concentrations and
Effects. Concentration-response relationship between PM 2.5 concentrations and effect estimates was examined for singleday and multiday lags using metaregression model. Figures  6 and 7 showed that the percent increase in respiratory mortality kept constant with increased PM 2.5 concentrations for     single-day estimates but fell off for multiday estimates, which suggested a potential tend-to-flat pattern in the cumulative effects of PM 2.5 on respiratory mortality at high air pollution levels observed in China. The pattern recalled the findings from previous studies that lower RR appeared with high PM 2.5 concentration with the concentration-response curve turning flat at extremely high PM 2.5 levels [18,19,22,23]. The 95% confidence intervals of coefficients for mortality increase per 10 g/m 3 of PM 2.5 are "−0.055, 0.055" and "−0.333, 0.071" for single-day lags and multiday lags, respectively.

Nonlinear Relationship between PM 2.5 Concentrations and
Effects. Due to the limited amount of studies, nonlinear relationship between PM 2.5 concentration and percent increase in respiratory mortality was examined using metaregression analysis with two linear splines. None of the linear splines was statistically significant (fitted splines were shown in Figures  6 and 7) and therefore no nonlinear concentration-response relationship was found across the studies. statistically significant in meta-regression analysis, difference in combined estimates was found between the northern cities and the southern cities. Combined effect estimates are relatively higher in southern cities than those in northern cities (0.46 versus 0.24 and 0.94 versus 0.64 for singleday lags and multiday lags, resp.; Table 2). Such difference suggested that PM 2.5 might pose higher relative risk on respiratory mortality in the south region than in the north region in China. While the average daily respiratory mortality was generally lower in the south region than that in the north (Table 1), it is likely that a small increase in count of deaths would result in higher variance in mortality. The most recently published nationwide analysis in 272 Chinese cities also revealed a significant heterogeneity across different regions of China on associations between PM 2.5 and daily mortality [28]. Limited by the small number of the studies and few cities included in our study, it is noteworthy to further investigate the possible factors behind the varying levels of PM 2.5 effects on cause-specific mortality in different geographic regions. No statistically significant linear or nonlinear relationship was found between the observed PM 2.5 effects and concentrations range across the studies, which cannot provide sufficient evidence for a threshold of currently observed PM 2.5 concentrations posing fatal respiratory effects in China. Although the results from our metaregression analysis were not statistically significant, the result is in line with previous research findings of lower RR with high PM 2.5 concentration with the exposure-response curve turning flat at extremely high PM 2.5 level [18,19,22,23,28].

Implication from This Research.
To the best of our knowledge, this research is the first meta-analysis that specifically looked into the concentration-response relationship between PM 2.5 and respiratory mortality in China. It provided an exhaustive screening of currently available literature and synthesized population-based information regarding combined risk estimate as percent increase in respiratory mortality per 10 g/m 3 increase in PM 2.5 concentrations. The assessment of fatal respiratory outcomes from real PM 2.5 data in China at high concentration levels also filled in the gaps between previously projected estimates only based on extrapolated data from western countries [11,20]. The study also responded to the evidence from a nationwide multicity investigation in China [28] as well as recommendations from similar studies conducted in India particularly on the role of fine PM and its effect on respiratory health among megacity residents [49]. By stratifying lag-day effects and regional divisions in the subgroup analysis, the research approach excluded potential confounding factors for estimates on respiratory effects. The findings from this study also recalled a positive association between PM 2.5 exposure and lung cancer mortality as well as respiratory diseases in a Japanese cohort, where the hazard ratio for lung cancer mortality associated with a 10 g/m 3 increase in PM 2.5 concentrations was 1.24 (95% CI: 1.12-1.37) [21]. Though focusing on different outcomes, the Japanese study regarding PM 2.5 and hazard ratio for mortality is a comparable reference for concentration-response associations in Asian populations.
As no specific thresholds were identified for fatal respiratory effects at observed PM 2.5 concentrations in China from this study and with uncertainties and little evidence for damage thresholds from previous studies on a global scale [29,50], notes should be taken that China's National Ambient Air Quality Standards [13] and WHO's AQGs values [10] may not guarantee the complete protection against adverse or fatal respiratory effects of PM 2.5 . We also hope that such findings could draw attention to public health strategies and environmental policies and call for more evidence-based decision making to address AAP at both global and local levels.

Strengths and Limitations.
Our systematic synthesis about the associations between PM 2.5 and respiratory mortality provides solid quantitative evidence for the evaluation and refinement of air quality guidelines and interim targets in consideration of country-specific situations and localized priorities. The study provides insights into future studies for the estimation of cause-specific mortality trends from observed and projected PM 2.5 levels as well as implications of fatal health consequences with deteriorating AAP in LMICs. In addition to categorizing lag-day structures which were distinctive across studies, we addressed the geographical difference between the south and north regions for a more thorough interpretation of the results. Based on the findings from available literature, we noticed a lack of data availability especially in regard to geographic variations in China for the recorded study period. Nevertheless, with limited amount of available studies at current stage, we made the first step investigating the concentration-response pattern among studies with a wider variation, which shed light on future investigations of concentration-response relationship between PM 2.5 and cause-specific mortality.
Although the meta-analysis took account of the influence of potential confounders and publication biases, the subgroup categorizations were based on prespecified confounding characteristics with limited data reported. During data extraction, we also attended to other confounders including temperature, humidity, and copollutants across studies; however, due to the limited studies, we were unable to perform further subgroup analysis. For subgroup analysis, the combined effect for single-days was not stable when considering publication bias, which warrants a cumulative meta-analysis to minimize the bias in the future. For geographic areas, in the northern cities, the data were mostly derived from the studies conducted in Beijing, while, for the southern cities, the data were mainly from Shanghai and Guangzhou. The limited cities could not let us extrapolate our findings to the larger geographic areas. We should also notice that only three included studies (Table 1) were conducted after PM 2.5 data were systematically measured and officially published in China from 2012. Limitations on consistency of PM 2.5 data measurement during different study period (before and after 2012) may lead to potential underestimation on risk effect associations and add more complexity for interpretation. To attend to broader data coverage as well as potential risk factors and confounders and address research gaps across existing and forthcoming literatures on PM 2.5 , future studies are needed to present a more comprehensive analysis and possibly derive a more conclusive association between PM 2.5 and its associated health effects. We have to admit that, with the small number of analysed studies in our analysis, the possible methods for detecting publication bias are underpowered and the estimation for standard error might be poor. Therefore, the nonparametric "trim-and-fill" method developed by Duval and Tweedie was applied to account for publication bias in meta-analysis. In our study, the results for multiday lags are quite similar between trimand-fill method and non-trim-and-fill method (Table 2). However, notable difference for single-day lags was found between the two methods (Table 2). Therefore, the bias from small number of studies cannot be ignored in our analysis. The trim-and-fill method is a rank-based data-imputation technique, which formalizes the use of funnel plots, estimates the number and outcomes of missing studies, and adjusts the meta-analysis to incorporate the imputed missing studies [51]. There exist a number of methods to estimate the number of missing studies, model the probability of publication, and provide an estimate of the underlying effect size. However, these methods are complex and highly computer-intensive to run and thus have failed to find acceptance in meta-analysis. Trim-and-fill analysis is a simple technique that seems to meet the objections to other methods and is effective and consistent with other adjusted adjustment methods [47].
In conclusion, although no statistically significant concentration-response trend was found in our metaanalysis, our study confirms and quantifies the negative association between PM 2.5 and respiratory mortality in China. PM 2.5 associated relative risk of respiratory mortality might be higher in certain southern cities in China. Further studies are needed to investigate the concentration-response effect of PM 2.5 exposure on fatal health outcomes in China as well as other LMICs where AAP has been one of the major public health threats.