Health burden assessments of air pollution commonly consider the impact of poor air quality on the risk of premature death, and a subset of these analyses also estimate chronic and acute effects.1–4 The causal relation between short-term (a few days up to several weeks) exposure to fine particulate matter (PM2.5) and acute morbidity endpoints (including hospital and emergency room visits for respiratory and cardiovascular diseases) is well established.5 A smaller, but growing, literature also finds associations between both short- and longer-term (months to years) exposures to PM and the incidence of chronic cardiovascular diseases including cerebrovascular disease outcomes (such as stroke).6–8 The sequelae to stroke may greatly influence the number of years a patient lives with a disability and contribute greatly to lost productivity1,9,10.
The most recent Integrated Science Assessment for Particulate Matter from the US Environmental Protection Agency (EPA)5 and a recent American Heart Association literature review11 conclude that short- and long-term exposure to PM2.5 are causally related to cardiovascular disease (CVD), including strokes.12–19 The literature yields few systematic reviews and meta-analyses of PM2.5-related strokes in particular. This study is distinct from recently published meta-analyses in 2 ways: first, it uses a novel statistical approach and second, it is designed to directly inform air pollution risk assessments.18
The goals of this article are 2-fold. First, we evaluated the current evidence regarding both short-term and long-term exposure to fine particulate air pollution and the incidence of ischemic stroke (ICD-9 433–444), hemorrhagic stroke (ICD-9 430–432), and cerebrovascular disease (ICD-9 430–438). We assessed the degree to which the literature supports an association between PM2.5 exposure and stroke.
Second, we drew upon the epidemiologic evidence to derive quantitative estimates of the risks for each type of stroke that may be included in air pollution risk assessments. We performed a quantitative meta-analysis that proceeds in 2 stages. In the first stage, we evaluate the strength of the epidemiologic evidence supporting the relation between PM2.5 and cerebrovascular disease by performing a random-effects meta-analysis to estimate pooled concentration-response relations.20 In the second, we reflect scientifically based conclusions of causality on the epidemiologic evidence by asserting a nonnegative prior. We followed this 2-stage approach because (1) it accounts for our belief, supported by the overall evidentiary base, that PM2.5 is unlikely to decrease the risk of stroke, and (2) not imposing this assumption may yield pooled estimates that include a negative lower confidence interval. In this latter case, health impact assessments applying these results will generally also report negative lower confidence intervals—an implausible result that is not consistent with the overall literature and is challenging to characterize.21
Identifying Cerebrovascular Outcomes
Evidence from clinical and toxicological studies supports a causal relation between exposure to PM2.5 and ischemic stroke, hemorrhagic stroke, and cerebrovascular disease.5 Ischemic stroke (ICD-9 433–444) characterized by a blood vessel blockage, accounts for about 80% of all cases; hemorrhagic strokes (ICD-9 430–432), characterized by bursting of blood vessels, account for the remaining 20% of cases.22 Cerebrovascular events (ICD-9 430–438) encompass these stroke outcomes and other transient events, as well as sequelae to stroke, including effects on speech and use of limbs.
The evidence for these latter effects is not as strong or consistent as that for cardiovascular disease, especially regarding long-term exposure to PM2.5. The reasons for this disparity are not well understood, but some evidence suggests that responses to PM2.5 may be modified by differences in exposures, exposure measurement errors, composition of PM2.5, and underlying population susceptibility, including use of statin drugs which can offset inflammatory responses.12 Indeed, many epidemiologic studies are prone to mischaracterize time of stroke onset and hence misclassify exposure and report a null result.11 To the extent that there is a relation between exposure to fine particles and cerebrovascular outcomes, these studies are likely to underestimate the health risks attributable to air pollution.
Pulmonary oxidative stress and systemic inflammation offer a plausible biological pathway describing the relation between long- and short-term PM exposure and stroke (Figure 1).12,13 PM2.5 may initiate a systemic inflammatory response even in the case of mild pulmonary inflammation.5 The recent Integrated Science Assessment by the US EPA finds that a number of other biological responses can mediate the pathway from systemic inflammation to the onset of stroke, including atherosclerosis, plaque rupture, pro-coagulation effects, and thrombosis.5 Determining whether PM and stroke are causally related should account for clinical and toxicological evidence, but developing quantitative risk functions requires epidemiologic literature. The effects of short- and long-term exposures to PM2.5 may be complementary, with longer-term exposures exacerbating susceptibility to shorter term PM2.5 elevations.23
We conducted searches of epidemiologic studies in the medical literature in Medline and PubMed, using the terms “particulate matter” or “air pollution” and “cerebrovascular” or “stroke,” and also reviewed lists of studies included in the US EPA Integrated Science Assessment for Particulate Matter.5 This initial search yielded 1801 studies (Figure 2). We did not identify any additional studies from alternate sources. From this set of 1801 studies, 1,574 remained after excluding duplicates and those published before 1990. We excluded 1443 studies that that did not include fine particulate matter (but rather PM10, total suspended particulate, or black smoke), focused only on second-hand smoke exposures, did not evaluate air pollution, or were not epidemiologic studies. We focused our review on the fine particle fraction because the epidemiologic, clinical and toxicological evidence finds the strongest relation between exposure to this size fraction and adverse health outcomes.5 In addition, the current standards for particulate matter in the U.S. and many other countries are based on PM2.5 concentrations.
Of the 131 remaining studies, we excluded 111 that lacked quantitative effect estimates (eg, risk ratios, beta coefficients) or that assessed fatal (rather than incident) stroke, leaving 20 studies that were the subject of the quantitative meta-analysis. (See eTable 1 for a list and description of these studies.) These 20 remaining studies reported at least the minimum level of detail regarding the study population, risk estimates, unit change in PM2.5, and type of stroke that would enable a quantitative meta-analysis; 1 cross-sectional study reported these minimum data, but upon further investigation the risk ratios proved not to be valid due to insufficient variation in PM2.5 across the study area. Certain attributes—including temperature, monitoring data used to quantify population exposure, and the measures the authors used to validate the stroke diagnosis—were reported inconsistently across studies. Other literature has underscored the importance of reporting such data to support quantitative meta-analyses and risk assessments.24
We used 2 statistical pooling approaches to reflect the 2 goals of this analysis. In the first procedure, we performed a traditional random-effects meta-analysis (ie, the frequentist approach), using the risk estimates reported in each study to characterize the overall strength of the evidence regarding the risk of PM2.5-related stroke.25 The random-effects meta-analysis also allowed us to evaluate between-study variation in the association between PM2.5 exposure and various cerebrovascular outcomes. However, when the number of study estimates pooled is small, this procedure too often fails to reject the null hypothesis of no heterogeneity—thus yielding an unrealistic characterization of uncertainty attributed completely to sampling error.26
To address this limitation, we introduce a Bayesian random-effects meta-analysis with 2 models. The first Bayesian approach treats the unknown overall risk and heterogeneity both as random variables; this is a typical model with a normal prior for the overall risk and an inverse-gamma prior for the heterogeneity. We favor this approach mainly due to its computational ease. However, the dispersion of the uncertainty distribution could become unrealistically large when only a small number of studies are available for analysis and can include negative values, implying a probability that increases in PM2.5 may decrease incidence of stroke. Risk analyses generally develop quantitative risk functions that permit both negative and positive risk estimates, regardless of the biological plausibility for such an outcome; the result can be health impact estimates whose quantitative bounds include substantial negative tails, implying that decreases in air pollution result in increases in strokes. To the extent that this is not biologically plausible, this would be a misleading result and not useful in informing policy decisions.
For these reasons, we propose the second model, which is a new meta-analytic method that combines features of both the frequentist and Bayesian approaches by adding our prior belief to the data.27 Specifically, we believe that the overall evidence supports a positive uncertainty risk distribution, reflecting the biological implausibility of stroke incidence decreasing as PM2.5 exposures increase. Thus, we assume a gamma prior with positive support to characterize the uncertainty distribution. In addition, the estimate of the heterogeneity from a noninformative prior is too imprecise to support pooling risk estimates; to overcome this, we use an empirical prior. The dispersion of the estimated uncertainty distribution is bounded above by the observed variation in study-specific risk estimates, a property of the frequentist approach. In this respect, the model blends classical and Bayesian approaches. For both pooling approaches, we drew from the literature review described above.
In our primary analysis, we preferentially selected risk estimates associated with distributed or cumulative lags in days of ambient PM2.5 exposure. Where this lag structure was unavailable and the author reported risk estimates associated with 2 or more lag periods, we selected the largest risk estimate available; we took this approach under the premise that the lag associated with the greatest effect estimate was capturing the critical window of exposure.28 Put differently, the unknown true lag structure between exposures to onset of strokes may not be in days but in hours.11 For example, if the true lag were 18 hours (ie, 0.75 day), the highest risk estimate would be observed for PM2.5 with 1 day lag. The question of lag structure is not relevant to long-term studies, which generally detect differences in risk between locations rather than over time. To the extent that a study reported risk estimates stratified by copollutant, we selected risk estimates associated with single-pollutant models, as the studies specified this model most frequently; by doing so we maximized the number of study-specific estimates available for pooling. After this approach to selecting risk estimates yielded 4 risk ratios from long-term studies and 221 risk ratios from short-term studies.
We evaluated the sensitivity of the results to several study attributes. We first characterized the sensitivity of the pooled risk estimate to the selection of risk estimates from single-pollutant models by performing a mixed-effects analysis in which we pooled risk estimates within each copollutant. We also applied a mixed-effects model to pool effect estimates by geographic area and age. Finally, using the random-effects meta-regression models, we attempted to predict risk levels based on study attributes, including the years in which the study was performed, as well as PM2.5 and temperature levels. Due to the limited number of studies with sufficient information on covariates, the meta-regression models were not sufficiently powered to detect whether any variables modified the PM2.5-stroke relation (results not shown).
As briefly discussed above, in the second pooling approach, we specified that the true uncertainty distribution of risk follows a gamma distribution, implying that risk must be positive. This model specification reflects our understanding, which is based on the overall PM2.5 health-effects literature, that exposure to PM2.5 cannot be protective and that negative estimates are biologically implausible. The observed study-specific logarithm of the risk estimate is assumed to be normally distributed with a gamma-distributed mean. The focus of the Bayesian meta-analysis is to estimate the mean and variance of the gamma distribution given the observed risks and underlying assumptions on the distributions of the unknown parameters. As part of the Bayesian approach, prior distributions of the mean and variance of the gamma distribution need to be specified, and a numerical value of the mean and variance of these prior distributions also needs to be provided. Typically we have no direct information on the value of these quantities; we thus supply a large (noninformative) value. However, when the number of studies examined is small, noninformative values yield very large, and often unreasonable, estimates of the gamma distribution variance. To bound this variance, we select the variance of the prior variance distribution to be no larger than the observed variation in the study-specific risk estimates, a property of the frequentist approach (S.H.H., unpublished data, 2013). In this manner, we are combining features of both the Bayesian and frequentist approaches to meta-analysis. Details of the pooling methodology are provided in the eAppendix (http://links.lww.com/EDE/A823).
After removing articles focused on second-hand smoke, mortality, and other measures of PM, we identified 20 studies investigating a relation between ambient PM2.5 and nonfatal stroke or cerebrovascular disease.6,29–45 Of these 20 cerebrovascular disease or stroke studies, 4 investigated the effects of long-term PM2.5 exposures,6,36,37,44 whereas the remaining 16 investigated short-term exposures. Some studies estimated the risks for strokes and total cerebrovascular disease separately; in others, only a combined cerebrovascular disease risk estimate was provided. In instances where both are available, we focus on the ischemic or hemorrhagic stroke risk estimates (and not cerebrovascular) because they are more clearly defined health outcomes and because they exclude transient and reversible effects. The risk estimates for each of the 20 studies are provided in Figure 3 and basic study attributes as shown in eTable 1 (http://links.lww.com/EDE/A823). Because of important differences between the long- and short-term studies of population exposure in the distributions, we pool these 2 sets of studies separately.
Four cohort studies characterized the relation between long-term PM2.5 exposures and cerebrovascular disease. The first is the Women’s Health Initiative (WHI) study, which estimated the effect of PM2.5 on both stroke and total cerebrovascular disease among postmenopausal women and examined time to onset of stroke as the outcome measure.6 The WHI study found a strong significant association between long-term PM2.5 exposures and first onset of stroke. The second is an analysis of the California Teacher’s Cohort comprising current and former female public school teachers; it assessed the risk of cerebrovascular disease due to PM2.5 exposure.44 This study found particularly strong risks among postmenopausal women. The third study followed a cohort of patients receiving treatment from general practices in England, using modeled air quality data to predict population exposure.45 This study found weak relations between long-term PM2.5 exposure and the risk of cerebrovascular disease. Finally, Kloog and coauthors36 used a land-use regression model in concert with remote sensing techniques to estimate exposure among a cohort of Medicare recipients, finding an increased risk of stroke from short- and long-term exposures.
Using the frequentist approach, pooling the 4 risk estimates from these 4 studies yields a pooled risk ratio estimate of 1.06 (95% confidence interval = 1.00–1.13) (Figure 4). A funnel plot analysis provides little evidence of asymmetry (PEgger = 0.25). The trim-and-fill method imputed 2 hypothetically missing studies, slightly attenuating the risk. However, because the overall number of long-term studies is small, the funnel plot analysis is not highly informative and so we did not include a figure here.
In addition to the classical frequentist approach, we used a Bayesian approach to the meta-analysis to account for all potential uncertainty by exploring 2 prior assumptions, normal or gamma distribution, for the unknown true study-specific risk.20,46 For the 4 estimates, the gamma prior returns the posterior mean of 1.05 (95% posterior interval [PI] = 1.02–1.10), whereas the normal prior yields the posterior mean of 1.08 (0.96–1.27). Note that the normal prior used noninformative distributions for all parameters involved in the model, assuming no previous knowledge on the study-specific risk and the mean risk. However, the gamma prior used a semi-informative distribution that required the risk to be positive. More details on statistical differences between normal and gamma priors can be found in the eAppendix (http://links.lww.com/EDE/A823).
Of the 16 short-term studies, we selected 221 risk estimates (202 of which were drawn from the Dominici et al32 multi-city study) that met the inclusion criteria we noted above. Of these 221 risk estimates, 141 were positive; 12 estimates were negative and statistically significant, whereas 23 were positive and statistically significant. Of the 16 studies examining short-term exposures, 10 were conducted in North America, 2 in Asia, 2 in Europe, and 1 in Australia. Although 14 of the studies used time-series or case-cross over approaches, 1 study followed a 2-stage modeling technique in which the authors first estimated city-specific risks and then pooled estimates across cities.30 Single-city estimates were available in the report by Dominici and colleagues,32 and so we pooled the 202 single-city estimates from this article with the single-city estimates from the remaining 15 studies. We discuss this procedure further below.
We first pooled the individual city time-series and case-crossover study estimates across the stroke endpoints, generating a pooled estimate (risk ratio) of 1.007 (95% confidence interval = 1.003–1.01) (Figure 5A). In our mixed-effects model, where we pooled within each stroke outcome, we find that the pooled ischemic stroke risk ratio is 1.04 (1.01–1.07), the cerebrovascular estimate was 1.006 (1.002–1.01), and the hemorrhagic estimate was 1.012 (0.92–1.11). In another mixed-effects model, where we pooled studies according to the continent in which they were performed, we generate a positive and significant pooled estimate (risk ratio) of 1.008 (1.004–1.013) for North America, a negative estimate for Europe, and a negative estimate for Asia (results not shown). The funnel plot analysis (Figure 5B) provides some evidence of asymmetry (PEgger = 0.0607).47 The trim-and-fill method imputed 18 hypothetically missing study estimates, slightly attenuating the risk estimate (results not shown).
Pooling the 221 estimates across all 3 health endpoints using the Bayesian approach, we obtained a posterior mean risk ratio of 1.008 (95% posterior interval = 1.004–1.013) from the gamma prior and 1.008 (1.003–1.013) from the normal prior (Figure 6). For all cerebrovascular disease combined (213 estimates), the Bayesian approach again returned very similar results: the posterior mean of the risk ratio 1.007 (1.004–1.012) from the gamma prior and 1.007 (1.002–1.012) from normal prior. Cerebrovascular disease is the dominant cause of stroke and covers more than 95% of all estimates—and for this reason, the results for total cerebrovascular disease (n = 213) and all stroke types (n = 221) do not differ substantially (Figure 6).
However, for ischemic and hemorrhagic strokes, the risk ratio was much wider than for cerebrovascular disease. The posterior mean risk ratio of ischemic stroke was 1.05 (95% posterior interval = 1.01–1.09) from the gamma prior and 1.05 (0.99–1.14) from the normal prior. The risk ratio of hemorrhagic stroke was a bit lower but with wider credible interval than ischemic stroke: the posterior mean of 1.02 (1.00–1.06) from the gamma prior and 1.01 (0.84–1.25) from the normal prior. Figure 7 displays the difference over cause (ischemic versus hemorrhagic strokes) and over prior distribution (normal versus gamma distributions).
To assess the influence of the 202 city risk ratios from the study by Dominici et al,32 we excluded these values and then pooled across the remaining study risk estimates (n = 19) and for the cerebrovascular endpoint alone (n = 11). The results are displayed in eFigure 1 (http://links.lww.com/EDE/A823). Both prior distributions, normal and gamma, returned comparable posterior medians (represented by dots), but the normal prior returned much wider posterior intervals that cover unrealistic negative risk values.
After examining 3 types of stroke—cerebrovascular, ischemic, and hemorrhagic—we conclude that the evidence supports a causal relation between PM2.5 exposure and cerebrovascular disease (strokes), particularly ischemic strokes associated with short-term exposure to PM2.5. Our conclusions are generally consistent with several other recent reviews.19 Both pooling approaches—frequentist and Bayesian—yield small, nonzero short- and long-term risk estimates. The results of the short-term risk were fairly consistent across the 2 pooling methods; with both the frequentist and Bayesian approaches, there are increases in excess risk ratio, but they are small. The frequentist and Bayesian techniques each report a small positive estimate for ischemic stroke. Although the frequentist and Bayesian gamma models generate a small pooled estimate for ischemic stroke, the Bayesian normal prior does not. Taken together, these results suggest a stronger relation between short-term PM2.5 exposure and ischemic stroke than the other 2 strokes. Both the long- and short-term studies demonstrate a limited degree of funnel plot asymmetry, suggesting that these results are not greatly influenced by publication bias. The fact that about 90 percent of the short-term estimates came from a single study is a source of bias, as it reduces the level of between-study heterogeneity. However, the population in that study of those estimates were distributed throughout the United States, were composed of multiple ethnic backgrounds, and were exposed to a range of PM2.5 levels.
Differences in the pooled risk estimates for the short- and long-term studies may be attributed to the time periods caused by these studies. For example, the long-term studies integrated the effects of the most recent exposure (in hours or days), as well as chronic exposures to air pollution that affect the underlying cardiovascular pathologies, which in turn increase a person’s propensity to suffer a stroke; these studies will also account for stroke events triggered by other acute causes on days in which air quality is good. For these reasons, long-term studies tend to observe much larger effects if air pollution causes chronic cardiovascular pathologies, such as atherosclerosis. Given that the long- and short-term studies are observing different effects, it would be inappropriate for risk assessors to use both pooled effect estimates in the same health impact analysis—doing so would likely incorrectly estimate effects.
This article demonstrates how the frequentist and Bayesian approaches may be applied in a complementary manner to produce pooled risk estimates that may inform air pollution risk assessments. In this analysis, we applied the frequentist approach as a first step, probing the extent to which PM2.5 exposure was associated with various stroke outcomes; we found a positive relation between long- and short-term exposure and stroke. These findings—and the strong evidence that exposure to PM2.5 could not be health-protective—then informed our prior belief that the PM2.5-related risks of stroke may be positive, but not zero or negative, arguing for the use of the Bayesian model. Such an approach may prove useful in future air pollution meta-analyses—particularly those in which there are a small number of estimates or those for which the estimates are substantially skewed. However, given the somewhat inconsistent support for a positive, nonzero effect, the use of the proposed Bayesian gamma model may place too much weight on strictly positive risk estimates, and thus future approaches should explore the feasibility of applying both zero and gamma prior models.
This Bayesian approach exhibits 2 key strengths that make it particularly well suited to generating risk distributions that inform air pollution health impact assessments. First, it offers an opportunity to inform the shape of the risk distribution with prior knowledge about the biological plausibility of air pollution affecting health without distorting that distribution. Conversely, alternative approaches to adjusting the uncertainty distribution—for example, ignoring negative values or centering the distribution on zero—distort the distribution. Second, this approach ensures that if the evidence was not sufficient to support a strictly positive distribution, then the model would not converge.
We thank H. Ross Anderson for help in developing the conceptual basis for this article, Ana Rappold for developing the funnel plots, and Matthew Strickland for feedback on an earlier version of the abstract.
1. Lim SS, Vos T, Flaxman AD, et al. A comparative risk assessment of burden of disease and injury attributable to 67 risk factors and risk factor clusters in 21 regions, 1990–2010: a systematic analysis for the Global Burden of Disease Study 2010. Lancet. 2012;380:2224–2260
2. Anenberg SC, Horowitz LW, Tong DQ, West JJ. An estimate of the global burden of anthropogenic ozone and fine particulate matter on premature human mortality using atmospheric modeling. Environ Health Perspect. 2010;118:1189–1195
3. Fann N, Lamson AD, Anenberg SC, Wesson K, Risley D, Hubbell BJ. Estimating the national public health burden associated with exposure to ambient PM(2.5) and ozone. Risk Anal. 2012;32:81–95
4. Murray CJ, Vos T, Lozano R, et al. Disability-adjusted life years (DALYs) for 291 diseases and injuries in 21 regions, 1990–2010: a systematic analysis for the Global Burden of Disease Study 2010. Lancet. 2012;380:2197–2223
6. Miller KA, Siscovick DS, Sheppard L, et al. Long-term exposure to air pollution and incidence of cardiovascular events in women. N Engl J Med. 2007;356:447–458
7. Bauer M, Moebus S, Möhlenkamp S, et al.HNR Study Investigative Group. Urban particulate matter air pollution is associated with subclinical atherosclerosis: results from the HNR (Heinz Nixdorf Recall) study. J Am Coll Cardiol. 2010;56:1803–1808
8. Künzli N, Jerrett M, Garcia-Esteban R, et al. Ambient air pollution and the progression of atherosclerosis in adults. PLoS One. 2010;5:e9096
9. Ganz DA, Kuntz KM, Jacobson GA, Avorn J. Cost-effectiveness of 3-hydroxy-3-methylglutaryl coenzyme A reductase inhibitor therapy in older patients with myocardial infarction. Ann Intern Med. 2000;132:780–787
10. Mittleman MA, Wellenius GA Air Pollution and Stroke
. PowerPoint lecture presented to EPA STAR Applied Science Webinar Series. 18 July 2012.
11. Brook RD, Rajagopalan S, Pope CA 3rd, et al.American Heart Association Council on Epidemiology and Prevention, Council on the Kidney in Cardiovascular Disease, and Council on Nutrition, Physical Activity and Metabolism. Particulate matter air pollution and cardiovascular disease: an update to the scientific statement from the American Heart Association. Circulation. 2010;121:2331–2378
12. Wellenius GA, Burger MR, Coull BA, et al. Ambient air pollution and the risk of acute ischemic stroke. Arch Intern Med. 2012;172:229–234
13. Brook RD. Cardiovascular effects of air pollution. Clin Sci (Lond). 2008;115:175–187
14. Brook RD, Franklin B, Cascio W, et al.Expert Panel on Population and Prevention Science of the American Heart Association. Air pollution and cardiovascular disease: a statement for healthcare professionals from the Expert Panel on Population and Prevention Science of the American Heart Association. Circulation. 2004;109:2655–2671
15. Maitre A, Bonneterre V, Huillard L, Sabatier P, de Gaudemaris R. Impact of urban atmospheric pollution on coronary disease. Eur Heart J. 2006;27:2275–2284
16. Franchini M, Mannucci PM. Short-term effects of air pollution on cardiovascular diseases: outcomes and mechanisms. J Thromb Haemost. 2007;5:2169–2174
17. Mills NL, Donaldson K, Hadoke PW, et al. Adverse cardiovascular effects of air pollution. Nat Clin Pract Cardiovasc Med. 2009;6:36–44
18. Peters A. Particulate matter and heart disease: evidence from epidemiological studies. Toxicol Appl Pharmacol. 2005;207(2 suppl):477–482
19. Li XY, Yu XB, Liang WW, et al. Meta-analysis of association between particulate matter and stroke attack. CNS Neurosci Ther. 2012;18:501–508
20. Whitehead A Meta-Analysis of controlled Clinical Trials. 2002 Chichester, West Sussex, England John Wiley & Sons, Ltd
22. Warlow CP. Epidemiology of stroke. Lancet. 1998;352(3 suppl):S1–S4
23. Künzli N, Medina S, Kaiser R, Quénel P, Horak F Jr, Studnicka M. Assessment of deaths attributable to air pollution: should we use risk estimates based on time series or on cohort studies? Am J Epidemiol. 2001;153:1050–1055
24. Fann N, Bell ML, Walker K, Hubbell B. Improving the linkages between air pollution epidemiology and quantitative risk assessment. Environ Health Perspect. 2011;119:1671–1675
25. Borenstein M, Hedges L, Higgins J, Rothstein H Comprehensive Meta-Analysis. 2005 New Jersey Wiley
26. Huedo-Medina T, Sanchez-Meca J, Marin-Martinez F, Botella J Assessing Heterogeneity in Meta-Analysis: Q Statistic or I2 Index? Center for Health, Inervention, and Prevention Report 6-1-2006. 2006 University of Connecticut Available at: http://digitalcommons.uconn.edu/chip_docs/19/
. Accessed 12 November 2013
27. Mehta S, Shin H, Burnett R, North T, Cohen AJ. Ambient particulate air pollution and acute lower respiratory infections: a systematic review and implications for estimating the global burden of disease. Air Qual Atmos Health. 2013;6:69–83
28. Lokken RP, Wellenius GA, Coull BA, et al. Air pollution and risk of stroke: underestimation of effect due to misclassification of time of event onset. Epidemiology. 2009;20:137–142
29. Anderson HR, Bremner SA, Atkinson RW, Harrison RM, Walters S. Particulate matter and daily mortality and hospital admissions in the west midlands conurbation of the United Kingdom: associations with fine and coarse particles, black smoke and sulphate. Occup Environ Med. 2001;58:504–510
30. Bell ML, Levy JK, Lin Z. The effect of sandstorms and air pollution on cause-specific hospital admissions in Taipei, Taiwan. Occup Environ Med. 2008;65:104–111
31. Chan CC, Chuang KJ, Chien LC, Chen WJ, Chang WT. Urban air pollution and emergency admissions for cerebrovascular diseases in Taipei, Taiwan. Eur Heart J. 2006;27:1238–1244
32. Dominici F, Peng RD, Bell ML, et al. Fine particulate air pollution and hospital admission for cardiovascular and respiratory diseases. JAMA. 2006;295:1127–1134
33. Haley VB, Talbot TO, Felton HD. Surveillance of the short-term impact of fine particle air pollution on cardiovascular disease hospitalizations in New York State. Environ Health. 2009;8:42
34. Halonen JI, Lanki T, Yli-Tuomi T, Tiittanen P, Kulmala M, Pekkanen J. Particulate air pollution and acute cardiorespiratory hospital admissions and mortality among the elderly. Epidemiology. 2009;20:143–153
35. Jalaludin B, Morgan G, Lincoln D, Sheppeard V, Simpson R, Corbett S. Associations between ambient air pollution and daily emergency department attendances for cardiovascular disease in the elderly (65+ years), Sydney, Australia. J Expo Sci Environ Epidemiol. 2006;16:225–237
36. Kim SY, Peel JL, Hannigan MP, et al. The temporal lag structure of short-term associations of fine particulate matter chemical constituents and cardiovascular and respiratory hospitalizations. Environ Health Perspect. 2012;120:1094–1099
37. Kloog I, Coull BA, Zanobetti A, Koutrakis P, Schwartz JD. Acute and chronic effects of particles on hospital admissions in New-England. PLoS One. 2012;7:e34664
38. Lippmann M, Ito K, Nádas A, Burnett RT. Association of particulate matter components with daily mortality and morbidity in urban populations. Res Rep (Health Eff Inst). 2000:5–72
39. Lisabeth LD, Escobar JD, Dvonch JT, et al. Ambient air pollution and risk for ischemic stroke and transient ischemic attack. Ann Neurol. 2008;64:53–59
40. Metzger KB, Tolbert PE, Klein M, et al. Ambient air pollution and cardiovascular emergency department visits. Epidemiology. 2004;15:46–56
41. Moolgavkar SH. Air pollution and hospital admissions for diseases of the circulatory system in three U.S. metropolitan areas. J Air Waste Manag Assoc. 2000;50:1199–1206
42. Villeneuve PJ, Johnson JY, Pasichnyk D, Lowes J, Kirkland S, Rowe BH. Short-term effects of ambient air pollution on stroke: who is most vulnerable? Sci Total Environ. 2012;430:193–201
43. Villeneuve PJ, Chen L, Stieb D, Rowe BH. Associations between outdoor air pollution and emergency department visits for stroke in Edmonton, Canada. Eur J Epidemiol. 2006;21:689–700
44. Lipsett MJ, Ostro BD, Reynolds P, et al. Long-term exposure to air pollution and cardiorespiratory disease in the California teachers study cohort. Am J Respir Crit Care Med. 2011;184:828–835
45. Atkinson RW, Carey IM, Kent AJ, van Staa TP, Anderson HR, Cook DG. Long-term exposure to outdoor air pollution and incidence of cardiovascular diseases. Epidemiology. 2013;24:44–53
46. Carlin BP, Louis TA Bayesian Methods for Data Analysis. 20093rd ed Florida Taylor & Francis Group, LLC
47. Viechtbauer W. Conducting meta-analyses in R with the metafor package. J Stat Softw. 2010;36:1–48 Available at: http://www.jstatsoft.org/v36/i03/
. Accessed 5 April 2014