Introduction Tuberculosis (TB) causes an estimated 2 million deaths per year, the majority of which occur in the developing world. Many studies conducted over the past 60 years have found an association between tobacco smoking and TB, as manifested by a positive tuberculin skin test (TST) or as active disease and its sequelae. A smaller number have found that indoor air pollution from biomass fuels (IAP) and passive smoking are also risk factors for TB and its sequelae. Tobacco smoking has increased substantially in developing countries over the past three decades, with an estimated 930 million of the world's 1.1 billion smokers currently living in the low-income and middle-income countries [1,2]. Approximately half of the world's population uses coal and biomass, in the form of wood, animal dung, crop residues, and charcoal as cooking and heating fuels especially in Africa and Asia. Given the persistent or growing exposure to both smoking and IAP in regions where TB poses a major health risk, it is essential to delineate the role of these environmental factors in the etiology and epidemiology of TB. Previous reviews have addressed qualitatively the epidemiologic and biologic link between tobacco smoke and TB, but have not systematically reviewed the epidemiologic data on this association [3,4]. We therefore undertook to quantitatively assess the association between smoking, passive smoking, and IAP, and the risk of infection, disease, and death from TB. We have considered smoking, passive smoking, and IAP together because these sources result in exposure to common set of respirable pollutants, and because their effects are currently or increasingly found in the developing countries. Methods Data Source We searched the PubMed via the NCBI Entrez system (1950 to February 1, 2006) (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi) and the EMBASE via Ovid (1988 to 2003) (http://www.ovid.com) for studies of the association between smoking, passive smoking, and indoor air pollution and TB infection, disease, and mortality. We also searched bibliographies of identified reports for additional references. Our search strategy is described in Box 1. Study Selection We limited our search to studies published in English, Russian, and Chinese. Studies were included if they involved human participants with TB or at risk from TB. We included studies if a quantitative effect estimate of the association between ever, former, or current tobacco smoking, passive smoking, or IAP, and TST positivity, clinical TB disease, or TB mortality was presented or could be estimated from the data provided in the paper or through contact with the authors. Studies were included in the review if they were full-length peer-reviewed reports of cohort studies, case-control studies, or cross-sectional studies, if they controlled for possible confounding by age or age group, and if they screened for the presence of TB among exposed and unexposed study participants in the same way. For analyses of the effect of passive smoking on TB outcomes, we excluded studies if they did not restrict the population under study to nonsmokers. If multiple published reports from the same study participants were available, we included only the one with the most detailed information for both outcome and exposure. Box 1. Search Strategy and Terms Used to Identify Studies on Smoking and TB MeSH term search 1. “tuberculosis” 2. “smoking” 3. “air pollution, indoor” 4. “biomass” 5. “fuel oils” 6. “(1) AND (2)” OR “(1) AND (3)” OR “(1) AND (4)” OR “(1) AND (5)” Direct keyword search: 7. “tuberculosis” 8. “smoking” 9. “indoor air pollution” 10. “cooking fuel” 11. “biomass” 12. “(7) AND (8)” OR “(7) AND (9)” OR “(7) AND (10)” OR “(7) AND (11)” 13. (6) OR (12) Data Extraction and Quality Assessment For every eligible study, we collected detailed information on year and country of study, study design, study population, sample size, choice of controls, definition and measurement of tobacco smoking or IAP, type of TB outcome, confounders adjusted for, effect sizes and 95% confidence intervals (CIs), and dose-response relationships. Since TB disease and death are relatively rare events, even in high-incidence areas, we assumed that odds ratios (ORs), risk ratios, and rate ratios all provided an equivalent estimate of risk and therefore reported them as ORs . Although latent TB infection is not a rare event, each of the studies of latent TB infection estimated ORs and we therefore reported ORs for this outcome as well. Data were extracted independently by two of the investigators (HL and MM), and differences were resolved by discussion with a third (ME). Data Synthesis We performed separate analyses for each exposure-outcome association that had been studied. Within each subanalysis we further stratified on different study designs. When more than one study used a specific study design, we assessed heterogeneity using the I 2 statistic described by Higgins et al. . Because of the significant heterogeneity and different study designs within subgroups, we did not compute pooled effect measures . Instead, we graphically presented each of the weighted point estimates and 95% CIs of effect estimates for individual studies within subanalyses. For the subanalysis in which we found no significant heterogeneity, effect estimates were given a weight equal to the inverse variance of the study (fixed effects model). For those subanalyses in which we noted significant heterogeneity, we used a random effects model to assign the weight of each study according to the method described by DerSimonian and Laird . In order to assess the effect of study quality on the reported effect estimates, we conducted sensitivity analyses in which we compared pooled effect estimates for subgroups stratified on quality-associated study characteristics including study design (cohort, case-control or cross-sectional), type of control selection (population based or other), adjustment for important potential confounder (alcohol and socioeconomic status), and outcome classification (microbiological or other). We considered studies to be of higher quality if they (1) were cohort studies, (2) were case-control studies using population-based controls, (3) adjusted for important confounders, (4) classified the outcome on the basis of microbiological findings, and (5) restricted the outcome to pulmonary TB. As above, pooled estimates were calculated using a fixed effects model if there was no significant heterogeneity and a random effects models for those subanalyses in which we found heterogeneity. We tested for possible publication bias using Begg's and Egger's tests and by visual inspection for asymmetry of a plot of the natural logarithms of the effect estimates against their standard errors according to method described by Begg [9,10]. Several large studies on smoking and TB mortality had highly variable results and thus fell outside the lines of the funnel plot. Therefore, we conducted a sensitivity analysis in which we repeated the funnel plot excluding all of the mortality studies. All statistical procedures were carried out in Intercooled Stata Version 8.2 (Stata, http://www.stata.com). Results We identified and screened 1,397 papers by titles and abstracts. We excluded 1,340 papers because they were judged not to be related to smoking, IAP, and TB. The remaining 57 articles were obtained for detailed review; 19 of these were excluded because the same studies were published in different journals [11,12], the effect sizes and CIs of interest were not reported or could not be estimated [13–24], there were severe flaws in study design [25–27], or the article was not original [28,29]. Thirty-eight papers were included in the final analysis. Figure 1 delineates the exclusion process and Table 1 summarizes the studies that were included in the final analysis. Tobacco Smoking and Latent TB Infection Figure 2 shows the risk of latent TB among smokers compared with nonsmokers in six studies [30–35] on tobacco smoking and latent infection. The studies were conducted in five countries: the US, Spain, South Africa, Pakistan, and Vietnam. Although the timing of smoking (current, former, and ever) in relation to the study varied, we did not differentiate between these reported exposures, because the actual time of TB infection was unknown. There was only one case-control study; for the five cross-sectional studies that were included, we found minimal heterogeneity (I 2 = 0%). We also stratified studies that used different cutoffs for the TST; among those analyses that used induration size of 5 mm as the cutoff for a positive test [32,33], the pooled OR for latent TB was 2.08 (95% CI, 1.53–2.83), while among those that used a 10 mm cutoff [30,31,34,35], the pooled OR was 1.83 (95% CI, 1.49–2.23). When we stratified on other quality-associated study characteristics, we found that ORs for TB infection were lower among studies that adjusted for alcohol (Table 2), but that a positive effect of smoking on latent TB remained. Tobacco Smoking and Clinical TB Disease The 23 studies that we identified on the association between tobacco smoking and clinical TB disease were conducted in 12 countries: China/Hong Kong, India, The Gambia, Guinee Conakry, Guinea Bissau, US, UK, Australia, Malawi, Estonia, Spain, and Thailand [2,36–57]. Figures 3–5 shows the risk of clinical TB among current, former, and ever smokers, respectively, compared to nonsmokers for the individual studies. Given the significant heterogeneity among each of these effect estimates, we do not report pooled estimates within each of these three categories; rather, we stratified on important study characteristics within each category for the purpose of sensitivity analysis (Table 3). These analyses show that there was a significantly increased risk of clinical TB among smokers regardless of outcome definition (pulmonary TB versus any TB), adjustment for alcohol intake or socioeconomic status, type of study, or choice of controls. Although stratification by these study-specific variables did not fully explain the variability between studies, heterogeneity was partially accounted for by outcome (pulmonary versus any TB) and by adjustment for alcohol intake. As might be predicted on the basis of biological plausibility, we found a higher risk of clinical TB among smokers when we restricted the analyses to studies that included only cases of pulmonary disease. However, the differences between the effect estimates for pulmonary TB and those for any TB were not statistically significant. Tobacco Smoking and TB Mortality We identified five studies on tobacco smoking and TB mortality in adults [2,58–61], conducted in India, South Africa, and China/Hong Kong. Although all of the studies found a positive association between smoking and TB mortality (Figure 6), there was substantial heterogeneity (I 2 = 98.5% among case-control studies) and a five-fold difference between the most extreme effect estimates. We therefore do not report a pooled estimate for this analysis. A dose-response relation was noted in the two [59,60] studies that stratified on dose. When we conducted a sensitivity analysis excluding the study conducted in India where TB may have been differentially overdetected among smokers [2,61], heterogeneity was markedly reduced (I 2 = 38.6%). Other sensitivity analyses are demonstrated in Table 4. Passive Smoking and TB We identified five studies on passive smoking and TB, of which four were case-control studies assessing the risk of clinical TB [50,53–55,62,63] and one a cross-sectional study on the risk of latent infection . Two studies did not exclude active smokers while assessing passive smoking and were, therefore, not included in the analysis of passive smoking and TB [50,53]. Figure 7 shows the individual effect measures for the studies on active disease; each found a positive association between passive smoking and TB. The heterogeneity among the studies was largely explained by the age of the participants; the risk of TB among children exposed to passive smoking was significantly higher than it was among adults (p = 0.002), and there was no remaining heterogeneity within the subgroups stratified by age. The single study examining the risk of latent TB infection among those exposed to passive smoking reported an OR of 2.68 (95% CI, 1.52–4.71) . Sensitivity analyses for these estimates are given in Table 5. A dose response was found in both of the two studies that stratified on exposure intensity; one found that TB risk increased with the number of cigarettes smoked by the family per day , and the other found that close and very close contact with smoking household members was strongly associated with TB (adjusted OR 9.31 [95% CI, 3.14–27.58]), while distant contact was not (adjusted OR 0.54 [95% CI, 0.25–1.16]) . IAP and Clinical TB Disease Only five studies of IAP and TB were identified (Figure 8) [36,42,48,65,66]. Of these, only two studies adjusted for tobacco smoking [42,66] while three others did not [36,48,65]. In each of the studies, IAP was assessed by questionnaire on cooking and heating with biomass fuels (wood or dung). Although three of the five studies reported a positive association between biomass use and TB disease, there was significant heterogeneity among the studies (I 2 = 74.1% in case-control studies) (Figure 8). We noted that in one study, houses were reportedly well ventilated and therefore the impact of IAP might have been attenuated . The sensitivity analyses are presented in Table 6. Publication Bias When we plotted the natural logarithms of the effect estimates against their standard errors using the methods described by Begg (Figure 9A) , we detected some slight asymmetry of effect estimates among small studies. We also noted that several large studies fell outside the projected lines of the funnel plot, indicating substantial variability among studies with small standard errors. When we repeated this analysis excluding the five mortality studies, we found that the studies with small standard errors clustered within the funnel plot (Figure 9B). We found no evidence for substantial publication bias by either the Begg's test (p = 0.256) or the Egger's test (p = 0.977). Discussion This analysis shows that exposure to tobacco smoke is consistently associated with TB, regardless of the specific types of exposures and specific TB outcomes. Compared with people who do not smoke, smokers have an increased risk of a positive TST, of having active TB, and of dying from TB. Although there were fewer studies for passive smoking and IAP from biomass fuels, those exposed to these sources were found to have higher risks of active TB than those who are not exposed. An important finding of this study is the suggestion that the risk of TB among those exposed to passive smoking is especially high in children who are not normally at high risk for active disease. These findings support the hypothesis that exposure to respirable pollutants from combustion of tobacco and biomass fuels increases the risk of both TB infection and TB disease. In addition to the positive association demonstrated here, multiples lines of evidence support a causal relationship between combustion smoke and TB. A dose–response relationship has been demonstrated in most of the studies that have stratified on dose; in this meta-analysis, we found that the risk of TB increases with both daily dose of cigarettes and duration of smoking. There is also accumulating evidence for the biological plausibility of this association. Chronic exposure to tobacco as well as to a number of environmental pollutants impairs the normal clearance of secretions on the tracheobronchial mucosal surface and may thus allow the causative organism, Mycobacterium tuberculosis, to escape the first level of host defenses, which prevent bacilli from reaching the alveoli . Smoke also impairs the function of pulmonary alveolar macrophages (AMs), which are not only the cellular target of M. tuberculosis infection but also constitute an important early defense mechanism against the bacteria; AMs isolated from the lungs of smokers have reduced phagocytic ability and a lower level of secreted proinflammatory cytokines than do those from nonsmokers . Recent work has suggested a novel mechanism for this effect; nicotine is hypothesized to act directly on nicotinic acetylcholine receptors on macrophages to decrease intracellular tumor necrosis factor-α production and thus impair intracellular killing of M. tuberculosis . Wood smoke exposure in rabbits has also been shown to negatively affect antibacterial properties of AMs, such adherence to surfaces, ability to phagocytize bacteria, and intracellular bactericidal processes . Boelaert and colleagues  have also proposed an alternative explanation for the impaired ability of macrophages from smokers to contain M. tuberculosis infection. These investigators noted that AMs from smokers have an markedly elevated iron content and that macrophage iron overload impairs defense against intracellular microorganisms through reduced production of both tumor necrosis factor-α and nitric oxide. The available data support a causal link between smoke exposure and either an increased chance of acquiring TB or progression of TB to clinical disease. Our study shows that the risk of latent TB among smokers is quantitatively similar to their risk of active disease, which would suggest that much of the impact of smoking takes place during infection. At the same time, one case-control study selected TST-positive controls, thereby comparing patients who were TST positive and had clinical TB to people who were also TST positive but had not progressed to clinical TB ; that study also found a strong association between smoking and disease, suggesting that smoking may induce progression or reactivation disease in those infected. We included the outcome TB mortality in this study in order to investigate the association between smoke and TB occurrence rather than the association between smoke and TB treatment outcomes. The risk of death from TB among smokers was found to be somewhat higher than the risk of latent infection or disease, possibly because smoking has been identified as a risk factor for poor TB treatment outcomes among those undergoing therapy [57,72,73]. There are several potential limitations to this study. First, our findings are based on the results of observational studies; we cannot, therefore, exclude the possibility of confounding by variables that may be associated with each of the exposures. The issue of confounding is particularly a concern in a meta-analysis of observational studies when effect sizes are relatively small, as was the case in the studies considered in this analysis . We therefore performed a stratified analysis to explore the degree to which potential confounders may have influenced the findings. Among possible confounders, alcohol use is a known risk factor for TB and is closely associated with tobacco use in many populations. Those studies that adjusted for alcohol intake in a multivariable model found that the effect of smoking was reduced, but not eliminated. Those studies that controlled for the effect of alcohol were also less heterogeneous as a group than those that did not, a finding which suggests that some of the variability may have resulted from differences in alcohol consumption. Other risk factors that may confound the association between smoking, passive smoking, and IAP and TB include socioeconomic status, gender, and age. Although it is impossible to fully exclude bias introduced by residual confounding, we found that the effects the exposures on TB remained after adjustment for these factors. More than half of the studies in our review are case-control studies. These used different approaches to the selection of controls, including sampling from hospitals and clinics, from household members, and from the community. Since smoking is associated with a wide range of diseases, the choice of hospital- or clinic-based sampling may lead to over-representation of smokers among the controls, thereby biasing the results toward the null. Similarly, since people dwelling in the same household may share behavioral risk factors, controls chosen from households of smoking TB patients may have been more likely to smoke than would the general population . When we compared the effect estimates for studies stratified on the basis of the control selection strategy, we found that studies that had not used population-based controls tended to report lower effect estimates, consistent with our expectation of a bias toward the null among studies that used hospital- and household-based controls. Other potential sources of bias include possible misclassification of both exposure and outcome status. The assessment of tobacco smoking relied on self-reported behavior, which may not have been accurate especially among those who consider smoking to be stigmatizing, such as women in some cultural settings. The exposure “current smoking” may also have been subject to reverse causation. Patients are often diagnosed with TB months or more after having first experienced symptoms of the disease, which may cause some patients to quit smoking. This is consistent with the finding of several studies that “former” smoking to be a stronger risk factor for TB than current smoking [34,42,48]. Nonetheless, since “former” smoking also included very distant smoking, both current and former smoking may underestimate the effect of smoking that had occurred just prior to the onset of disease. Similarly, misclassification of passive smoking and IAP may have introduced a bias toward the null in our analysis. The classification of passive smoking among children, for example, relied on parent reports, which may have been influenced by guilt or shame at having exposed the child to an agent suspected of causing disease. Most problematic among exposures was the classification of IAP; this usually relied on the proxy “use of biomass cooking fuel,” which probably only coarsely captured the actual exposure to inhaled smoke. For example, one study that found no association between biomass fuel use and TB noted that houses in the area were well ventilated, and thus actual exposure to inhaled smoke among those using biomass fuels was probably lower. Misclassification of outcome may have also introduced bias into this analysis. For example, we included a large mortality study conducted in India in which the odds of death among urban male smokers was 4.5 times that of nonsmokers. Since diagnosis of TB in India relies heavily on radiographic findings, TB may be overdetected, especially among patients with pulmonary lesions—such as malignancies—that may be causally linked to smoking . When we repeated our analysis excluding the two Indian mortality studies, the heterogeneity among the remaining studies was reduced. Similarly, when the mortality studies were excluded from the funnel plot, there was much less variability among the studies with the smallest standard errors. Another possible source of outcome misclassification was suggested by Plant and colleagues , who noted that the frequency of small induration sizes among TSTs was higher among smokers than nonsmokers, suggesting that smokers may be less capable than nonsmokers of eliciting a vigorous skin test reaction and that latent TB infection in smokers may thus be underdetected when the 10 mm cutoff is used. Despite this possible limitation, we found that the two studies of latent infection that used 5 mm cutoffs for the TST [32,33] reported effects that were not statistically different from those that used 10 mm [30,31,34,35]. Finally, the diagnosis of TB in children is notoriously difficult; if children exposed to passive smoke were more likely to be successfully diagnosed with disease than those who were not, this might have introduced a bias that would explain the strong positive association between passive smoking and TB. Although our evidence suggests that tobacco smoking is only a moderate risk factor in TB, the implication for global health is critical. Because tobacco smoking has increased in developing countries where TB is prevalent, a considerable portion of global burden of TB may be attributed to tobacco smoking (see Text S1 for an illustrative calculation of population-attributable fraction and attributable deaths in different regions of the world). More importantly, this association implies that smoking cessation might provide benefits for global TB control in addition to those for chronic diseases. Despite heterogeneity in design, measurement, and quantitative effect estimates among the studies included in this analysis, we found consistent evidence for an increased risk of TB as a result of smoking, with more limited but consistent evidence for passive smoking and IAP as risk factors. These findings suggest that TB detection might benefit from information on exposure to respirable pollutants from sources such as smoking and biomass use, and that TB control might benefit from including interventions aimed at reducing tobacco and IAP exposure, especially among those at high risk for exposure to the infection. Supporting Information Text S1 Population-Attributable Fraction and Attributable Death Due to Tobacco Smoking on TB Mortality in Different Regions of the World (33 KB DOC) Click here for additional data file. Alternative Language Abstract S1 Translation of the Article into Chinese by Hsien-Ho Lin (64 KB PDF) Click here for additional data file.