Lung Cancer Risk in Painters: A Meta-Analysis

Objective We conducted a meta-analysis to quantitatively compare the association between occupation as a painter and the incidence or mortality from lung cancer. Data sources PubMed and the reference lists of pertinent publications were searched and reviewed. For the meta-analysis, we used data from 47 independent cohort, record linkage, and case–control studies (from a total of 74 reports), including > 11,000 incident cases or deaths from lung cancer among painters. Data extraction Three authors independently abstracted data and assessed study quality. Data synthesis The summary relative risk (meta-RR, random effects) for lung cancer in painters was 1.35 [95% confidence interval (CI), 1.29–1.41; 47 studies] and 1.35 (95% CI, 1.21–1.51; 27 studies) after controlling for smoking. The relative risk was higher in never-smokers (meta-RR = 2.00; 95% CI, 1.09–3.67; 3 studies) and persisted when restricted to studies that adjusted for other occupational exposures (meta-RR = 1.57; 95% CI, 1.21–2.04; 5 studies). The results remained robust when stratified by study design, sex, and study location and are therefore unlikely due to chance or bias. Furthermore, exposure–response analyses suggested that the risk increased with duration of employment. Conclusion These results support the conclusion that occupational exposures in painters are causally associated with the risk of lung cancer.


Review
Lung cancer is the most common cancer diagnosis worldwide and is the major cause of cancer mortality, particularly among men. The International Agency for Research on Cancer (IARC) estimated that there were > 900,000 new cases of lung cancer each year among men and > 330,000 among women (IARC 2001(IARC , 2003. Approximately 90% of the lung cancer burden in developed coun tries is attributed to smoking, which acts either independently or synergistically with other occupational, lifestyle, or hereditary risk fac tors (Boffetta and Trichopoulos 2002;Peto et al. 1994). Several agents encountered in the occupational setting, such as asbestos, poly cyclic aromatic hydrocarbons, arsenic, beryl lium, cadmium, chromium(VI), and nickel compounds, are established carcinogens that target the lung (IARC 2008).
An increased incidence and mortality from lung cancer has been observed in painters, an occupation that employs several million people worldwide (IARC 1989). This has led IARC to classify occupational exposure as a painter as "carcinogenic to humans" (Group 1) (IARC 1989, in press;Straif et al. 2007). Painters are exposed to many known and suspected lung carcinogens through inhalation or dermal con tact (IARC 1989;Siemiatycki et al. 2004), such as talc containing asbestos fibers, chromium VI compounds, chlorinated solvents, and cadmium compounds (IARC 1987(IARC , 1995(IARC , 1999Straif et al. 2009), although the specific caus ative agents have not yet been identified.
Cohort and record linkage studies demon strating a relatively consistent increased inci dence and mortality from lung cancer among painters [Alexander et al. 1996;Boice et al. 1999;Dubrow and Wegman 1984;Dunn and Weir 1965;Enterline and McKiever 1963;Gubéran et al. 1989;Guralnick 1963;Hrubec et al. 1995;Logan 1982;Menck and Henderson 1976;Office of Population Censuses and Surveys (OPCS) 1958, 1971, 1978, 1995Petersen and Milham 1980;Pukkala 2009;van Loon et al. 1997;Whorton et al. 1983] have supported the IARC Group 1 classification, although potential confound ing by tobacco smoking could not be ruled out in several of these studies. (Here we refer to record linkage studies as a subset of cohort studies where two databases are linked, such as a cohort of painters derived from census data and national mortality data, with only minimum demographic information available for the cohort.) Case-control studies have also shown that occupational exposure as a painter is a risk factor for lung cancer (Bethwaite et al. 1990;Bouchardy et al. 2002;Breslow et al. 1954;De Stefani et al. 1996;Finkelstein 1995;Milne et al. 1983;Pohlabeln et al. 2000;Wynder and Graham 1951), albeit some what less consistently (Baccarelli et al. 2005;Morabia et al. 1992;Muscat et al. 1998;Vineis et al. 1988;WünschFilho et al. 1998), and the increased risk persisted after adjusting for the potential confounding by smoking (Brüske Hohlfeld et al. 2000;Coggon et al. 1986;Decouflé et al. 1977;Houten et al. 1977;Jahn et al. 1999;Kjuus et al. 1986;Lerchen et al. 1987;Richiardi et al. 2004;Ronco et al. 1988;Viadana et al. 1976;Williams et al. 1977).
To assess the risk of lung cancer associated with occupational exposure as a painter, we conducted a metaanalysis of cohort, record linkage, and case-control studies to quantita tively compare the results of the different study designs and the potential confounding effect of smoking (by restricting to neversmokers), as well as other analyses to support the causal associa tion. A thorough discussion of the indi vidual studies included in the metaanalysis is not presented here but was summarized in the IARC Monographs (IARC 1989, in press). All of the studies reviewed, including the new studies published since the IARC Monographs, are summarized in Supplemental Material, Tables 1-3, available online (doi:10.1289/ ehp.0901402.S1 via http://dx.doi.org/).

Materials and Methods
Selection criteria. All epidemiologic studies included in the previous IARC Monographs were considered (IARC 1989, in press). Further, we searched PubMed (National Center for Biotechnology Information 2009) for articles in any language describ ing lung cancer in painters referenced in or published since the previous IARC Monograph (IARC 1989) through 24 August 2009, using the following search terms [by text word (tw), MeSH heading (mh), or publication type (pt)]: "paint*[tw]" or "varnish*[tw]" or "lacquer*[tw]"; and "can cer" or "neoplasms[mh]"; and "casecontrol study[mesh]" or "cohort study[mesh]" or "metaanalysis[mh]" or "review[pt]" or "risk factors[mh]" or "neoplasms/epidemiology" or "neoplasms/etiology" or "neoplasms/CI" or "occupational diseases/etiology" or "occupa tional diseases/epidemiology" or "occupational diseases/CI" or "occupational diseases/MO" or "occupational exposure/adverse effects" or volume 118 | number 3 | March 2010 • Environmental Health Perspectives "death certificates[mh]" or "epidemiologic methods[mh]"; and "lung." We identified 121 publications after restricting results to studies in humans. From the PubMed search, 69 studies were excluded because they were not epidemiologic studies, did not include original data (they were review articles), did not assess occupation as a painter, or lung can cer was not the outcome. The reference lists of pertinent publications were also reviewed to capture relevant data sources that may not have been identified with the search criteria.
Th e definition of painter varied between studies and often included other occupations exposed to paints such as plasterers, glaziers, wallpaper hangers, artists, decorators, French polishers, and aerographers [see Supplemental Material, Table 4 (doi:10.1289/ehp.0901402. S1) for definitions]. It is likely that paper hangers and other aforementioned occupa tions work in the same job environment as painters or may also paint; therefore, we con sidered this category as painters (Carstensen et al. 1988).
To be included in this metaanalysis, studies had to report estimates of the rela tive risk (RR), odds ratio (OR), standardized incidence ratio (SIR), standardized mortality ratio (SMR), proportionate mortality ratio (PMR), or proportional registration ratio with corresponding 95% confidence intervals (CIs) for everversusnever occupation as a painter or have provided enough informa tion that allowed for their computation. For studies that did not report the everversus never painter category, we estimated the risk estimates and 95% CIs for these categories. For studies that reported only point estimates without corresponding CIs, pvalues, or stan dard errors, or did not report the distribution of data to allow for computation of relative risks and CIs (also for non overlapping popu lations), we made conservative assumptions to estimate RRs and 95% CIs from the data provided on a studybystudy basis. These conservative assumptions under estimated the relative risk (toward the null) and over estimated the width of the CI (i.e., by dou bling the variance to approximate a 95% CI adjusted for multiple factors).
For example, overlapping lung cancer cases among AfricanAmerican (black) men was identified by Morabia et al. (1992) and Muscat et al. (1998). We accounted for this population overlap by approximating the pro portion of black male participants (cases and controls) based on distributions presented in other publications detailing this population, applying this proportion to the distribution presented by Morabia et al. (1992) (for black and whites combined) to determine the num ber of overlapping subjects, and subtracting the overlapping subjects from the distribution presented in Muscat et al. (1998).
Studies were excluded if estimation was impossible. In Supplemental Material, Tables 1-3 (doi:10.1289/ehp.0901402. S1), we use brackets to indicate the RRs and 95% CIs we calculated. For studies with over lapping populations, we included only the publication with the most complete study population. Further comments on study qual ity and any exclusions made are presented in detail in Supplemental Material, Tables 1-3. In total, we included in the metaanalysis 17 cohort and record linkage studies, 29 casecontrol studies, and 12 proportionate mortal ity analyses.
Data abstraction. All articles were assessed independently by three reviewers (A.A., F.M., N.K.S.) who extracted data that included authors, publication date, country of ori gin, characteristics of the study population including sex, and any details on the defini tion of painters, incidence versus mortality, lung cancer histology, observed and expected cancer cases (for cohort and proportionate mortality studies), number of exposed cases and controls (for case-control studies), yes/no adjustment for smoking or other occupational carcinogens, relative risks with corresponding 95% CIs, and results on exposure-response [see Supplemental Material, Tables 1-3 (doi:10.1289/ehp.0901402.S1)]. If adjusted and unadjusted results were reported, the most valid point estimate (i.e., adjusted for smoking and other variables) was abstracted. Any discrepancies in data collection were resolved by two other reviewers (N.G., K.S.).
Summary statistics calculated for inclusion in the meta-analysis. For cohort and record linkage studies, relative risk estimates (SIR and SMR) were computed by dividing the observed number of cases by the expected number, based on an external reference population. The corresponding 95% CIs were estimated using the PAMCOMP program (Taeger et al. 2000). If only subgroup results (e.g., by sex or duration of exposure) were reported, fixed effects models were used to combine stratum specific data into one summary estimate [see Supplemental Material, Tables 1 and 2 (doi:10.1289/ehp.0901402.S1)].
To allow for inclusion in the meta analysis, we calculated 95% CIs if they were not presented in the original paper. If a 90% CI was presented and if the upper limit (UL) and lower limit (LL) were proportionally sym metric around the risk ratio (for RR and OR; i.e., if UL/RR = RR/LL), an estimate of the standard error (SE) was calculated by SE = (ln UL -ln LL/3.29), where 3.29 = 2 × 1.645 for 90% CIs. If only a pvalue for the null hypothesis was presented, then a testbased SE was estimated using SE = (ln RR)/Z p , where Z p is the value of the standard normal test statistic corresponding to the pvalue using a twotailed test. The UL and LL of the 95% CI were estimated by RR ± 1.96 (SE), where Z p = 1.96 if p = 0.05 using a twotailed test (Rothman et al. 2008). A 95% CI cor responding to an unadjusted RR was used in the metaanalysis if a paper did not present enough data to allow for estimation of the adjusted CI.
Statistical analysis. Because cancer inci dence data are often more accurate than mortality data, we used SIRs in the analyses instead of SMRs whenever both were pre sented. However, mortality data for lung can cer are a very reasonable proxy for incidence because of the high fatality of lung cancer and the good quality of data from death cer tificates (Schottenfeld and Fraumeni 2006). We performed a separate metaanalysis for proportionate mortality studies. The PMRs were, however, not included in the overall metaanalyses because of their often lower quality of exposure assessment and their additional potential for bias. Assuming that the different effect estimates (e.g., SMR, SIR, RR, OR) represent the relative risk, the data were combined for all of the cohort, record linkage, and case-control studies. Subanalyses were also performed by stratifying on study design.
Many of the cohort and record linkage studies used an external reference population to calculate the expected cases. The use of an external reference population may result in a healthy worker effect, so that incidence or mortality rates of cancer in the exposed cohort may spuriously appear lower than in the general population. When the external reference rates used to calculate the expected cases are usually assumed to be known with out error, an estimate of the exposure coef ficient in a regression could be obtained by a weighted linear regression of the natural log of the adjusted SMR on exposure (Sutton et al. 2000). The risk estimates from nested case-control studies were included with the analysis of cohort studies because, essentially, this design can represent a more efficient way to analyze cohort studies and does not suffer from the problems associated with control selection in a case-control study. Summary ORs (metaORs) were obtained separately from the metaanalysis of case-control stud ies. Subgroup analyses were performed strati fied by sex, study region, study design, types of adjustment, and duration of employment.
The I 2 statistic quantifies the extent of inconsistency among the studies (Higgins and Thompson 2002). I 2 values of 25-50% indicate moderate inconsistency, whereas val ues > 50% reflect large inconsistencies among studies. We present the I 2 values instead of the Cochran's Qstatistic because the Qstatistic informs about the presence or absence of het erogeneity but does not quantify the extent (HuedoMedina et al. 2006). We used both random and fixedeffect models, with weights equal to the inverse of the variance, to calcu late a summary risk estimate (DerSimonian and Laird 1986). Results from randomef fects models, which account for heterogeneity among studies, are presented.
We conducted sensitivity analyses by drop ping one study at a time and examining its influence on the summary effect estimates. Forest plots were used to graphically display the data (Lewis and Clarke 2001). Publication bias was visually assessed using Funnel plots (Deeks et al. 2005). We performed all statistical analy ses using STATA (version 10.0; StataCorp, College Station, TX, USA), employing the "metan" command for the metaanalyses (Bradburn 2004).

Results
We reviewed 74 reports published since 1951 assessing the relationship between occupa tion as a painter and the risk of lung can cer [see Supplemental Material, Tables 1-3 (doi:10.1289/ehp.0901402.S1)]. The estimates of the relative risk reported in 47 indepen dent studies ranged from 0.60 to 5.76, with 43 studies reporting an RR > 1.0 (Tables 1  and 2). The combined analysis of 18 cohort and record linkage studies (metaRR = 1.36; 95% CI, 1.29-1.44; I 2 = 76.4%, p = 0) and 29 case-control studies (metaOR, 1.35; 95% CI, 1.22-1.51; I 2 = 48.4%, p = 0.002), including > 11,000 incident cases and/or deaths from lung cancer among painters, demonstrated a significantly increased risk overall in persons who had ever reported occupation as a painter (metaRR = 1.35; 95% CI, 1.29-1.41; I 2 = 63.6%, p = 0) ( Figure 1). Although the results of 13 proportionate mortality studies were not included in the combined analysis, they also demonstrated a significantly increased risk of lung cancer in painters (metaPMR, 1.22; 95% CI, 1.17-1.28). The Forest plot ( Figure 1) shows that there was no obvious trend in risk (at least no obvious trend toward a reduction in risk) over time. An influence analysis showed that dropping individual stud ies did not significantly alter the results (data not shown).
There appeared to be no evidence of publication bias among cohort and record linkage studies (data not shown). However, visual inspection of the funnel plot for 30 independent case-control studies demon strated some evidence of publication bias: the plot was slightly skewed with a deficit of smaller non positive studies (represented by large SEs) (Figure 2). When restricting the analysis to the larger case-control studies that showed both positive and negative results, the metaOR remained significantly elevated (metaOR, 1.31; 95% CI, 1.18-1.45; I 2 = 51.6%, p = 0.003). There was little difference in the results of case-control studies stratified by hospitalbased controls (metaOR, 1.37; 95% CI, 1.09-1.74; I 2 = 59.3%, p = 0.002) or populationbased controls (metaOR, 1.34; 95% CI, 1.18-1.51; I 2 = 25.9%, p = 0.16), although the populationbased studies were less heterogeneous.

Discussion
Previous studies demonstrating an increased risk of lung cancer in painters have allowed IARC to classify occupation as a painter as carcinogenic to humans (Group 1) (IARC 1989, in press). This metaanalysis supports the IARC Group 1 classification by demon strating a 35% increased risk of lung can cer in painters after adjusting for smoking (metaRR = 1.35; 95% CI, 1.21-1.51; I 2 = 41.2%, p = 0.01). This association was stron ger for populationbased case-control stud ies (metaOR, 1.34; 95% CI, 1.18-1.51; I 2 = 25.9%, p = 0.16) or studies that adjusted for other potentially confounding occupa tional exposures (metaRR = 1.57; 95% CI, 1.21-2.04; I 2 = 0%, p = 0.68). Furthermore, exposure-response analyses suggested that the risk increased with duration of employ ment. Although paint composition or the painting environment could have differed by major geographic region, the results did not vary much when stratified by region (North America, Europe, Asia, and South America). This is the first metaanalysis that demonstrates a relative increase in incidence/ mortality from lung cancer in persons occupa tionally exposed as painters when restricted to neversmokers (and also non smokers), as well as demon strating a statistically significant, positive duration-response relationship.
It is important to note that the inter pretation of a metaSMR (or metaSIR) for the cohort and record linkage studies is dif ficult because different reference populations were used in each study for the calculation of expected cases or deaths (Rothman et al. 2008). Although the cohort studies of paint ers could assess possibly higher exposures from longer periods of followup, exposure assessment in many of the record linkage stud ies was often crude: Occupation as a painter was usually assessed at a single time point in a census and then linked to death registries. Although there can be relatively poor cor respondence between occupation recorded on death certificates and in census records (Dubrow and Wegman 1984;Enterline and McKiever 1963;Guralnick 1963;OPCS 1971OPCS , 1978 and there is a chance of false positive results due to multiple testing of occupations in record linkage studies, the SMRs were remarkably consistent between individual studies, generally ranging between  Ronco et al. 1988, Italy, 1976 126 men 384 men who died from causes other than from smoking-related or chronic lung diseases Lifetime occupational history from interview with next of kin Painter 5 1.33 (0.43-4.11) Age, year of death, smoking, other employment in suspect high-risk occupations continued next page 1.10 and 2.57. This also suggested that the significant results were not likely due to chance. Thus, the approach to combine the cohort and record linkage study SMRs for cal culating a metaSMR seemed to be justified.
In case-control studies, painters may only form a small proportion of the study popula tion, but the full occupational history and additional information on lifestyle factors allowed several studies to adjust for tobacco smoking and some for other occupational carcinogens. An increased lung cancer risk associated with painting was consistently demonstrated in the case-control studies, sug gesting that occupation as a painter is a risk factor for lung cancer. Populationbased casecontrol studies may be less subject to selection biases than hospitalbased case-control studies (Rothman et al. 2008) because there is gener ally no concern about the appropriate source population if indeed the general population is represented. However, if response rates are low in population controls, this could result in a lack of comparability with cases and therefore be prone to selection biases. A sub analysis comparing the metaOR of hospital based and populationbased case-control studies showed similar results.
Estimates of the PMR may be biased if the population under study does not share the same distribution of mortality as the standard population used to compute the proportions for categories other than the ones studied (Rothman et al. 2008). However, the propor tionate mortality analyses also showed signifi cantly elevated relative risks for lung cancer in painters within the same range of effect as the analyses overall and in cohort studies, further suggesting that these results remained robust to these biases.
Smokingadjusted estimates were avail able for 23 of 29 case-control studies and in only 4 of 18 cohort and record linkage studies. The robustness of the summary estimates after adjusting for tobacco use, and the higher rela tive risk in neversmokers, suggest that residual confounding by tobacco use is unlikely and that occupation as a painter is independently associated with the risk of lung cancer.
In women, the metaRR was similar for all studies (metaRR = 2.04; seven studies) (Jahn et al. 1999;Muscat et al. 1998;OPCS 1958OPCS , 1971Pronk et al. 2009;Pukkala 2009;Zeka et al. 2006) and for studies restricted to neversmokers (metaRR = 2.00; three stud ies) (Kreuzer et al. 2001;Pronk et al. 2009;Zeka et al. 2006), further strengthening the evidence that the results are not confounded by smoking. However, female painters (and neversmoking females) may not actually have a higher risk of lung cancer compared with male painters (metaRR = 1.37; 39 studies). The relative risk in women is higher, which  Morabia et al. 1992 and thus some estimations were used to eliminate the overlap in men and the estimated variance was doubled to approximate an adjusted CI. e Calculated using a fixed-effects model. f Variance was doubled to approximate an adjusted 95% CI. g Included in the analysis restricted to case-control studies but excluded from the combined meta-analysis because of possible overlap with OPCS 1986. h The CI was estimated by applying the ratio of reduced/total controls to the observed cell counts reported for the total control group.
volume 118 | number 3 | March 2010 • Environmental Health Perspectives may be due to the fact that women have a lower background lung cancer risk than men (Schottenfeld and Fraumeni 2006). The robustness of the results is also indi cated by the presence of a duration-response relationship, with higher RRs seen for expo sure over ≥ 10 years (metaRR = 1.95) and ≥ 20 years (metaRR = 2.00) compared with those with < 10 and < 20 years of exposure, respectively (the reference category was no exposure).
Some painters (e.g., in the construction industry) could have been exposed to asbes tos. Indeed, a number of studies have shown an increased risk of mesothelioma in painters (Brown et al. 2002;Peto et al. 1995), which is most likely due to occupational asbestos exposure. However, taking into account that the exposure-response relationship for pleu ral mesothelioma is very different from that for lung cancer, potential asbestos exposure cannot explain all of the increase in lung can cer. Therefore, other suspected carcinogens to which painters are exposed, such as chlori nated solvents, chromium VI compounds, and cadmium compounds (IARC 1987(IARC , 1995(IARC , 1999Straif et al. 2009), may also partially explain the increased risk of lung can cer. Very few studies reported results for spe cific suspected causa tive agents. van Loon et al. (1997) reported a positive exposure-response relationship with paint dust and Siemiatycki et al. (1987) found a suggestive association with mineral spirits, whereas Alexander et al. (1996) did not find an increased risk of lung cancer in a cohort of painters and other employees in the aerospace industry exposed to chromium VI compounds.

Conclusion
There is great variability and complexity in painting environments, which complicates the interpretation of epidemiologic studies of lung cancer risks in painters. Painters are exposed to a wide variety of chemical mix tures, with compositions that change over time. In more recent decades, a number of hazardous chemicals-including benzene, some other solvents, phthalates (plasticiz ers), and lead oxides-have been reduced or replaced in paint, although these chemicals are still used in some countries. This trend in reducing exposures to hazardous chemicals in paint has been promoted by the increas ing use of waterbased paints and powder coatings. New formulations may also contain lowertoxicity solvents, neutralizing agents (e.g., amines), and biocides (IARC 1989, in press). However, this has not yet resulted in lower relative risks for lung cancer in painters, as reported in the more recent observational epidemiologic studies. The elevated risk of lung cancer may also be partly due to the role that other substances may play in increasing the risk of lung cancer among painters.
Although there was not enough informa tion in the studies provided to assess the asso ciation of lung cancer with specific chemical agents encountered in painting, the robustness of the estimates in the subgroup analyses (by sex, region, study design, and controlling for smoking and other occupational exposures) and the stronger associations seen in specific subgroups (by duration of exposure) support the conclusion that occupational exposures in painters are causally associated with the risk of lung cancer. Because several million people are employed as painters worldwide and because lung cancer is the most common cancer in painters, even a modest increase in Figure 1. Meta-analysis of all studies assessing lung cancer among persons with occupation as a painter, stratified by study design. Weights are from random-effects analysis. The relative risk estimate for each study is represented by a black diamond, and the horizontal line shows the corresponding 95% CI. The dashed line marks the combined estimate, and the vertical solid line represents no association.  Figure 2. Begg's funnel plot with pseudo-95% CIs to assess publication bias in case-control studies of lung cancer among persons reoporting occupation as a painter.

SE of logRR
1 the relative risk is remarkable. It is important for cancer control and prevention to design studies with better exposure assessment to identify the underlying carcinogenic agents encountered in painting.