Meta-analysis of silicosis and lung cancer

Objectives This study examined the association between silicosis and lung cancer in a systematic review (and meta-analysis) of the epidemiologic literature, with special reference to the methodological quality of observational studies. Methods We searched Medline, Toxline, BIOSIS and Embase (1966–May 2004) for original articles published in any language and systematically reviewed the bibliographies of the retrieved articles. Observational studies (cohort and case–control studies) were selected if they reported a measure of association [standardized mortality ratio (SMR), relative risk or odds ratio] relating lung cancer to silicosis. Results Thirty-one studies (27 cohort studies, 4 case–control studies) met the inclusion criteria of the meta-analysis. Without any adjustment for smoking, the meta-analysis of the cohort studies indicated that the common SMR was 2.45 [95% confidence interval (95% CI) 1.63–3.66; homogeneity P<0.0001]. When the results of the cohorts for which mortality data were adjusted for smoking were pooled, the common SMR was 1.60 (95% CI 1.33–1.93; homogeneity P=0.52). In a “dose–response” analysis, the profusion of small and large opacities found in chest X-rays correlated with the risk of death from lung cancer. Overall, the case–control studies were more conservative in their conclusions. and from a “dose–response” analysis, that silicosis and lung cancer are associated.

Silicosis is a parenchymal lung disease caused by the inhalation of crystalline silicon dioxide, or silica (1). Such exposure occurs in a wide variety of occupations, among which foundry workers, miners, quarriers, and sandblasters are the most at risk (2). In jurisdictions in which occupational exposure standards have been specified, the incidence of silicosis has dramatically diminished (3). Nevertheless, currently recommended exposure limits for occupational exposure to silica are still associated with a significant risk of silicosis (4). In the province of Quebec (Canada), silicosis remains the third most common cause of compensation for work-related respiratory disorders, following occupational asthma and asbestos-related diseases.
The association between silicosis and lung cancer has long been suggested by clinical observations and case series (5)(6)(7). However, more recent epidemiologic studies have often provided conflicting results. The interpretation of this epidemiologic evidence has been hampered by shortcomings that include the noncomparability of reference groups, detection bias, and the confounding effect of other carcinogenic risk factors, such as cigarette smoking and exposure to other known occupational carcinogens (8).
We sought to reexamine the epidemiologic evidence regarding the association between silicosis and lung cancer through a systematic review (and meta-analysis) of the epidemiologic literature, with special reference to the methodological quality of the observational studies. The methods that we used are in accordance with the Meta-analysis of Observational Studies in Epidemiology (MOOSE) Group's recommendations (9).

Literature search
We searched Toxline, BIOSIS, Embase and Medline

Study selection
Observational studies (cohort and case-control studies) were selected if they reported a measure of association [standardized mortality ratio (SMR), relative risk or odds ratio (OR)] relating lung cancer to silicosis. Proportional mortality studies were excluded because of their potential for a systematic overestimation of risk (11). To limit selection bias, we also excluded autopsy studies (12). In addition, narrative reviews, letters to the editor, clinical commentaries, case series, and case reports were disregarded.
Two reviewers (YL and SM) successively applied these criteria to the titles and abstracts of all the citations obtained. If the title of an article or, when available, its abstract suggested any possibility that it might be relevant, the paper was retrieved and independently assessed by the same reviewers for a final decision about its inclusion into the meta-analysis. Throughout this process, the reviewers were blinded to the authors' names, the journal name, and the year of publication of the papers. Those published in languages other than English and French were translated into French. Any disagreement was resolved by consensus or by consulting a third reviewer (MD). When we identified studies that had been reported in multiple papers, we limited our analysis to the most recent report, unless the necessary data had appeared only in an earlier paper. Agreement between coders was measured using quadratic weighted kappa statistics (13). We kept a log of the reasons for rejecting citations identified from the searches. Publication bias was investigated by visual inspection of a plot of the magnitude of risk in a study versus the number of silicotics in the study (14). We determined a priori that the effect of publication bias should be slight if this plot showed a rough, symmetric funnel shape.

Study evaluation for methodological quality
We evaluated study validity by systematically considering the following three important sources of bias in observational studies: (i) selection bias, which stems from the absence of comparability between the groups being studied; (ii) information bias, which results from an incorrect (or differential) determination of exposure or outcome; and (iii) confounding bias, which is likely when the results can be accounted for by the presence of a factor associated with both the exposure and the outcome but not directly involved in the causal pathway (15).

Information extraction
Two reviewers (YL and SM) abstracted information from all of the selected papers for inclusion in the metaanalysis. The abstracted information included (i) the study design, (ii) the industry and country where the occupational exposure to silica that led to silicosis had occurred, (iii) the period during which new cases of silicosis were included in the cohort and the length of follow-up, (iv) the record source, (v) whether confounders were accounted for, (vi) whether the comparison group was from the general population or consisted of any other group of workers, and (vii) the actual data reported in the paper. In case of missing data, we did not attempt to contact any of the authors for additional information.

Analysis of cohort studies
For each study, we calculated the SMR from the observed and expected number of deaths from lung cancer among the silicosis patients and the unexposed (SMR=observed/expected) and computed the corresponding 95% confidence interval (16). Since the lung cancer incidence rate approaches its mortality rate (17), standardized incidence ratios (SIR) were analyzed similarly and pooled with the SMR values. In addition, we noted that several studies had reported the SMR value after adjustment for smoking according to the method described by Axelson (18). For these studies, we computed the expected number of deaths from lung cancer from the number of deaths actually observed and the adjusted SMR reported in the study. In addition, we noted that several authors controlled for smoking by restricting the analysis to silicosis patients who had never smoked. We therefore conducted three separate metaanalyses, one with the unadjusted SMR, another with the adjusted SMR according to Axelson's method, and a third one of the SMR restricted to never-smokers. The SMR values were weighted by the inverse of their variance and combined according to a random-effects model (19). Homogeneity was tested by the method described by Fleiss (19). Subgroup analyses were indicated when significant heterogeneity was found among the primary study results. Statistical significance was set at P<0.05.
Subgroup analyses and a priori hypotheses explaining heterogeneity among studies. In meta-analyses that did not meet the criteria for homogeneity, we conducted subgroup analyses in an effort to identify the source of heterogeneity according to the following hypotheses: (i) studies of silicosis patients who acquired their lung disease from underground mining (with potential exposure to cocarcinogens such as radon daughters) result in higher risks of lung cancer, (ii) the longer the follow-up period, the higher the risk of lung cancer, and (iii) studies of compensation registries result in higher risks of lung cancer.
Dose-response analysis. In conducting this review, we realized that several authors reported SMR values according to the International Labour Organization's (ILO) radiographic classification of pneumoconioses (20). In this classification, cases of silicosis are categorized according to the presence and profusion of small (≤1 cm) and large (>1 cm) opacities. Categories 1, 2, and 3 represent an increasing profusion of small opacities, as defined by standard radiographs. Categories A, B, and C are defined in terms of the dimension of large opacities. The comparison of the risks of lung cancer among silicosis patients was appealing since it allowed an "exposure-response analysis" in which workers with high exposure were compared with workers with low exposure, both groups presumably sharing similar smoking habits and clinical characteristics otherwise (21). Because there was no clear evidence from the literature that smoking predisposes to the development or the progression of silicosis (22), this analysis does not seem to be confounded by smoking. This analysis was conducted according to the method for the meta-analysis of epidemiologic dose-response data described by Berlin et al (23). In this method, a weighted least-square estimate of the regression slope (β) of the logarithm of the risk versus exposure (ILO category) is computed for each study (log [risk]=β · exposure). The inverse of the variance of log [risk] is used as the regression weight. The regression slopes are then combined according to a random-effects model (19). Finally, we converted back the common logarithm into natural units expressed in terms of relative risk. Homogeneity was also tested by the method described by Fleiss (19).

Analysis of case-control studies
The case-control studies that met the inclusion criteria of the meta-analysis used different statistical methods to report the risk of lung cancer among silicosis patients. We restricted our statistical analysis to unadjusted data. The unadjusted odds ratios were weighted by the inverse of their variance and combined according to a randomeffect model (19). Homogeneity was tested by the method described by Fleiss (19). Statistical significance was set at P<0.05. We reported the results of the stratified or logistic regression analyses in a narrative fashion.

Validity assessment
Cohort studies. With one exception (37), the cohort studies included in the meta-analysis presented SMR (or SIR) values computed from national mortality rates for lung cancer. In two studies, separate analyses were conducted using comparison groups that were not from the general population (8,38). Since the general population rates were used for calculating expected deaths and the smoking prevalence rate among blue collars is higher than in the general population (52,53), the unadjusted SMR values should be interpreted cautiously. Accordingly, in unadjusted analyses, lung cancer SMR values tend to be overestimated for smoking and underestimated when the analysis is limited to nonsmokers. Fourteen cohorts were from compensation registries (12, 24-26, 29-31, 33, 34, 36, 39, 45-47). Compensation for silicosis largely depends on disability (selection bias) (38,40,54,55) and, therefore, on smoking (confounding bias). Another difficulty in interpreting several of the available epidemiologic investigations was the combined exposure with other lung carcinogens, such as radon daughters in underground mining and polycyclic aromatic hydrocarbons in foundries (confounding bias). Most of the studies reported on silicosis patients whose diagnosis was made in the 1960s; this approach increased the likelihood of misdiagnoses among patients with pre-existing diffuse lung disease (information bias). In all of the studies, misclassification may have resulted in the inclusion of other pneumoconioses in silicosis patients (information bias). In addition, periodic surveillance in cohorts of compensated silicosis patients may have increased the likelihood of a lung cancer diagnosis for these patients (information bias). Overall, these potential sources of bias favor the finding of a positive association between silicosis and lung cancer.
Case-control studies. One case-control study (49) was hospital-based and at risk of Berkson's (or admission) bias, which comes into play whenever there are differences in admission rates between exposed (ie, silicosis patients) and unexposed persons (selection bias) (56). Otherwise, the selection of controls and data collection Lacasse et al were appropriate, without clear indication of a differential determination of exposure or outcome. In all of the studies, smoking was accounted for, either in stratified or logistic regression analyses.

Analysis of cohort studies
A total of 23 305 silicosis patients contributed to the analysis of the cohort (table 2). We did not find any clear indication of publication bias from the visual inspection of the funnel plot (figure 1). Without any adjustment for smoking or any other co-carcinogen, the pooled SMR value was 2.45 (95% CI 1.63-3.66). However, significant heterogeneity was found among the study results (P<0.0001). We could not explain this heterogeneity by excluding the cohorts of underground miners from the analysis (table 3). It is also noteworthy that excluding studies of compensation registries did not result in a lower risk of lung cancer (table 3). In addition, we could not find any correlation between the length of follow-up and the overall risk of lung cancer.
When the results of the cohorts were pooled in which mortality data were adjusted for smoking according to Axelson's method, the pooled SMR value was 1.60. The results of the studies that contributed to this analysis were homogeneous. When the results of the cohorts restricted to never-smokers were pooled, the SMR value was 1.52. The results of the studies that contributed to this analysis were also homogeneous. In the latter analysis, because the unexposed populations included smokers, this result probably represents an underestimate of the risk of lung cancer among silicosis patients.
In the three studies that conducted analyses using comparison groups that were not from the general population, the risk of lung cancer among silicosis patients was increased. In Amandus & Costello's study (8), the age-adjusted lung cancer risk for silicosis patients was 1.56 (95% CI 0.91-2.68) times higher than that of nonsilicotic metal miners. In the study by Dong et al (37), the lung cancer risk of silicosis patients was 2.10 times higher (P<0.01) than that of workers from rough rolling steel mills (37). Finally, in Finkelstein's study (38), after adjustments were made for mining sectors and cumulative radon exposure, the miners with radiographic abnormalities suggestive of silicosis were found to be at substantially greater risk of lung cancer than those with normal radiographic findings (OR 6.88, 95% CI 1.89-25.00) (38).
"Dose-response" analysis. Five studies reported the risk of lung cancer according to the radiological severity of silicosis (29,31,37,45,47). One was excluded from the analysis because of a lack of data (42) 2).

Analysis of case-control studies
The authors of the four case-control studies that met the inclusion criteria of the meta-analysis used different statistical methods to report the risk of lung cancer among silicosis patients. The raw data are given in  (50) found no statistically significant association between silicosis and lung cancer after matching cases and controls by age and tobacco consumption. Forastière et al (51) reported an increased risk of lung cancer for silicosis patients (Mantel-Haenszel rate ratio 3.9, 95% CI 1.8-8.3) after control for age and smoking. After adjusting for smoking, Mastrangelo et al (49) found a twofold increase in lung cancer risk for workers compensated for silicosis (Mantel-Haenszel rate ratio 1.8, 95% CI 1.1-2.8). Fu et al (48) found that the presence of silicosis did not contribute to the prediction of risk for lung cancer independently of the years spent underground.

Discussion
Because of biases inherent to observational studies, it is likely that the SMR values reported in the cohort studies that met the inclusion criteria of this meta-analysis represent overestimates of the real risk of lung cancer among silicosis patients. The results of four case-control studies were more conservative. There is nevertheless evidence, from data restricted to never-smokers, from cohort studies using appropriate comparison groups, and from a "dose-response" analysis, that silicosis and lung cancer are truly associated.  Figure 1. Study of publication bias (funnel plot). The two outliers are the studies by Chiyotani et al (28) and Chen et al (27). See the text for discussion.

Lacasse et al
conducted by independent reviewers (61). Sources of discordance include differences in study selection, data extraction, assessment of the ability to combine studies,  (24)

Relation with previous studies
Our systematic review differs from that published in 1997 by Tsujda et al (57). In their meta-analysis, all types of pneumoconiosis (with the exception of asbestosis) were included. However, our results are in agreement with previous meta-analyses published by Smith et al (58)  An important difference between the two studies is that these authors could not find any trend between the radiographic category of silicosis and the risk of lung cancer. Such differences are not unusual in meta-analyses and statistical methods for data synthesis (61). In the "dose-response" analysis, we used a set of studies that differs from the one in the Japanese review. Neither the radiographic classification of silicosis nor the statistical methods were the same in both studies. Nevertheless, Kurihara & Wada (60) concluded that silicosis increases lung cancer risk.

Publication bias
We limited our analysis of publication bias to a visual inspection of the funnel plot because the statistical methods used to judge its symmetry have been questioned (14). We found no clear indication of publication bias from the visual inspection of the funnel plot (figure 1). Publication bias would have modified the shape of the funnel. If small statistically nonsignificant studies had been omitted, a "bite" would have been taken out of the display for effects near zero (ie, SMR 1.00). The two outliers are the study by Chiyotani et al (28) on hospitalized patients and that by Chen et al (27) on underground miners. The authors of the latter study commented that their results may have been confounded by exposure to radon daughters. The two studies accounted for less than 7% of the total weight of the studies that we included in our analysis. Both contributed only to the unadjusted analyses (figure 1 and table 3) and were not included in the dose-response analysis. The exclusion of both studies from the meta-analysis did not have a significant impact on its results (data not shown).

Clinical and medicolegal implications
Our finding of a "dose-response gradient" among silicosis patients deserves further comments. Whether this result reflects a real dose-response relationship between silicosis and lung cancer remains uncertain. On one hand, lung fibrosis (as seen in idiopathic pulmonary fibrosis and asbestosis) increases the risk of lung cancer (62,63). On the other hand, whether the association between silicosis and lung cancer is due to the effect of the fibrotic process or to the effect of quartz dust itself is unclear (64). Some data suggest that the progression of silicosis may be determined by total accumulated silica dust exposure (65). If silica was a true carcinogen, our finding would rather be interpreted as an indication of the dose-response relationship between silica exposure and lung cancer (66). Such a distinction is, however, irrelevant for compensatory boards in their assessment of patients with silicosis and lung cancer. It is however relevant for the compensation of patients with occupational exposure to silica (without silicosis) and lung cancer.

Concluding remarks
We conclude that silicosis increases the risk of lung cancer. This association does not necessarily imply that silica is a lung carcinogen. A better and modern understanding of the relationship between silica exposure and lung cancer requires additional investigations and doseresponse analyses, including exposures in the range of currently recommended exposure limits.