Occupational exposure to silica dust and risk of lung cancer: an updated meta-analysis of epidemiological studies

We searched MEDLINE and EMBASE databases from January 1982 through 29 April 2016 using the search terms “lung cancer”, “silica”, “silicosis”, “risk”, “incidence” and “mortality” with variation in term construct to identify epidemiological studies published in the literature which evaluated the relationship between silica exposure and lung cancer in workers, irrespective of their silicotic status (Additional file 1). Reference lists of the identified articles were also screened for potentially eligible studies.

The following inclusion criteria were used for the analysis:

Reviews, autopsy studies, comments, editorials, studies with insufficient quantitative data required for the analysis (no risk estimate, no confidence intervals) and those overlapping with studies which were already considered, were excluded.

When a particular study was reported in several papers, the most recently-published reference was used unless the required data was reported in a previous paper and not in the latest-published one.

For each study, the following data was extracted: geographical location, year of publication, industrial setting, study design, total number of subjects, exposure assessment (including level and duration of exposure to silica dust), outcome examined, study period (including the start date, end date and duration of follow-up), person-years of follow-up, covariates adjusted for, potential occupational carcinogens including radon, arsenic, asbestos, diesel, polycyclic aromatic hydrocarbons (PAH), talc, cadmium and amphiboles, number of lung cancer cases, total number of deaths and number of deaths due to lung cancer, measure of association and effect estimates with corresponding 95 % confidence interval (CI). Two authors worked independently for study selection and data extraction. Any disagreement was resolved after a team discussion. The list of included studies was made in consensus.

We used the Newcastle-Ottawa Assessment Scale (NOS) for assessing the methodological quality of observational studies [15]. The scale consists of three main categories including selection of study population, comparability of subjects and ascertainment of exposure for case-control studies or ascertainment of outcome for cohort and proportional mortality studies. Scores of 0–3, 4–6 and 7–9 were assigned to low, moderate and high quality studies respectively.

Studies were pooled together according to the outcome examined which could be incidence or mortality, the study design which could be cohort or case-control or proportional mortality study design and the measure of association. When a study reported results stratified by race, gender, industrial setting and silicotic status, they were treated as two separate reports for analysis. Studies which gave risk estimates according to silicotic status were grouped and analyzed separately. Statistical analysis was performed on the natural logarithm (ln) of the risk estimate so as to approximate its sampling distribution to a normal one. The difference between the upper and lower limits of the confidence intervals was transformed to the log scale and the standard error was calculated by dividing the transformed interval by 3.92 [16]. Random effects model was used to calculate the pooled effect estimates since a high level of heterogeneity was expected. Heterogeneity between studies was quantified by two methods namely the chi-squared test (Q test) for heterogeneity, reported by its p value, and the variability due to heterogeneity (I² statistic), reported as a percentage in this paper [17]. We performed subgroup and meta-regression analyses to try to explain any observed between-study heterogeneity. In subgroup analysis, the studies were categorized into subgroups based on the predefined covariates. In meta-regression, we investigate the relationship between the covariates and the observed outcome [18]. The proportion of variance explained (R²) was used to quantify the amount of heterogeneity accounted for by each covariate. It was calculated as the percentage of ratio of variance explained to the total amount of variance. Both univariate and multivariate meta-regression models were used to try to lower the variability due to heterogeneity (I²) to the minimum level and to bring the p value of Q close to 1. Heterogeneity should be completely absent (I² is 0 and p value of Q is 1) for an ideal comparison [19]. Differences in exposure assessment, study design and quality, data collection processes, outcome assessment, selection of subjects and definition of confounding factors often account for significant between-study heterogeneity [16]. Based on this statement, the covariates considered were year of publication, presence of at least one confounding factor, adjustment for smoking, industrial setting, geographical location, NOS score, cumulative silica dust exposure level, duration of exposure, concentration of silica dust, person-years of follow-up, number of subjects and total number of deaths. For sensitivity analyses, we assessed the influence of individual studies on the pooled estimate by omitting each study in turn (leave-one-out analysis). Publication bias was assessed graphically by means of funnel plots and quantitatively by Egger’s linear regression method [20]. For the exposure-response analyses, we used the average cumulative silica dust exposure as covariate and the risk estimate of the corresponding study as the effect. No imputation was made in relating the effect estimate to the exposure level. Statistical analysis was done using R software version 3.1.2 (2014-10-31) [21] with the ‘metafor’ package version 1.9-5 [22].

The levels of significance for all statistical tests were assumed to be equal to or less than 0.05, except in the case of heterogeneity testing whereby the level of significance was assumed to be equal to or less than 0.10 [17].

The PRISMA flowchart for the selection of studies is shown in Fig. 1. The initial search criteria yielded 227 citations from the databases. 58 additional records were further identified from references of related articles. After removing duplicates, we were left with 273 records. Preliminary screening of abstracts eliminated 158 studies. Of the remaining 115 articles, 30 were excluded for the following reasons: 6 articles were found to be either reviews, editorials, comments or autopsy studies, 11 papers had no risk estimate data, 2 articles did not give the confidence intervals of the effect estimate, 2 articles did not report lung cancer as outcome and 9 articles had overlapping populations with selected studies. After exclusion of these 30 studies, 85 articles were left and included in the final main quantitative synthesis [9, 10, 23‐104]. The study of Puntoni et al. [105], which was excluded from the main synthesis due to overlap with the study cohort of Merlo et al. [66], was included in the list of silicotic studies since it contained the risk estimate based on silicotic status whereas the study by Merlo et al. had the risk estimate of the whole cohort and not according to silicotic status.

After categorization of the studies by outcome assessed, study type and measure of association there were 63 cohort studies reporting mortality due to lung cancer in the form of SMR as measure of association, 19 cohort studies reporting incidence of lung cancer in the form of SIR, 1 cohort study reporting incidence in the form of RR, 3 case-control studies reporting mortality in the form of MOR, 9 case-control studies reporting incidence in the form of OR, 5 case-control studies reporting mortality in the form of OR and 2 proportional mortality studies or PMR. The characteristics of all included studies are shown in Additional file 2.

Most studies comprised males only and a mere few included both males and a small proportion of females (around 10 %). Only exceptions were the studies of Zhang et al. [104] and Smailyte et al. [84] with 26 and 31 % women respectively. Nine papers reported 2 or more results stratified by industrial setting, sex, silicotic status and racial background. Forty-one studies were conducted in European countries, 18 in the United States, 21 in Asian countries, 9 in Canada, 3 in Australia and 1 in South Africa. The industries of concern were mining, foundry, pottery and ceramic, refractory brick and diatomaceous earth processing, granite which included sand and quarry, cement production and construction. The total number of studies available for analysis according to silicotic status was 34. The characteristics of silicotic and non-silicotic studies are described in Table 1.

The results of the study quality assessment are presented in Additional file 3. Ten articles were deemed to be of low quality, 49 articles were found to be of medium quality and 26 articles were shown to be of high quality. The median score for all 85 articles was 5.3 out of a maximum of 9.

As demonstrated graphically by the funnel plots in Additional files 4, 5 and 6, there was evidence of publication bias for studies reporting mortality in the form of SMR (p = 0.024 for Egger’s regression test) but no evidence of publication bias for studies reporting incidence in the form of SIR (p = 0.238) and OR (p = 0.457).

Using the random effects model, the pooled estimate was 1.55 (95 % CI 1.38–1.75) for SMR studies, 1.68 (95 % CI 1.45–1.96) for SIR studies, 1.10 (95 % CI 0.89–1.36) for PMR studies, 1.69 (95 % CI 1.26–2.26) for MOR studies, and 1.34 (95 % CI 1.24–1.46) for case-control studies reporting incidence as outcome and 1.82 (95 % CI 1.25–2.66) for case-control studies reporting mortality as outcome. The risk estimate in each category was statistically significant (p < 0.05) except in the category of PMR studies (p = 0.38). The results of the SMR, SIR and OR studies with incidence as outcome are illustrated in forest plots in Figs. 2, 3 and 4 respectively. Significant between-study heterogeneity was observed in SMR, MOR and SIR studies with I² of 96 %, 87 % and 75 % respectively. PMR and OR studies with mortality as outcome showed lower between-study heterogeneity (I² 62 % and 51 % respectively), which was statistically insignificant (p value for Q test >0.10 for both). No heterogeneity was observed in the meta-analysis of OR studies with incidence as outcome. Studies conducted in silicotic subjects yielded a significantly higher pooled SMR of 2.32 (95 % CI 1.91–2.81) and SIR of 2.49 (95 % CI 1.87–3.33) as compared to non-silicotic studies which gave a resulting estimate of 1.78 (95 % CI 1.07–2.96) for SMR studies and 1.18 (95 % CI 0.86–1.62) for SIR studies. Between-study heterogeneity was statistically significant in silicotic and non-silicotic studies with SMR as risk measure (I² = 94 % with p < 0.0001 and I² = 74 % with p = 0.013 respectively) but they were found to be lower and statistically insignificant in silicotic and non-silicotic studies with SIR as measure of association (I² = 25 % with p = 0.377 and I² = 41 % with p = 0.192) respectively.

Subgroup analyses were carried out individually for SMR and SIR studies. We could not perform similar subgroup analysis for the few remaining studies reporting other measures of association due to their limited number. Since the level of between-study heterogeneity was found to be 0 % in the group of OR studies reporting incidence as outcome, we did not perform any further analysis to explore heterogeneity. Subgroup analysis for SMR studies showed a positive association between silica dust exposure and lung cancer in all subgroups except in the subgroup of cement industries which had a pooled risk estimate of 0.87 (95 % CI 0.42–1.82). Heterogeneity became non-significant (p > 0.10) in the subgroups of potteries (I² = 0 % and p = 0.273), construction industries (I² = 0 % and p-0.656) and in the subgroup including studies done in Australia (I² = 20 % and p = 0.265). In all other subgroups, between-study heterogeneity remained significant. Out of 63 SMR studies, only 2 adjusted for smoking [28, 102] and the effect measure in this subgroup was 1.83 (95 % CI 1.51–2.22). In the subgroup of studies without adjustment for smoking, the pooled estimate was 1.55 (95 % CI 1.37–1.75). Thirteen SMR studies having none of the other potential occupational carcinogens mentioned in the Methods Section yielded an estimate of 1.32 (95 % CI 1.14–1.54). The positive association between silica and lung cancer became weaker with increasing quality of the included studies, from 2.56 (95 % CI 1.57–4.19) among SMR studies with an NOS score of 1–3 to 1.24 (95 % CI 1.01–1.52) in those with an NOS score of 7–9. A similar trend was observed among SIR studies.

Subgroup analysis for SIR studies showed a positive relation between occupational silica exposure and risk of lung cancer in all subgroups, with statistically significant risk estimates in all subgroups except that including studies conducted in Canada (p = 0.494). Much of the between-study heterogeneity could be explained by sub-grouping the SIR studies and in most of the subgroups, it became statistically insignificant (p > 0.1). None of the SIR studies adjusted for smoking. Detailed main statistical analysis, subgroup analyses for SMR and SIR studies and statistical analysis of silicotic and non-silicotic studies are shown in Table 2.

Separate meta-regression analyses were performed for SMR studies, SIR studies and silicotic studies reporting SMR as risk measure. These were done using both univariate and multivariate models. In the univariate meta-regression analysis of SMR studies, NOS score was the most important covariate accounting for 19 % of heterogeneity while in the multivariate analysis, the combination of industrial setting, year of publication, geographical location and number of subjects accounted for the maximum amount of heterogeneity (R² = 37 %). As for the SIR studies, univariate analysis showed that total number of deaths accounted for the highest amount of heterogeneity (R² = 100 %) and a combination of number of subjects and NOS score corrected the maximum amount of heterogeneity (R² = 15 %) in multivariate analysis. Regarding the silicotic studies, the combination of year of publication and total number of deaths corrected 43 % of between-study heterogeneity. Detailed results of meta-regression analyses are shown in Table 3.

For the exposure-response analysis of the relationship between silica dust and risk of lung cancer, reference was made to the subgroup analysis of SMR studies by average level of cumulative silica dust exposure. Nineteen studies were included. It was found that the risk of lung cancer increased with rising exposure level (risk estimate rose from 1.19 (95 % CI 1.02–1.39) in the first quartile to 1.36 (95 % CI 0.87–2.13) in the fourth quartile). However, the p value of the estimate was statistically insignificant for the second, third and fourth quartiles (p > 0.05). A high level of between-study heterogeneity was also noted especially with rising quartiles of cumulative silica dust (p < 0.0001).

Sensitivity analyses showed that omission of any study did not significantly influence the pooled estimates.