Global and regional burden of cancer in 2016 arising from occupational exposure to selected carcinogens: a systematic analysis for the Global Burden of Disease Study 2016

Objectives This study provides a detailed analysis of the global and regional burden of cancer due to occupational carcinogens from the Global Burden of Disease 2016 study. Methods The burden of cancer due to 14 International Agency for Research on Cancer Group 1 occupational carcinogens was estimated using the population attributable fraction, based on past population exposure prevalence and relative risks from the literature. The results were used to calculate attributable deaths and disability-adjusted life years (DALYs). Results There were an estimated 349 000 (95% Uncertainty Interval 269 000 to 427 000) deaths and 7.2 (5.8 to 8.6) million DALYs in 2016 due to exposure to the included occupational carcinogens—3.9% (3.2% to 4.6%) of all cancer deaths and 3.4% (2.7% to 4.0%) of all cancer DALYs; 79% of deaths were of males and 88% were of people aged 55 –79 years. Lung cancer accounted for 86% of the deaths, mesothelioma for 7.9% and laryngeal cancer for 2.1%. Asbestos was responsible for the largest number of deaths due to occupational carcinogens (63%); other important risk factors were secondhand smoke (14%), silica (14%) and diesel engine exhaust (5%). The highest mortality rates were in high-income regions, largely due to asbestos-related cancers, whereas in other regions cancer deaths from secondhand smoke, silica and diesel engine exhaust were more prominent. From 1990 to 2016, there was a decrease in the rate for deaths (−10%) and DALYs (−15%) due to exposure to occupational carcinogens. Conclusions Work-related carcinogens are responsible for considerable disease burden worldwide. The results provide guidance for prevention and control initiatives.


Estimating the proportion of the population exposed
The data available for estimating the exposed population were, for each GBD region, population estimates by sex and age group (Pi for age group i); estimates of economically active proportions (pEAi) by age and sex; estimates of the proportions employed in nine industry groups (pWj for industry group j); and data from CAREX of the proportions of workers within these industry groups actually exposed to the carcinogenic agents of interest (pwE jk for carcinogenic agent k). These estimates were combined to produce a point (prevalence) estimate of numbers exposed by sex, industry group (j), and agent (k) in each region as: njk = Σi(Pi * pEAi) *pWj * pwE jk (1) It was also assumed that for high-income regions, proportions exposed at high and low levels were 0.1 and 0.9, respectively, and for LMI country regions the proportions were 0.5 and 0.5,based on information about exposure prevalence in high-income countries (countries in the Australasia, highincome North America, Western Europe, and high-income Asia Pacific regions) and low-and middleincome (LMI) countries (all other countries) from identified relevant cohort studies 1-7 .
The turnover methodology used to estimate numbers ever exposed is based on the estimation method used in the CRA 2000 study 8 . For estimating population attributable fraction (PAF), the proportion of the population exposed to a workplace carcinogen is estimated as the numbers who have ever been exposed for at least a year during the risk exposure period (REP) and are still alive in the target year (say 2010), divided by the numbers who were of working age during that risk period and are still alive in that year. I.e., it is the proportion of the population at risk of developing this occupational cancer at the specific site in the target year. For these estimates, occupational turnover was estimated by age group, leading to separate age-specific PAF estimates that were applied to age-group-specific total cancer deaths (or DALYs).
The period during which exposure occurred that was relevant to the development of the cancer in the target year, the risk exposure period (REP), is defined by cancer latency, assumed to be 10-50 years for solid tumours (giving a REP for 2010 of 1961-2000, for example) and 0-20 years for haematopoietic neoplasms (giving a 2010 REP of 1991-2010) 9 . Numbers ever exposed were estimated taking into account annual staff turnover, life expectancy, age at recruitment (15-44 assumed) and age at retirement (65 for men and women), using the "turnover equation" for age group a and agent k: k c j exposed estimate, rather than taking account of the current age structure of the economically active population. For consistency, the same entry age assumptions have been used across all regions, with entry into the exposed workforce assumed to be distributed uniformly between the ages of 15 and 44. Using separate life expectancy estimates for countries representing each of the GBD regions takes some account of differences in age profiles of workers across the regions. (1) The age-specific OT factors are a product of (1) the multiplier which gives the estimate of overall numbers based on REP length, staff turnover, and numbers missing by the target year due to exceeded life expectancy, which varies with entry cohort age assumptions, and (2) the proportion of Ne(REP) falling in the specific age group by the target year, which is determined mainly by age at recruitment assumptions, plus a minor contribution from the entry cohort age profile assumption. (2) Element (1), which is the "all age" factor, remains constant across age groups. To use this equation, the assumption was made that where the exposures overlapped in the working population, the exposures were independent and their joint effect on initiating or promoting cancer was multiplicative, i.e., RR(exp 1 and exp2) = RR(exp1) * RR(exp2). (Some bias is introduced where the assumptions are unmet, although this can be estimated. This is also an appropriate equation to use for combining disjoint exposure PAFs, as it introduces less bias into the results compared with, for example, direct summing of the PAFs (Hutchings -personal communication). To estimate overall (all-age) PAFs, for individual agents or combined across agents, the summed attributable numbers are divided by the cancer numbers summed across age groups.

Estimating cumulative exposure to asbestos
The relative risk of lung cancer resulting from asbestos exposure was based on an estimate of cumulative exposure to asbestos. The relevant cumulative exposures to use were estimated for this study by first estimating the duration of exposure for each age group and using this information to estimate the average cumulative exposure, assuming all were exposed at the relevant United States Permissible Exposure Limit (PEL) in operation at the relevant time.
To estimate average duration of exposure by age group, the age at which workers born from 1920 onward enter the exposed cohort was allocated at random between a minimum 15 and maximum age 44. The workers' age in the target year (2010), their maximum possible duration of exposure up to retirement, and a (lognormally distributed) random duration of exposure 1 applied to the time between recruitment and the target year or retirement were used to estimate the average duration of exposure in five-year age groups (using 90 (year) *1000 Monte Carlo runs). The results are shown in Table A2. Recruitment from the 1920 birth cohort has been used in this update of the 2000 CRA methodology to allow all ages (to 105) into the entry cohort to avoid cutting off long durations of exposure of up to 50 years for entry at 15 and retirement at 65. Along with a cutoff at age 95, this accommodated annual turnover recruitment in, for example, a 15-44 age range, without causing durations to be cut off prematurely at higher ages.
Then cumulative exposure, for example, to asbestos was the product of the number of years of work (average duration of exposure) and the average exposure level during that period. The average exposure level was estimated using the relevant United States Permissible Exposure Limit (PEL) to asbestos in operation during the relevant years and taking (a multiple of) the weighted average 1 A number between zero and 1 obtained as the inverse minus 1 of a lognormal cumulative distribution function of x, where ln(x) is normally distributed with mean = ln(1.5) and standard deviation = (ln(2)-ln(1))/6. based on the age of the person in the target year and thus the PEL that was in operation during the years the person could have worked, in fibres/ml (f/ml). The results are in Table A2. Multiples of 2.0 (high) and 0.5 (low) for men and 1 (high), 0.2 (low) for women were used for all regions, based on an assessment taking into account relevant exposure data in published literature.

Relative risk for asbestos-related lung cancer
The majority of the relative risks were based on "exposed versus non-exposed" results. However, for lung cancer arising from exposure to asbestos, separate relative risks were calculated for high and low cumulative exposure levels (described in the Supplementary Matter). a relative risk for cumulative exposure at high level (assuming a United States Permissible Exposure Level (PEL) multiplied by two for men and one for women) and low level (PEL*0.5 for men, PEL*0.2 for women) was estimated from the study by Lenters and colleagues 12 , based on a meta-analysis of the five most methodologically sound studies, using their formula: KL= the slope of increase in the RR per unit of cumulative exposure (CE) to asbestos (in fibre-years per ml) = 0.0036, and CE = (average level of exposure) * (average years of exposure) * PEL multiplier = 13.6 * 5.93 * PEL multiplier (see Table A2 in Appendix 1).

Introduction
This example provides a step-by-step description of the approach to estimating PAFs, deaths, and DALYs, using diesel engine exhaust (DEE) exposure in males aged 55-64 years in Eastern Europe as the focus. This is based on data from GBD 2010 rather than GBD 2016.

Risk factor -exposure evidence
Diesel engine exhaust is a Group 1 IARC carcinogen, with sufficient epidemiological evidence that DEE causes lung cancer in humans 13 .

Exposure
The proportion of workers in each industry came from the ILO database 14 for the year for which the burden estimates were being made (2010 in the current example) (Table A3 -Column 2). The same estimates were used for all ages, but different estimates were available by sex. The proportion of persons within an industry who were exposed came from CAREX 15 (Table A3 -Column 3). The same estimates were used for all ages and both sexes.
The proportion of persons in the workforce who were exposed was estimated by multiplying the proportion of workers in each industry by the proportion of persons within an industry who were exposed (Column 2 * Column 3) (Table A3 -Column 4).
The number of persons in the current workforce aged 55-64 and estimated to have been exposed in 2010, stratified by the industry in which the persons worked was then calculated. This number was estimated by multiplying the proportion of persons in the workforce who were exposed (Column 4) by the total number of people in the workforce (for Eastern Europe in 2010, the number of male workers in the workforce was 48,740,000, which came from workforce data not shown in Table A3) The total number of people ever exposed (and so at risk) by 2010 will be higher than the number currently exposed because people remain at risk after they are exposed. That is, people who have left the workforce remain at risk for some period of time (often decades) afterwards. Using estimates of workforce turnover and life expectancy (as described in Section A1.4), the number of people ever exposed over a period of 10-50 years prior to 2010 and still alive in 2010, as a proportion of the currently exposed workforce, was estimated, both overall and by age group.  Column 2 = Proportion of workers who are currently (in 2010) employed in each industry sector (from ILO database). Column 3 = Proportion of the workers in each industry who are currently (in 2010) exposed to DEE (from CAREX). Column 4 = Proportion of the total workforce who are in each industry and who are currently (in 2010) exposed to DEE (= Column 2 * Column 3). Column 5 = Number of workers currently (in 2010) exposed (= Column 4 * total number of people in the workforce) (from ILO database). Column 6 = Number of persons ever (in the period 1960 to 2000) exposed through work and still alive in 2010: by industry (= Column 5 X age-specific turnover factor in Table A4). Column 7 = Proportion of persons ever (in the period 1960 to 2000) exposed through work and still alive in 2010: by industry (=Column 6 / total number of people in the workforce in 2010). Column 8 = Proportion of persons ever (in the period 1960 to 2000) exposed at high level through work and still alive in 2010: by industry (= Column 7 * 0.5). Column 9 = Proportion of persons ever (in the period 1960 to 2000) exposed at low level through work and still alive in 2010: by industry (= Column 7 * 0.5). Column 10 = Proportion of persons ever exposed (in the period 1960 to 2000) at high level through work and still alive in 2010 multiplied by ((RR for high exposures) -1): by industry. Column 11 = Proportion of persons ever exposed (in the period 1960 to 2000) at low level through work and still alive in 2010 multiplied by ((RR for low exposures) -1): by industry. For Eastern Europe, this overall proportion was 2.72 (Table A4). That means that the total number of persons ever exposed and still alive in Eastern Europe in 2010 was estimated to be 2.72 times the number of persons in the workforce who were currently exposed to diesel engine exhaust in 2010. These everexposed persons were of various ages. The number of persons in a given age group now (i.e., 2010, the year we are estimating the burden for) who had ever been exposed and who were still alive was estimated by multiplying the number of persons currently exposed (Column 5) by the relevant turnover proportion for that age group (which for the 55-64 year age group was 0.65 - Table A3) (Table A3, Column 6).

Supplementary material
The proportion of the population who had ever been exposed, by age group, was estimated by dividing the number ever exposed in a given age group (Table A3 -Column 6) by the total number of persons in that age group (which came from population data, not shown in Table A3) (Table A3, Column 7).
The proportion ever exposed was divided into high-exposed and low-exposed on the basis of 50:50 (high:low) in LMI countries and 10:90 (high:low) in high-income countries. For this analysis, Eastern Europe was considered to be a LMI country region, so the proportions with high exposure were estimated by multiplying Column 7 in Table A3 by 0.5 (Table A3, Column 8), and the proportions with low exposure were also estimated by multiplying Column 7 by 0.5 (Table A3 -Column 9).

Risk estimate
The high exposure RR for lung cancer arising from DEE was 1.47 (85% CI 1.29-1.67). This estimate came from a meta-analysis by Lipsett and Campleman, 1999 16 . The low exposure RR was set to 1.0 (i.e., equal to background risk).

Calculating age-specific PAFs
To estimate the PAF (Exposed*(RR-1)/[Exposed*(RR-1)+1]), the quantity [Exposed*(RR minus one)] was calculated. To do this, the proportion ever exposed, by age group, was multiplied by (RR minus one) for each industry and summed. For high exposure, the RR was 1.47. So, the proportion ever exposed (Table A3, Column 8) was multiplied by 0.47 (Table A3, Column 10). The same occurred for low exposure. However, the low exposure RR was 1.0 (this was true for diesel, but not for all exposures), which meant the proportion ever exposed at low levels was multiplied by zero (Table A3, Column 11).

Calculating deaths, DALYs, and overall PAF
The age-specific PAF (Table A5, Column 3) was multiplied by the number of deaths from lung cancer estimated by the GBD 2010 project for males aged 55-64 years in Eastern Europe (Table A5, Column 2).
This provided the estimate of the number of deaths from lung cancer due to DEE in males aged 55-64 years in Eastern Europe (757: Table A5, Column 4). This approach was repeated for all age groups.
The age-specific attributable lung cancer deaths for males were added, to produce the total number of deaths from lung cancer due to DEE in males in Eastern Europe (1,919: Table A5, Column 4, bottom row). This total was divided by the total number of deaths from lung cancer in Eastern Europe (658,977: Table A5, Column 2, bottom row) to produce the all-age PAF for DEE and lung cancer in Eastern Europe (2.9%: Table   A5, Column 3, bottom row). The same approach was used for both sexes and all regions to estimate the relevant number of deaths and the PAFs by age and sex and overall.
DALYs were calculated by multiplying the relevant age-sex-specific PAF by the number of DALYs from lung cancer estimated by the GBD 2010 project for the relevant age and sex groups in the region (calculations not shown).

Supplementary material
Occup

Risk factors
Exposure to second-hand smoke is now well-recognised as an important risk factor for lung cancer and for several decades has been the focus of efforts to decrease exposure [17][18][19] . As for asbestos, the results from this study primarily reflect the result of exposures before many of the current control measures were implemented. They also serve to show the effect of not having appropriate control measures in place, as is commonly still the case in many LMI countries 20 and for some occupations in high-income countries 17,18 .
Silica is a common exposure in a range of occupational situations 10,21,22 . There has been extensive debate as to its human carcinogenicity 6,23 , but silica appears to be responsible for a considerable burden due to occupational exposures.
Diesel engine exhaust has many similarities to silica. Like silica, it is a common exposure in the occupational environment and its potential harm has probably not been fully appreciated 24 , diesel engine exhaust having only been formally recognized as a definite human carcinogen in 2012 13 . Although the relative risk of lung cancer associated with diesel engine exhaust is not high, the high prevalence of past and current exposure means the cancer burden resulting from it is significant.
Our primary analysis did not include any IARC Group 2A exposures ("probably carcinogenic to humans"), or cancer sites with limited (as determined by IARC) epidemiological evidence of a causal connection to included exposures. This decision was based on an assessment by the GBD central team that there was insufficient evidence to allow other exposures and outcomes to be included, given the evidence requirements for exposures and outcomes in the overall GBD study. However, an additional analysis using similar methods was able to be undertaken to provide an indication of the potential contribution from these additional exposures and outcomes. This identified an additional 97,000 deaths (about 28% more than the official GBD estimate), in particular resulting from second-hand smoke (due to cancer of the larynx and pharynx), asbestos (cancer of the stomach and colon), and lead (cancer of the stomach and lung).

Supplementary material
Occup Based on these results, the highest PAFs apart from mesothelioma were for cancer of the larynx (30%), pharynx (23%), and lung (19%) (Tables S5 and S6). Even these estimates do not include some widespread and therefore potentially very important occupational exposures such as UV exposure from sunlight (associated with skin cancers) shift work (associated with breast cancer) and the herbicide glyphosate and insecticides malathion and diazinon (associated with non-Hodgkin's lymphoma). For example, in the UK Burden of Cancer study, breast cancers arising from shift work were estimated to be responsible for over 50% of the female cancer burden arising from occupational exposures 25 . As further epidemiological evidence becomes available, and molecular epidemiologic studies provide insights into the carcinogenic mechanisms of Group 2A exposures (as was the case, for example, with trichloroethylene 26,27 and diesel 28 ), IARC's classifications are likely to be re-evaluated. Also, the GBD decisions regarding exclusion are likely to be modified for some of these 2A exposures, potentially increasing the burden attributed to occupational exposures. Finally, no account is made of the apparent causal association between COPD, which can be related to occupational exposures, and lung cancer 29 ).

Cancers
Cancer deaths occurred mainly in men and in persons aged 55 to 79 years. Rates in men were about three to six times higher than in women for all but the youngest ages, and rates increased with age. Men are more likely to be employed in tasks that entail exposure to carcinogens and probably more likely to undertake tasks that result in exposure to higher levels of carcinogens 22,30 . In addition, due to latency and the fact that cancer risk increases with age, most people are retired before they develop cancer as a result of work exposures, which explains the high rates at older ages.
Lung cancer was the predominant type of cancer, responsible for about 300,000 deaths. Malignant mesothelioma (primarily pleural in origin) was responsible for another 27,500 deaths. These results reflect the fact that the respiratory tract, and the lung in particular, is the primary route of entry for all airborne carcinogens, and that airborne exposure is the main route of occupational exposure to carcinogens.

Supplementary material
Occup Changes over time (Tables 3 and S4) and differences between regions (Figures 1 and S2) are affected by changes and differences in exposure to occupational carcinogens. For lung cancer they are, however, also affected by time trends or regional differences in background rates due to varying prevalence of tobacco smoking. For non-asbestos carcinogens, the changes over time reflect changes in employment distribution between different industries and the resultant different probability of exposure. There was insufficient information available to allow differences in exposure probability or level over time to be taken into account in this analysis but it is hoped this will be possible in future analyses.

Comparison with other studies
The CRA 2000 study was the first attempt to comprehensively assess the contribution of occupational carcinogens to cancer 31 , but it considered a smaller number of carcinogens and cancer sites than were considered here. The CRA 2000 study estimated that in 2000 there were 152,000 deaths from cancer, compared to the 349,000 estimated for the current study. Looking only at the cancer outcomes included in the CRA 2000 study, the current study estimated about 198,000 (190%) more deaths from lung cancer, 15,000 (36%) fewer deaths from malignant mesothelioma, and 4,500 (64%) fewer deaths from leukaemia.
The main reasons for the differences appear to be that in this study additional exposures and outcomes were included (and IARC Group 2A exposures included in the CRA 2000 study were not included), many of the risk measures were different (due to updates in the literature), the estimates were made for the 2016 world population rather than the 2000 population, and the approach to estimating the population at risk was more sophisticated in the current study than was possible for the CRA 2000 study. The differences in the estimates for mesothelioma additionally arise from the different methodologies used in the two studies to develop these estimates.
Other global or national estimates of burden arising from occupational carcinogenic exposures have been published. The most recent and comprehensive of these is that by Rushton and colleagues, who estimated the burden arising from exposure to occupational carcinogens in Britain, providing PAF estimates of 8.2% for men and 2.3% for women for deaths from cancer 25,32 . The higher PAF estimates compared to those in

Supplementary material
Occup the current study arise primarily from the inclusion in that study of a much wider range of carcinogens and cancer types, including most relevant IARC Group 2A exposures, and some outcomes with limited epidemiological evidence regarding causality. This is the general approach used in several recent occupational cancer burden and exposure studies 22,[33][34][35][36][37][38] .

Methodological considerations and limitations
Methodological issues relevant to the overall study are considered in detail in the occupational risk factors overview paper 39 . The main aspects relevant to the carcinogen analysis are considered in more detail here.
Assumptions regarding latency, turnover, and at-risk period were made when estimating the population at risk. There is reasonable evidence to support these assumptions, but there is a general lack of information on the latency of specific cancers and uncertainty about variation in turnover worldwide, and the risk functions are likely to be complex and not known for most exposure-cancer pairs. CAREX provided estimates of carcinogen exposure prevalence. That study was focussed on 1990-1993 for Western Europe (and incorporated estimates from North America). It thus specifically reflects exposure circumstances from that time and those regions. This has the advantage that it reflects exposures in the past rather than in the present, which is appropriate for the current analysis due to latency considerations.
However, the information is not separately available for different sexes, different ages, or other regions.
The logic of the approach used was that in terms of chemical substances, the type of carcinogens to which workers would be exposed are likely to be reasonably consistent within a given industry regardless of the country involved. Similarly, it was assumed the proportion of workers with any exposure is likely to be reasonably consistent within a given industry regardless of the country involved. It appears likely that for most carcinogens, exposures would have been less well controlled over the relevant exposure period in LMI countries than in high-income countries (for example, see references 1-7,40 ). To attempt to take this into account, the same exposure prevalences were used for all regions, but in LMI countries a higher proportion of the exposed persons were assumed to be exposed at a high level than was the case in high-income countries. For most carcinogens, the relative risk used for low exposure was one. This means that the lowexposed persons were considered the same as those who were not occupationally exposed, giving a zero attributable burden, and that in practice the burden for these risk factors was based only on persons with high exposure.
The method used for estimation of relative risk for lung cancer from asbestos exposure required an estimate of the absolute cumulative exposure level for asbestos. This in turn required estimates of years of exposure, and of the levels of exposure during those years. With few published data to provide guidance, the developed method estimated the number of years of work, based on age, using cohort modelling. The estimates of lung cancer deaths were developed using exposure prevalences estimated using the AIR approach, which is based on mesothelioma rates. Since the latency for mesothelioma is longer than for lung cancer, our approach may overestimate current lung cancer risk for countries in which asbestos use declined long before 1990-1993 (e.g., Sweden 41 ). The same issue arises for burden from laryngeal and ovarian cancer due to asbestos exposure.
The relative risk estimates came primarily from working cohorts in high-income countries, from a range of time periods, involving a range of exposures and varying follow-up. Therefore, some mismatch between the relative risk estimates used and the exposure circumstances to which they have been applied is likely.
Nevertheless, the measures used were considered the most appropriate available.
The prolonged period of risk following exposure for most included carcinogens suggests that a considerable number of previous workers aged 80 and over would remain at risk. However, for non-asbestos exposures, RRs were set to 1.0 for ages 80 and over. This is likely to have led to a moderate underestimate of burden arising from non-asbestos-related cancers (for example, 34% of all asbestos-related cancer deaths were in persons 80 years or older, although the longer latency of asbestos-related cancers suggests the underestimate for other cancers would be lower than this). The curtailing of risk at age 80 arose from a programming error and will be corrected in subsequent GBD iterations.

Supplementary material
Occup Our study estimates for high-income countries appear to have reasonable agreement with published data on mesothelioma occurrence 42,43 . Nevertheless, there are reasons to consider that the estimates of mesothelioma in the current study are low for some LMI regions, particularly South Asia and China. The allcause estimates of cause of death are heavily influenced by reported cases. In the relevant countries, there is evidence that mesothelioma is considerably underreported, notwithstanding that there may be an effect from competing causes, with a larger proportion of the population dying at a younger age, before mesothelioma has developed, than occurs in high-income countries. There is no reliable independent information on mesothelioma numbers in these regions to allow comparison to the estimates in this study, and much of the asbestos use probably occurred more recently than in many other countries 44 ; however, recent estimates based on asbestos consumption suggest our estimates are lower than would be expected 42 . Balancing this somewhat is the fact that the PAF calculation for mesothelioma assumes that all mesothelioma is a product of occupational asbestos exposure, which could potentially overestimate burden due to occupational asbestos exposure in populations with high non-occupational asbestos exposure.
Finally, no explicit account is made of possible interactions between risk factors in people exposed to multiple risk factors, such as that known to occur in persons exposed to both asbestos and tobacco smoke 45 . The data were not available to allow potential synergistic (or antagonistic) interactions to be included in the estimates, and to the extent these occur, the presented results would probably be underestimates.

Supplementary material
Occup

Table S1
Number of workers exposed per 100,000 workers -risk factor by industry 1