Secondary Logo

Mortality Among Hardmetal Production Workers: US Cohort and Nested Case-Control Studies

Marsh, Gary M. PhD; Buchanich, Jeanine M. PhD; Zimmerman, Sarah MS; Liu, Yimeng MPH, MS; Balmert, Lauren C. PhD; Esmen, Nurtan A. PhD; Kennedy, Kathleen J. PhD

Journal of Occupational and Environmental Medicine: December 2017 - Volume 59 - Issue 12 - p e306–e326
doi: 10.1097/JOM.0000000000001075
ORIGINAL ARTICLES
Free

Objectives: To evaluate total and cause-specific mortality among hardmetal production workers with emphasis on lung cancer.

Methods: Subjects were 7304 workers ever employed in one of eight US plants from 1952 to 2008. Vital status through 2012 was determined for 97% of subjects and cause of death for 98.3% of 1087 deaths. We computed standardized mortality ratios (SMRs) and evaluated exposure–response via relative risk regression analysis.

Results: We observed overall deficits in deaths for total mortality, all cancers, and lung cancer and found no evidence of any exposure–response relationships for lung cancer.

Conclusions: We found no evidence that exposure to tungsten, cobalt, or nickel, at levels experienced by the workers examined, increases lung cancer mortality risks. We also found no evidence that work in the US hardmetal industry increases mortality risks from any other causes of death.

Center for Occupational Biostatistics and Epidemiology, Department of Biostatistics, Graduate School of Public Health, University of Pittsburgh, Pittsburgh, Pennsylvania (Dr Marsh, Dr Buchanich, Ms Zimmerman, Ms Liu, Dr Balmert); Division of Environmental and Occupational Health Sciences, School of Public Health, University of Illinois at Chicago, Chicago, Illinois (Dr Esmen, Dr Kennedy).

Address correspondence to: Gary M. Marsh, PhD, Department of Biostatistics, Graduate School of Public Health, University of Pittsburgh, Pittsburgh, PA 15261 (gmarsh@pitt.edu).

Clinical Significance Statement: This article provides estimates of total and cause of death specific mortality risks associated with employment in the US hardmetal production industry with emphasis on lung cancer and may be of interest to physicians who treat or monitor patients employed in this industry.

Sponsorship: This project was sponsored by grants from the Pennsylvania Department of Health (PADOH) and by a sponsored research contract between the University of Pittsburgh (UPitt) and the International Tungsten Industry Association (ITIA). The design, conduct, analysis, and conclusions of the study are exclusively those of the authors.

Conflicts of Interest: The authors have no conflicts of interest to declare.

Tungsten carbide (WC) is the most common hardmetal, a material formed by binding or cementing metallic carbides with a soft and ductile metal binder,1 usually cobalt (Co) or nickel (Ni). Hardmetal (WCCo), or tungsten carbide (WC) with a cobalt (Co) binder, was deemed a probable human carcinogen (Group 2A) by the International Agency for Research on Cancer (IARC) in 2006.1 IARC based this decision primarily on epidemiologic studies of European workers in hardmetal facilities that provided some evidence of increased lung cancer risk with exposure to hardmetal dust.2–5 The risk was strongest among those who worked with unsintered rather than sintered product and among those with longer durations of employment. Some members of the working group supported an evaluation of Group 1 (carcinogenic to human) based on the epidemiologic evidence and strong mechanistic evidence. However, the majority of members felt that there was need “for either sufficient evidence in humans or strong mechanistic evidence in exposed humans.” In particular, members felt that the previous epidemiologic investigations had some limitations, including use of national, and not regional, mortality comparisons, inability to adjust for confounding by smoking, and use of external, and not both external and internal, rate comparisons.1

To address these limitations, the International Tungsten Industry Association (ITIA) initiated a large international occupational epidemiologic investigation of workers employed in the WC industry. The study included eight US sites, three German sites, three Swedish sites, two United Kingdom (UK) sites, and one Austrian site, each independently conducted under the direction of country-specific occupational epidemiology experts. The study, which included three companies, multiple manufacturing sites and processes and a comprehensive, quantitative exposure assessment, is larger, more robust and more definitive than any hardmetal epidemiology study done to date. Results of the country-specific components, the exposure assessment, and the results of a pooled cohort analysis are presented in the same volume of this journal as a series of companion papers.6–10 We present here the results of the US component of the international WCCo study.

Back to Top | Article Outline

METHODS

Study Sites and Subjects

In 2006, ITIA sponsored a Phase 1 feasibility study conducted by an independent contractor to determine the availability and accessibility of company records needed for an international WCCo epidemiology study. University of Pittsburgh (UPitt) and University of Illinois at Chicago (UIC) conducted an extended and enhanced Phase 2 feasibility study under support from the ITIA. In Phase 2, conducted from October 2007 through October 2008, we developed and applied several criteria for a candidate site to be included in the international study, including a minimum size of 100 or more employees, producing WCCo or WC products since at least 1982 to allow an adequate latency period for disease development, and having detailed work history (WH) information available for all employees who ever worked at the facility. We identified nine US sites using these criteria. Subsequently, one company chose to withdraw its two facilities from the study. To maintain study size and statistical power, investigators assessed the possibility of including additional US sites with slightly relaxed criteria. The early data collection stages of the final Phase 3 component of the epidemiology study were funded by two grants from the Pennsylvania Department of Health. The remaining components of the study were performed under contract with ITIA.

The 12 eligible plant sites are shown in Table 1 and consist of 8436 employees from 10 Company A and two Company B sites. Four plants were included using the relaxed criteria (Plants 4, 6, 8, and 9), but were ultimately determined to have incomplete WH records and were not included in analyses (n = 1,132 employees). Table 1 also shows the earliest year of hire for each plant and the start of person years (earliest observation year). The earliest year of hire and earliest start of person years were in Plant 2 (1941 and 1952, respectively). December 31, 2008 is the last date of common work history collected. Several plants had workers hired before the start of follow-up. These were either employees working at another in-study plant who transferred or employees who spent time in non-study plant(s) (“other” plant). Person time for employees who initially worked at a non-study plant started on their first day working at a study plant (n = 86). For Plants 5 and 13, we adjusted the start of person time to account for a period when cohort completeness could not be confirmed; 21 employees in Plant 5, and 28 employees in Plant 13 were removed from the cohort because they did not work in the study period.

TABLE 1

TABLE 1

Back to Top | Article Outline

Cohort Enumeration and Validation

We enumerated the study cohort from multiple sources identified through discussions with company human resources personnel and exhaustive searches of the company's archives and facilities. Study-relevant records were identified in hard copy personnel records and from company-held human resource databases. Hard copy records were scanned and study variables were abstracted from the digitized images; data abstraction was verified by crosschecking random samples of files.

We conducted standard internal checks of cohort completeness by examining patterns in work history characteristics relevant to record filing, for example, name, date of hire and termination. We produced histograms of relevant variables for review and approval by facility personnel. We also compared our final cohort to active employee listings provided by the companies to ensure complete enumeration. This process led to the exclusion of Plants 4, 6, 8, and 9 as noted above.

Back to Top | Article Outline

Vital Status Tracing and Cause of Death Ascertainment

We used a modified version of our standard two-stage vital status tracing protocol to identify deaths among cohort members with unconfirmed vital status (not known from company-held records to be alive as of the study end date, December 31, 2012).11 The modified Phase 1 consisted of comparing all the names of cohort members not known to be alive against the Social Security Master Death file. Phase 2 consisted of sending subjects identified as deceased before 1979 to the corresponding state health department to obtain a death certificate. We coded death certificates to the underlying cause of death by a National Center for Health Statistics nosologist using the International Classification of Diseases (ICD) rules in effect at time of death. Also in Phase 2, we sent subjects not known to be deceased and those deceased after 1978 to the National Death Index-Plus to obtain the coded cause of death. We could not determine cause of death for 19 deaths, 15 of which occurred before 1979.

Employees without a valid social security number and without a date of death were considered lost to follow up (n = 239). We stopped person-time for those not lost to follow up at the end of study (December 31, 2012), the employee's 90th birthday, or death, whichever came first.

Back to Top | Article Outline

Exposure Estimation

The details of the exposure assessment conducted are reported elsewhere.6 In brief, quantitative job-exposure matrices (JEMs) consisting of job classes and exposure estimates were constructed for tungsten, cobalt, and nickel. Job classes were created from knowledge of production processes, discussions with plant personnel, and a review of anonymized WH record job combinations (derived from job and department titles, job and department codes, and other relevant identifying information) before generation of the exposure estimates. Job classes were then assigned to all WH record lines by a combination of programmatic and manual methods to provide linkage to the JEMs.

Exposure measurements were sorted by job class and year using a hierarchical approach that generated year categories for each class with less than or equal to 50.0% censoring (values below the limit of detection). Medians were calculated for each year category using censored data analysis (a maximum likelihood estimation of censored data as described by Cohen12) and the job classes were tested for the presence of time trends via linear regression. Measurements for job classes without significant trends, and those without sufficient data for time trend testing, were combined and a median calculated for the class that applied to all study years. The resulting job class and median combinations were categorized into one-half order of magnitude exposure intervals, and the midpoints of the intervals were applied to the JEMs and used in the epidemiologic analyses. Despite the large number of measurements available, there were some job classes with limited or completely censored data, which necessitated exposure interval assignment using professional judgment.

Job classes and their corresponding exposure intervals are presented in Kennedy et al.6 Overall, the exposures for cobalt, tungsten, and nickel were similar to or lower than those previously reported for the hardmetal industry over the time period covered by the reconstruction (1952 to 2008).6

Back to Top | Article Outline

Construction of Worker Exposure Indices

We considered both quantitative and qualitative indices of exposure in our evaluation of mortality risks from lung cancer. These analyses included all but 45 workers with unknown work history information.

Back to Top | Article Outline
Quantitative Indices

For each study plant, we linked job classes and corresponding time-specific quantitative exposure estimates for tungsten, cobalt, and nickel to the detailed subject work histories to enable the construction of working lifetime exposure profiles for each subject. The individual subject exposure profiles were used to compute three summary measures for each subject: duration of exposure (Dur) = the sum of the days spent in jobs (in years); cumulative exposure (Cum) = number of days in each job and the estimated average daily exposure (in unit-years); and average intensity of exposure (AIE) = the ratio of cum to dur (in units). The notation used to describe these summary measures is Agent_MeasureX, where X denotes the number of categories if so formed. For example, Tungsten_AIE4 refers to the average intensity of tungsten exposure based on four categories.

Alternative characterizations of the Agent_Metrics were also computed using a 15-year lag period as described in detail by Youk et al.13 Lagging attempts to characterize only the most etiologically relevant exposures and is particularly relevant for examining chronic diseases, such as cancer, associated with long latency periods. A 15-year lag is commonly used for solid tumors such as lung cancer, the focus of this investigation (eg, Hauptmann et al,14).

Back to Top | Article Outline

Occupational Exposure Limits

In the US the legally enforceable exposure limits for chemical agents are the permissible exposure limits (PELs) promulgated by the Occupational Safety and Health Administration (OSHA). The PELs are not often reviewed or updated, therefore it is common to reference the threshold limit values (TLVs) established by the American Conference of Governmental Industrial Hygienists (ACGIH) when evaluating workers’ exposures. The current TLVs for cobalt (adopted 1995), nickel (adopted 1998), and tungsten (adopted 1969) are 0.02 mg/m3 (total aerosol), 1.5 mg/m3 (inhalable fraction), and 5.0 mg/m3 (total aerosol), respectively.15,16

Back to Top | Article Outline
Qualitative Indices

An underlying interest in this study was to compare workers exposed only to WC with those exposed only to WCCo. Our exposure assessment revealed, however, that such a comparison was available only for workers at Plant 3. Using a job class-based scheme (Job classes 00-08 [background and support operations] were assigned 0 due to low overall Co exposure relative to Co-containing production operations. Job class 34 [Thermit, a Plant 3 WC production process] was assigned 0 due to absence of Co; all other production-related job classes were assigned 1 due to presence of Co), we categorized the 883 Plant 3 workers with known exposures into three mutually exclusive groups: WC only (n = 589), WCCo only (n = 138), and mixed WC/WCCo exposure (n = 147), then evaluated lung cancer mortality within each group.

Back to Top | Article Outline

Nested Case-Control Study of Lung Cancer

We obtained lifetime smoking history and other more detailed subject data via a nested, matched case-control study of lung cancer. Cases were defined as all study members who died from lung cancer from 1952 to 2012. We matched each case on exact age (at the date of death of the case), sex, and year of birth (within 1 year with a few exceptions where a larger caliper was needed) to two controls from the remaining living and deceased members of the cohort. We also selected several replacement controls for each case that were used only if all attempts to locate an original control failed. All controls were selected randomly from pools of eligible controls. To avoid possible overmatching on occupational exposure, we did not match cases to controls by study plant. Extensive efforts were made to locate and contact each case and control or a knowledgeable informant (usually a surviving family member for cases and deceased controls), and if contacted, to obtain their consent and administer a structured telephone interview. The interview sought information on lifetime smoking history, occupational history and exposures within and outside the hardmetal industry, and sociodemographic factors.

Back to Top | Article Outline

Statistical Methods

Analysis of General Mortality Patterns, External Comparisons

We examined the total and cause-specific mortality experiences of the US cohort from January 1, 1952, to December 31, 2012. Cohort analyses were performed using the Occupational Cohort Mortality Analysis Program (OCMAP-PLUS).17 Using a modified life-table procedure, the person-years at risk contributed by each study member were jointly classified by plant, race, sex, age group, calendar time, year of hire, duration of employment (DOE), and time since first employment (TSFE). Some analyses were classified by worker type, as the mortality experience of long-term workers may be less confounded by the unhealthy behavior and lifestyle patterns often observed among short-term workers.18,19 We defined short-term workers as those whose total DOE was less than 1 year (worker type 1) or less than 5 years (worker type 5), and long-term workers having a total DOE of 1 or more or 5 or more years, respectively. In the analysis by worker type, the first 1 or 5 years of employment for long-term workers were excluded from the corresponding short-term worker person-year counts.

Expected numbers of deaths were computed by multiplying average annual race-, sex-, age-, and time- specific standard population death rates by the person-years at risk in the corresponding race–sex–age–time intervals. Person-years and observed deaths for subjects of unknown race were assigned to white or non-white categories in proportion to the factor-specific distributions of study members with known race using the Proportion Allocation Method in OCMAP-Plus.17 Expected numbers of deaths were computed using as standard populations the total United States and the local plant areas (defined as aggregates of counties) from which the plant workforces were largely drawn. Appendix A shows the county groups used for each of the eight study sites. Population-weighted county rates were obtained from the Mortality and Population Data System (MPDS).17 Due to MPDS data limitations (mortality rates for all causes combined and nonmalignant causes limited to 1962 to 2012), expected numbers of deaths for these cause of death categories during the years 1952 to 1961 were based on standard rates for 1962 to 1964. Appendix B shows the cause of death categories and corresponding ICD codes used in our analyses.

To account for geographic variability, the analysis of general mortality patterns focused primarily on the local county comparisons. Local county comparisons also usually provide the most valid external mortality comparisons by helping to adjust for the social, cultural, and economic factors related to disease, and at some level geographic differences in smoking habits. Standardized mortality ratios (SMRs) and their 95% confidence intervals (CI) were computed for the total US cohort and selected sub-groups. Statistically significant deviations of the SMR below and above 1.00 were identified using Poisson distribution probabilities.20

Back to Top | Article Outline

Analysis of Lung Cancer Mortality, Internal Comparisons

We used relative risk (RR) regression modeling to investigate the dependence of the internal cohort mortality rates (modeled as time to death) for lung cancer on combinations of the exposure variables considered, with adjustment for potential confounding factors. For each cause of death, risk sets were explicitly constructed from the cohort data using the RISKSET program module in OCMAP-PLUS.17 Risk sets were matched on exact age at death (event time), sex, and on year of birth (±1-year caliper) to control for cohort effects. Time-dependent exposures were evaluated for each individual at each event time they were at risk.

Multiplicative Cox RR regression models of the form λ(t) = λ0(t) expx ([t ]β) were fitted to the internal cohort rates21,22 and either asymptotic or exact (depending on number of cases) conditional logistic regression programs in Stata 1223 were used to estimate β (beta) from the explicitly constructed risk sets. To parallel the descriptive SMR analysis of mortality in relation to exposure, categorized forms of the covariates were considered. The statistical significance of each main effect (expressed as a global P value) was assessed with a likelihood ratio statistic. For the quantitative exposure variables, a test for linear trend was also conducted (expressed as a trend P value). Using the continuous form of the exposure metrics, we also evaluated main effects with a test of the β (linear slope) coefficient derived from the model. We considered both unlagged and 15-year lagged versions of all exposure metrics. All tests on RRs and SMRs were done at the two-sided 0.05 significance level, and no adjustment was made for multiple comparisons.

Back to Top | Article Outline

RESULTS

Cohort Description

Table 2 shows the distribution of the US cohort by selected study factors. The final cohort consisted of 7304 persons (71.1% male) with approximately 167,000 person years. The majority of employees with known race were white; however, 46.5% of the cohort had no race information in company records. More than 70% of the cohort worked longer than 1 year and nearly half worked longer than 5 years. The largest plant in the study was Plant 13 (n = 1588) and the smallest was Plant 5 (n = 467). Plant 7 had the highest proportion of females (58.2%). About 42% of the cohort was born after 1960, making them 50 years old or younger at the end of study, and nearly one-third of the cohort was hired at age 35 or older. Almost half the cohort was hired after 1989. At the end of the study period, 5973 subjects (82.8%) were assumed alive (0.4%), 30 subjects were 90 years old or greater (not shown, n = 30, 0.4%), and 1087 (14.9% were known to be dead). Cause of death was ascertained for 1068 or 98.3% of the deaths. Only 3.3% of the cohort was untraceable. Nearly 20% of the cohort was employed for 20+ years and over half had 20 or more years’ TSFE. While 30.3% of the cohort had more than 5 years’ employment with more than 20 years of follow-up, only 16.8% of the cohort had more than 10 years of employment and more than 30 years of follow-up.

TABLE 2

TABLE 2

Back to Top | Article Outline

Exposure Assessment

Table 3 shows selected descriptive statistics for the Agent_Measures used in the study computed across all workers with known exposures. Figure 1 shows the mean Agent_AIEs by study plant; Fig. 2 shows the mean Agent_AIEs by study time periods. Mean Tungsten_AIE and Nickel_AIEs were highest in Plant 3, whereas the highest mean Cobalt_AIE occurred in Plant 5. Across the time period of the study, mean Agent_AIEs increased during the earliest time period, then decreased at varying rates across time.

TABLE 3

TABLE 3

FIGURE 1

FIGURE 1

FIGURE 2

FIGURE 2

Table 3 and Fig. 1 also show that for each plant the median and maximum Tunsten_AIE (mg/m3) and median and maximum Nickel_AIE (mg/m3), calculated using the exposure interval midpoints, were well below the current TLVs. For cobalt, the median Cobalt_AIE (mg/m3) was below the current TLV for workers in each study plant, whereas for all plants some workers had maximum Cobalt_AIE values that exceeded the current TLV.

Back to Top | Article Outline

Nested Case-Control Study

Table 4 provides a breakdown of our participation rates by case-control status. While the overall participation rate for the 102 identified lung cancer cases was moderate (47.1%), it was low for the 190 matched controls (33.7%) (controls for seven recently identified cases were not selected). The less than ideal participation rates stemmed mainly from difficulties (for cases and controls, respectively) locating the subject or a knowledgeable informant (63.7% and 55.7%) rather than from difficulties contacting (95.3% and 82.0%) or interviewing (77.4% and 73.5%). We attribute our difficulties locating respondents to modern day issues including: lack of a national cell phone database, increasing use of call blocking and caller ID, restrictions on uses of databases other than public records, and heightened awareness of privacy invasion and identity theft. Such problems are increasingly encountered in research as described in recent commentaries and editorials.24–26

TABLE 4

TABLE 4

Table 5 shows selected demographic characteristics of cases and controls with interview data. About one-third of cases and one-half of controls were from Plant 2. Most (97.9%) interviews for cases were obtained from a spouse, child, or sibling, whereas control interviews were largely (64.1%) self-interviews. Information on cigarette smoking history (ever/never smoking) was obtained for all but one case. Ever smoking was reported in 87.5% of cases compared with 64.1% of controls. While a larger percentage of “unknown” responses were associated with information collected on detailed smoking habits and other occupational exposures, the available data on the frequencies of years smoked, cigarettes smoked per day, and their product, pack-years, indicate that control smokers generally smoked less than the case smokers. Case respondents also reported higher frequencies of exposures to known or suspected lung carcinogens, asbestos, soot/diesel exhaust, and roof/asphalt fumes than their control counterparts did.

TABLE 5

TABLE 5

Back to Top | Article Outline

General Mortality Patterns: External Comparisons

Table 6 shows, for the total US cohort during the entire 1952 to 2012 study period, observed deaths and SMRs for all causes of death combined and for specific nonmalignant neoplasm cause of death categories with at least one observed death. We observed 1087 deaths overall yielding statistically significant (P < 0.01) 11% and 16% overall deficits in mortality compared with the United States and local county experiences, respectively. This included 97 deaths due to nonmalignant respiratory disease (NMRD), including 50 from chronic obstructive pulmonary disease (COPD) (bronchitis, emphysema, and asthma), and 36 from the residual category “Other NMRD.” The NMRD categories generally yielded deficits in the infectious disease categories and excesses in the chronic disease categories, based on both United States and local county rates; however, none was statistically significant. While not shown, the other NMRD category, with a local rate-based SMR = 1.00, included six deaths coded to categories including or suggestive of occupational pneumoconioses: two from Plant 2 coded as “coal workers’ pneumoconiosis” (ICD9 = 500 and ICD10 = J60) and four coded as “other interstitial pulmonary disease with fibrosis” (ICD10 = J84.1) from Plant 3, Plant 12, and Plant 13 (two deaths). The latter category may include deaths from hardmetal disease, which is associated with exposure to fume or dust from WCCo. The small numbers precluded a more detailed evaluation of these six deaths.

TABLE 6

TABLE 6

The category “asthma” in Table 6 yielded a local rate based SMR of 1.77, however, this was based on only three deaths and was not statistically significant. Overall deficits in mortality were also observed for most of the other non-malignant cause of death categories examined, and many of these were statistically significant. We observed elevated SMRs for some cause of death categories, however, many were based on small numbers of deaths and none was statistically significant. SMRs based on local county rates were generally similar but somewhat lower than SMRs based on US rates.

Table 7 shows observed deaths and SMRs for all malignant neoplasms combined and selected subcategories. Overall cancer mortality in the US cohort was slightly less than that expected in both the US and local county comparison populations (SMRs = 0.94 and 0.91, respectively). We observed 105 respiratory system cancer deaths, of which 102 were cancers of the bronchus, trachea, or lung (termed “lung cancer”) and three were cancers of the larynx. These numbers of observed deaths were almost identical to those expected based on both US and local country rates, producing SMRs very close to 1.00. With the exception of two statistically significant deficits in mortality (“cancer of bladder and other urinary organs” and the residual category, “all other malignant neoplasms”), none of the other cancer categories shown in Table 7 revealed remarkable or statistically significant excesses or deficits in deaths. As in Table 6, SMRs in Table 7 based on local county rates were generally similar but somewhat lower than SMRs based on US rates.

TABLE 7

TABLE 7

Table 8 shows observed deaths and SMRs for lung cancer by plant and sex. Here and in all tables that follow, SMRs were based only on the local county comparison. For males and females combined, lung cancer SMRs vary from 0.41 for Plant 1 (based on only two deaths) to 1.96 (based on six deaths) for workers from Plant 10. None of the elevated SMRs was statistically significant. Across all plants, we observed a not statistically significant 32% excess in lung cancer among females based on 24 deaths, compared with a 10% deficit in males based on 78 deaths. While the observed numbers of female lung cancer deaths were very small, producing imprecise SMRs for most plants, 12 deaths were observed among Plant 7 females yielding a borderline, not statistically significant 91% excess in deaths (SMR = 1.91, 95%CI = 0.99 to 3.34). We conducted additional cohort and case-control analyses of lung cancer mortality among Plant 7 females. SMRs showed 1.5 to 5-fold excesses at all levels of all factors examined, however, these were based on small numbers of deaths, none was statistically significant, and there was no consistent evidence of an association with occupational factors (data not shown). We also observed excesses in male lung cancer deaths in Plant 10 and Plant 13, however, neither exceeded 2-fold nor were statistically significant.

TABLE 8

TABLE 8

Back to Top | Article Outline

Lung Cancer Mortality in Relation to Study Factors: External and Internal Comparisons

Table 9 focuses on lung cancer mortality and shows observed deaths, local rate-based SMRs and RRs by selected study factors. RRs are not shown for sex, age at risk, and time period at risk as these factors were used to match lung cancer cases to controls in the internal comparisons. We also did not compute RRs by race category due to the large number of workers with unknown race whose race was imputed for the SMR analysis (same imputation method not possible for internal comparisons). None of the category-specific SMRs for the factors considered was statistically significant. In addition, we observed no trends or patterns in SMRs for factors that related more directly to working experience, such as DOE and TSFE. In fact, among workers within the four categories of longer DOE and TSFE, where an occupational association is most likely to be detected, we found slight 3% excesses or deficits in lung cancer deaths. Also of note is that short-term workers (based on less than 1 and less than 5 years) had small excesses in deaths (SMRs = 1.28 and 1.13, respectively) compared with deficits for the corresponding long-term workers (SMRs = 0.89 and 0.87, respectively).

TABLE 9

TABLE 9

Among the study factors considered in the internal comparisons, only worker type 1 and plant were statistically significant predictors of lung cancer mortality (global P values = 0.047 and 0.028, respectively). This finding for worker type 1 reflected a borderline statistically significant 36% deficit in deaths for long-term versus short-term workers (73 deaths, RR = 0.64, 95%CI = 0.41 to 0.99). For plant, using Plant 3 as the arbitrary baseline, this finding reflected the large variability in RRs ranging from 0.64 for Plant 1 and Plant 2 to a statistically significant 2.67-fold excess (six deaths, RR = 2.67, 95%CI = 1.03 to 6.89) for Plant 10. The plant heterogeneity was also evident in the SMRs in Table 9, as previously shown in Table 8. Similar to the SMRs in Table 9, we observed no evidence of elevated risks or trends in lung cancer RRs associated with DOE or TSFE or combinations of these two work experience-related factors.

Back to Top | Article Outline

Lung Cancer Mortality in Relation to Occupational Exposure

Table 10 shows univariate RR regression models relating internal cohort rates for lung cancer to unlagged cumulative and average intensity of exposure to tungsten, cobalt, and nickel. Because this analysis included only exposed workers (45 workers with unknown exposure were excluded), duration of exposure applied to all three agents. We categorized exposures at approximate percentiles of the distribution for the lung cancer cases. To provide complementary external comparisons and to evaluate lung cancer mortality risks in the baseline categories of the RRs, we also computed SMRs for exposure category.

TABLE 10

TABLE 10

The global test P values and continuous form slope estimates P values in Table 10 indicate that no Agent_Metric was a statistically significant predictor of lung cancer risk. Moreover, we observed elevated lung cancer risks in certain exposure categories, including every level of each Agent_AIE, but with the exception of Nickel_AIE4, none was statistically significant. For Nickel_AIE4, we observed statistically significant slightly greater than 2-fold risks in the three non-baseline categories, however, these elevated risks stemmed from an inordinately low, statistically significant deficit in lung cancer deaths in the baseline category used for the RRs (14 deaths, SMR = 0.58, 95%CI = 0.32 to 0.97). This large deficit in the baseline category led to more than 2-fold elevations in RRs even though the non-baseline SMRs were all close to 1.00. Consistent with this finding is that for RRs for Nickel_Cum4, in which the baseline SMR was close to 1.00, we observed deficits in lung cancer deaths for each exposure category. Parallel analyses based on a 15-year lagged version of the same Agent_Metrics, using two forms of the baseline category (one using low exposure and one using unexposed [here, exposure lagging resulted in some workers having zero exposure]) yielded similar results (data not shown).

Because the variables worker type 1 and plant were statistically significant predictors of lung cancer risk in Table 9, we refitted the unlagged RR regression models in Table 10 including these two factors alone and together as potential confounding factors. The results for the univariate and multivariate models are shown in Table 11. Similar to Table 10, none of the Agent_Metrics in the adjusted models were statistically significant predictors of lung cancer risk and we observed no evidence of any positive exposure–response relationships. The worker type 1 adjustment increased the RRs for most factors, while the plant adjustment decreased the RRs. Inclusion of both worker type 1 and plant led to RRs approaching 1 for all factors. The only statistically significant elevations in category-specific RRs appeared again for Nickel_AIE (only in worker type 1 adjusted model) and again stemmed from large deficits in deaths in the baseline category.

TABLE 11

TABLE 11

We also fit the univariate RR regression models and models adjusted for plant on the subcohort of long-term (1+ years employed) alone. Short-term workers, who comprise 27.7% of the US cohort, revealed a statistically significant elevation in lung cancer risk relative to long-term workers (Table 9) that is not likely due to occupational exposure in the hardmetal industry. The results for these models are shown in Table 12 and illustrated in Figs. 3 and 4 for the AIE and CUM metrics. Generally, the pattern of findings for long-term workers was similar to those for the total cohort, although RRs for the highest levels Tungsten_Cum4 increased from below 1 to 1.41. Neither the unadjusted or plant-adjusted models revealed statistically significant main effects or evidence of exposure–response relationships. For each Agent_Metric, plant-adjusted, category-specific RRs were generally lower than unadjusted RRs and none, including Nickel_AIE, were statistically significant.

TABLE 12

TABLE 12

FIGURE 3

FIGURE 3

FIGURE 4

FIGURE 4

Back to Top | Article Outline

Lung Cancer Mortality Risk Estimates for Workers Exposed to Tungsten Carbide Without Cobalt Binder

This qualitative analysis of workers from Plant 3 revealed that 23 of 27 lung cancer deaths among workers with known exposures occurred among the WC only group yielding an overall not statistically significant SMR of 1.13 (95%CI = 0.71 to 1.69) (data not shown). Of these deaths, 17 occurred among workers who were employed less than 1 year and no trends were evident by DOE. Three of the remaining 27 deaths were among workers with mixed WC/WCCo exposure (SMR = 0.96, 95%CI = 0.20 to 2.81).

Back to Top | Article Outline

Lung Cancer Mortality Risk Estimates Adjusted for Smoking

Despite our extensive efforts described above, participation rates for cases and matched controls in our nested case-control study were below 50% and this yielded too few complete risk sets (case and at least one matched control with interview data) for robust statistical analysis. We also found differential participation rates for cases and controls (or their proxies) relative to key demographic and work history variables (data not shown), further limiting the utility of these data. Thus, we chose not to use the limited smoking data obtained via interview to adjust lung cancer mortality risk estimates directly for potential confounding by smoking and other possible risk factors for lung cancer.

Table 13 shows the available data on smoking history and “other” occupational exposures obtained from the case-control study. Using an unmatched unconditional logistic regression analysis, we estimated odds ratios (ORs) for categories of two smoking variables and five other exposure variables. Table 13 shows that the risk of lung cancer among ever smokers was 4.71 (95%CI = 1.64 to 13.58) and 10.51 (95%CI = 2.83 to 39.09) among workers who smoked 20 to 39 pack-years. While statistically significant, these risk estimates are well below the at least 10-fold risk among ever smokers and 35-fold risk for heavy smokers (compared with non-smokers) found in earlier studies,27 reflecting the low participation rates and/or possible misclassification of smoking histories among respondents. Table 13 also shows statistically significant main effects for lung cancer in relation to reported other occupational exposures (ever vs. never exposed) to ionizing radiation, arsenic and soot/diesel exhaust, although these variables include a substantial number of unknown responses with the highest estimates of risk.

TABLE 13

TABLE 13

Alternatively, we applied the indirect method proposed by Miettinen28 and described by Axelson and Steenland29 and Steenland and Greenland30 to adjust our external mortality comparisons (overall and plant-specific lung cancer SMRs) for potential confounding by smoking. Due to small numbers of lung cancer deaths, our plant-specific analysis excluded Plant 1 and was limited to male workers in all but Plant 7, which was limited to female workers. The overall estimate was based on all workers.

The basic approach to adjusting SMRs indirectly requires first determining the confounding risk ratio (CRR), which is a function of the estimated RR for smoking and lung cancer among the cohort and the prevalence of ever smoking among cohort members and members of the corresponding external standard populations. We used US Department of Health and Human Services (USDHHS) estimates of lung cancer risk in relation to smoking (RR = 10.0) as an alternative to the underestimated cohort-based risks.27 We estimated smoking prevalence among cohort members using smoking data from the case-control study. We treated cases and controls as a stratified sample of the total cohort and formed a combined weighted estimate that accounted for the probability of selection and non-response. We also used post-stratification weights to make the samples representative of the total cohort. Because all controls represented a much larger subcohort than all cases (who were treated as a 100% sample), the prevalence of smoking estimate was driven mainly by the controls. Interviewed controls were representative of the total cohort with respect to race and sex, but were older, longer-term employees. Thus, while our estimates of smoking prevalence were heavily weighted by earlier birth cohorts associated with higher smoking prevalence, we found that smoking prevalence did not vary materially across the 3 years of birth cohorts (data not shown). Our final estimate of smoking prevalence for the total cohort was formed as a weighted average of the estimates from the 3 years of birth cohorts.

We used the 1995 Behavioral Risk Factor Surveillance System Survey31 to determine smoking prevalence for the states that included each study plant. The “all plants” smoking prevalence was found as a weighted average of the 1995 US male and female smoking prevalence rates.31 Analytical details of our indirect method for smoking adjustment are presented elsewhere.32

Table 14 shows two versions of CRRs and indirectly adjusted lung cancer SMRs for the total cohort and by plant. Version 1 is based on smoking prevalence data estimated for each plant; Version 2 uses for each plant the smoking prevalence estimated for all plants combined and thus is the more robust approach. Under Version 2, the smoking prevalence among workers from all plants is greater than the corresponding rate in the standard population, which indicates that the unadjusted SMRs were positively confounded by smoking. This led to CRRs greater than 1.00 and smaller adjusted SMRs for each plant and overall. In particular, the near 2-fold lung cancer excesses noted among females for Plant 7 (SMR = 1.91) and Plant 10 (SMR = 1.98) are reduced to SMRs of 1.32 and 1.87, respectively. The adjusted lung cancer SMRs for all plants combined was reduced from 0.90 to 0.70.

TABLE 14

TABLE 14

We were also concerned about potential confounding by smoking impacting the results of our internal mortality comparisons, including the exposure–response analyses. In the latter case, smoking would confound the exposure–response only if smoking prevalence was related to exposure level. To evaluate this possibility, we computed ever smoking prevalence for each of three categories of AIE to tungsten, cobalt, and nickel and found no evidence of an association that would indicate potential confounding by smoking (data not shown).

Back to Top | Article Outline

DISCUSSION

Our findings of overall deficits in mortality for all causes of death combined and for most of the cause of death categories examined are consistent with those found in other occupational groups.33 These favorable mortality patterns, especially for the long-term chronic disease categories, are probably influenced in part by the “healthy worker effect,” the relative absence of deleterious employment-related health risks, the positive health effects of continuing employment, and better health care access. Although we standardized all risk estimates for age, our study population was relatively young, explaining why only 14.9% of study members died before the end of the observation period (December 31, 2012). Moreover, many members did not reach the older age groups for which mortality rates for many cause of death categories increase dramatically. Thus, continued follow-up of this cohort is indicated, as it may reveal different patterns and rates of mortality for some of the cause of death categories associated with older age and/or longer latency periods.

Lung cancer was the health endpoint of primary interest in this study, and with the exception of elevated mortality risks among male workers in Plant 10 and female workers in Plant 7, both based on small numbers of deaths, our study revealed no consistent evidence of elevated lung cancer mortality risks overall or among the many demographic and exposure-based subgroups examined. Of particular note were consistent deficits in lung cancer mortality based on internal comparisons of workers with the greatest potential for risk (eg, employed more than 5 or more than 10 years, and followed for 20 or more or 30 or more years) with workers without these risk potentials (Table 9). We also observed small excesses in lung cancer mortality among short-term workers compared with long-term workers (using both 1 and 5 years cut points) who revealed deficits in lung cancer mortality. These findings in short-term workers are unlikely due to occupational factors in the hardmetal industry but rather differences in lifestyle and behavior (eg, higher rates of smoking) that could impact lung cancer risk.33

We observed considerable heterogeneity in lung cancer risk estimates across the eight study plants, and in some cases between males and females within a given plant (eg, Plant 3). Discerning reasons for this heterogeneity is difficult as we used local county rates to generate expected deaths, but the numbers of deaths were small yielding imprecise risk estimates with CIs that were all consistent with null risk. While exposure levels to the three measured agents, tungsten, cobalt, and nickel, also varied by plant (Fig. 1), mean exposure levels were well below recommended standards and plants with the highest mean AIEs were among those with the lower lung cancer risks and vice versa (Table 8). The plant heterogeneity in lung cancer risk may be due in part to differences in plant-specific smoking prevalence as evident in Table 14.

Tungsten is not considered to be a known or suspected carcinogen.1,34 Nickel has been deemed a Group 1 carcinogen (carcinogenic in humans).1 However, the nickel exposures in this study are considerably lower than those found in cohort studies of nickel refinery workers.35–39 IARC1 classified cobalt metal with WC, or hardmetal (WCCo), as Group 2A (probably carcinogenic), and cobalt metal without WC as Group 2B (possibly carcinogenic). There was some support among IARC working group members to designate hardmetal as Group 1 (carcinogenic) based on the earlier epidemiologic studies. The IARC evaluation identified the highest Co exposures in weighing, grinding, and finishing.

In this study, Co exposures were highest in presintering and powder production operations. However, cobalt exposure levels here were equal to or lower than those reported for hardmetal operations in other studies of workers in the industry.6 We were able to evaluate qualitative WCCo exposure in one plant (Plant 3). However, due to the small numbers of deaths it was not possible to evaluate exposure–response relationships comprehensively.6

Our findings are not consistent with earlier studies of workers in the hardmetal industry. The studies of French workers on which the IARC evaluation of cobalt was primarily based3–5 found overall SMRs for lung cancer from 1.303 to 2.135. The earliest French study, Lasfargues et al,5 found increased risks for lung cancer in those with the longest DOE, longest TSFE, and highest degree of exposure to WCCo. Our study did not find similar evidence of lung cancer risk in any of those groups. However, the Lasfargues study did not perform a quantitative exposure assessment. Similarly, Wild et al4 did not quantify exposures, but instead used level of WCCo exposure and an evaluation of work with exposure to pre-sintered versus sintered product. The lung cancer SMR for pre-sintering was 1.42 (95%CI = 1.10 to 4.59) based on 10 deaths; the lung cancer SMR for sintered product workers (based on five deaths) was 1.28 (95%CI = 0.41 to 2.98). We did not evaluate pre- versus post-sintering processes in that manner in the US study, but did so in our pooled analysis of the US and EU cohort studies.40

Moulin et al3 performed a cohort study of 10 French hardmetal plants with a nested case-control study. A semi-quantitative exposure matrix was developed and measured cobalt concentrations were used to validate exposures. At the cohort level, there was evidence that lung cancer risk increased with TSFE. The nested case-control study found excess risk with increasing DOE and cumulative exposure, especially in pre-sintering operations. Adjustment for smoking or other potential confounders did not change the risk estimates. A Swedish cohort study also found evidence of increased risk of lung cancer for the total cohort, among workers with more than 10 years DOE, and among those with more than 20 years’ TSFE.2

The earlier cohort studies of workers in the hardmetal industry suffered from several limitations: the use of national rates only, and not local rates, for SMRs; no SMR adjustment for confounding by smoking; a lack of quantitative exposure estimates; incomplete work histories; and no internal analyses. Our study has improved upon these limitations primarily by producing both national and local rate-based SMRs, and performing internal RR analyses. Our results showed that using local comparisons, which adjusted for the demographic and social characteristics of the populations around the facilities, produced lower SMRs than national rates. We found some evidence of increasing RR by exposure, but the low to very low SMRs in some baseline groups precluded meaningful interpretation. Moulin et al3 adjusted ORs for smoking and found little evidence of confounding. Our indirect adjustment for confounding by smoking produced uniformly lower risk estimates, indicating the possibility of positive confounding by smoking.

Our comprehensive quantitative exposure assessment enabled us to evaluate rigorously lung cancer mortality risks in relation to the duration, average intensity, and cumulative exposure to tungsten, cobalt, and nickel. Our exposure–response analyses of these Agent_Metrics, based primarily on RR regression modeling that included the cofactors, worker type 1 and plant, which we found to be potential confounding factors in our univariate RR models. The adjusted, multivariable models indicated that none of the agents examined was a risk factor for lung cancer mortality. We corroborated this finding by a parallel exposure–response analysis limited to long-term (1+ years) workers adjusted for plant. These findings are consistent with the observation that the median AIE to each agent (calculated across all workers) was well below the current agent TLV and indicate that workers exposed to these levels of tungsten, cobalt, and nickel are not at an increased risk of lung cancer.

Our initial study design included a nested case-control study of lung cancer mortality aimed at providing individual level data on smoking histories for adjusting lung cancer risk estimates for potential confounding by smoking. Unfortunately, despite extensive efforts to achieve participation rates that would minimize non-response bias, we deemed our final participation rates for cases and controls too low for robust statistical analysis. Alternatively, we used the available smoking history data from the case-control study to estimate overall and plant-specific “ever” smoking prevalence and along with other data indirectly adjusted lung cancer SMRs for potential confounding by smoking. While the general absence of elevated lung cancer risks in our study population lessened the need to adjust for potential positive confounding by smoking (ie, falsely inflated risks), we applied adjustments to rule out the possibility of negative confounding by smoking (ie, falsely decreased risks). Our estimates of smoking prevalence indicated that workers tended to smoke more than their corresponding general populations (a usual observation when comparing blue collar working populations to general populations33). Subsequent indirect adjustments led to reduced lung cancer SMRs in all study plants indicating the presence of moderate positive confounding by smoking.

Another health endpoint potentially of interest in our study was NMRD, especially the occupational pneumoconiosis hardmetal disease, which can be caused by dust or fume exposures from hardmetal objects containing cobalt. The mortality endpoints were included in our cause of death category, “other NMRD,” which revealed no evidence of statistically significant excess mortality risks overall or in any study plant. Our qualitative examination of specific deaths within this category yielded only four deaths coded as “other interstitial pulmonary disease with fibrosis” that could have included deaths from hardmetal disease. The small numbers of deaths and unspecific categorization precluded further evaluation.

Aside from our problems with the nested case-control study noted above, a limitation of our study was the need to exclude 1132 workers from four prospective study plants, although these were among the smallest plants that yielded relatively small numbers of deaths. The study cohort was also young in age with only 14.9% of the total cohort dying during the observation period. Moreover, many of the study members did not reach the older age groups associated with higher mortality rates from many chronic diseases including lung cancer. Thus, our study may not be fully representative of the long-term mortality experience of workers from this industry, arguing for continued follow-up of this cohort. The extensive exposure assessment provided quantitative estimates for tungsten, nickel, and cobalt, but, aside from the qualitative exposures for Plant 3, we were unable to assess exposure to WC or WCCo (the primary exposure of interest) either quantitatively or qualitatively for the entire cohort.

Our historical cohort study of US hardmetal workers has many methodological strengths. It includes a moderately large cohort representing two companies, eight manufacturing sites, and substantial geographic variability. Our study included a comprehensive, quantitative individual-level assessment of exposure to tungsten, cobalt, and nickel that enabled robust RR regression modeling of exposure–response with control for potential confounding factors. Our study also included a nested case-control study of lung cancer that ultimately enabled the estimation of smoking prevalence used in our indirect adjustments for smoking. We conducted a systematic verification of cohort completeness that led to the exclusion of four original study plants due to evidence of incomplete work histories. We also attained excellent mortality follow-up and cause of death ascertainment rates that covered a long 60-year (1952 to 2012) observation period. The focus of our study was to determine whether hardmetal production workers are at an increased risk for lung cancer. We observed 102 lung cancer deaths that yielded excellent statistical power to detect important excesses in deaths overall and in many of the subcategories examined.

Back to Top | Article Outline

CONCLUSIONS

Our historical cohort and nested case-control studies of US workers in hardmetal production provides no consistent evidence that work in this industry is associated with an increased risk of lung cancer, as suggested in the earlier French and Swedish epidemiologic studies. We also found no evidence that exposure to tungsten, cobalt, or nickel, at levels experienced by the workers examined, increases lung cancer mortality risks. We also found no evidence that work in the US hardmetal industry increases mortality risks from any other causes of death. Our favorable results should help guide risk management efforts for workers exposed to hardmetals.

Back to Top | Article Outline

Acknowledgments

The authors would like to acknowledge the cooperation and assistance of the representatives from the ITIA and its member companies and employees that provided assistance throughout the study. In addition, they acknowledge the data processing and management work done by DataBanque of Pittsburgh, the computer programming work of Charles Alcorn and Michael Lann, the data collection efforts of Michael Cunningham, the project management assistance of Terri Washington, the case-control study tracing and interviewing efforts of Terri Washington and Juley Stragand, the telephone interviewing work by Daniesha Hunter and Chanelle Labash, the statistical analysis support of Dr. Ada Youk, and the death certificate coding performed by June Pearce. They also acknowledge the faculty and staff of the UPitt Evaluation Institute for Public Health who assisted in survey instrument design, subject recruitment, and data collection under the supervision of Dr. Todd M. Bear. The research proposal was approved by the Institutional Review Boards (IRB) of the University of Pittsburgh and the University of Illinois at Chicago.

Back to Top | Article Outline

Appendix A. County Groups Used for the Construction of Local Comparison Mortality Rates, by Plant

Table

Table

Back to Top | Article Outline

Appendix B. Causes of Death Categories and Revision-Specific International Classification of Disease (ICD) Codes Used in Mortality Analysis

Table

Table

Table

Table

Table

Table

Back to Top | Article Outline

REFERENCES

1. IARC Working Group on the Evaluation of Carcinogenic Risks to Humans. Formaldehyde, 2-butoxyethanol and 1-tert-butoxypropan-2-ol. IARC Monogr Eval Carcinog Risks Hum 2006; 86:1–478.
2. Hogstedt C, Alexandersson R. Mortality among Hard-metal workers. Arbete Halsa 1990; 21:1–26.
3. Moulin JJ, Wild P, Romazini S, et al. Lung cancer risk in hard-metal workers. Am J Epidemiol 1998; 148:241–248.
4. Wild P, Perdrix A, Romazini S, Moulin JJ, Pellet F. Lung cancer mortality in a site producing hardmetals. Occup Environ Med 2000; 57:568–573.
5. Lasfargues G, Wild P, Moulin JJ, et al. Lung cancer mortality in a French cohort of hard-metal workers. Am J Ind Med 1994; 26:585–595.
6. Kennedy KJ, Esmen NA, Buchanich JM, Zimmerman S, Sleeuwenhoek AJ, Marsh GM. Mortality among hardmetal production workers: occupational exposures. J Occup Env Med 2017; 59:e297–e305.
7. McElvenny DM, MacCalman LA, Sleeuwenhoek A, et al. Mortality among hardmetal production workers: UK cohort and nested case–control studies. J Occup Env Med 2017; 59:e275–e281.
8. Morfeld P, Groß JV, Erren TC, et al. Mortality among hardmetal production workers: German historical cohort study. J Occup Env Med 2017; 59:e288–e296.
9. Wallner P, Kundi M, Moshammer H, Zimmerman SD, Buchanich JM, Marsh GM. Mortality among hardmetal production workers: a retrospective cohort study in the Austrian hardmetal industry. J Occup Env Med 2017; 59:e282–e287.
10. Westberg H, Bryngelsson I-L, Marsh G, et al. Mortality among hardmetal production workers: the Swedish cohort. J Occup Env Med 2017; 59:e263–e274.
11. Buchanich JM, Dolan DG, Marsh GM, Madrigano J. Underascertainment of deaths using social security records: a recommended solution to a little-known problem. Am J Epidemiol 2005; 162:193–194.
12. Cohen A. Truncated and Censored Samples: Theory and Applications. 1991; New York, NY: Marcel Dekker, Inc, 96–105.
13. Youk AO, Marsh GM, Stone RA, Buchanich JM, Smith TJ. Historical cohort study of US man-made vitreous fiber production workers: III. Analysis of exposure-weighted measures of respirable fibers and formaldehyde in nested case-control study of respiratory system cancer. JOEM 2001; 43:767–778.
14. Hauptmann M, Lubin JH, Stewart PA, Hayes RB, Blair A. Mortality from solid cancers among workers in formaldehyde industries. Am J Epidemiol 2004; 159:1117–1130.
15. American Conference of Governmental Hygienists. Threshold Limit Values for Chemical Substances and Physical Agents and Biological Exposure Indices. Cincinnati, OH: American Conference of Governmental Industrial Hygienists; 1986–2016.
16. American Conference of Governmental Hygienists. Documentation of the Threshold Limit Values and Biological Exposure Indices. 6th ed.Cincinnati, OH: American Conference of Governmental Industrial Hygienists; 1991.
17. Marsh GM, Youk AO, Stone RA, Sefcik S, Alcorn C. OCMAP-PLUS: a program for the comprehensive analysis of occupational cohort data. J Occup Environ Med 1998; 40:351–362.
18. Gilbert ES. Some confounding factors in the study of mortality and occupational exposures. Am J Epidemiol 1982; 116:177–188.
19. Kolstad HA, Olsen J. Why do short term workers have high mortality? Am J Epidemiol 1999; 149:347–352.
20. Breslow NE, Day NE. The Design and Analysis of Cohort Studies. Statistical Methods in Cancer Research. Lyon, France: Oxford University Press; 1987.
21. Cox DR. Regression models and life-tables. J R Stat Soc B (Methodological) 1972; 34:187–220.
22. Cox DR. Partial likelihood. Biometrika 1975; 62:269–276.
23. StataCorp. Stata Statistical Software: Release 12. College Station, TX: StataCorp LP; 2011.
24. Galea S, Tracy M. Participation rates in epidemiologic studies. Ann Epidemiol 2007; 17:643–653.
25. Keeble C, Law G, Barber S, Baxter P. Choosing a method to reduce selection bias: Researchers. Open J Epidemiol 2015; 5:155–162.
26. Bauman A, Phongsavan P, Cowle A, et al. Maximising follow-up participation rates in a large scale 45 and up study in Australia. Emerg Themes Epidemiol 2016; 13:1–7.
27. U.S. Department of Health and Human Services. The Health Consequences of Smoking: 50 Years of Progress. A Report of the Surgeon General. 2014; Altlanta, GA: U.S. Department of Health and Human Services, Centers for Disease Control and Prevention, National Center for Chronic Disease Prevention and Health Promotion, Office on Smoking and Health, 17.
28. Miettinen OS. Components of the crude risk ratio. Am J Epidemiol 1972; 96:168–172.
29. Axelson O, Steenland K. Indirect methods of assessing the effects of tobacco use in occupational studies. Am J Ind Med 1988; 13:105–118.
30. Steenland K, Greenland S. Monte Carlo sensitivity analysis and Bayesian analysis of smoking as an unmeasured confounder in a study of silica and lung cancer. Am J Epidemiol 2004; 160:384–392.
31. US Department of Health and Human Services. Behavioral Risk Factor Surveillance System, Survey Data. Rockville, MD: National Center for Chronic Disease Prevention and Health Promotion, Center for Disease Control and Prevention; 1995.
32. Marsh GM, Buchanich JM, Youk AO. Historical cohort study of US man-made vitreous fiber production workers: VI. Respiratory system cancer standardized mortality ratios adjusted for the confounding effect of cigarette smoking. J Occup Environ Med 2001; 43:803–808.
33. Checkoway H, Pearce N, Kriebel D. Research Methods in Occupational Epidemiology. New York: Oxford University Press; 2004.
34. Chen W, Liu Y, Wang H, et al. Long-term exposure to silica dust and risk of total and cause-specific mortality in Chinese workers: a cohort study. PLoS Med 2012; 9:e1001206.
35. Grimsrud TK, Berge SR, Haldorsen T, Andersen A. Exposure to different forms of nickel and risk of lung cancer. Am J Epidemiol 2002; 156:1123–1132.
36. Grimsrud TK, Berge SR, Haldorsen T, Andersen A. Can lung cancer risk among nickel refinery workers be explained by occupational exposures other than nickel? Epidemiology 2005; 16:146–154.
37. Grimsrud TK, Berge SR, Martinsen JI, Andersen A. Lung cancer incidence among Norwegian nickel-refinery workers 1953–2000. J Environ Monit 2003; 5:190–197.
38. Grimsrud TK, Berge SR, Resmann F, Norseth T, Andersen A. Assessment of historical exposures in a nickel refinery in Norway. Scand J Work Environ Health 2000; 26:338–345.
39. Grimsrud TK, Peto J. Persisting risk of nickel related lung cancer and nasal cancer among Clydach refiners. Occup Environ Med 2006; 63:365–366.
40. Marsh GM, Buchanich JM, Zimmerman S, et al. Mortality among hardmetal production workers: US cohort and nested case-control studies. J Occup Env Med 2017; 59:e306–e326.
Copyright © 2017 by the American College of Occupational and Environmental Medicine