Lung cancer mortality among workers in the European production of man-made mineral fibers--a Poisson regression analysis.

P, ZOCCHETTI C. Lung cancer mortality among workers in the Euro pean production of man-made mineral fibers - a Poisson regressionanalysis. Scand J Work Environ Health 1992;18:279-86. One hundred and eighty-onelung cancer deaths among workers during 301 085 person-yearsin European man-made mineral fiber production between 1930-1955 and 1982wereana lyzedaccording to Poisson regressionmodelsincluding age, calendar period, country, and exposurevari ables. Timesincefirst employmentwasthe variablemost stronglyassociated with lungcancerrisk in both the rock-slag wool and glass wool subcohorts. Workers in the early technological phase were at higher risk than those in other categories, particularly in rock-slag wool production. No clear trend with dura tion of employmentwas suggested. No major changes occurred in the interpretation of the results when workers with lessthan one year of employment or lessthan 20 years since first exposure wereexcluded. The original results, based on analysesfor standardized mortality ratios, were confirmed, and workers with a short duration of employmentor a short timesince first employment did not need to be excluded from the analysis.

Reprint requests to: Dr P Boffetta, Unit of Analytical Epidemiology, International Agency for Research on Cancer, 150 cours Albert-Thomas, F-69372 Lyon cedex 08, France.
culated on the basis of national reference rates with the use of regional correction factors. The SMR analysis showed an excess mortality from lung cancer (189 observed and 151.2 expected deaths). There was an increase in lung cancer risk with time since first employment but not with duration of employment. The excess of lung cancer was concentrated among rock-slag wool workers, but not glass wool workers, employed in the " early technological phase" (12).
Among the problems arising in the interpretation of these results was the difficulty in disentangling the effect of the different exposure variables, given the small number of deaths in the groups with the highest expo sure. Moreover, analyses of trend on the basis of SMR values involve a lack of mutual compara bility between the ratios (13,14). An alternative approach to the calculation of SMR values is multivariate modeling on the basis of a Poisson regression, in which the number of deaths occurring in different cells of exposure variables and covariates such as age and calendar period are regarded as random Poisson variables and fitted to a multiplicative model (15). This approach allows the reciprocal confounding effect of exposure variables to be investigated and also allows the problem of comparability among SMR values based on subgroups of the cohort and between the cohort and the reference populations to be overcome. On the other hand, this approach is relatively unin-formative if the variability of exposure within the cohort is small. The present paper presents the results of a Poisson regression analysis based on the data of the last follow-up of the IARC MMMF study.

Subjects
The base of the study was the experience of workers ever employed (with at least one year of employment in England and Sweden) in the 13 factories enrolled in the study (table 1). Individuals were identified from factory records and were followed for mortality from the year work in production began (ranged between 1933 and 1950) to 1982-1983 (table 1). Cancer incidence data were collected for individuals from the countries covered by a national cancer register. However, they were not used in this analysis. The original cohort included 24 609 individuals. After women and office workers, which represented 17.4 and 10.70/0 of the overall cohort, respectively, were excluded, there were 18753 male production workers left, who provided 301 085 person-years of observation.
The factories were divided according to production process into rock-slag wool, glass wool, and continuous filament (table 1). Factory 14 changed type of production during the study period, and its population was divided into two subcohorts of workers employed respectively during the production of continuous filament only and of glass wool and continuous filament, the latter group being analyzed together with the remaining glass wool factories. Six individuals were employed in factory 14 as office workers during the period of glass wool production and moved subsequently to continuous filament production. They were excluded from this analysis.
The subjects were classified into the following occupational groups: (i) production and preproduction, (ii) secondary processes, (iii) maintenance occupations, (iv) other jobs, such as internal transport driver, storeman, (v) unspecified or mixed manual job, and (vi) unknown job. The 329 individuals employed in both manual and office jobs were excluded from this analysis.
Individuals contributed to the person-years between the beginning of employment and the end of followup or exit (death, emigration, or last date of employment, for subjects lost to follow up). This time period was also used to classify subjects according to time since first employment, divided into four levels (:59, 10-19,20-29, and~30 years). For the analysis by duration of employment, only employment in the aforementioned production processes was taken into account. There were 219 individuals who had at least one period of employment of unknown duration. They were excluded from the analysis by duration of employment, which was based on five levels ( < 1, 1-4, 5-9, 10-19, and~20 years).

280
All of the analyses were repeated after all of the subjects with less than one year of employment were excluded, as well as after the subjects with less than 20 years since first employment were excluded. A third approach was a lag-time analysis, in which the personyears and the mortality experience were referred to the exposure experience that took place 5 or 20 years earlier, respectively (16, pp 153-155).
The cohort was analyzed as a whole, and most of the analyses were repeated for the rock-slag wool and glass wool subcohorts separately. The continuous filament subcohort was too small to allow a separate analysis. The outcome of interest was death from cancer of the trachea, bronchus, and lung [referred to as lung cancer, International Classification of Diseases (ICD), 7th revision, 162-163, and ICD 8th and 9th revisions, 162] (17).

Estimation of exposure to man-made mineral fibers
The period of MMMF production in each factory was divided into an early, intermediate, and late technological phase (12). Airborne levels of MMMF were estimated to be highest when no dust-suppressing agents were used or a batch process involving labor intensive and hand-operated production methods was in operation. These aspects characterized the early technological phase. The late technological phase corresponded to the periods when oil and resin binders were in use with modern mechanized production methods. The periods in between were classified as intermediate phases. However, such a classification was not applied to two factories, and for five other factories only two phases were identified (12). The workers were allocated according to the phase within which their date of first exposure fell.

Statistical analysis
Relative risks (RR) and confidence intervals were estimated by fitting Poisson regression models to country-, age-, time-, and exposure-specific rates according to a multiplicative model (15). No external mortality rates were used. All of the regression models included age (usually on eight levels), calendar period (six levels), and country. Different combinations of the exposure variables were tested, along with the first-level interaction terms. The results presented in the tables were generally based on models including all of the relevant exposure variables and no interaction terms.
The tests for linear trend were performed by introducing a single exposure term to the model as a continuous variable with values corresponding to the mid point of each interval or to a unitary scale. The program PERSON-YEARS was used to estimate the individual contribution to each stratum (18), and the GUM (general linear interactive modeling) statistical package was used for the multivariate analysis (19).

Results
During the study period there were 181 lung cancer deaths . Nine of them occurred among the subjects excluded from the analysis for the aforementioned reasons, and 172 deaths were left for the analysis. There were 79 lung cancer deaths in both the rock-slag wool and the glass wool subcohorts, the remaining 14 deaths having occurred among continuous filament workers. When the analysis was restricted to individuals with at least one year of employment , the number of lung cancer deaths fell to 131. After the subjects with less Scand J Work En viron Health 1992, vol 18, no 5 than 20 years since first employment were excluded also, there were 61 lung cancer deaths left , of which 26 had occurred in the rock-slag wool subcohort and 35 in the glass wool subcohort. There were no continuous filament workers with more than 20 years since first employment. Table 2 shows the results according to time since first employment, duration of employment, technological phase, and type of produ ction for the rock-slag wool and glass wool cohorts combined. The results specific for type of produ ction are presented in table 3. Workers in continuous filament production were excluded from table 2 since they were  Albeit the relat ive risks of lung cancer increased in subsequent categori es of time since first employment , as shown in tables 2 and 3, the slop e of the regression 282 line fitted to a multiplicative model (ie, on the assumption of an exponential increa se in relative risk) did not reach the conventional level of statistical significance. Duration of employment was not associated with lung cancer risk when the other variables, and specifically time since first employment, were included in the model. The relati ve risks of lung cancer for durat ion of employment, not controlled for time since first employment (all individuals, all subcohorts), were 1.05 (95% CI 0.64-1.73),1.13 (95% CI 0.65-1.96),1.40 (95% CI 0.82-2.37), and 0.96 (95% CI 0.49-1.95) for 1-4, 5-9, 10-19, and :2: 20 years of employment , respectively.
The risk of lung cancer in the glass wool subcohort was 0.56 that of the rock-slag wool subcohort when the other variables were controlled (table 2). After .workers with less than 20 years since first employment were excluded, this difference was smaller. The relative risk of the continuous filament workers, as compared with the rock-slag wool workers, was 0.64 (95070 CI 0.24-1.73) for all durations of employment and 0.44 (95% CI 0.15-1.30) for one or more years of employment.
In the analysis for all individuals (table 3), workers first employed in the early technological phase in both the rock-slag wool and glass wool subcohorts showed a higher lung cancer risk than individuals employed later _ However, when the analysis was restricted to workers with at least 20 years since first employment (and at least one year of employment), only the rockslag wool subcohort showed higher lung cancer risks in the early phase, as compared with the late phase.
The reciprocal confounding effect of the exposure variables was investigated by fitting regression models including different combinations of covariates. There Scand J Work Environ Health 1992, vol 18, no 5 was no evidence of strong confounding effects, not even in the case of variables with some degree of colinearity, such as time since first employment and technological phase at first employment. Table 4 presents the results for time since first employment and duration of employment for the rockslag wool workers who were first employed in the early technological phase. These are the workers most likely to have been exposed to high levels of MMMF. The power of this analysis was strongly limited by the small number of lung cancer deaths (nine deaths overall and only six among the workers employed for~1 years), and the variables had to be categorized into few levels. However, there was a suggestion of an association between lung cancer risk and both time since first employment and duration of employment.
The results of the analysis by department of employment for the workers having at least 20 years since first employment are shown in figure 1. In most cases, the small number of lung cancer deaths caused impre-   cise risk estimates. No department was consistently at higher risk than the remaining ones. The relative risk for all other specified jobs (ie, secondary processes, maintenance, and other jobs) combined, as compared with primary production, was 0.70 (95070 CI 0.43-1.13) for all of the workers. The analysis by department was repeated including also workers with less than 20 years since first employment (

Discussion
Theoretically, interpreting trends based on SMR values could lead to erroneous conclusions, given the inherent lack of comparability of risk estimates obtained via indirect standardization with an external population (14,20). It is well recognized , however , that in most instances the difference in the structure of the populations from which the SMR values were calculated were not large enough to cause severe problems (16, pp 126-127). The results of the present multivariate analysis agree with this conclusion. In fact, the risk estimates which were derived from the Poisson regression and which were adjusted only for age, calendar period, and country (ie, the variables used in the indirect standardization) were remarkably similar to the ratios of the corresponding SMR values. As an example, figure 2 shows lung cancer risk estimates for time since first emplo yment. (The difference in the risk ratios in figure 2 and those shown in tables 2 and 3 are due to the inclusion of other covariates in the latter models.) All of the major results based on SMR values have been confirmed by the present multivariate analysis. A trend of lung cancer risk is shown for time since first employment, but not for duration of employment, and workers employed in the early technological phase and in the rock-slag wool sub cohort had a higher risk of lung cancer than workers in the remaining categories. The application of a 5-or 20-year lag time further added to the consistency of the results according to time since first employment. Moreover, this analysis allowed some insight into the relationship between the exposure variables, and the results suggest that time since first employment is the variable most strongly associated with lung cancer risk, but that an effect of employment during the early technological phase is still suggested after adjustment for time since first employment and duration of employment. The only interaction terms which significantly improved the goodness of fit were age-duration and time since first employment-duration, but their inclusion did not allow any material change in the interpretation of the results .
It has been argued that short -term workers should not be included in epidemiologic studies on occupational cancer since they are likely to be different from workers with longer employment with respect to social or behavioral factors possibly related to cancer risk, such as education or tobacco smoking (21). This criticism has been recently applied to studie s on the mortality of MMMF workers (20). When duration of employment was analyzed in this study, workers employed for less than one year in the rock-slag wool subcohort were at lower risk of lung cancer than workers employed for longer periods, whereas the opposite pattern was found for the glass wool cohort (table 3). All of the analyses shown in tables 2-5 have been repeated after workers employed for less than one year were excluded, and the results were not modified. As an example, the relative risks for lung cancer according to time since first employment were (for the categories 10-19, 20-29, and~30 years, respecti vely) 1.28, 1.54, and 2.33 for the rock-slag wool cohort and 1.35, 1.21, and 2.33 for the glass wool cohort. Therefore, it does not seem necessary to exclude short-term workers from the analys is of this population. Two further arguments strengthen this conclusion. Employment in MMMF production, especially in the past, did not require high technical skill, and therefore no major difference in social class may be expected between short-and long-term workers. In addition short-term workers might well have been employed in most dusty jobs, such as the removal of fibers from the blowing chamber when batch production was in use, and, therefore , they might represent a category of particularly high exposure to MMMF. Some problems, however, remain in the interpretation of the results. In particular, the possible confounding effect of lung carcinogens in the raw material used in the production of MMMF, such as arsenic in slag and asbestos, was not controlled in this analysis. However, previous analyses based on SMR values did not suggest that confounding by arsenic-containing slag, asbestos, or formaldehyde was likely to have played an important role (22). On the other hand, it is unlikely that confounders such as tobacco smoking or socioeconomic status were responsible for the differences in lung cancer risk found in this analysis, which was entirely based on internal comparisons.
Further developments of the IARC study include an updating of the follow-up, which will provide a larger contribution of person-years, in particular in the category of workers with the longest latency, and a pilot phase of a nested case-referent study of lung cancer in the rock-slag wool cohort, aimed to assess the feasibility and reliability of abstracting for selected individuals of the cohort, more detailed job descriptions, and the tracing of next-of-kin to obtain data on smoking and occupational exposures outside the MMMF production industry. Finally, the possibility of applying model-based quantitative estimates of MMMF exposure in the rock -slag wool plants (23) is being currently investigated.