Reanalysis of updated mortality among vinyl and polyvinyl chloride workers: Confirmation of historical evidence and new findings

Population under study

The present update of the follow-up included 1658 employees (42253 person-years) who had worked in a large petrochemical plant located in Porto Marghera (Venezia, Italy) for any period of time between July 1950 and July 1985. The follow-up period extended from January 1972 to December 1999; 248 deaths were registered. The classification and coding of previous studies were maintained, both for job category (28 plant sectors, used as surrogates of homogenous exposures) [1, 7] and for cause of death (ICD-9th revision [9]).

Workers were classified according to their job categories in four large, but reasonably homogeneous, groups: autoclave workers (n = 210, 5446 person-years), PVC baggers (n = 198, 4978 person-years), PVC compound workers (i.e., workers employed in the production of PVC granules from resins, n = 404, 10310 person-years) and technicians and clerks (n = 202, 5303 person-years). The above workers comprised 61% of the total workforce. An additional, more heterogeneous group, categorized as other blue collar workers (n = 644; 16216 person-years), included the plant employees in the remaining 24 sectors.

As technicians and clerks, thanks to environmental workplace sampling, were found in a previous study to be a group with low (or null) exposure to toxic and carcinogenic agents, they were used as an internal reference in the present analysis [10].

This reanalysis considered all causes of death previously reported as being associated to an increased risk [3, 4].

An analysis of changes in job category revealed that only 15 workers transferred from an exposed to an unexposed category, and that only nine moved from an unexposed to an exposed category. Overall, 81% of workers never changed their initial occupation. For this reason, the job category at hiring was used to split the workforce into different categories.

Statistical Analysis

Considering overall mortality among all exposed subjects (i.e., the total blue collar workforce), the study had an estimated power of 0.80 to detect a univariate relative risk (RR) of 1.50 as significantly increased at a 0.05 level (two-sided).

An internal comparison, using the above mentioned reference group, was carried out through a Poisson regression model for main causes of death.

Potential confounding factors considered in the analysis were age (five-year classes, 25–75+), calendar period of follow-up (1972–1974, 1975–1999 by five-year periods), age at hiring (2 levels: before age 25, 25+), year of hiring (before 1973, 1973 or later), duration of exposure (i.e., duration of employment; 3 levels: 0–4; 5–9; 10+), and latency (i.e., time since first hiring; 2 levels: 0–9; 10+). Data were grouped into cells according to the level of the explanatory variables described above.

The model used was:

ln(Ok) = ln(Ek) + α + β i X i + ε _k

where O _k are the observed deaths for a specific cause in the k ^th cell, and E _k are the expected deaths in the k ^th cell according to age, calendar year and cause specific mortality rates of the Italian male population (ln() are their natural logarithms); exp(α) is the Standardized Mortality Ratio (SMR) compared to the Italian population for persons falling in the baseline category (i.e., with all the variables set at the reference level); X _i is a matrix of indicator variables that specify the levels of the explanatory variables defining each cell; β _i is a vector of unknown parameters to be estimated from the data, such that exp(β _i ) represents the Relative Risk (RR) related to each explanatory variable with respect to its own reference level; ε _k is the random error. Each job category RR estimated by this model represents the ratio between the SMR of the exposed group and the SMR of the unexposed one for each specific cause of death. As suggested by Breslow and Day, the linkage of the method of indirect standardization to the multivariate regression analysis improves the efficiency of the statistical modeling [11].

The impact of the inclusion of each possible confounder in the regression models was tested through the likelihood-ratio test and by evaluating the change in the estimated RRs and their standard errors.

In order to avoid biased estimation of the standard errors of the parameters and to improve the goodness-of-fit of the selected model, when the ratio between the residual deviance (DV) and its degrees of freedom (DF) was far from 1 (i.e., in presence of over- or underdispersion), we refitted the model by introducing the following scale parameter:

DV _MAX /DF _MAX

where DV _MAX and DF _MAX were obtained by fitting a maximal model, i.e., a model with a high level of interactions, as suggested by Aitkin et al. [12].

Due to the low number of cases of liver and brain tumors and for liver cirrhosis, Bayesian estimates of the mean of the posterior distribution of RR (RR_pm) for these diseases were obtained by means of the Monte Carlo – Markov Chain method [13].

Confidence intervals at the arbitrary levels of both 95% (95% CI) and 95% Bayesian credibility interval (95% BCI) between the 2.5^th and 97.5^th percentile of the posterior distribution of RR, were estimated as appropriate. To assess the goodness-of-fit of the statistical models, standard diagnostic procedures were performed. Poisson models were fitted to the data using STATA statistical software [14].

where RR _e is the RR of the exposed (sub)group and d is the number of observed deaths.

Attributable number of deaths

The total number of deaths attributable to employment as a blue collar worker in a VC-PVC production plant was estimated to be 83.2 (out of 160) and 81.3 (out of 229) when the follow-up was updated to 1995 and 1999, respectively [23]. The consistency of these values is evidence of the reliability of results.

The longer follow-up did not lead to an increase in the total number of attributable deaths. Actually, while the RR was estimated to be 2.09 after the previous update (1972–1995), it was estimated to be 0.97 for the most recent 5-year period (1995–1999), suggesting that the effect of the exposure to noxious substances – or exposure itself – may have been greater in the past and lower in more recent years. On the other hand, as expected for diseases with a long latency (i.e., tumors), attributable deaths from all tumors increased from 28.1 to 34.9 when the duration of follow-up was increased.

Limitations, strengths and remarkable results

The present epidemiological study, like many other studies on occupational cohorts, lacked information about measures of individual occupational exposures over time during the employment period. Job category and length of employment were used as surrogate indicators of exposure type and cumulative exposure, respectively. Two studies on the same cohort already cited [19, 20] revealed a drawback in the cohort database stemming from a misclassification of both exposure and disease (information bias). In any case, such a misclassification could have led to an underestimation of the risk among exposed workers. A lack of fund prevented a more recent follow-up update, nevertheless the study had a power large enough to detect significant risks. Finally, it could be prudent to bear in mind that the study suffered from some limitation concerning the control of possible confounders, even if it unlikely could have lead to a significant modification of the results, as it will be made clearer afterward.

Despite these limitations, this analysis was, in our opinion, grounded on a reasonable approach and achieved some noteworthy results. First of all, the present reanalysis sought to reduce the comparison bias by using as a reference group – instead of the general population – a working population with limited or null exposure to occupational carcinogenic agents and with hiring criteria and duration of employment similar to those of the exposed group(s) under study. In addition, we tried to avoid another possible source of underestimation of the true risk, i.e. the inclusion of unexposed (or less exposed) workers in the exposed group (misclassification of exposure). In this way, the HWE and the related underestimation of the health risks should have been eliminated or at least controlled. Even if it was not possible to assess whether and how technicians and clerks might differ from the exposed workers other then the exposures of interest, the comparison seemed appropriate, and the socio-economic status of the exposed and unexposed workers should be fairly comparable. For this reason and as it was found to be at low (or null) exposure to toxic and carcinogenic agents [10], this group of workers (i.e., technicians and clerks) was the best available comparison group, while other blue collar workers included workers employed in different tasks not specialized and/or not confined in a specific workplace, such maintenance staff, cleaning operators, and other workers that handled, or were exposed to, most of the toxic substances listed in table 1. A recent study tried to quantify the possible amount of the bias due to the joint effect of the dilution effect (DE) and the comparison bias (CB) in occupational cohort studies. DE and CB respectively occur when exposed and unexposed workers in a cohort are pooled together and an internal referent group is not used. The study confirmed that "results from occupational cohort studies without internal referent groups may be unreliable" [24]. Then, internal comparison and separating blue collar workers into subgroups with homogenous exposure reduced the possible DE and resulting estimates were likely not biased by confounding due to smoking or to other factors. As suggested by "good epidemiological practice" guidelines [25], RR estimates obtained using an internal reference are more reliable than those based upon both national and local reference, and, as it was observed, "internal analyses of "dose-response" in cohort studies are unlikely to be seriously confounded by smoking habits" [26]. Within the same cohort analyzed in the present study, Mastrangelo et al. [19] found a lung cancer risk among baggers comparable to our estimates, but using a case-control approach, that by definition takes smoking (as other confounders) into account. Moreover, a clear dose-response relationship between length of exposure and lung cancer was detected among PVC bagger workers. In fact, not exposed workers and PVC bagger workers with a length of employment – used as a proxy of length of exposure – of 0–5, 6–10 and 10+ years, showed the following SMRs: 64.79, 108.35, 216.45 and 355.86, respectively. As this trend (p = 0.054) was observed among workers belonging to the same job category, it was likely unaffected by confounding effects, including those generated by smoking status. Finally, applying the Axelson approach to indirect adjustment for a confounder, it was clear that smoking unlikely played a major role when lung cancer RR was greater than 1.66 [27, 28].

The two procedures used in this study (i.e. internal comparison and separating blue collar workers into subgroups with homogenous exposure) revealed the pattern of diseases, likely due to exposures to specific occupational risk factors, typical of each subgroup of workers.