Lung cancer risk among workers in the construction industry: results from two case–control studies in Montreal

Data for the present study come from two case–control studies conducted in the greater Montreal area. Details of subject ascertainment, data collection and exposure assessment were previously described [26‐29].

The first study (Study I) was a population-based case–control study including 21 sites of cancers conducted between 1979 and 1986. Cases were histologically confirmed and were restricted to males aged between 35 and 70 years. Population controls were selected from electoral lists and frequency matched by age. The response rates among eligible subjects were 79 % among cases and 72 % among controls. From Study I, 857 male lung cancer cases and 533 male controls were available for these analyses.

The second study (Study II) was a population-based case–control study focused on lung cancer and included cases diagnosed between 1996 and 1998. Cases were histologically confirmed and included both males and females aged between 35 and 75 years. Controls were randomly sampled from population based electoral lists, stratified by sex and age to the distribution of cases. The response rates were 84 % among cases and 69 % among controls. From Study II, 736 cases and 894 controls were available for these analyses. Due to the small number of females in the construction industry, analyses were carried out among males only.

This research was approved by the ethics committees of the Institut national de recherche scientifique-IAF and McGill University. Ethics approval was also received from the following Montreal-area hospitals in which subjects were recruited: Montreal Chest Hospital Centre, Centre Hospitalier de Verdun, Cité de la santé, Hôpital Général Fleury, Hôtel Dieu de Montréal, Hôpital Jean-Talon, Hôpital Maisonneuve-Rosemont, Hôpital Notre Dame, Hôpital Sacré-Coeur, Hôpital St-Luc, Jewish General Hospital, Lakeshore Hospital, Montreal General Hospital, Queen Elizabeth Hospital, Reddy Memorial Hospital, Royal Victoria Hospital, Hôpital Santa Cabrini, Hôpital Ste Jeanne-d’Arc and St Mary’s Hospital. All participants provided informed consent.

A standardized questionnaire was administered by a trained interviewer to each subject or to a proxy respondent in case of death or if the subject was too ill to answer himself. The questionnaires were virtually identical in the two studies and entailed two parts: i) a structured section related to socio-demographic characteristics and lifestyle habits including detailed history of smoking; and ii) a semi-structured section related to employment history. For each job held for at least 6 months during the working lifetime history, information was solicited by means of probing questions on several characteristics such as company name, occupation title, general description of the work environment, main tasks performed, equipment and products used, use of protective equipment, and other information. To aid the interviewers to ask focused and technically appropriate questions, specialized questionnaires were developed for complicated occupations including brick and stone masons, concrete and terrazzo workers, carpenters, welders, braizers and solderers, plumbers and pipefitters, insulation workers, construction painters and cabinet makers [30, 31]. These were used whenever an interviewer was confronted with a subject who had been in one of these occupations.

Each recorded job was coded according to the Canadian Classification and Dictionary of Occupations of 1971 (CCDO) [32] and the Canadian Standard Industrial Classification (CSIC) of 1970 for study I [33] and 1980 for study II [34].

A team of chemists and industrial hygienists examined each completed questionnaire and translated each job into a list of potential exposures using a checklist of 294 agents that included several substances of interest, such as crystalline silica, asbestos, asphalt, and diesel emissions.

Approximately 25,000 jobs were evaluated in both studies combined. The team of experts spent about 40 person-years on this project, including helping to develop the methodology, monitoring the quality of the interviewing, conducting background research on exposures in different occupations, coding the individual participants’ files, and recoding after the initial complete round of coding was finished. The final exposures attributed to a subject were based on consensus among the experts, who were blinded to the subject’s case or control status. For each substance considered present in each job, the coders noted three dimensions of information, each on a three-point scale: degree of confidence that the exposure had actually occurred (possible, likely, definite), average frequency of exposure per workweek (<5 %, 5 to 30 %, >30 % time of the week), and relative intensity of concentration of the agent (low, medium, high). Concentration levels were established with reference to benchmark occupations in which exposure to the substance is present. We identified certain workplace situations a priori which correspond to low, medium and high exposure for each substance, and the experts rated each reported job against these benchmarks. Non-exposure was interpreted as exposure up to the level that can be found in the general environment. When subjects self-reported that they had been exposed to some agent, the information was noted but not accepted at face value, unless our experts agreed that such exposures occurred. The exposure assessment was based not only on the worker’s occupation and industry, but also on individual characteristics of the workplace and tasks reported; an illustrative example is in the Appendix of Parent et al. [35]. Periodic tests showed a satisfactory degree of reproducibility among our experts and between our experts and others [36‐38]. There was no evidence that cases provided more complete or more valid job histories than controls, as judged by the numbers of jobs reported per subject and by the interviewers’ rating of the quality of interviews. More detail can be found in Siemiatycki et al., 1991 [39].

First, we had to operationalize the designation of “construction worker” and we had to designate an appropriate comparison population. For ease of comparability and interpretation, a construction worker was defined as a subject who held at least one job in the construction industry (as it is referred to in the Canadian Standard Industrial Classifications of 1970 and 1980). In previous studies of construction workers (mostly cohort studies), the most commonly used comparison population had been the general population of the country in which the cohort was located. In designating a comparison population, there are several considerations. First, we would like the comparison population to be as similar as possible to the construction worker population, without the specific occupational conditions of construction workers. Second we would like to maximize the generalizability of the comparison. And third, we would like to maximize the statistical power of the comparison. These various desiderata lead to different choices and indeed we used two different strategies for a comparison population: i) the general population, using the entire case–control dataset (which enhances generalizability and power); and ii) non-construction blue-collar workers in the case–control dataset (which enhances comparability). The latter required an operational definition of “blue-collar worker”. We defined a blue-collar worker as one for whom over 50 % of the entire number of years of employment were spent in blue-collar occupations, as proposed by Ahrens et al. [40].

Odds ratios (ORs) were estimated by unconditional logistic regression models adjusted for age, ethnicity, socioeconomic status (SES) measured by the median family income for census tract of residence and by education level, and smoking history as synthesized by the comprehensive smoking index (CSI). The CSI integrates into a single index information on smoking duration, smoking intensity, and time since cessation and it has been shown to be useful for the purpose of adjusting for smoking in lung cancer studies [41]. We first performed analyses for each study separately, and then pooled the studies using models in which a binary variable for study was added to the covariates. To address the question of whether construction workers have an excess risk of lung cancer, subjects who held at least one occupation in the construction industry (as previously defined) were compared to those who never held an occupation in the sector. First we included all study subjects in an analysis, and second we restricted the study base to blue-collar workers in the study sample. These analyses were further stratified by large subdivision of the construction industry and by duration of employment in construction.

To examine which agents commonly found in the construction industry environment carry an excess risk of lung cancer, we restricted the study base to subjects who had ever worked in the construction industry. That is, for each agent, the reference category consisted of those construction workers not exposed to that agent.

We first established a list of agents to consider. Among the 294 agents in the database of agents coded by our exposure experts, we selected those that satisfied the following criteria: the lifetime prevalence of exposure to the agent among construction workers was more than twice that among non-construction workers, at least ten cases or ten controls were exposed among workers who had the Construction industry as the industry code, and the lifetime prevalence of exposure to the agent among controls was greater than 5 %. Twenty agents satisfied these criteria.

For each agent we derived an exposure index, based on the reliability of exposure, the frequency of exposure, the total duration of exposure and the relative concentration of exposure. The index has three categories: i) non-exposure; ii) non-substantial exposure; and iii) substantial exposure. Non-exposed is self-explanatory. Substantial exposure comprised subjects who had been exposed with a likely or definite reliability to medium or high concentrations for more than 5 % of their workweek and for at least 5 years [29]. Non-substantial exposure comprised those with less than substantial exposure.

Differences in ORs between the two studies were tested by introducing an interaction term between the binary indicator for study and the exposure variable of interest.

Because of the relatively high number of proxy respondents, we conducted a whole set of analyses, parallel to the main ones, but restricted to subjects who responded for themselves.

There is an ongoing debate as to whether it is appropriate to adjust for markers of socioeconomic status (SES) in occupational studies [42‐44], with some arguing that SES is a confounder to be adjusted and others that it is a collider to be omitted from statistical models. To examine whether inclusion of SES has the potential to bias the association between construction exposures and lung cancer, we conducted a sensitivity analysis in which we compared results from two models, one with adjustment for SES markers (income and education), and one without such adjustment. The other core covariates remained in both models.

To explore whether the selected cut-point of 50 % of working life in blue-collar occupations as a definition of a blue-collar worker influenced the results unduly, we also carried out entire sets of analyses using alternative cut-points, ranging from 25 to 75 %.

For the analyses of occupational chemical agents, we carried out the main set of analyses using as covariates various non-occupational covariates (age, socioeconomic status, education level, ethnicity and smoking history as synthesized by the CSI). In sensitivity analyses, we added to the models some other occupational exposure covariates that might confound the exposure-cancer associations. These covariates were chosen with two criteria in mind: i) classified by IARC as a Group 1 carcinogen for lung cancer [45] and ii) prevalence of exposure of at least 5 % in our study population. The agents thereby included were: asbestos, crystalline silica, and diesel engine emissions. The reason for not considering this as the main set of results is that it could obscure real associations if multiple agents, all measured with error, and probably with correlated error, are allowed to “soak up” some of the explanatory power of the agent being assessed. In this regard we prefer to avoid the possibility of over-adjustment, though we show the alternative results in Supplemental Digital Content Tables.

A complete set of parallel analyses was conducted among self-respondents, i.e., excluding proxy respondents. Additional file 1: Table S2 is analogous to Table 2, but restricted to self-respondents. As expected, compared with the results among all respondents, those restricted to self-respondents entail wider confidence intervals. There was some indication that ORs were higher for construction workers in the analyses restricted to self-respondents, but little indication that ORs for specific exposures differed between the analyses among all respondents and those among self-respondents.

Some analyses were restricted to blue-collar workers. In order to assess whether the OR estimates in Table 2 were sensitive to the particular cut-point of 50 % used to define blue-collar workers, we also implemented analogous analyses using cut-points ranging from 25 to 75 %. Additional file 1: Table S3 shows results analogous to those shown in Table 2, but using 25 and 75 % as cut-points, as well as 50 %. The OR estimate for ever employed in the construction industry among blue-collar workers varied from 1.16 to 1.13 as we varied the blue-collar career cut-point across the range from 25 to 75 %; in other words it had little impact.

Additional file 1: Table S4 is analogous to Table 3, but it contrasts the results using non-occupational confounder covariates with those using both non-occupational and occupational ones. Except for some slight variations, inclusion of the three occupational covariates did not materially change the ORs.

Additional file 1: Table S5 is analogous to Additional file 1: Table S4, but it is restricted to self-respondents. It shows that, apart from decreased precision of the estimates, the restriction to self-respondents has no important impact.

Additional file 1: Tables S6 and S7 are analogous to Tables 2 and 3 and show results of a sensitivity analysis in which we excluded and included two SES variables (median family income of the census tract and education level) from the statistical models. There was not really much impact of inclusion or exclusion of these covariates on the ORs between the various construction exposures and lung cancer. Hardly any of the OR pairings differed by more than one decimal point.

Table 4 shows the results of an analysis in which we explored potential effect modification by smoking for selected occupational agents. Namely, we selected those agents that, in Table 3, showed an OR ≥1.5 and sufficient numbers to support stratified analyses. Six agents were selected. Since there were very few never smokers among cases, the non-smokers category was supplemented with lifetime low intensity smokers. Operationally, we defined lifetime low intensity smokers as individuals having a CSI value below the 25th percentile on this scale. Smokers with CSI values above the 25th percentile were considered medium/heavy smokers. To evaluate if the ORs between occupational exposures and lung cancer differed between the two strata of smokers, we carried out an analysis based on all subjects including the two variables, smoking status and exposure to occupational chemical agent, by testing their cross-product term. The continuous CSI variable was maintained as a covariate in the models to avoid any residual confounding within the smoking status strata. There was no evidence of interaction in Table 4 except for results on silica. It appeared that there was a stronger effect of silica on lung cancer among non and low smokers than among medium-heavy smokers.