The joint cohort analysis of male uranium workers confirms a statistically significant linear exposure–response relationship between low cumulative radon exposures and lung cancer mortality (ERR/WLM = 0.022; 95% CI 0.013–0.034) based on 408 lung cancer deaths and 394,236 person-years of follow-up from 1953 to 1999. The exposure–response relationship was statistically significant and consistent across the three cohorts.
An important result of this analysis is the apparent lack of an effect below 10 WLM cumulative radon exposures, where the risk estimates were around one. The results at this exposure range are imprecise, but may suggest a non-linear relationship. Although the upper confidence limits are compatible with an increased risk, predicted by a linear non-threshold model, the lower confidence limits are equally compatible with a reduced risk at low exposures model (i.e., hormesis). However, a conclusion of no effect or any effect is not possible because of very low statistical power at these exposures. Larger pooled studies would be useful to provide some answers.
Strengths and limitations
The main strengths of this study were the large sample size, the relatively good quality, measured radon exposure data during the study periods, and good long-term cohort mortality ascertainment.
The joint cohort’s sample size substantially increased the statistical power of the individual cohorts (Tomasek et al.
2008; Lane et al.
2010; Vacquier et al.
2011; Rage et al.
2012). Higher and more precise risk estimates were found in this joint analysis than in the previous analyses of low radon exposures that were based on earlier time periods and included estimated or extrapolated radon exposure estimates (National Research Council
1999).
Harmonizing the cohorts by time periods of radiation protection measures, high number and quality of ambient radon measurements, and individual monitoring through PADs substantially increased the likelihood that radon progeny exposure measurements were of high quality (Tomasek et al.
2008; Lane et al.
2010). However, measurement error may still exist (Stayner et al.
2007). Substantial reductions in radon exposures from means of > 20 to < 5 WLM/year occurred during the time periods under study, largely due to mechanical ventilation systems and regulatory exposure limits (Tomasek et al.
2008; Lane et al.
2010). The past time periods of low mean annual radon exposures reflect modern occupational exposures although the exposures were still higher than current mean annual exposures (< 0.25 WLM/year from 2005 to 2015) (National Dose Registry, special tabulations,
2016-11-15).
The joint analysis had long-term and high-quality mortality follow-up of uranium miners to 1999. The Czech and Beaverlodge cohorts had almost complete ascertainment of lung cancer mortality. Histology was available for at least 80% of lung cancers in the Czech cohort. Of those who died; only 0.4% of the Czech cohort had missing causes of death (Tomasek et al.
2008). Causes of death were obtained for ~ 89% in the Eldorado cohort (1950–1999) (Zablotska et al.
2013). This percentage was likely greater for Beaverlodge workers employed after 1965 due to improvements in the quality of the national mortality database over time and the use of the SIN which substantially improved record linkage (CNSC
2004). Although the French national mortality database did not exist before 1968; only ~ 3% of overall miners had missing causes of death. Correcting for missing causes of death did not have a large impact on the exposure–response relationship, because most French workers were still alive in 1968. Thus, the number of lung cancer deaths before 1968 was small (Tomasek et al.
2008; Laurier et al.
2004).
The main limitation of this study was that only grouped person-year data, not individual data, were available for this analysis. Thus, analytic decisions such as choice of categorical variable cutoffs were limited. For time-varying factors, workers contributed to the appropriate category as time progressed. All workers, regardless of their final cumulative exposures, contributed person-years to the exposure data set. However, deaths would only reflect those of workers with < 100 WLM (or < 50 WLM, ≤ 5.0 WL), since any worker with a greater exposure would have died at a higher exposure level. Alternative analytic methods would have been possible if individual-level data were available. In general, grouped and ungrouped data provide equivalent results when modelled identically (Richardson et al.
2004; Loomis et al.
2005).
Had individual-level data been available, only those individuals with lifetime exposure < 100 WLM would have been analyzed. However, because only grouped data were available, the decision to exclude person-year strata at > 100 WLM from the primary analysis could have introduced a bias through censoring follow-up. Thus, individuals’ earlier and lower exposures might have causally affected the risk of lung cancer death, despite that they accumulated more exposure and eventually developed and died from lung cancer at a later time. However, an alternative approach of not excluding person-year strata at > 100 WLM and instead only reporting excess relative risks estimates in the range of 0–100 WLM would have potentially allowed bias from measurement error at higher levels of exposure, which were predominately accumulated in the earlier time periods. In fact, the excess relative risk was estimated when strata at < 100 WLM (Table
5, ERR/WLM = 0.022; 95% CI 0.013–0.034) was similar to that estimated without making this exclusion (ERR/WLM = 0.020; 95% CI 0.015–0.027).
We addressed potential healthy worker survivor bias (Buckley et al.
2015; Bjor et al.
2015; Picciotto and Hertz-Picciotto
2015) several ways. Radon exposures were lagged by 5 years to address any changes in a worker’s exposure due to lung cancer. Short-term workers were excluded, since they have higher mortality rates than long-term workers unrelated to occupational radon exposures (Buckley et al.
2015; Bjor et al.
2015). Although short-term Beaverlodge workers (< 6 months) did not have higher crude mortality rates, they had unique characteristics: higher proportion of person-years with age at risk 19–29 years (22.7% versus 4.8%) and cumulative exposure < 5 WLM (87.9% versus 45.3%) compared to long-term workers (≥ 6 months). For these reasons, we excluded them to be consistent with the Czech and French cohorts. Workers with less strenuous jobs (open pit miners, mill workers), independent of duration of work, were included, since we found no evidence that they were less healthy than underground miners (Lane et al.
2010; Vacquier et al.
2009). Workers with 0.0 WLM were included as the reference population, since they were not systematically different for those with higher exposures and all job types had some workers with 0.0 WLM.
The start of follow-up began after either 1 or 4 years of employment in the Czech and French cohorts. Unfortunately, the exclusion of miners employed for < 6 months (74 deaths, 97,617 person-years) introduced immortal time bias in the Beaverlodge cohort, because it was not possible to exclude the first 6 months of at risk person-time from person-year tabulations of Beaverlodge workers employed for ≥ 6 months (195 deaths, 150,964 person-years). Start of follow-up began the first day of employment rather than after the first 6 months had passed. This overestimated the person-years of Beaverlodge workers by 6 months and, therefore, could have resulted in underestimation of the radon–lung cancer relationship. However, the potential healthy worker survivor bias of short-term workers likely would have had a more important impact than the immortal time bias. Workers employed < 6 months had a skewed person-year distribution of age at risk 16–29 years and low cumulative radon exposures (mean 9.95 WLM). This would have affected the harmonization of the Beaverlodge cohort with the Czech and French cohorts.
Only miners who were “incident hires”, who were first employed on or after 1953 or 1956, respectively, were included in the Czech and French cohorts. However, “prevalent” hires were introduced into the Beaverlodge cohort, since the use of person-year tables meant that workers who were first employed before 1965 were still retained, but their person-time contributions before 1965 were excluded. About 13% (37,265 person-years) of total person-time occurred from 1950 to 1964.
Sensitivity analyses of the Beaverlodge cohort suggested that risk estimates may be slightly overestimated, but there was not a big influence of prevalent hires on the exposure–response relationship. The impact of prevalent and incident hires on the potential healthy worker survivor bias, based on the workers’ date of hire before or after the start of follow-up, has been assessed in two recent occupational cohort studies (Costello et al.
2011; Applebaum et al.
2007). Despite the loss of statistical power and a restricted exposure range, decreasing the relative proportion of prevalent to incident hires reduced healthy worker bias, resulting in stronger evidence for a dose–response between occupational exposures and cancer mortality.
We used two different sensitivity analyses to assess the magnitude of the confounding effect of tobacco smoking. The first sensitivity analysis varied miners’ presumed smoking prevalence and the strength of the smoking–lung cancer mortality relationship to assess the sensitivity of the results under different plausible tobacco smoking scenarios. While our approach assumed that smoking-related risk estimates were generalizable to uranium miners, the meta-analysis (Gandini et al.
2008) included different populations and different smoking status definitions than those used in this study (Hunter et al.
2013). Smoking intensity may be a better measure than smoking status. Unfortunately, studies of radon-exposed miners (Villeneuve et al.
2007; Schubauer-Berigan et al.
2009) assessed the impact of smoking intensity only at > 100 WLM. We assumed that the effect of radon was constant across levels of smoking status; however, the joint effect of tobacco smoking and radon is between an additive and multiplicative interaction (National Research Council
1999; Kreuzer et al.
2018; Hunter et al.
2013; Leuraud et al.
2007,
2011; Schubauer-Berigan et al.
2009; Tomasek
2011). Nested case–control studies and the 1960 + sub-cohort of German uranium miners suggest that estimates of parameters in relative risk models are relatively close between estimates of parameters when smoking is adjusted for and when smoking is ignored. Likewise, these estimates correspond to the risk in smokers (majority of cases). The overall risk (in smokers + never smokers) is somewhat higher, because risk coefficients in never smokers are higher by a factor of ~ 2–3 (National Research Council
1999; Kreuzer et al.
2018).
The second sensitivity analysis likely better reflects the bias due to unmeasured smoking, since it was derived from nested case–control studies and a sub-cohort of uranium miners similar to those in our study (Kreuzer et al.
2018; Hunter et al.
2013; Leuraud et al.
2007,
2011; Schnelzer et al.
2010; Tomasek
2011,
2013; L’Abbé et al.
1991). The European joint nested case–control study at < 100 WLM (Hunter et al.
2013) and restricted to later time periods (Leuraud et al.
2011) are likely most reflective of our joint cohort analysis, since they include the French and Czech cohorts, low radon exposures and time periods of quality exposures. The larger size of the joint nested case–control study (Hunter et al.
2013; Leuraud et al.
2011) and the 1960 + sub-cohort of the German cohort study (Kreuzer et al.
2018), compared to the cohort-specific nested case–control studies (Schnelzer et al.
2010; Tomasek
2011,
2013; L’Abbé et al.
1991; Leuraud et al.
2007) provided more statistical power to assess the risk of lung cancer mortality at low radon exposures, adjusting for tobacco smoking. Nonetheless, measures of tobacco smoking status were crude in the sensitivity analysis, so had implications for residual confounding.
We assessed the impact of tobacco smoking on the radon–lung cancer relationship, because confounding can play a larger relative role when evaluating small effect sizes. Nonetheless, examples of substantial confounding are rare in studies of occupational exposures and lung cancer (Blair et al.
2007); tobacco-adjusted RRs are rarely appreciably different from unadjusted estimates. Our tobacco smoking sensitivity analyses support this outcome.
Gamma radiation, long-lived alpha radionuclides, residential radon, arsenic, silica, and diesel exhaust are human carcinogens (IARC
2012a,
2012b,
2013) and were reviewed as potential confounding factors for this study (Lane et al.
2010; Rage et al.
2012,
2015; Tomasek
2013; Walsh et al.
2010; Leuraud et al.
2011; Vacquier et al.
2011; Amabile et al.
2009). Many findings on other well-established human carcinogens indicate that confounding in occupational studies of lung cancer is rare and is not likely to be an explanation for positive study findings (Bruske-Hohlfeld et al.
2000; Lubin et al.
2000; Richiardi et al.
2005). If tobacco use does not confound lung cancer risks in occupational studies, it is even less likely that those more modest risk factors for lung cancer, with no known association with the occupational radon exposure of interest, would have a substantial effect (Blair et al.
2007).