tox 5

Risk Analysis, Vol. 35, No. 4, 2015 DOI: 10.1111/risa.12315

Diesel Engine Exhaust and Lung Cancer Mortality: Time-Related Factors in Exposure and Risk

Suresh H. Moolgavkar,1,∗ Ellen T. Chang,2,3 Georg Luebeck,4 Edmund C. Lau,2 Heather N. Watson,5 Kenny S. Crump,6 Paolo Boffetta,7 and Roger McClellan8

To develop a quantitative exposure-response relationship between concentrations and dura- tions of inhaled diesel engine exhaust (DEE) and increases in lung cancer risks, we examined the role of temporal factors in modifying the estimated effects of exposure to DEE on lung cancer mortality and characterized risk by mine type in the Diesel Exhaust in Miners Study (DEMS) cohort, which followed 12,315 workers through December 1997. We analyzed the data using parametric functions based on concepts of multistage carcinogenesis to directly estimate the hazard functions associated with estimated exposure to a surrogate marker of DEE, respirable elemental carbon (REC). The REC-associated risk of lung cancer mortality in DEMS is driven by increased risk in only one of four mine types (limestone), with statisti- cally significant heterogeneity by mine type and no significant exposure-response relationship after removal of the limestone mine workers. Temporal factors, such as duration of exposure, play an important role in determining the risk of lung cancer mortality following exposure to REC, and the relative risk declines after exposure to REC stops. There is evidence of effect modification of risk by attained age. The modifying impact of temporal factors and effect modification by age should be addressed in any quantitative risk assessment (QRA) of DEE. Until there is a better understanding of why the risk appears to be confined to a single mine type, data from DEMS cannot reliably be used for QRA.

KEY WORDS: Diesel exhaust; lung cancer; multistage carcinogenesis parametric models; quantitative risk assessment

1Exponent Inc.–Health Sciences Practice, 15375 SE 30th Place, Suite 250, Bellevue, WA, USA.

2Exponent, Inc. – Health Sciences Practice, 149 Commonwealth Drive, Menlo Park, CA, USA.

3Division of Epidemiology, Department of Health Research and Policy, Stanford University School of Medicine, HRP Redwood Building, Stanford, CA, USA.

4Fred Hutchinson Cancer Research Center, Seattle, WA, USA. 5Exponent, Inc. – Data and Statistical Sciences Practice, 149 Com- monwealth Drive, Menlo Park, CA, USA.

6Independent Consultant, 2220 S. Vienna, Ruston, LA 71270, USA.

7Institute for Translational Epidemiology and the Tisch Cancer Institute, Icahn School of Medicine at Mount Sinai, One Gustave L. Levy Place, New York, NY, USA.

8Toxicology and Human Health Risk Analysis, 13701 Quaking Aspen Place NE, Albuquerque, NM, USA.

∗Address correspondence to Suresh Moolgavkar, Exponent Inc., 15375 SE 30th Place, Suite 250, Bellevue, WA 98007, USA; [email protected]. The copyright line for this article was changed on May 28, 2015 after original print publication.

1. INTRODUCTION

Concern that exposure to diesel engine exhaust (DEE) might cause adverse health effects developed soon after the technology was discovered and intro- duced more than a century ago. With rapid expansion of the use of diesel engines post-World War II, this concern intensified, with attention focused on the po- tential for prolonged exposure to DEE to cause lung cancer.(1,2) The approach to studying this issue has been multifaceted, including the conduct of epidemi- ological investigations, laboratory animal bioassays, and mechanistic studies using various models. Of these approaches, the epidemiological studies have received particular attention because the results are directly applicable to human populations without ex- trapolation across species or from in vitro settings.

Two influential papers describing the results of a cohort and a nested case-control analysis of the Diesel Exhaust in Miners Study (DEMS) recently

663 0272-4332/15/0100-0663$22.00/1 C© 2015 Society for Risk Analysis This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.

664 Moolgavkar et al.

reported a statistically significant increase in lung cancer mortality among both underground and sur- face miners exposed to DEE.(3,4) Based on these re- sults and others, the International Agency for Re- search on Cancer (IARC) concluded in June 2012 that there is sufficient evidence of carcinogenicity for DEE in humans and upgraded its classification of DEE from “probably carcinogenic” (Group 2A) to “carcinogenic to humans” (Group 1).(5)

No quantitative risk assessment (QRA) has been conducted after the publication of the results of DEMS, and the issue of the suitability of these results for QRA remains open. Prior to 2012, that is, before the DEMS results had been published, only one gov- ernment agency, the California Air Resources Board (CARB), had conducted a QRA for DEE.(6–10) It is noteworthy that three federal agencies, the Na- tional Institute of Occupational Safety and Health (NIOSH),(11) the Mine Safety and Health Admin- istration (MSHA),(12) and the U.S. Environmental Protection Agency (EPA),(13) evaluated the epidemi- ological literature available pre-2000 and deemed it inadequate for QRA. However, all three agencies is- sued guidance or regulations intended to limit expo- sure to DEE.

Despite regulations and technological advances in diesel technology, there remains substantial inter- est from government agencies and the private sector in the potential for conducting QRA for DEE from all technology diesel engines. QRA is of keen inter- est to multiple government agencies, including U.S. EPA, NIOSH, and MSHA, as well as the private sec- tor. This article analyzes the DEMS cohort data us- ing methods that complement the approach used by Attfield et al.(3) We hope that the results of our anal- yses will contribute to the understanding of the re- lationship between DEE and lung cancer in DEMS, and to the use of this important study for QRA for DEE.

Two related issues have recently received atten- tion in the epidemiologic literature. First, although summary measures of exposure, such as cumulative exposure, have been widely used in epidemiology and have led to important insights into risk factors for human disease, it is becoming increasingly clear that such summary measures of exposure cannot capture the impact of complex temporal patterns of exposure on disease risk.(14,15) The second, related issue has to do with the measure of effect. Recent papers(16,17)

have emphasized the limitations inherent in focusing on hazard ratios in epidemiologic studies. For cohort

data, both issues can be addressed simultaneously by using parametric methods to estimate hazard func- tions instead of relative hazards. Tammemagi et al.(18)

used a purely statistical approach to estimate hazard functions in cohort data. They used logistic regres- sion models with time-related factors being modeled by cubic splines modifying the impact of pack-years of smoking on the risk of lung cancer. A more bi- ologically motivated approach is based on concepts of multistage carcinogenesis.(19–21) These ideas have a long history.(22–27) In this approach, parametric hazard functions derived from multistage models of carcinogenesis that explicitly incorporate patterns of exposure are used to directly estimate the hazard functions. In an early application to DEE, Dawson and Alexeeff(10) used the multistage model of Ar- mitage and Doll(28) to estimate unit risks for DEE- associated lung cancer in a cohort of railroad work- ers. Statistical approaches that modify the effect of cumulative exposure on the odds ratio have been proposed for analyses of case-control data by Lu- bin et al.(29–31) and more recently by Vlaanderen et al.(32)

With detailed information on patterns of ex- posure and a relatively large data set, the DEMS provides an excellent opportunity to examine the impact of patterns of exposure to REC on lung can- cer risk. In this article, we extend the Attfield et al.(3)

analyses of the DEMS cohort data by exploring the impact of temporal factors, such as duration of ex- posure and time since cessation of exposure, on lung cancer mortality following exposure to DEE. To assure ourselves that we were analyzing identi- cal data sets, we first reproduced the Cox propor- tional hazards analyses reported in Attfield et al.,(3)

including the results in their Supplementary Materi- als. We then explored time-related factors in DEMS by using a version of the multistage clonal expansion model(33–41) to directly estimate hazard functions in- stead of hazard ratios.

2. METHODS

The data used in our analyses are those ac- quired by the DEMS team, including their estimates of the respirable elemental carbon (REC) exposure metrics. The DEMS effort was conducted by inves- tigators associated with NIOSH and the National Cancer Institute (NCI) and initiated in the early 1990s. The DEMS research is reported in five papers that describe the exposure assessment,(42–46) a paper

Diesel Engine Exhaust and Lung Cancer Mortality 665

describing a cohort study,(3) and a paper and let- ter describing a nested case-control study.(4,47) A key feature of DEMS, as noted by the DEMS team, is the attention given to constructing retrospective es- timates of workers’ exposure to REC as a primary surrogate indicator of exposure to DEE.

The DEMS cohort study, which has been de- scribed in detail by Attfield et al.,(3) included 12,315 workers from eight nonmetal mines and associated facilities, as summarized in Table I. The cohort in- cluded individuals who worked for at least one year in the mines or related facilities, excluding adminis- trative and managerial employees.

The eight mines and related facilities handled four types of ore: one limestone mine in Missouri, three potash mines in New Mexico, one salt (halite) mine in Ohio, and three trona mines in Wyoming. These nonmetal mines were all considered to pro- duce ore that was thought not to present a carcino- genic hazard on inhalation. The seven potash, salt, and trona mines all had generally similar operations. The ore in those mines was extracted from seams of various heights (on the order of 10–15 ft) at various depths below the surface with the ore transferred, typically by conveyers, to central vertical shaft loca- tions where it was lifted to the surface for further processing. The processing facilities varied in com- plexity at each mine, dependent on the intended use of the ore and purified and sized products. All seven of these mines had substantial mechanical ventila- tion supplying large quantities of air to minimize air- borne dust concentrations and, in the case of the trona mines, to minimize the build-up of methane, an explosive gas. The mining operations made substan- tial use of electric-powered equipment and selective use of diesel-powered equipment for various opera- tions, including underground transport of miners.

The limestone mining operation was quite dif- ferent from that in the seven other mines. It involved mining of limestone on a single plane with ore hauled by large diesel-powered haul units to surface openings at approximately the same elevation as the underground mining operations. The areas where ore was mined were high-ceiling caverns (approximately 40 ft in height) with intermittent support columns of unmined ore. The limestone mine primarily used natural ventilation, with air flowing up or down vertical shafts between the surface and the mining operations.

Demographic and work history information was abstracted from facility personnel records by DEMS study investigators. Mortality follow-up was conducted from the year of introduction of

diesel-powered equipment in each facility (1947– 1967) through December 31, 1997, using the National Death Index and the Social Security Administration death files.

The analyses presented in this article are lim- ited to the data, including exposure information, pro- vided by NIOSH under a data-use agreement. We had no access to the original records underlying the data, nor could we use alternative measures of expo- sure. Each miner’s time-varying historical exposure to DEE was quantitatively estimated as REC based on work histories. Fig. 1 shows histograms for the dis- tributions of estimated cumulative REC exposure in the various mine types. It is important to recognize that these histograms are for all workers, both on the surface and underground. The cohort of limestone mine workers clearly had the lowest estimated cumu- lative exposures. Table 2 from Attfield et al.(3) shows that the estimated REC concentration was also low- est for the workers (surface and underground) in the limestone mining operation.

2.1. DEMS Data

In response to our request for the data set for the DEMS internal cohort analysis, NIOSH provided three de-identified data files with demographic, oc- cupational, and death outcome data, which we com- bined into a single data set. The death outcome file identified miners who died of lung cancer or esophageal cancer as an underlying or contributing cause, but did not provide specific cause-of-death codes for any other deceased miners in the cohort. By the end of the mortality follow-up period, 2,185 deaths were ascertained, including 200 deaths from lung cancer as the underlying cause and an additional 12 deaths from lung cancer as a contributing cause. Cigarette smoking histories were not available for analysis of the cohort data. In addition, at the time our analyses were conducted, we had not been pro- vided accurate radon exposure data on the mines and workers.

To accommodate the time-varying nature of the exposure, we constructed a data set with monthly av- erage REC intensity and cumulative REC starting from the date of first exposure until the date of last known follow-up or death, with exposure estimates indexed at the end of each month. Lagged exposure on a given date was calculated by retrieving the ex- posure during the same month a set number of years earlier. If no work record existed at that time, then the lagged exposure was set to zero. Age attained at the beginning and end of each month was used

666 Moolgavkar et al.

Table I. Number of Miners and Lung Cancer Deaths by Worker Location and Mine Type in the DEMS Cohort

Ever-Underground Workers

Underground- Surface and Complete Surface-Only Workers Only Workers Underground Workers Cohort

Mine Type Miners Deaths Miners Deaths Miners Deaths Miners Deaths Limestone 730 15 123 12 823 10 1,676 37 Potash 1,293 38 1,951 46 1,327 18 4,571 102 Salt 50 <5 208 9 289 <5 547 <19 Trona 1,935 23 1,798 15 1,788 11 5,521 49

Entire Cohort 4,008 <81 4,080 82 4,227 <44 12,315 200

Fig. 1. Distribution of cumulative respirable elemental carbon (REC) exposures in the various mine types. Note that the cumulative REC exposure is shown on a logarithmic scale.

as the primary time scale. An event was recorded if a worker died from lung cancer during a particular monthly period. Other causes of death and loss to follow-up were represented as right-censoring times.

2.2. Standardized Mortality Ratio (SMR) Analyses and Cox Proportional Hazards Regression

We replicated the results in Attfield et al.(3) and performed additional analyses as described in the

Supporting Information. In particular, we performed extensive new analyses to explore the shape of the exposure-response relationship and the role of tem- poral factors in determining risk.

2.3. Multistage Carcinogenesis Parametric Models

Dawson and Alexeeff(10) estimated unit risks for DEE-associated lung cancer in a cohort of railroad workers using the multistage model of Armitage and

Diesel Engine Exhaust and Lung Cancer Mortality 667

Doll.(28) In this article, we used two versions of mul- tistage clonal expansion models, the two-stage clonal expansion model (2SCE) and the three-stage clonal expansion model (3SCE), to analyze the DEMS co- hort data. Note that because we set the first two “mutation” rates equal in the 3SCE model, we esti- mate the same number of parameters in the 2SCE and 3SCE models. Both models yielded very sim- ilar results, but the 3SCE had a higher likelihood model. Therefore, we present only the results using the 3SCE model. The two main advantages of us- ing multistage models are direct estimation of hazard functions rather than relative hazards, and explicit consideration of temporal patterns of exposure, such as ages at exposure initiation and cessation and time- dependent patterns of exposure.

A pictorial depiction of the main features of this model is shown in Fig. 2. Details can be found in the published literature.(21,33–36) The parameters of the model are treated as functions of the ex- posures of interest, in this case the concentration of REC. The hazard function, the probability of lung cancer by age t, denoted P(t), and the sur- vival function, S(t) = 1 − P(t), are therefore also functions of REC concentration. In terms of these quantities, the likelihood function is given as fol- lows. Assuming independence between individ- ual response probabilities, the cohort likelihood is the product of individual likelihoods over all subjects j, L = ∏ L j . Individual likelihoods L j = L j (s j , t j , (· · ·)) depend on age at entry into the study sj, censoring or failure age tj, and detailed expo- sure histories in conjunction with general exposure- response models for the biological parameters in the TSCE model, and possibly on a lag time or distribu- tion of lag times.

The individual likelihoods for cases and sur- vivors, accounting for left-truncation, are given by:

L j (t j , s j ) = ⎧ ⎨ ⎩

P′(t j )/S(s j ) if diagnosed with cancer,

S(t j )/S(s j ) otherwise. (1)

Furthermore, let hm(u) represent the individual TSCE model hazard and Sm(u) represent the TSCE model survival at some age u. For a fixed lag time from the first malignant cell to cancer death, tlag, indi- vidual likelihoods in Equation (1) are calculated us- ing S(tj) = Sm(tj − tlag); S(sj) = Sm(sj − tlag); P′(tj) = hm(tj − tlag)Sm(tj − tlag). Here we used lag = 0 or lag = 15, in accordance with the approach used by Attfield et al.(3) Left-truncation requires calculation

of the survival probability, S(sj), at entrance into the study.

Identifiability of parameters of multistage clonal expansion models is discussed in Heidenreich et al.(48) and Little et al.(33) For the analyses here, the identifiable parameters are “slope” = μ0μ1(g/α); “promotion” = g = α − β − μ2; and μ2, which is a mutation rate and therefore very small, on the order of 10−7; therefore, g � α − β. The interpretation of these parameters in terms of the shape of the cancer incidence curve is described in recent papers.(34,37)

In addition, a dose-response parameter for exposure to DEE is estimated, as described later. Had smok- ing and other covariate information been available, it could have been included in the analyses by mod- eling its effects on these parameters. Explicitly, each of the identifiable parameters can be made a time- dependent function of the estimated concentrations of REC. We found that the REC intensity impacted only the promotion parameter, g. We used the fol- lowing dose-response relationship for g through its dependence on the cell division rate α:

α(REC) =a0 × [1 + log(1 + a1 xREC)], where a0 is the background rate of initiated cell division and REC is the age-dependent concentra- tion of REC in μg/m3. Confidence intervals for the parameters are based on the Wald statistic and also on Markov chain Monte Carlo (MCMC) meth- ods. These two approaches yielded similar results; however, we believe that the MCMC results are more reliable. We used the Metropolis-Hastings algorithm for our MCMC samples. We ran eight independent chains, each for several thousand cycles. We con- cluded that the cycles had converged when there was little evidence of drift in means and variances of the parameter estimates. We conducted these analyses for the entire cohort and also separately for each of the mine types. We used likelihood-based methods to test for equality of exposure-response parameters across mine types.

3. RESULTS

We successfully replicated the analyses de- scribed by Attfield et al.,(3) yielding results that were virtually identical to those reported in the main ar- ticle and Supplementary Materials (data not shown; results of these and other analyses described herein are available upon request if not shown). In doing so, we confirmed that the data set that we analyzed was essentially the same as that used by Attfield et al.(3)

668 Moolgavkar et al.

Fig. 2. Three-stage clonal expansion model. The model assumes that initiation requires the biallelic inactivation of a (tumor sup- pressor) gene. The parameters are X, the number of susceptible stem cells; μ0, the first mutation rate; μ1, the second muta- tion rate; α, the division rate of initiated cells; β, the death rate of initiated cells; and μ2, the rate of malignant conversion. The time between the first malignant cell and death from tumor is modeled as a constant lag (here, 0 or 15 years, following Attfield et al.3).

Table II. Log-Likelihoods for Fits of the Three-Stage Clonal Expansion Model to the Entire Cohort and to Each Mine Type Separately

Null Max Log-Likelihood Dose-Response Max Log-Likelihood 2 × LLR

Mine Type Limestone −258.6 −254.7 7.8; p = 0.005 Potash −700.2 −700.2 0 Salt −79.9 −79.9 0 Trona −370.7 −369.2 3.0; p = 0.08

Entire Cohort −1,418.0 −1,415.4 5.2; p = 0.02

Note: Maximum log-likelihoods are reported for both the null model (no dose-response parameter for respirable elemental carbon) and the model with dose response on promotion (see text for details). LLR = log of the likelihood ratio. 2 × LLR > 3.84 is significant at the 0.05 level. Significant results are shown in bold font (limestone mine and the entire cohort).

The results of fitting the 3SCE model to the over- all DEMS cohort and to individual mine types are shown in Tables II and III and in Figs. 3–6. Table II shows that the impact of REC is statistically signifi- cant in the limestone mine and in the entire cohort. Table III shows the parameter estimates together with the Wald 95% confidence intervals and the 95% credibility intervals based on 2,500 MCMC cycles. Fig. 3 indicates that the 3SCE model describes well the observed hazard function for lung cancer mortal- ity in the DEMS cohort. Fig. 4 shows that the hazard function for lung cancer mortality after exposure to REC stops approaches the hazard function among the unexposed, as is the case for the hazard func- tion among ex-smokers. Fig. 5 shows that the relative risk (RR) associated with a given cumulative expo- sure depends upon how the exposure was accumu- lated, as is the case for smoking and lung cancer.(14,15)

Fig. 6 shows the RR as a function of age for contin- uous exposure to REC starting at birth. The strong effect modification of the RR by age is of particu- lar note. This figure shows also that, for the same ex- posure concentration, the RR in the limestone mine is considerably higher than in the combined cohort, with the excess relative risk (ERR) associated with a 1-μg/m3 exposure to REC starting at birth being about five-fold higher in the limestone mine than in the entire cohort at age 45. In these analyses, the RR and therefore the ERR are age-dependent for a given continuous exposure concentration. The RR of cigarette smoking on lung cancer risk is similarly modified by age.(15,20,49)

Results were similar when the model was fit to ever-underground miners, whereas among surface- only workers, REC exposure was not associated with lung cancer mortality (data not shown). A model

Diesel Engine Exhaust and Lung Cancer Mortality 669

Fig. 3. Fit of the three-stage clonal expan- sion model to the observed hazard rates from the entire DEMS cohort. The observed haz- ard rates were estimated by dividing the ob- served number of lung cancer deaths in five- year age groups by the total person-years of risk in the age group. The predicted haz- ard function was calculated by summing up the predicted hazard functions for each in- dividual in the data at any given age. The graph shows three distinct fits. First, the null model assumes that respirable elemental car- bon (REC) has no impact on risk of lung can- cer mortality. The hazard function is shown by the black solid line. The dashed red line shows the hazard function for the model that incorporates a dose-response (DR) for REC (see text for details). The red solid line shows the hazard function from this model, but with the dose-response parameter (a1) set to 0. Although introduction of a dose- response parameter significantly improves the fit as judged by the likelihood (p = 0.02; Table II), this figure shows that the hazard functions for the model with a dose-response parameter and the null model are virtually identical.

Table III. Results of Fitting the Three-Stage Clonal Expansion Model with Dose-Response on Promotion to the Entire DEMS Cohort and to the Cohort of Workers Employed in the Limestone Mine

Mine MLE (95% MCMC Median Type Parameter Wald-Based CI) (MCMC 95% CI)

Entire cohort Max log-likelihood = −1,415.4

g0,u 0.254 (0.163, 0.374) 0.235 (0.166, 0.356) g0,s 0.297 (0.183, 0.443) 0.275 (0.192, 0.427)

a1 × 104 0.556 (0.020, 1.092) 0.544 (0.112, 1.091) Slope × 103 0.183 (0.127, 0.263) 0.169 (0.130, 0.269)

μ2 × 107 4.926 (0.490, 35.28) 7.970 (0.583, 29.75) Limestone

Max log-likelihood = −254.7 g0,u 0.207 (0.100, 0.380) 0.191 (0.117, 0.297) g0,s 0.239 (0.097, 0.478) 0.231 (0.112, 0.387)

a1 × 104 0.226 (0.0, 0.451) 0.237 (0.050, 0.717) Slope × 103 0.186 (0.082, 0.422) 0.151 (0.068, 0.486)

μ2 × 107 6.55 (0.128, 79.28) 16.75 (1.39, 47.10)

Note: The parameters are defined as follows: g0,u, background underground promotion rate; g0,s, background surface promotion rate; g1, dose-response promotion parameter as defined in the text; slope, identifiable parameter as defined in the text. MLE, maximum likelihood estimate; MCMC, Markov chain Monte Carlo; CI, confidence interval.

with a 15-year lag fit substantially better than a model with no lag. In both the overall cohort and ever-underground miners, a small but statistically significant impact of REC was observed on the promotion of initiated cells, resulting in increased lung cancer mortality among miners exposed to REC. The background hazard in the absence of

exposure to REC was higher in surface-only than ever-underground workers, but worker location did not modify the relationship between REC exposure and lung cancer mortality.

Separate analyses by mine type show that REC is significantly associated with lung cancer mortality in only the limestone mine (Table II). We tested the

670 Moolgavkar et al.

Fig. 4. Age-specific lung cancer mortality curves generated by the three-stage clonal expansion model for the entire DEMS cohort. The solid curves show, respectively, the background age-specific mortality curve among those not exposed to respirable elemental carbon (REC), and the age-specific mortality curve for a 50-μg/m3 continuing exposure starting at age 20 years. The dashed curve shows the age-specific mortality curve if the exposure, that started at age 20 years stopped at age 40 years. Note that the mortality among those whose exposure stopped at age 40 years approaches that among those not exposed, just as in the case of cigarette smoking and lung cancer.

hypothesis that the exposure-response parameter for the entire cohort is driven by the limestone mine by removing the limestone mine and analyzing the other mine types together. The result was highly insignifi- cant; that is, the null hypothesis that the exposure- response parameter in the entire cohort minus limestone is zero cannot be rejected (likelihood ra- tio test for the hypothesis that the dose-response parameter in the entire cohort minus limestone is zero = 1.86 on 1 degree of freedom [df]; p = 0.17). We analyzed the entire cohort with a single exposure-response parameter and with two exposure-response parameters, one for the limestone mine and the second for the other mine types. The likelihood ratio test rejected the hypothesis that the two exposure-response parameters were equal (likelihood ratio test = 4.56 on 1 df; p = 0.03). Finally, in the model with two exposure-response parameters, we tested the hy- pothesis that the parameter for the non-limestone mine types was equal to 0. This null hypothesis was not rejected (likelihood ratio test = 3.0 on 1 df; p = 0.08).

As we report in the Supporting Information, the Cox proportional hazards regression analysis also demonstrated that the risk in the entire cohort is driven by the increased risk in the limestone mine.

4. DISCUSSION

Because it is now abundantly clear that that tem- poral patterns of exposure are important in deter- mining the risk of lung cancer,(14,15,19,21,36) and in light of the valuable opportunity to investigate this issue in the DEMS cohort data, we conducted extended analyses of this cohort using both time-dependent proportional hazards models and biologically-based multistage clonal expansion models.

We did not deal with the fundamental question of whether the back-extrapolated exposure estimates used in DEMS are reliable estimates of actual expo- sure to DEE, which has been questioned.(50–53) We could not use the smoking history data used in the case-control analysis of Silverman et al.(4) Therefore, the analyses reported here focused on other vital issues, namely, the temporal patterns of exposure- response relationships and the dependence of these relationships on specific mining facilities.

We achieved three principal goals in the analy- ses of these data. First, we successfully reproduced the results reported in Attfield et al.(3) Second, we explored temporal aspects of exposure and the evo- lution of lung cancer risk following exposure more carefully by focusing on hazard functions, rather than hazard ratios, using parametric models based on

Diesel Engine Exhaust and Lung Cancer Mortality 671

Fig. 5. Relative risk of lung cancer mortality as a function of age in the entire DEMS cohort for the identical cumulative exposure (50 μg/m3·year) accumulated in different ways starting at age 20 years: 5 μg/m3 over 10 years; 2.5 μg/m3 for 20 years; 1 μg/m3 for 50 years. It is quite clear that the temporal pattern of the relative risk depends strongly on how total dose is accumulated. All exposures were lagged 15 years.

concepts of multistage carcinogenesis. Finally, we re- peated these analyses for individual mine types, and performed a variety of sensitivity analyses and ex- tended analyses using alternative statistical models, as detailed in the Supporting Information.

The extended analyses using multistage models confirmed that time-related exposure patterns are of critical importance in the DEMS cohort data. Fig. 3 shows the fit of the 3SCE model to the entire DEMS cohort. This figure shows also that, although the im- pact of REC is statistically significant in this data set, the inclusion of REC makes only a small difference in the estimated hazard function. We could not con- struct a similar figure for the limestone mine because, with only 37 lung cancer deaths in that subgroup, we could not estimate the observed age-specific hazard function. Fig. 4 shows that after exposure to REC stops, the hazard function approaches that among the unexposed, similar to the pattern seen among ex- smokers after they quit. Fig. 5 shows that the impact of REC on lung cancer mortality depends on how ex- posure is accumulated, and does not remain constant after exposure stops. Fig. 6 shows age-related effect modification of the RR associated with continuous lifetime exposure of 1 μg/m3 both for the full DEMS cohort (Fig. 6a) and the limestone cohort (Fig. 6b). Clearly, these dependencies should be explicitly considered when using the DEMS cohort data for a

QRA. Finally, these analyses show that the risk in the entire cohort is driven by the limestone mine.

For regulatory purposes, there is considerable interest in developing a QRA for DEE. Hitherto, only CARB has estimated a unit risk for DEE, as mentioned above. The U.S. EPA initiated its prepa- ration of a health assessment for DEE soon after IARC concluded its first review in 1988. The EPA report went through five drafts that were subjected to external review before a final assessment was published.(13) A major contentious issue during the prolonged review process was whether the epidemi- ological evidence was sufficiently robust to use for a QRA. The final EPA assessment(13) concluded that DEE was a probable human carcinogen, but deemed the data insufficient for QRA. Thus, the DEMS data now have a pivotal role in any QRA for DEE, and it is therefore imperative that the association be- tween DEE and lung cancer in these data be care- fully and critically evaluated using multiple analytic techniques.

In light of the growing evidence that time-related factors are important in determining risk, they should not be ignored in the development of a unit risk. Ef- fect modification by age should be accounted for in any unit risk estimation using life tables. The results from the 3SCE parametric analyses suggest one ap- proach to the development of a unit risk for lung

672 Moolgavkar et al.

Fig. 6. Relative hazard of lung cancer mortality as a function of age generated by the three-stage clonal expansion model for continuous exposure to 1 μg/m3 respirable elemental carbon (REC) starting at birth. The red solid line represents the relative hazard generated by the maximum likelihood estimates (MLE) of the parameters. The blue solid line represents the point-wise median of the relative hazards generated by 2,500 Markov chain Monte Carlo (MCMC) cycles. The red dashed lines shown the point-wise 95% credible intervals for the relative hazards generated by the MCMC runs. Individual outputs of the MCMC runs are shown in gray. Note the strong effect modification of the relative hazard by age. (a) Lung cancer mortality in the entire DEMS cohort. (b) Lung cancer mortality in limestone mine workers.

cancer mortality in association with DEE exposure. An analysis that effectively addressed effect modifi- cation could yield a result that differs substantially from the Vermeulen et al.(54) meta-regression risk estimate. The development of a unit risk may be premature, however, given the many uncertainties surrounding both exposure and analysis issues. In

particular, we need a better understanding of why the risk in DEMS is driven by a single mine type. Only af- ter resolving these complex issues can the exposure- response relationship between DEE and lung cancer be quantified reliably.

Finally, in considering the results of the DEMS analyses as reported by Attfield et al.,(3) Silverman

Diesel Engine Exhaust and Lung Cancer Mortality 673

et al.,(4) and now in this article, it is important to rec- ognize the continuous improvements in diesel tech- nology that occurred from 1947, the earliest year of dieselization for any DEMS mine, to the present time.(2) Estimates of lung cancer risk associated with DEE from engines through the end of DEMS are probably inappropriate for evaluating risks asso- ciated with exposure to DEE from contemporary sources.

In summary, our results show that time-related factors have a crucially important influence on lung cancer mortality in the DEMS cohort data, and ig- noring them may yield misleading conclusions. Our analyses show that the RR of lung cancer mortality following DEE exposure is modified by age. Further- more, in both the Cox model (see Supporting Infor- mation) and the 3SCE analyses, a significant posi- tive association of REC with lung cancer mortality is based on a single mine type (limestone), with no ob- vious explanation for this pattern, given that both cu- mulative REC levels and average REC levels in that mine were substantially lower than in all other mines studied. Until important questions raised by these findings are adequately addressed, the DEMS data cannot reliably be used for QRA.

ACKNOWLEGMENTS

The analyses reported in the article are funded by a coalition of trade organizations working through the Truck and Engine Manufacturers Association (EMA). Along with EMA, the organizations that co- funded this work include the American Petroleum Institute (API), European Automobile Manufac- turers Association (ACEA), American Trucking Association (ATA), International Organization of Motor Vehicle Manufacturers (OICA), Alliance of Automobile Manufacturers (Alliance), European Research Group on Environment and Health in the Transport Sector (EUGT), Association of Equip- ment Manufacturers (AEM), Association of Amer- ican Railroads (AAR), and European Association of Internal Combustion Engine Manufacturers (EU- ROMOT).

The analyses of the Diesel Exhaust in Miners Study (DEMS) data reported in this article utilized data provided to the analytical team by the Na- tional Institute of Occupational Safety and Health (NIOSH). Only individual authors who had signed the confidentiality agreement with NIOSH were pro- vided with access to the raw data. The analytical ap- proach, conduct of the analyses, interpretation of the

results, and conclusions drawn are exclusively the professional work product of the authors and may not necessarily be those of organizations that funded the research. Those organizations had no input into the interpretation of the results or the writing of this article.

NIOSH reviewed an earlier version of this article to ensure compliance with all provisions of the agree- ment under which the DEMS data were released to the analytical team. The review by NIOSH was not a review of the technical contents of the article or the scientific conclusions drawn.

REFERENCES

1. Hesterberg TW, Long CM, Bunn WB et al. Health effects re- search and regulation of diesel exhaust: An historical overview focused on lung cancer risk. Inhalation Toxicology, 2012; 24(1):1–45.

2. McClellan RO, Hesterberg TW, Wall JC. Evaluation of car- cinogenic hazard of diesel engine exhaust needs to consider revolutionary changes in diesel technology. Regulatory Toxi- cology and Pharmacology, 2012; 63(2):225–258.

3. Attfield MD, Schleiff PL, Lubin JH et al. The Diesel Exhaust in Miners Study: A cohort mortality study with emphasis on lung cancer. Journal of the National Cancer Institute, 2012; 104(11):869–883.

4. Silverman DT, Samanic CM, Lubin JH et al. The Diesel Ex- haust in Miners Study: A nested case-control study of lung cancer and diesel exhaust. Journal of the National Cancer In- stitute, 2012; 104(11):855–868.

5. IARC. IARC monographs on the evaluation of carcinogenic risks to humans. Vol. 105. Diesel and Gasoline Engine Ex- haust and Some Nitroarenes. Lyon, France: International Agency for Research on Cancer (IARC), 2013.

6. California EPA (Environmental Protection Agency). Reso- lution 98-3-5 on particulate emissions from diesel engines, agenda item no. 98-8-1, August 27, 1998.

7. California EPA (Environmental Protection Agency). Report to the Air Resources Board on the Proposed Identification of Diesel Exhaust as a Toxic Air Contaminant, Exposure Assess- ment (Part A), Health Risk Assessment for Diesel Exhaust. Sacramento, CA: Office of Environmental Health Hazard As- sessment, California Environmental Protection Agency.

8. California EPA (Environmental Protection Agency). Health Risk Assessment for Diesel Exhaust: Proposed Identifica- tion of Diesel Exhaust as a Toxic Air Contaminant (Part B), Health Risk Assessment for Diesel Exhaust. Sacramento, CA: Office of Environmental Health Hazard Assessment, California Environmental Protection Agency.

9. California EPA (Environmental Protection Agency). Find- ings of the Scientific Review Panel on the Report on Diesel Exhaust as Adopted at the Panel’s April 22, 1998 Meeting. Available at: http://www.Arb.Ca.Gov/toxics/dieseltac/de-fnds. Accessed July 29, 2008.

10. Dawson SV, Alexeeff GV. Multi-stage model estimates of lung cancer risk from exposure to diesel exhaust, based on a U.S. railroad worker cohort. Risk Analysis, 2001; 21(1):1–18.

11. NIOSH (National Institute of Occupational Safety and Health). Carcinogenic effects of exposure to diesel exhaust. Current Intelligence Bulletin 50, DHHS (NIOSH), Publica- tion No. 88-116. Cincinnati, OH: NIOSH.

12. MSHA (Mine SaHA. Diesel particulate matter exposure of underground metal and non-metal miners: Final rule, January 19, 2001. Federal Register, 2001; 66(13):5706.

674 Moolgavkar et al.

13. U.S. EPA (Environmental Protection Agency). Health As- sessment Document for Diesel Engine Exhaust, EPA/600/8- 90/057f. Washington, DC: U.S. EPA, Office of Research and Development, National Center for Environmental Assessment.

14. Peto J. That the effects of smoking should be measured in pack-years: Misconceptions 4. British Journal of Cancer, 2012; 107(3):406–407.

15. Thomas DC. Invited commentary: Is it time to retire the “pack-years” variable? Maybe not! American Journal of Epi- demiology, 2014; 179(3):299–302.

16. Hernan MA. The hazards of hazard ratios. Epidemiology, 2010; 21(1):13–15.

17. Poole C. On the origin of risk relativism. Epidemiology, 2010; 21(1):3–9.

18. Tammemagi CM, Pinsky PF, Caporaso NE et al. Lung cancer risk prediction: Prostate, lung, colorectal and ovarian cancer screening trial models and validation. Journal of the National Cancer Institute, 2011; 103(13):1058–1068.

19. Hazelton WD, Clements MS, Moolgavkar SH. Multistage car- cinogenesis and lung cancer mortality in three cohorts. Cancer Epidemiology Biomarkers & Prevention, 2005; 14(5):1171– 1181.

20. Moolgavkar SH. Multistage carcinogenesis and epidemiologic studies of cancer. In Almudevar AL, Hall WJ, Oakes D (eds). Modeling and Inference in Biomedical Science—In Memory of Andrei Yakovlev. Institute of Mathematical Statistics Col- lections Series, 2014.

21. Richardson DB. Lung cancer in chrysotile asbestos workers: Analyses based on the two-stage clonal expansion model. Can- cer Causes Control, 2009; 20(6):917–923.

22. Brown CC, Chu KC. Implications of the multistage theory of carcinogenesis applied to occupational arsenic exposure. Jour- nal of the National Cancer Institute, 1983; 70(3):455–463.

23. Day NE, Brown CC. Multistage models and primary preven- tion of cancer. Journal of the National Cancer Institute, 1980; 64(4):977–989.

24. Knoke JD, Shanks TG, Vaughn JW et al. Lung cancer mor- tality is related to age in addition to duration and inten- sity of cigarette smoking: An analysis of CPS-I data. Can- cer Epidemiology Biomarkers & Prevention, 2004; 13(6):949– 957.

25. Moolgavkar SH, Holford TR, Levy DT et al. Impact of re- duced tobacco smoking on lung cancer mortality in the United States during 1975–2000. Journal of the National Cancer Insti- tute, 2012; 104(7):541–548.

26. Thomas DC. Models for exposure-time-response relationships with applications to cancer epidemiology. Annual Review of Public Health, 1988; 9:451–482.

27. Thomas DC. Statistical Models in Environmental Epidemiol- ogy. Oxford, UK: Oxford University Press, 2009.

28. Armitage P, Doll R. The age distribution of cancer and a multi-stage theory of carcinogenesis. British Journal of Can- cer, 1954; 8(1):1–12.

29. Lubin JH, Caporaso N, Wichmann HE, Schaffrath-Rosario A, Alavanja MC. Cigarette smoking and cancer risk: Modeling total exposure and intensity. American Journal of Epidemiol- ogy, 2007; 166(4):479–489.

30. Lubin JH, Caporaso N, Wichmann HE et al. Cigarette smok- ing and lung cancer: Modeling effect modification of total ex- posure and intensity. Epidemiology, 2007; 18(5):639–648.

31. Lubin JH, Caporaso NE. Cigarette smoking and lung cancer: Modeling total exposure and intensity. Cancer Epidemiology Biomarkers & Prevention, 2006; 15(3):517–523.

32. Vlaanderen J, Portengen L, Schuz J et al. Effect modification of the association of cumulative exposure and cancer risk by intensity of exposure and time since exposure cessation: A flexible method applied to cigarette smoking and lung can- cer in the synergy study. American Journal of Epidemiology, 2014; 179(3):290–298.

33. Little MP, Heidenreich WF, Li G. Parameter identifiability and redundancy in a general class of stochastic carcinogene- sis models. PLoS One, 2009; 4(12):e8520.

34. Luebeck EG, Curtius K, Jeon J, Hazelton WD. Impact of tu- mor progression on cancer incidence curves. Cancer Research, 2013; 73(3):1086–1096.

35. Luebeck EG, Moolgavkar SH. Multistage carcinogenesis and the incidence of colorectal cancer. Proceedings of the National Academy of Science, USA, 2002; 99(23):15095–15100.

36. Meza R, Hazelton WD, Colditz GA, Moolgavkar SH. Analy- sis of lung cancer incidence in the nurses’ health and the health professionals’ follow-up studies using a multistage carcinogen- esis model. Cancer Causes & Control, 2008; 19(3):317–328.

37. Meza R, Jeon J, Moolgavkar SH et al. Age-specific incidence of cancer: Phases, transitions, and biological implications. Pro- ceedings of the National Academy of Science, USA, 2008; 105(42):16284–16289.

38. Moolgavkar SH, Krewski D, Schwartz M. Mechanisms of car- cinogenesis and biologically-based models for quantitative es- timation and prediction of cancer risk. In Moolgavkar SH, Krewski D, Zeise L, et al. (eds), Quantitative Estimation and Prediction of Cancer Risk. Lyon, France: IARC Scientific Publications, 1999.

39. Moolgavkar SH, Luebeck EG. Multistage carcinogenesis: Population-based model for colon cancer. Journal of National Cancer Institute, 1992; 84(8):610–618.

40. Moolgavkar SH, Luebeck EG, Krewski D, et al. Radon, cigarette smoke, and lung cancer: A re-analysis of the Col- orado plateau uranium miners’ data. Epidemiology, 1993; 4(3):204–217.

41. Moolgavkar SH, Luebeck G. Two-event model for carcino- genesis: Biological, mathematical, and statistical considera- tions. Risk Analysis, 1990; 10(2):323–341.

42. Coble JB, Stewart PA, Vermeulen R, Yereb D, Stanevich R, Blair A, Silverman DT, Attfield M. The Diesel Exhaust in Miners Study: II. Exposure monitoring surveys and develop- ment of exposure groups. Annals of Occupational Hygiene, 2010; 54(7):747–761.

43. Stewart PA, Coble JB, Vermeulen R, Schleiff P, Blair A, Lu- bin J, Attfield M, Silverman DT. The Diesel Exhaust in Miners Study: I. Overview of the exposure assessment process. Annals of Occupational Hygiene, 2010; 54(7):728–746.

44. Stewart PA, Vermeulen R, Coble JB, Blair A, Schleiff P, Lu- bin JH, Attfield M, Silverman DT. The Diesel Exhaust in Min- ers Study: V. Evaluation of the exposure assessment methods. Annals of Occupational Hygiene, 2012; 56(4):389–400.

45. Vermeulen R, Coble JB, Lubin JH, Portengen L, Blair A, At- tfield MD, Silverman DT, Stewart PA. The Diesel Exhaust in Miners Study: IV. Estimating historical exposures to diesel exhaust in underground non-metal mining facilities. Annals of Occupational Hygiene, 2010; 54(7):774–788.

46. Vermeulen R, Coble JB, Yereb D, Lubin JH, Blair A, Porten- gen L, Stewart PA, Attfield M, Silverman DT. A The Diesel Exhaust in Miners Study: III. Interrelations between res- pirable elemental carbon and gaseous and particulate compo- nents of diesel exhaust derived from area sampling in under- ground non-metal mining facilities. Annals of Occupational Hygiene, 2010; 54(7):762–773.

47. Silverman DT, Lubin JH, Blair AE, Vermeulen R, Stewart PA, Schleiff PL, Attfield MD. Re: The Diesel Exhaust in Min- ers Study (DEMS): A nested case-control study of lung cancer and diesel exhaust. Journal of National Cancer Institute, 2014; 106(8). DOI: 10.1093/jnci/dju205.

48. Heidenreich WF, Luebeck EG, Moolgavkar SH. Some prop- erties of the hazard function of the two-mutation clonal expan- sion model. Risk Analysis, 1997; 17(3):391–399.

49. Burns DM, Shanks TG, Choi W, Thun MJ, Heath CW, Jr., Garfinkel L. The American Cancer Society Cancer Prevention Study I: 12-year followup of 1 million men and women. Pp. 113–304 in Burns DM, Garfinkel L, Samet JM (eds). Smoking

Diesel Engine Exhaust and Lung Cancer Mortality 675

and Tobacco Control, Monograph 8, NIH Publication No. 97- 4213. Washington, DC: National Institutes of Health, 1997.

50. Borak J, Bunn WB, Chase GR, Hall TA, Head HJ, Hester- berg TW, Sirianni G, Slavin TJ. Comments on the Diesel Ex- haust in Miners Study. Annals of Occupational Hygiene, 2011; 55(3):339–342; author reply 43-6.

51. Crump K, Van Landingham C. Evaluation of an exposure as- sessment used in epidemiological studies of diesel exhaust and lung cancer in underground mines. Critical Reviews in Toxi- cology, 2012; 42(7):599–612.

52. Hesterberg TW, Long CM, Valberg PA. Re: The Diesel Ex- haust in Miners Study: A nested case-control study of lung cancer and diesel exhaust and a cohort mortality study with emphasis on lung cancer. Journal of National Cancer Institute, 2012; 104(23):1841; author reply 8–9.

53. McClellan RO. Re: The Diesel Exhaust in Miners Study: A nested case-control study of lung cancer and diesel exhaust, a cohort mortality study with emphasis on lung cancer, and the problem with diesel. Journal of National Cancer Institute, 2012; 104(23):1843–1845; author reply 7–9.

54. Vermeulen R, Silverman DT, Garshick E, Vlaanderen J, Portengen L, Steenland K. Exposure-response estimates for diesel engine exhaust and lung cancer mortality based on data from three occupational cohorts. Environmental Health Per- spectives, 2014; 122(2):172–177.

SUPPORTING INFORMATION

Additional Supporting Information may be found in the online version of this article at the publisher’s website:

Table SI. Results of Proportional Hazards Re- gression Analysis for Lung Cancer Mortality in Association with 15-Year Lagged Cumulative Ex- posure to Respirable Elemental Carbon (REC) and Average Intensity for Underground-Only (i.e., Exclusively Underground) Workers: Quartiles, Expanded Categories, and Continuous Modeling Results

Table SII. Results of Proportional Hazards Regres- sion Analysis for Lung Cancer Mortality in As- sociation with Log of 15-Year Lagged Cumulative Exposure to Respirable Elemental Carbon (REC), that is, Log(1 + REC-Lagged 15 years) for the En- tire Cohort, the Entire Cohort Minus the Limestone Mine, and Each Mine Type Separately

Table SIII. Results of a Proportional Hazards Re- gression Analysis for the Cohort of Miners with Total Exposure <1,280 μg/m3·year; All the Co- variates in this Model are Highly Statistically Significant Fig. S1. Exposure-response relationship between cu- mulative respirable elemental carbon (REC; μg/m3- yr) and relative hazard in the entire cohort of ever- underground miners

Copyright of Risk Analysis: An International Journal is the property of Wiley-Blackwell and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.