In our study, job-specific exposure levels to DE were estimated from personal REC measurements collected in 1998–2001 during the DEMS surveys (Coble et al., 2010
; Stewart et al., 2010
). However, almost no REC or other EC monitoring data were available prior to these surveys, prohibiting us from estimating past DE levels based on EC measurements. As a consequence, the historical estimation of REC for underground jobs relied on back-extrapolation of the REC estimates from the DEMS data using predictive time trend models. CO was chosen to estimate relative changes in historical REC levels because of its frequent use in the past as a proxy of DE exposure (Pronk et al., 2009
) and because it was the most frequently measured DE component in our study facilities EC and CO are both produced by incomplete combustion of diesel fuel. This decision was supported by our finding that CO area concentrations were correlated with REC, that CO increased approximately linearly in log-log space with REC area concentrations in the DEMS surveys, and that it loaded on the same factor as EC and DE gases (Vermeulen et al., 2010
There were, however, several limitations to the use of CO to predict historical DE levels. First, except for four measurements from 1972 for one mine (G), CO measurements were only available after 1975 while we needed to predict exposures from the date of dieselization, which varied by facility from 1947 to 1967, to the reference year, which varied from 1998 to 2001. The prediction for all years, including prior to 1976, was achieved by applying the derived model parameter estimates for Ln(ADJ HP/CFM), Ln(ADJ HP1990+), and the other variables to the determinant data available for all years. This procedure, because the model was developed from measurement data after 1975, assumed that the parameter estimate for Ln(ADJ HP/CFM) was applicable to the time period prior to 1976. This assumption seemed reasonable since engine and ventilation technology did not differ substantially from the 1960s to the 1990s and there was little diesel equipment prior to the 1960s (ADJ HP ranged from <1 to 10% of the maximum ADJ HP in the five facilities that were in existence before 1960).
A second possible limitation was that CO might have arisen from sources in underground mining other than diesel engines. Another source of CO was likely from the explosives used to blast the face (Douglas and Beaulieu, 1983
; Jacobsen et al., 1988
). This source, however, was unlikely to have influenced our results as we only used CO face measurements, of which ~95% were from inspection data. MSHA inspectors did not routinely monitor at the face immediately after a blasting because of the possible high exposures to the blasting fumes. Thus, the inspectors would not have entered the face areas until the blasting fumes were exhausted and diluted. This assumption was further supported by additional statistical analyses of the MIDAS data that did not find any indication of higher CO air concentrations shortly after the time of blasting compared to other times during the day (data not shown). Therefore, these data suggest that CO was a specific proxy for DE in underground mining operations in our study facilities.
We acquired CO measurement data from several data sources. Of these sources, MIDAS was the only source with truly longitudinal data, and it contributed >95% of the CO data that were used in the modeling. MIDAS measurements taken at the production face found average CO concentrations generally ranging from 1.2 to 1.6 ppm in 1975–1979, 0.9–3.1 ppm in the 1980s, and <1 ppm in the 1990s. The major exception was Facility A where concentrations increased over time. The difference for this facility was consistent with the increasing size (and therefore HP) and number of diesel equipment, particularly haulage trucks, used in the underground operations over time (Coble et al., 2010
; Stewart et al., 2010
). Although the year of introduction of the first diesel equipment varied considerably between the underground operations of the facilities (1947–1967), in general, the largest increase in diesel equipment usage occurred between 1960 and 1980 in all facilities (when the facilities rapidly dieselized their underground mining equipment) and use peaked in the early 1980s (Facility A being the exception). After the mid-1980s, diesel usage generally changed little or decreased slightly. Increases in exhausted airflow rates followed the increase in HP in almost all operations to control gaseous contaminant levels and rose or remained unchanged as HP remained the same or decreased, thus resulting in lower exposure levels. We also estimated the amount of HP that was introduced after 1990, the period which corresponded to the introduction of cleaner direct injection engines and cleaner fuels (Haney and Saseen, 2000
). At the time of the DEMS surveys, the ADJ HP of the newer engines ranged between 32 and 77% of the total ADJ HP, per facility, suggesting that for most of the mines this additional parameter was important.
Our facility-specific CO models used ADJ HP, total exhaust airflow rates (CFM) and ADJ HP1990+
as predictors and CO concentration measurements as the dependent variable. These a priori
selected predictors were similar to the main parameters of the deterministic model developed by MSHA to estimate diesel particulate matter (DPM) exposure at the faces of underground operations (Haney and Saseen, 2000
). The MSHA model had basically three components: the quantity of exhaust emissions (i.e. estimated based on engine HP, engine DPM emission rates, the number of engines, and the length of the work shift), efficiency of exhaust control emissions (i.e. fuel properties and the efficiency of applied control technology), and the quantity of air exhausted from the face. A difference between the MSHA deterministic model and our predictive models is that we lacked airflow rate or ADJ HP data at each operating face of our facilities. For each face, however, a minimal set of equipment was generally necessary, so that the types and amount of equipment used for the production operations were not likely to differ substantially between faces within a facility at a given point in time. Support for this assumption was also found in the fact that the REC measurements on the jobs, which were taken in different areas, were relatively homogeneous within each facility, as were the area measurements, which also were taken in different places (Coble et al., 2010
). It seemed reasonable, therefore, to use mine-wide ADJ HP estimates and total exhaust airflow rates to model area-specific CO concentrations.
Ln(ADJ HP/CFM) was significantly associated with area CO concentrations in six of the seven facility-specific models. The median parameter estimate of Ln(ADJ HP/CFM) in the seven facilities was ~1 (range: 0.74–2.72). Except for Facility I, the 95% confidence intervals of the parameter estimates for all facilities included 1. The global P
value for the test of homogeneity of the Ln(ADJ HP/CFM) parameter estimates for all facilities was 0.03, while if Facility I was excluded, the P
value was 0.64, indicating that the coefficient was indeed similar across the facilities except for Facility I. A parameter estimate of 1 indicates that the change for CO and ADJ HP/CFM are directly proportional and equal, i.e. that a doubling in the ratio ADJ HP/CFM corresponds to a doubling of the CO concentration. This is consistent with the MSHA model where direct proportionality is assumed for HP and an inverse proportionality is assumed for airflow rate (Haney and Saseen, 2000
). The explanation for the large parameter estimate for Ln(ADJ HP/CFM) in Facility I is unclear and might indicate incorrect data for HP or airflow rate in 1980–1990.
The negative estimates for Ln(ADJ HP1990+
) in all facility-specific models where this parameter could be fitted indicates that the increase in CO concentrations relative to HP was lower for the newer engines compared to older engines. The global P
value for the test of homogeneity of the Ln(ADJ HP1990+
) parameter estimates for all facilities was 0.08, suggesting that this parameter estimate was not statistically different across the facilities. The observed reduced emission of newer engines in combination with cleaner fuels is in agreement with other reports (Haney and Saseen, 2000
), which indicated that engines introduced after 1990 resulted in lower DE emissions (including CO) per unit of HP.
Because we found such similar results for the main model parameters [i.e. Ln(ADJ HP/CFM) and Ln(ADJ HP1990+)] in the various facilities, it can be assumed that the widely varying HP, loads, efficiency of the engines, and other characteristics tended to cancel those characteristics that affect the emission characteristics of an individual engine.
Since the CO models were developed to predict REC levels over time, the absence of temporal variations in the residuals was an important confirmation of internal model validity. We observed significant deviations in residuals for seven 5-year time periods in 35 facility/5-year time period combinations. In two instances, the models predicted higher levels than the measurements, and in five instances, the models predicted lower levels than the measurements. Thus, we found that the CO predictions were free of any marked temporal bias.
This finding was further corroborated by our alternative set of models using the 5-year average CO concentration time trends that showed temporal patterns similar to our primary CO models, indicating the good temporal fit of our models. We also used the CO models to extrapolate the DEMS average CO concentrations back to 1976–1977 to compare with average face concentrations measured during the MESA/BoM 1976–1977 survey. This comparison showed a median relative difference of 29%, which varied between 49 and -25% by facility. These differences were close to what would be expected if side-by-side measurements were taken (Zey et al., 2002
). Noteworthy is the difference estimate of Facility J (46%). This facility was closed in 1993 and could not be measured in the DEMS surveys, so the model parameter estimates of Facility B were used. The relative difference seen for this facility was within the range for the other facilities, indicating that this approach for exposure estimation in Facility J was satisfactory.
Few industry-based studies of chemical agents have been able to evaluate historical quantitative exposure estimates in comparison to external measurement data. Investigators of a study of ethylene oxide workers found a relative difference of 24% when comparing predictions with measurements over a 6-year period (Hornung et al., 1994
). In an acrylonitrile study, the relative difference for two estimation methods was −17% and −66% when compared to measurement data over an 11-year period (Stewart et al., 2003
). A relative difference of −2% was found in a study of textile workers between the estimates and the measurements over a 15-year period (Astrakianakis et al., 2006
). Investigators of an asphalt paving study found relative differences of −70 and −51% for bitumen fume and benzo(a
)pyrene, respectively, when comparing measurements to estimates for up to 12 years back in time (Burstyn et al., 2002
). The difference seen in our evaluation (median of 29%) is comparable to these earlier studies, even though our comparison covered over 20 years, whereas the other studies covered 15 years or less.
The main variables in our models were Ln(ADJ HP/CFM) and Ln(ADJ HP1990+), where the effect estimate of ADJ HP was conditioned on CFM. We therefore explored alternative model specifications in which Ln(ADJ HP), Ln(ADJ HP1990+), and Ln(CFM) were introduced as separate variables in the regression models. As previously indicated, the inclusion of CFM and ADJ HP as separate determinants in the models proved to be difficult for some of our facility-specific models (D, G, and H), due to collinearity between these two variables, resulting in implausible estimates of the parameters (i.e. increasing exposure levels with increasing airflow rates). However, for the models where collinearity was not a problem, the results were quite similar to the results based on the primary model specifications (data not shown). In addition, we explored models in which we modeled Ln(ADJ HP1990−/CFM) and Ln(ADJ HP1990+/CFM) separately (where ADJ HP1990− designates ADJ HP of equipment acquired before 1990). These models resulted in very similar fits and predictions as our primary models (data not shown). Based on these results, we concluded that our facility-specific models are reasonably robust toward differences in model specification.
The model coefficients were subsequently used to predict the relative CO concentrations from the reference year to the date of dieselization. These relative trends were used to back-extrapolate 1998–2001 personal exposure estimates by assuming that a relative change in historical CO levels could be directly translated to an identical change in REC over all the years of the study. A study by Yanowitz et al. (2000)
suggests that indeed the CO-REC relationship probably changed little from 1976 to 1997. In that study, emission data based on laboratory tests of various engines with varying model years (1976–1997) were analyzed using regression analyses on emissions per mile of grams of CO and of grams of DPM. The authors found that CO and DPM emissions increased back through time at only slightly different rates [parameter estimate difference = −0.003, i.e. DPM increased slightly less than CO (US EPA, 2002
)]. A comparable relationship was also likely to hold true for EC, given that it is a significant proportion of DPM (Birch and Noll, 2004)
However, in our cross-sectional DEMS surveys we observed that the relation between CO and REC might not be strictly proportional because the regression of Ln(REC) on Ln(CO) rendered a parameter estimate of 0.58, indicating that REC concentrations increased with CO concentration to the power of 0.58 (Vermeulen et al., 2010
). We previously noted that, given the cross-sectional nature of this survey, it is possible that the observed parameter estimate might not apply longitudinally to past conditions (Vermeulen and Kromhout, 2005
; Vermeulen et al., 2010
). However, to assess the sensitivity of the epidemiologic findings to our decision of making the REC changes over time directly proportional to temporal changes in CO, we developed an alternative set of facility-specific models using the 0.58 power parameter to modify the relative CO concentrations. As expected, this change resulted in lower historical REC levels but had little effect on the ranking of the subjects’ cumulative exposure (Pearson correlation > 0.9) (Stewart et al., 2010
To estimate mining facility/department/job/year-specific REC exposure levels, the facility-specific relative changes in CO were subsequently multiplied with the reference REC exposure estimates of the underground jobs. In Stewart et al. (2010)
, we showed that the hierarchical grouping strategy was successful in explaining the between-job variability of the REC measurements. Of course, this approach assumes that these grouping remained valid historically, which is reasonable as long as job locations did not differ significantly over time. For illustrative purposes, we presented the predicted REC levels for the mine operator. REC exposure levels for this job, one of the highest exposed jobs in the DEMS surveys, ranged between 100 and 600 μg m−3
in the 1970–1980s. There are few previously published data with which to compare the results of our study. However, recent studies have reported personal EC measurement levels between 20 and >500 μg m−3
(Ramachandran and Watts, 2003
; Backe et al., 2004
; Birch and Noll, 2004
; Adelroth et al., 2006
; Burgess et al., 2007
; Dahmann et al., 2007
; Noll et al., 2007
). The wide range among these studies is likely due to different types and amounts of diesel equipment, airflow rates, and work practices, just as these differences influenced exposure levels in our facilities. The levels reported in the literature suggest that our historical REC estimates of up to ~600 μg m−3
20 years ago are within plausible ranges of REC exposure levels.
In summary, the facility-specific models were developed based on the regression of diesel use, total exhaust airflow rates, and engine technology on historical CO area concentration measurements to predict relative changes in DE exposure levels. The models indicate substantial changes over time, with the highest DE exposure levels for most underground operations between 1970 and the early 1980s. Comparisons of the CO estimates from our time trend models with external CO measurement data showed that the estimates had a median relative difference of 29% in 1976–1977, a difference that is comparable to what other epidemiologic studies have found. We concluded that the predictions derived from the time trend models were plausible, resulting in time-varying REC exposure estimates that were in a range observed in other published monitoring studies. The subsequently derived time-varying REC job-specific exposure estimates were used in the investigation of exposure–response relationships in the epidemiologic evaluation.