Electronic Theses and Dissertations (PhDs)
Permanent URI for this collectionhttps://hdl.handle.net/10539/37983
Browse
Search Results
Item Preventing Coal Mine Dust Lung Disease: Application of Bayesian Hierarchical Framework for Occupational Exposure Assessment in The South African Coal Mining Industry(University of the Witwatersrand, Johannesburg, 2023-10) Made, Felix; Brouwer, Derk; Lavoue, Jerome; Kandala, Ngianga-BakwinBackground: The world's largest energy source is coal with nearly 36% of all the fuel used to produce power. South Africa is the world's top exporter and the seventh-largest producer of coal. In the upcoming years, it is expected that South Africa's coal production output rate will rise. Coal mine dust lung disease (CMDLD) is an irreversible lung disease caused by the production of coal, the emission of dust, and prolonged exposure to the dust. When conducting safety evaluation, exposure is typically reported as an eight-hour time-weighted average dust concentration (TWA8h). In occupational exposure contexts, occupational exposure limits (OEL) are often used as a threshold where workers can be exposed repeatedly without adverse health effects. The workers are usually grouped into homogenous exposure groups (HEGs) or similar exposure groups (SEGs). In South Africa, a HEG is a group of coal miners who have had similar levels and patterns of exposure to respirable crystalline silica (RCS) dust in the workplace. Several statistical analysis methods for compliance testing and homogeneity assessment have been put into use internationally as well as in South Africa. The international consensus on occupational exposure analysis is based on guidelines from the American Industrial Hygiene Association (AIHA), the Committee of European Normalisation (CEN), and BOHS British and Dutch Occupational Hygiene Societies' guidelines (BOHS). These statistical approaches are based on Bayesian or frequentist statistics and consider the 90th percentile (P90) and 95th percentile (P95), with- and between-worker variances, and the lognormal distribution of the data. The current existing practices in South Africa could result in poor or incorrect risk and exposure control decision-making. Study Aims: The study aimed to improve the identification of coal dust overexposure by introducing new methods for compliance (reduced dust exposure) and homogeneity (similar dust exposure level) assessment in the South African coal mining industry. Study Objectives: The objectives of this study were: 1. To compare compliance of coal dust exposure by HEGs using DMRE-CoP approach and other global consensus methods. 2. To investigate and compare the within-group exposure variation between HEGs and job titles. 3. To determine the posterior probabilities of locating the exposure level in each of the OEL exposure categories by using the Bayesian framework with previous information from historical data and compare the findings and the DMRE-CoP approach. 4. To investigate the difference in posterior probabilities of the P95 exposure being found in OEL exposure category between previous information acquired from the experts and the current information from the data using Bayesian analysis. Methods: The TWA(8h) respirable coal dust concentrations were obtained in a cross-sectional study with all participants being male underground coal mine workers. The occupational hygiene division of the mining company collected the data between 2009 and 2018. The data were collected according to the South African National Accreditation System (SANAS) standards. From the data, 28 HEGs with a total of 728 participants were included in this study. In objectives 1 and 2, all 728 participants from the 28 HEGs were included in the analysis. For exposure compliance, the DMRE-CoP accepts 10% exceedance of exposure above the OEL (P90 exposure values from HEGs should be below the OEL). The 10% exceedance was compared to the acceptability criterion from international consensus which uses 5% exceedance above the OEL (P95 exposure is below the OEL) of the lognormal exposure data. For exposure data to be regarded as homogenous, the DMRE-CoP requires that the arithmetic mean (AM) and P90 must fall into the same DMRE-CoP OEL exposure category. The DMRE-CoP on assessment of homogeneity was also compared with the international approaches which include the Rappaport ratio (R-ratio) and the global geometric standard deviation (GSD). A GSD greater than 3 and an R-ratio greater than 2 would both indicate non-homogeneity of the exposure data of a HEG. The GSD and DMRE-CoP criteria were used to assess the homogeneity of job titles exposure within a HEG. In objective 3 a total of nine HEGs which have 243 participants, were included in the analysis. To investigate compliance, a Bayesian model was fitted with a Markov chain Monte Carlo (MCMC) simulation. A normal likelihood function with the GM and GSD from lognormal data was defined. The likelihood function was updated using informative prior derived as the GM and GSD with restricted bounds (parameter space) from the HEGs' historical data. The posterior probabilities of the P95 being located in each DMRE exposure band were produced and compared with the non-informative results and the DMRE approach DMRE-CoP using a point estimate inform of the 90 percentiles. In objective 4, a total of 10 job titles were analysed and selected. The selection of the job titles was based on if they have previous year's data so it can be used to develop prior information in the Bayesian model. The same job titles were found across different HEGs, so to ensure the mean is not different across HEGs, the median difference of a job title exposure distribution across HEGs was statistically compared using the Kruskal-Wallis test, a non-parametric alternative to analysis of variance (ANOVA). Job titles with statistically non-significant exposure differences were included in the analysis. Expert judgements about the probability of the P95 located in each of the DMRE exposure bands were elicited. The IDEA (Investigate", "Discuss", "Estimate" and "Aggregate) expert elicitation procedure was used to collect expert judgements. The SHELF tool was then used to produce the lognormal distribution of the expert judgements as GM and GSD to be used as informative prior. A similar Bayesian analysis approach as in objective 3 was used to produce the probability of the P95 falling in each of the DMRE exposure bands. The possible misclassification of exposure arising from the use of bounds in the parameter space was tested in a sensitivity analysis. Results: There were 21 HEGs out of 28 in objectives 1 and 2 that were non-compliant with the OEL across all methods. According to the DMRE-CoP approach, compliance to the OEL, or exposure that is below the OEL, was observed for 7 HEGs. The DMRE-CoP and CEN both had1 HEG with exposures below the OEL. While the DMRE-CoP showed 6 homogeneous HEGs, however, based on the GSDs 11 HEGs were homogeneous. The GSD and the DMRE-CoP agreed on homogeneity in exposures of 4 (14%) HEGs. It was discovered that by grouping according to job titles, most of the job titles within non-homogenous HEGs were homogenous. Five job titles had AMs above their parent HEG. For objective 3, the application of the DMRE-CoP (P90) revealed that the exposure of one HEG is below the OEL, indicating compliance. However, no HEG has exposures below the OEL, according to the Bayesian framework. The posterior GSD of the Bayesian analysis from non-informative prior indicated a higher variability of exposure than the informative prior distribution from historical data. Results with a non-informative prior had slightly lower values of the P95 and wider 95% credible intervals (CrI) than those with an informative prior. All the posterior P95 findings from both non-informative and informative prior distribution were classified in exposure control category 4 (i.e., poorly controlled since exceeding the OEL), with posterior probabilities in the informative approach slightly higher than in the non-informative approach. Job titles were selected as an alternative group to assess compliance in objective 4. The posterior GSD indicated lower variability of exposure from expert prior distribution than historical data prior distribution. The posterior P95 exposure was very likely (at least 98% probability) to be found in exposure control category 4 when using prior distribution from expert elicitation compared to the other Bayesian analysis approaches. The probabilities of the P95 from experts' judgements and historical data were similar. The non-informative prior generally showed a higher probability of finding the posterior P95 in lower exposure control categories than both experts and historical data prior distribution. The use of different parameter values to specify the bounds showed comparable results while the use of no parameter space at all put the posterior P95 in exposure category 4 with 100% probability. Conclusions: In comparison to other approaches, the DMRE-CoP tend to show that exposures are compliant more often. Overall, all methods show that the majority of HEGs were non-compliant. The HEGs that suggest non-homogeneity revealed that the constituent job titles were homogenous. Application of the GSD criterion indicated that HEGs are more likely to be considered as homogeneous than when using the DMRE-CoP approach. When using the GSD and the DMRE-CoP guidelines, alternative grouping by specific job titles showed a greater agreement of homogeneity. The use of job titles showed that using HEGs following the DMRE-CoP current guidelines might not show high-exposure job titles and would overestimate compliance. Additionally, since job titles within a HEG may be homogeneous or have a different exposure to the parent HEG, exposure variability is not properly recorded when using HEGs. In compliance assessment, it is important to use the P95 of the lognormal distribution rather than the DMRE-CoP approach that use the empirical P90. Our findings suggest that the subgrouping of exposure according to job titles within a HEG should be used in the retrospective assessment of exposure variability, and compliance with the OEL. Our results imply that the use of a Bayesian framework with informative prior from either historical or expert elicitation may confidently aid concise decision making on coal dust exposure risk. Contrary to informative prior distribution derived from historical data or expert elicitation, Bayesian analysis using the non-informative uniform prior distribution places HEGs in lower exposure categories. Results from noninformative prior distributions typically show high levels of uncertainty and variability, so a decision on dust control would be reached with less confidence. The Bayesian framework should be used in the assessment of coal mining dust exposure along with prior knowledge from historical data or professional judgment, according to this study. For exposure, findings are to be reported with high confidence and for sound decisions to be reached about risk mitigation, an exposure risk assessment should be considered while using historical data to update the current data. The study also promotes the use of experts in situations where it is necessary to combine current data with historical data, but the historical data is unavailable or inapplicable.