Meta-analysis of Rare Diseases in Occupational Epidemiology.

At the outset of this research in the early 2000s, the application of meta-analysis in observational epidemiology, including occupational epidemiology was regarded as controversial because of the greater potential for bias in such studies compared with randomized controlled clinical trials. This remains true even in 2017. The overall aim for this research is to identify the best approaches to the use of meta-analysis in the occupational health setting, and to summarise through meta-analysis the evidence at to whether or not occupational exposure to formaldehyde is an occupational carcinogen. Chapter 2 of this thesis concluded that meta-analysis in occupational epidemiology was becoming increasingly popular, but that its limitations appear not to have been heeded in practice. Two principal issues were identified: the heterogeneity of exposures estimates and the pooling of standardized mortality ratios from different study populations with different characteristics including length of follow-up. By 2016, neither of these issues had been addressed in the literature. Chapter 3 contains a description of the published statistical methods available for meta-analysis up to 2001. Methodological developments continue within the discipline of randomized controlled clinical trials and recently include the advent of multivariate methods and network analysis. It has been recommended that any new methods should be backed up by simulations. Chapter 4 concludes that the default approach adopted by most statistical packages could not deal with such studies and excluded them from any calculations. In meta-analyses of rare diseases, this biases the meta-relative risks upwards. Approaches that avoided exclusion of such studies are considered, in particular analyses on the original untransformed scale rather than the log scale. Chapter 5 concludes that there remains insufficient evidence for an association. There was significant heterogeneity in the lung cancer results and so this is a random effects analysis; the analyses for nasopharyngeal and sinonasal cancer contains no such heterogeneity and so are fixed effect analyses. The final chapter concludes that, when studies with zero cases are not excluded, there is insufficient evidence of an increased risk of lung cancer, nasopharyngeal cancer or sinonasal cancer, and that further methodological developments are still required to deal with the pooling of occupational epidemiological studies in relation to study characteristics, exposure assessments and standardised mortality ratios.