Health impact assessment is often required to inform policy decisions, including policy interventions to reduce occupational exposures and diseases. There were three main aims of this project. The first was to identify the principal methods used to estimate future burden of occupational disease, and to assess their usefulness for testing of intervention scenarios. A review of the literature was undertaken to identify the methods used firstly to estimate current disease burden, and secondly to see how these had been extended to prediction of future disease burden, with an emphasis on disease due to occupation. Lifetime risk (LR) and age-period-cohort (A-P-C) regression methods have been used to predict future disease levels, but not to test the effect of intervention to reduce causative exposures, other than by measuring the effect of a 'counter-factual' change introduced at a single point in time. More recently the future burden population attributable fraction (PAF), the similar 'Prevent' model, and data-intensive 'generalised regression' ('G'-formula) methods, have been used to test intervention scenarios taking into account time dependency. The future burden PAF method employs the concept of a 'risk exposure period' (REP) projected forward in time. The 'G'-formula models require good quality exposure and confounder data collected usually from a single large-scale prospective epidemiological study with only specific applicability, whereas the PAF method uses readily available national data sources. The second aim was to assess the sensitivity of these future Burden of Disease (BOD) methods to the input data required and the uncertainties associated with this input data. A general approach to identifying sources of bias is described, and a simulation study was set up to examine model and data bias. A 'real world' (maximal) dataset was established to represent the 'true' exposure and outcome experience of a large cohort of workers exposed to respirable crystalline silica (RCS) in the construction industry, with lung cancer mortality as the outcome of interest. The data variables were then 'collapsed' to represent the data that would be available for three alternative future burden methods being tested, the PAF, LR and A-P-C methods. The difference between the estimated values for each method estimand and the 'true' value of that estimand indicate the direction and size of the bias, with variability measured via multiple simulations. Two principal sources of potential bias in the PAF could be addressed in the maximal dataset, the choice of relative risk (RR), and the overall bias due simply to the method itself (that is the equations and constituent variables). The results demonstrated that the most representative summarised risk estimate produces downward bias, and there is also a tendency towards downward bias due to the method itself. No bias was demonstrated for the LR and A-P-C methods, although this was linked to the way the 'true' value dataset had been constructed. In a second approach the simulated dataset was built from an initial simple form, unfixing variables one at a time to move towards the original 'real world' dataset. The idea was to identify which of the main variables, apart from RR, might contribute to the bias in the PAF method. In this analysis bias was again towards underestimation. Other general observations were that this bias increases where there are four rather than two exposure categories, and with higher risk estimates (RRs). In a third approach the alternative 'by age' (versus 'all age') PAF method of estimating future disease burden was examined using a new minimal simulated dataset designed to accommodate the manipulation of exposed numbers by age, plus theoretical exposed worker and population datasets. It included a detailed look at the equations used to establish the 'true' and method estimates, in order to identify the precise variables contributing to the bias. It was shown that while both the 'all age' and 'by age' PAF estimators are biased towards underestimation, the bias is greater in the 'all age' than in the 'by age' estimator. Also bias is very much reduced where an exposure is rare in the population, disappearing where the ratio of exposed to unexposed population is 1:100 or less for the 'by age' estimator, and remaining very low for the 'all age' estimator. The final aim of the project was to produce a 'toolbox' that can be used to guide the choice of appropriate method, based on the purpose of the investigation, type and quality of input data available, nature of the disease outcomes of interest, and the uncertainties associated with each method. A table indicating which method to use in which circumstances fulfils this final aim.
- Statistical methodology
- Occupational exposure
- Health impact assessment
- Future burden of disease
- Simulation study
AN INVESTIGATION OF METHODS USED TO ESTIMATE THE FUTURE BURDEN OF DISEASE DUE TO OCCUPATION
Hutchings, S. (Author). 5 Apr 2024
Student thesis: Phd