A statistical framework for the apportionment of particulate contaminants and their health effect determination

Air pollution is a complex mixture of diverse substances from anthropogenic and natural sources. These sources in combination with factors such as meteorology and chemical/biological transformations, determine the air pollution concentration and the variation in the physiochemical components across space and time. The identification of these sources is a key element for developing effective and efficient strategies to control and reduce pollution through targeted actions. In addition, air pollution is a major public health concern, being increasingly associated with risk of morbidity and mortality of human populations. Recent evidence points out that mixture of particles from different sources can have a different detrimental contribution on health; this makes the understanding of pollutant sources even more important in order for air quality managers to fully understand the potential health outcomes of pollutant mixtures. While knowledge of the main sources of pollution can be effectively obtained on temporal and/or geographical localised setting, the modelling and the understanding of some aspects of the dynamic and physiochemical processes remain a substantial challenge.

This project focuses primarily upon: (i) the development of a methodological approach for particle matter (PM) source apportionment (SA) which uses nonparametric processes with dependence on dynamic factors (e.g. meteorology) to model the underlying spatial or temporal structure and the distribution of contaminants to identify sources; (ii) the quantification of the impact of apportioned air contaminants upon vulnerable populations; and (iii) the translation of this methodological approach to real-life decision making through the predictions of the health outcomes under changing scenarios of pollution mix as a result of potential policy implementations.
The proposed approach will be tested against the state-of-the-art tools for SA of air pollution, using simulated examples. Evaluation of the adverse responses associated with air particulate sources is reached by comparing two-stage procedures vs joint models for SA and health-effect assessment. We will consider two real case studies: (i) to identify time-varying sources of particles (PM2.5) in Greater London and evaluate their acute effects on respiratory hospital admissions in vulnerable populations (0-14 years, 65+) in a time-series framework; (ii) to disentangle spatially-varying sources of particles (PM2.5) in South East England and evaluate their respiratory chronic effects in the same region in a small-area framework.

By using rigorous and innovative methodologies, we believe that the proposed research (i) will provide scientific evidence of the differential harmful effect of PM chemical components, (ii) will help understand the sources that can be controlled, and (iii) will have the potential to inform air pollution policy implementation and regulation to improve UK population health.

Particulate matter (PM) is a complex mixture of diverse substances from anthropogenic and natural sources, which vary widely across space and time. Despite being a major public health concern, most studies have linked total PM to health outcomes, while it is now understood that contaminant mixtures deriving from different sources may have differential health effects.
We propose a Bayesian nonparametric approach that, starting from measurements of a wide range of components, estimate sources of PM and evaluate their health effects, through probabilistic clustering. We will account for spatial or temporal and covariate dependency (e.g. meteorology) in the cluster allocation; the link between sources and health outcomes will be evaluated using (i) a two-stage model where the sources identification is separated from the epidemiological model and (ii) a joint model, allowing for the outcome to influence the apportionment of the sources.
Our approach will provide competitive advantages over standard receptor methods, as it incorporates: (i) temporal dynamics in time-series speciation data, (ii) spatial dependence structure in multi-site contaminant data, (iii) missing data accommodation, (iv) quantification of uncertainty associated with source profiles, (v) nonlinear relationship between sources and health outcomes.
We will build an extensive simulation study and consider two real case studies: (i) to identify time-varying sources of PM in Greater London and evaluate their acute effects on respiratory hospital admissions in vulnerable populations; (ii) to disentangle spatially-varying sources of PM in South East England and evaluate their respiratory chronic effects. We believe that the proposed research (i) will provide new scientific evidence of the differential health impact of air pollutant mixtures, (ii) will help understand the sources that can be controlled and (iii) will have the potential to inform policy implementation for improving UK population health.

The proposed research will develop a novel statistical approach to apportion chemical components of particulate pollution into sources and to evaluate their health effects. It will provide methodological advances: i) in the field of environmental exposure, providing competitive advantages over standard receptor methods as it explicitly accounts for the spatial or temporal dependences present in the source allocation; ii) in the field of environmental epidemiology, shifting the paradigm of the evaluation of health effects from total particulate matter (PM) to their sources, which will be more directly linked to policy implementations. The flexibility of the approach will ensure that it can be adapted to increased complexity in the data, as it could incorporate possible nonlinear relationship in the source profile exposure and health outcomes and distributed lags; a main characteristic of the framework regards the quantification of uncertainty associated with source profiles and its propagation into the health model.

The dynamicity (in space or time) of our approach will allow to disentangle the potential differential effect of sources at different locations and time points, which has never been attempted before. For instance, we will be able to identify point sources (e.g. factories) or diffuse contributions (e.g. urban traffic), which will lead to discussions about permits, fines, policies (e.g. ultra low emission zone) to directly control those sources. It is important to note that there might be large uncertainties about the extent and timing of pollution; a main advantage of our approach is the ability to account for and propagate these uncertainties throughout the model. We will communicate the level of uncertainty in the estimates visually through plots and maps.

We will build a toolbox with a friendly interface (using the R shiny web interface) which will allow users to upload their data, choose the model to run (depending on the space or time structure of the data) and to get the results in the form of a report with maps/plots and tables. Through the toolbox the research from the project will be freely available to a wide range of applied researchers, thus with the potential of becoming the state-of-the-art method to carry out source apportionment analyses and to link sources with health outcomes in epidemiological studies. We will advertise the toolbox on the MRC Centre for Environment and Health website (http://www.environment-health.ac.uk/) as well as on our personal webpages.

We will disseminate the results from our framework through the university press offices which will allow reaching a wide audience and being in contact with the media. Through the involvement of Public Health England in our research team and through the advisory board we will have direct access to policy makers across environmental and public health field. We have devoted a work package (WP5) to the translation of our methods and results to decision-making and will run two workshops (as outlined in the case for support and in the Communication plan) targeting specific stakeholders and policy makers, as well as the wider scientific community, to present the results and to discuss ways to inform policy implementation.
Finally, the developed framework will serve as a methodological platform which will be then transferable to the investigation of other environmental exposures. For instance, we could use the developed methodology for pollution in water or soil, or even for estimating probabilistic clusters of green and blue spaces and to evaluate their association with mental wellbeing.