Spatio-Temporal Data Modelling and Prediction of PFOA Drinking Water Contamination

Abstract

TPS 931: Water and foodborne chemicals, Exhibition Hall, Ground floor, August 28, 2019, 3:00 PM - 4:30 PM Perfluorooctanoic Acid (PFOA) is suspected to have various adverse effects on human health. For the general population of large parts of North Rhine-Westphalia (NRW), Germany, contaminated drinking water can be regarded as an important pathway for exposure to PFOA. We model its state-wide concentrations in the course of time, thereby handling missing data. Following a PFOA contamination incident near the city of Arnsberg prior to 2006, concentrations in drinking water have been derived from a monitoring programme of the NRW state environmental agency (LANUV) and by further acquisition from water supply companies. Samples have been drawn from both the water supply stations and the network of water supply areas. Apart from such comprehensive data, there are regions with no or few measurements, usually non-detects. In order to estimate and to predict PFOA concentration means and standard deviations, we formulate spatio-temporal models based on these data, some information about (non-)contamination and spatial correlations via rivers. The water proportions in the complex relationship of 417 stations supplying 451 areas are estimated based on their supplied and demanded amounts. These weights are used to mediate between station and area level modelling. The correlation of spatio-temporally adjacent measurements are handled by a Markov-random-field (MRF) representing the neighbourhood structure and by smoothing approaches. We define spatial distances based on the stations’ connectivity via rivers. Spatio-temporal distances from polluted sites, particularly from the Arnsberg contamination, are included as covariates. Monthly predictions from 2006 to 2016 are accomplishable, whereby a MRF turns out to be computationally demanding. Restriction to stations with higher water amounts is a practicable option. Regional conditions have to be respected, as smoothing is not appropriate in regions or periods with sparse data. For the mainly affected region, a contamination decline along the river Ruhr and over time is reflected in the models.