USING BAYESIAN HIERARCHICAL MODELLING TO PRODUCE HIGH RESOLUTION MAPS OF AIR POLLUTION IN THE EU Gavin Shaddick University of Bath RSS Avon Local Group.

Slides:



Advertisements
Similar presentations
Karin Larsson. Exposure, Disease and Risk. Presentation of some case studies using geographical information systems (GIS) Occupational and Environmental.
Advertisements

Combining Information from Related Regressions Duke University Machine Learning Group Presented by Kai Ni Apr. 27, 2007 F. Dominici, G. Parmigiani, K.
S Larssen: PM-PP-Stockholm-Oct-2003.ppt slide 1 PM in Europe - State and past trends Emissions and concentration levels Steinar Larssen Norwegian Institute.
BACKGROUND Benzene is a known carcinogen. Occupational exposure to benzene is an established risk factor for leukaemia. Less is known about the effects.
Halûk Özkaynak US EPA, Office of Research and Development National Exposure Research Laboratory, RTP, NC Presented at the CMAS Special Symposium on Air.
Fighting the Great Challenges in Large-scale Environmental Modelling I. Dimov n Great challenges in environmental modelling n Impact of climatic changes.
U.S. EPA Office of Research & Development October 30, 2013 Prakash V. Bhave, Mary K. McCabe, Valerie C. Garcia Atmospheric Modeling & Analysis Division.
Nicky Best and Chris Jackson With Sylvia Richardson Department of Epidemiology and Public Health Imperial College, London
Indicators for policy support of atmosphere related environmental problems Robert Koelemeijer National Institute for Public Health and the Environment.
PM mapping in Scotland, 2007 Andrew Kent. What are we presenting today? 1) Context to the work 2) Modelling process 3) Model results 4) Future work possibilities.
RAINS review 2004 The RAINS model: Health impacts of PM.
LINKING EUROPEAN, NATIONAL & CITY SCALES UK National Focal Centre for Integrated Assessment Modelling Helen ApSimon and Tim Oxley, Imperial College in.
WORKING GROUP I MONITORING DATA ANALYSIS AND INTERPRETATION TFMM Workshop, Paris, 2006, Nov 29 –Dec 1.
Statistics, data, and deterministic models NRCSE.
GIS in Spatial Epidemiology: small area studies of exposure- outcome relationships Robert Haining Department of Geography University of Cambridge.
Using ranking and DCE data to value health states on the QALY scale using conventional and Bayesian methods Theresa Cain.
Jenny Stocker, Christina Hood, David Carruthers, Martin Seaton, Kate Johnson, Jimmy Fung The Development and Evaluation of an Automated System for Nesting.
Near Real Time analyzed maps of air quality in Europe Cécile Honoré, Laurence Rouïl.
NERAM 2006 Matching the metric to need: modelling exposures to traffic- related air pollution for policy support David Briggs, Kees de Hoogh and John Gulliver.
Simulation of European emissions impacts on particulate matter concentrations in 2010 using Models-3 Rob Lennard, Steve Griffiths and Paul Sutton (RWE.
The ARTEMIS tools for estimating the transport pollutant emissions Artemis project - EC DG Tren COST346 - Heavy duty vehicles emissions M. André, INRETS,
Investigating the Links Between GMES & Emission Inventories - An Update Justin Goodwin: ETC-ACC Supported by EEA TFEIP Meeting Stockholm 3th May 2011.
Chris Jackson With Nicky Best and Sylvia Richardson Department of Epidemiology and Public Health Imperial College, London
Slide 1 Estimating Performance Below the National Level Applying Simulation Methods to TIMSS Fourth Annual IES Research Conference Dan Sherman, Ph.D. American.
Incorporating heterogeneity in meta-analyses: A case study Liz Stojanovski University of Newcastle Presentation at IBS Taupo, New Zealand, 2009.
Small Area Health Insurance Estimates (SAHIE) Program Joanna Turner, Robin Fisher, David Waddington, and Rick Denby U.S. Census Bureau October 6, 2004.
Network for the support of European Policies on Air Pollution The assessment of European control measures and the effects of non-linearities.
RAINS Review Review of the RAINS Integrated Assessment Model Contract with CAFE Dec Sept 2004 Presentation 27 Sep 2004.
17 May 2007RSS Kent Local Group1 Quantifying uncertainty in the UK carbon flux Tony O’Hagan CTCD, Sheffield.
P. Otorepec, M. Gregorič IVZ RS Use of rutinely collected air pollution and health data on local level for simple evaluation of health impact.
| Folie 1 Assessment of Representativeness of Air Quality Monitoring Stations Geneva, Wolfgang Spangl.
BACKGROUND Benzene is a known carcinogen. Occupational exposure to benzene is an established risk factor for leukaemia. Less is known about the effects.
Implementing the Analysis Information System IN 2004 In the sub Saharan region of Africa In the Northern Africa region WHY This difference of level? Overall.
Uncertainty assessment in European air quality mapping and exposure studies Bruce Rolstad Denby, Jan Horálek 2, Frank de Leeuw 3, Peter de Smet 3 1 Norwegian.
1 Understanding and Measuring Uncertainty Associated with the Mid-Year Population Estimates Joanne Clements Ruth Fulton Alison Whitworth.
Climate Change – Defra’s Strategy & Priorities Dr Steven Hill Department for Environment Food and Rural Affairs 22 nd May 2007 FLOODING DESTRUCTION AT.
Characterizing Rural England using GIS Steve Cinderby, Meg Huby, Anne Owen.
Analysis of station classification and network design INERIS (Laure Malherbe, Anthony Ung), NILU (Philipp Schneider), RIVM (Frank de Leeuw, Benno Jimmink)
M. Houssiau | EIONET AQ | Ljubljana – 5 October 2015 Cross analysis between urban system and air quality Provisional results ETC/Urban Land Soil systems.
Chemical pollution of atmospheric air OSIPOVA NINA The Lecturer of Geoecology and Geochemistry Department, Ph.D in Сhemistry; MATVEENKO IRINA The lecturer.
Exposure Assessment for Health Effect Studies: Insights from Air Pollution Epidemiology Lianne Sheppard University of Washington Special thanks to Sun-Young.
Arno Graff Susan Klose UBA-II 4.2 – Air Quality Assessment Some Aspects on Air Quality in Germany related to SOER 14th EIONET Meeting, Warsaw,
INTEGRATING SATELLITE AND MONITORING DATA TO RETROSPECTIVELY ESTIMATE MONTHLY PM 2.5 CONCENTRATIONS IN THE EASTERN U.S. Christopher J. Paciorek 1 and Yang.
DAPPLE Science Meeting 30 th November 2004 Department of Environmental Science & Technology Imperial College London, UK EXPOSURE ASSESSMENT S Kaur, M Nieuwenhuijsen,
Impact of Temporal Fluctuations in Power Plant Emissions on Air Quality Forecasts Prakash Doraiswamy 1, Christian Hogrefe 1,2, Eric Zalewsky 2, Winston.
DIAS INFORMATION DAY GLOBAL WATER RESOURCES AND ENVIRONMENTAL CHANGE Date: 09/07/2004 Research ideas by The Danish Institute of Agricultural Sciences (DIAS)
By Russ Frith University of Alaska at Anchorage Civil Engineering Department Estimating Alaska Snow Loads.
Markus Amann International Institute for Applied Systems Analysis Cost-effectiveness Analysis in CAFE and the Need for Information about Urban Air Quality.
1 Spatial assessment of deprivation and mortality risk in Nova Scotia: Comparison between Bayesian and non-Bayesian approaches Prepared for 2008 CPHA Conference,
Konstantina Dimakopoulou, Evangelia Samoli, Rob Beelen, Massimo Stafoggia, Zorana Jovanovic Andersen,Barbara Hoffmann etc. Air Pollution and Nonmalignant.
Chiara Badaloni Roma, 15 Oct EXPOSURE TO TRAFFIC AIR POLLUTION IN A CASE-CONTROL STUDY OF CHILDHOOD LEUKAEMIA.
17 th TFMM Meeting, May, 2016 EMEP Case study: Assessment of HM pollution levels with fine spatial resolution in Belarus, Poland and UK Ilia Ilyin,
The application of Models-3 in national policy Samantha Baker Air and Environment Quality Division, Defra.
Adam Butler, Stijn Bierman & Glenn Marion Biomathematics & Statistics Scotland CEH Bush, April 2008 ALARM: a statistical perspective.
Evaluation of pollution levels in urban areas of selected EMEP countries Alexey Gusev, Victor Shatalov Meteorological Synthesizing Centre - East.
Exposure Prediction and Measurement Error in Air Pollution and Health Studies Lianne Sheppard Adam A. Szpiro, Sun-Young Kim University of Washington CMAS.
1 Černikovský, Krejčí, Volná (ETC/ACM): Air pollution by ozone in Europe during the summer 2012 & comparison with previous years 17th EIONET Workshop on.
Skudnik M. 1*, Jeran Z. 2, Batič F. 3 & Kastelec D. 3 1 Slovenian Forestry Institute, Ljubljana, Slovenia 2 Jožef Stefan Institute, Ljubljana, Slovenia.
Joint EMEP/WGE meeting, Geneva, 2016 Evaluation of B[a]P pollution in the EMEP region: temporal trends and spatial variability Alexey Gusev, Olga Rozovskaya,
Study of PM10 Annual Arithmetic Mean in USA  Particulate matter is the term for solid or liquid particles found in the air  The smaller particles penetrate.
N Engl J Med Jun 29;376(26): doi: 10
Joint thematic session on B(a)P pollution: main activities and results
Predicting PM2.5 Concentrations that Result from Compliance with National Ambient Air Quality Standards (NAAQS) James T. Kelly, Adam Reff, and Brett Gantt.
Exposure to Air Pollution James Tate and Paul Seakins
Jan Horálek (CHMI) Peter de Smet, Frank de Leeuw (RIVM),
Impact of Land use on Air pollution in Austin
Steve Griffiths, Rob Lennard and Paul Sutton* (*RWE npower)
Alexey Gusev, Victor Shatalov, Olga Rozovskaya, Nadejda Vulyh
MSC-E: Alexey Gusev, Victor Shatalov, Olga Rozovskaya, Nadejda Vulykh
EMEP Case study: Assessment of HM pollution levels with fine spatial resolution in Belarus, Poland and UK Ilia Ilyin, Olga Rozovskaya, Oleg Travnikov.
Presentation transcript:

USING BAYESIAN HIERARCHICAL MODELLING TO PRODUCE HIGH RESOLUTION MAPS OF AIR POLLUTION IN THE EU Gavin Shaddick University of Bath RSS Avon Local Group October 2006

Air Pollution Modelling for Support to Policy on Health, Environment and Risk Management in Europe A PMoSPHERE APMoSPHERE is a thematic project, funded under the Global Monitoring for Environment and Security initiative, as part of the European Union’s Fifth Research Framework Programme. Its aim is to compile high resolution maps of air pollution across the EU, as a basis for scientific research and policy support.

Air Pollution Modelling for Support to Policy on Health, Environment and Risk Management in Europe A PMoSPHERE APMoSPHERE is a thematic project, funded under the Global Monitoring for Environment and Security initiative, as part of the European Union’s Fifth Research Framework Programme. Its aim is to compile high resolution maps of air pollution across the EU, as a basis for scientific research and policy support.

Partners A PMoSPHERE Department of Epidemiology and Public Health, Imperial College Institute for Risk Assessment Sciences, University of Utrecht Institute for Environmental Research and Sustainable Development, National Observatory of Athens Centre for International Climate and Environmental Research, Oslo Department of Mathematical Sciences, University of Bath AEA Technology Netcen

What APMoSPHERE will do Key objectives of APMoSPHERE are:  to produce a detailed (1km) inventory of atmospheric emissions by major sector for the EU  to develop and test a range of different methods for mapping air pollution on the basis of these emissions estimates, in combination with other routinely available data sets (including air pollution monitoring data)  using these various methods and data sets to generate detailed (1km) and updatable maps of air pollution, together with a set of policy-related indicators on potential ecological and health risks  based on these results, to provide an assessment of the air pollution situation in the EU, and implications for future air quality monitoring and policy The pollutants: Particulates (PM 10 and black smoke) Nitrogen oxides (NO x and NO 2 ) Carbon monoxide Sulphur dioxide Ozone

What APMoSPHERE will do Key objectives of APMoSPHERE are:  to produce a detailed (1km) inventory of atmospheric emissions by major sector for the EU  to develop and test a range of different methods for mapping air pollution on the basis of these emissions estimates, in combination with other routinely available data sets (including air pollution monitoring data)  using these various methods and data sets to generate detailed (1km) and updatable maps of air pollution, together with a set of policy-related indicators on potential ecological and health risks  based on these results, to provide an assessment of the air pollution situation in the EU, and implications for future air quality monitoring and policy Particulates (PM 10 and black smoke) Nitrogen oxides (NO x and NO 2 ) Carbon monoxide Sulphur dioxideSulphur dioxide Ozone

Geographic Information System Study Area EU15 + Norway Concentration data AIRBASE & EMEP * 1 km predictors Topography Meteorology Roads * Land cover * Light intensity * Modelled Emissions * Population data 1 km modelled population*

Aims Provide modelled exposures (and measures of uncertainty). Provide modelled exposures (and measures of uncertainty). Impute missing values Impute missing values Unmeasured locations Unmeasured locations Combine information from multiple sources Combine information from multiple sources Investigate the spatio-temporal modelling of pollutants. Investigate the spatio-temporal modelling of pollutants. Assessing the contribution of spatial, temporal and random variabilty. Assessing the contribution of spatial, temporal and random variabilty.

Data dependencies  Relationship with covariates Climate, e.g. temperature Climate, e.g. temperature Local emissions, e.g. land cover Local emissions, e.g. land cover Topography, e.g. altitude Topography, e.g. altitude Temporal dependencies. Temporal dependencies. Spatial dependencies. Spatial dependencies. Distance between monitoring sites. Distance between monitoring sites. Site type (e.g. background, traffic). Site type (e.g. background, traffic).

Model framework Bayesian Hierarchical Model. Bayesian Hierarchical Model. Pollutants (log) modelled as a function of the ‘true’ underlying level with unstructured error. Pollutants (log) modelled as a function of the ‘true’ underlying level with unstructured error. Incorporate covariate information Incorporate covariate information True underlying level is a function of the previous year’s level. True underlying level is a function of the previous year’s level. Missing values treated as unknown parameters within the Bayesian framework and can be estimated. Missing values treated as unknown parameters within the Bayesian framework and can be estimated.

Priors Information from previous studies or years Information from previous studies or years Expert opinion Expert opinion Physical science Physical science ‘vague’ ‘vague’

Posteriors and parameter estimates In simple cases, e.g. where both prior and likelihood are conjugate, exact expressions for the posterior distributions can be found In simple cases, e.g. where both prior and likelihood are conjugate, exact expressions for the posterior distributions can be found In more complex cases, the posterior may be intractable In more complex cases, the posterior may be intractable Can use simulation to ‘build up’ the posterior Can use simulation to ‘build up’ the posterior MCMC (WinBUGS) MCMC (WinBUGS)

Model stages Level 1 : Observed data stage. Level 1 : Observed data stage. Y t =  t + covariates + site effect + v t, v t ~ N(0,  v ) Y t =  t + covariates + site effect + v t, v t ~ N(0,  v ) Level 2(a) : Temporal/system stage. Level 2(a) : Temporal/system stage.  t = α  t-1 + w t, w t ~ N(0,  w )  t = α  t-1 + w t, w t ~ N(0,  w ) Level 2(b) : Spatial stage Level 2(b) : Spatial stage Site random effects modelled as multivariate normal with correlations proportional to the distance,d, between sites. Site random effects modelled as multivariate normal with correlations proportional to the distance,d, between sites. f(d) = exp(-  d) f(d) = exp(-  d) Site effects can be estimated at unmeasured locations conditional on the measured values. Site effects can be estimated at unmeasured locations conditional on the measured values. Level 3 : Hyperparameters. Level 3 : Hyperparameters. Assign prior distributions to covariate effects and variances. Assign prior distributions to covariate effects and variances.

Prior information For the spatial effect For the spatial effect Φ given a uniform (1.3-4) distribution Φ given a uniform (1.3-4) distribution Corresponds to correlations falling to between 0.13 and 0.52 at a distance of 50km Corresponds to correlations falling to between 0.13 and 0.52 at a distance of 50km Normal distributions for covariate effects Normal distributions for covariate effects Gamma distributions for (inverse of) variances [precisions] Gamma distributions for (inverse of) variances [precisions]

Results UK data for SO2, UK data for SO2,

Components of variation Random (unstructured) error – 26% Random (unstructured) error – 26% Temporal – 13% Temporal – 13% Spatial – 61% Spatial – 61% Random error

Posterior estimates – temporal components

Spatial effects Posterior median for Φ : 3.79, 95% CrI ( ) Posterior median for Φ : 3.79, 95% CrI ( )

Predictions for UK Overall mean + temporal (2001) effect + covariate effect + spatial effect

Extending methodology to EU level  Increased number of sites brings large computational burden Following analysis performed on NO2 in 2001 Following analysis performed on NO2 in % dataset (sites) used to build models 75 % dataset (sites) used to build models 25 % for validation 25 % for validation

Modelling at different scales Based on theoretical and empirical environmental models Based on theoretical and empirical environmental models Variograms Variograms Scales defined by site type and associated covariates Scales defined by site type and associated covariates Global (climate and topological) Global (climate and topological) Rural (transport, population density, agriculture) Rural (transport, population density, agriculture) Urban (transport, population density, urban greenery) Urban (transport, population density, urban greenery) Eases computational burden Eases computational burden

Covariates

Model stages Global model Global model Y Gs =  G + global covariates S + site effects + v Gs, v Gs ~ N(0,  2 G ) Y Gs =  G + global covariates S + site effects + v Gs, v Gs ~ N(0,  2 G ) Rural model Rural model (Y Rs – predicted(Y Rs ) ) =  R + rural covariates S + v Rs, v Rs ~ N(0,  2 R ) (Y Rs – predicted(Y Rs ) ) =  R + rural covariates S + v Rs, v Rs ~ N(0,  2 R ) Urban model Urban model (Y Us – predicted(Y Us ) ) =  U + urban covariates S + v Us, v Us ~ N(0,  2 U ) (Y Us – predicted(Y Us ) ) =  U + urban covariates S + v Us, v Us ~ N(0,  2 U ) Predictions were made using the global models for every one of the 1km x 1km cells ( ) Predictions were made using the global models for every one of the 1km x 1km cells ( ) additional effects of rural ( cells) additional effects of rural ( cells) urban (65662 cells) urban (65662 cells) used to create an further two sets of predictions which were then combined to create a composite map. used to create an further two sets of predictions which were then combined to create a composite map.

Results – global model Increases with distance Increases with distance from sea and for climate from sea and for climate variables 2 & 5 – areas variables 2 & 5 – areas with warm or hot summers with warm or hot summers  Decreases with altitude Posterior median for , 0.037, corresponds to fall in correlation to at 100km Posterior median for , 0.037, corresponds to fall in correlation to at 100km Without any geograpahical covariates,  much smaller (by factor of ten), indicating much more ‘spatial’ residual error Without any geograpahical covariates,  much smaller (by factor of ten), indicating much more ‘spatial’ residual error

Results – rural and urban Rural - significant effect of major roads Rural - significant effect of major roads Urban - clear overall increase (intercept term) Urban - clear overall increase (intercept term)  transport (major, minor roads) minor roads)  population density  negative association with altitude with altitude

Pollutant NO 2 Scale Composite of global, rural and urban background Time period 2001, annual average Geographic extent Excludes Norway and Sweden Statistics (ug/m 3 ) Min0.45 Max Mean12.47 Std dev5.64 Modeling method Bayesian Hierarchical Modelling Model

Pollutant NO 2 Scale Composite of global, rural and urban background Time period 2001, annual average Geographic extent Excludes Norway and Sweden Statistics (ug/m 3 ) Min1.66 Max Mean19.19 Std dev9.04 Length of 95% credible interval

Validation Performed at each scale Performed at each scale (global, rural, urban) (global, rural, urban)  RSME, MAbsE, R 2, etc…  Best results for NO 2, PM 10 and O 3 and O 3  Best results for urban scale (relationships scale (relationships with covariates) with covariates)  exception of O 3

Summary Applied spatial-temporal model to ca. 200 sites measuring SO 2 in UK ( ). Applied spatial-temporal model to ca. 200 sites measuring SO 2 in UK ( ). Assessed proportions of spatial, temporal and random variation Assessed proportions of spatial, temporal and random variation Applied spatial model to entire EU Applied spatial model to entire EU Produced predicted levels at 1km resolution for different scales Produced predicted levels at 1km resolution for different scales Produced composite maps with measures of uncertainty Produced composite maps with measures of uncertainty

Future work/considerations  Combined spatial models different site types modelling simultaneously different site types modelling simultaneously Computational aspects Computational aspects Estimation and (joint) prediction Estimation and (joint) prediction Sensitivity analysis (to priors) Sensitivity analysis (to priors) Conditional modelling Conditional modelling Neighbouring sites Neighbouring sites  Other pollutants  multi-pollutant models

More information on APMoSPHERE

Alternative approach – conditional modelling Problems handling large spatial matrices at such a high resolution. Problems handling large spatial matrices at such a high resolution. Define sites as having ‘neighbours’ (may include distance cut-off). Define sites as having ‘neighbours’ (may include distance cut-off). Allows feasibility of different resolutions to be examined. Allows feasibility of different resolutions to be examined. Can be much, much faster! Can be much, much faster! Prediction and estimation may performed together during the MCMC. Prediction and estimation may performed together during the MCMC.

Conditional model Y s ~ N(S s,v) Y s ~ N(S s,v) S s = β + W s S s = β + W s W s ~ N(ρ Σ i in δs W s /n s, n s τ) W s ~ N(ρ Σ i in δs W s /n s, n s τ) Where Σ i in δs W s /n s is the average of the neighbours of point s. Where Σ i in δs W s /n s is the average of the neighbours of point s. The number of points that constitute the neighbourhood can be varied The number of points that constitute the neighbourhood can be varied

A 100km resolution structure with 10 neighbours 372 unknown points

Predicted SO2

Variability

Higher resolutions Example of 50km resolution 418 known 1469 unknown points

Computational aspects 100,000 iterations with ca. 400 sites 100,000 iterations with ca. 400 sites Joint model – 5 days Joint model – 5 days Conditional model – 30 minutes Conditional model – 30 minutes Using 2.5GHZ PC with 1GB RAM Using 2.5GHZ PC with 1GB RAM Using conditional model with observed and prediction points together at 20km Using conditional model with observed and prediction points together at 20km 1 day (1000 iterations – 15 minutes) 1 day (1000 iterations – 15 minutes) Higher resolutions computationally feasible (but problems writing the file!) Higher resolutions computationally feasible (but problems writing the file!)