Statistical Methods and State of the Techniques in Exposure Modeling - - PowerPoint PPT Presentation

Statistical Methods and State of the Techniques in Exposure Modeling

Howard Chang

Department of Biostatistics and Bioinformatics Emory University howard.chang@emory.edu

Different Exposure Modeling Goals

Past

• Estimate exposure retrospectively
• Linkage with health data

Current

• Health + exposures measured together
• Dimension reduction

Future

• “What if” assessment
• Quantifying uncertainties

Representative Statistical Methods

Past

• Data fusion

Current

• Source apportionment

Future

• Climate projection bias-correction

• A. Data Fusion

Observation measurements Proxy Data Computer model output Satellite imagery Stochastic simulation Land use/meteorology

Assimilation, Integration, Calibration, Melding, Blending

• Long-term or short-term pollutant predictions
• Uncertainty measures

Computer Model versus Satellite Retrievals

Computer Model Satellite Pollutants Major pollutants PM2.5, NO2, SO2, CO, O3 Resolution 4 ~ 50km gridded 1 ~ 25km pixel Frequency Hourly 1 or 2 samples/day Why biased? Incorrect inputs Chemistry Discretization Not ground-level Retrieval error Missing data Availability Open Source Publically available Computation Heavy run time Post-processing

Statistical Data Fusion (Downscaling)

𝑷𝒄𝒕 𝒕, 𝒖 = 𝜸𝟏 𝒕, 𝒖 + 𝜸𝟐 𝒕, 𝒖 𝑸𝒔𝒑𝒚𝒛 𝒕, 𝒖 + 𝜻 (𝒕, 𝒖)

Space-time additive bias Space-time multiplicative bias

For day t at monitoring location s:

Data Fusion Example: CMAQ

Spatial Bias in CMAQ

Longitude Latitude

31 32 33 34 35 36 37

• 86
• 84
• 82
• 1.5
• 1.0
• 0.5

0.0 0.5 1.0 1.5

Longitude Latitude

31 32 33 34 35 36 37

• 86
• 84
• 82
• 0.15
• 0.10
• 0.05

0.00 0.05 0.10 0.15 0.20

Additive Multiplicative

Predicted Concentrations

Longitude Latitude

31 32 33 34 35 36 37

• 86
• 84
• 82

Raw

31 32 33 34 35 36 37

Calibrated

5 10 15 20 25

Longitude Latitude

31 32 33 34 35 36 37

• 86
• 84
• 82

Raw

31 32 33 34 35 36 37

Calibrated

2 4 6 8 10 12 14 16

2004-01-01 2004-07-01

Does Data Fusion Help?

CMS = Central Monitor PWA = Spatial Kriging + Population Weighting FSD = Data Fusion + Population Weighting Asthma Emergency Department Visits in Atlanta

• B. Source Apportionment

UNOBSERVED pollutant mixture OBSERVED single- pollutant concentrations (20+)

Both the sources and the profiles are unknown.

Biomass Burning Traffic Dust/Soil

Example Source Apportionment

Change in FEV1 associated with each interquartile (a) immediately after commutes (b) 1h after commutes (c) 2h after commutes (d) 3h after commutes Golen et al. In prep

Source Apportionment Methods

Chemical Mass Balance Assume profiles known Latent Variable Assume profiles unknown + label latent variables

Cannot empirically validate results!

Numerical optimization Gas-/marker constraints Absolute principle Component Positive matrix factorization Bayesian factor analysis

Variability Across SA Methods

5 10 15 20 25 30 1 2 3 4 5 6 7 8 9 10

July 2001 Source Impact (g m-3) SOC, July 2001, Equal Weighting

CMB-RG CMB-LGO PMF CMB-MM CMAQ

BIOMASS BURNING

15

Ensemble-Trained Source Apportionment

PMF CMAQ CMB-MM CMB-GC Balachandran et al. Env Sci Tech 2013; 47: 13511-13518 Gass et al. Am J Epidemiol 2015; 181: 504-512. Biomass Burning

16

Data Fusion for Source Apportionment

Winter Biomass Burning Original Adjusted

Adjusted - Original

Adjusted - Original Original Adjusted

Summer Coal Combustion

Ivey et al. Geosci Model Dev 2015; 8: 2153-2165.

17

• C. Bias-Correction for Climate Model

Climate Model Simulations

Air Quality Model Disease Transmission Model Hydrological Model

18

Quantile-Mapping

15 20 25 30 0.00 0.05 0.10 0.15 0.20 Density

𝐏𝐜𝐭𝐟𝐬𝐰𝐟𝐞 Temperature

𝐃𝐦𝐣𝐧𝐛𝐮𝐟 𝐍𝐩𝐞𝐟𝐦 𝐔𝐟𝐧𝐪𝐟𝐬𝐛𝐮𝐯𝐬𝐟

Hindcast Period

19

Future Projection (2041-2070)

20

Health Impacts with Bias-Correction

Bias-correction reduces between-model variability

Chang et al. Atmos Env 2014; 89: 290-297.

Advances in Statistical Data Fusion

• Multiple pollutants
• Proxy “features”
• Multiple proxies
• Stochastic proxies
• Methods for massive spatial datasets

22

Additional Remarks

What method to use depends on:

• Study design
• Health outcome
• Exposure of interest

New and better data products are being developed and often are publically available. Lots of research activities but limited comparison and synthesis.