Global climate change affects human health most notably by increasing the frequency and intensity of dangerous heat waves, wildfires and hurricanes. In addition to extreme weather events, climate change can also lead to a myriad of persistent environmental changes that impact public health. Health impact assessment refers to the analytic framework for evaluating how a policy or program affects population health. It is frequently applied in climate and public health research to quantify future health and economic burdens attributable to various consequences of climate change. Performing health impact assessment entails the integration of various data. For projecting future climate-related health impacts, analyses require three sources of information: (1) health effects of environmental exposures, (2) projections of future exposures, and (3) distributions of exposures and effects in the future population. Each information source is subject to uncertainty because of data availability and assumptions made for the future. Climate research is highly interdisciplinary, bringing together tremendous amount of data, theory, and modeling efforts to provide timely knowledge for one of the most pressing issues of our time. Statistical modeling techniques and probabilistic reasoning can plan an important role in ensuring these findings are informative, accurate, and reproducible. This presentation will discuss recent development in statistical methods for quantifying health impacts of climate change, as well as related open problems in environmental epidemiology and exposure assessment.

1. Linking Environmental and Health Data for
Epidemiologic Research
Howard Chang
howard.chang@emory.edu
Department of Biostatistics and Bioinformatics, Emory University
SAMSI CLM Opening Workshop
August 2017

2. Health Impact Assessment
Analytic framework for evaluating how a policy
or program affects population health.
Health
Data
Weather
Data
1. Health Effect
Estimation
Population
Projection
Climate
Simulations
2. Bias Correction
3. Health Impact
Projection

3. Health Effects of Environmental Risks
Exposures Health Outcomes
• Temperature
• Air pollution
• Heat waves
• Wildfire
• Heavy rainfall
• Pollen
• Mold
• Salt water intrusion
• Mortality
• Hospital admissions
• Emergency room visits
• Adverse birth outcomes
• Diarrheal diseases
• Vector-born diseases
What are the relevant spatial and temporal scale?

4. Challenges in Environmental Epidemiology
• Observational and retrospective.
• Small effects but ubiquitous exposures.
• Data are complex, incomplete, and messy.
• Both health and exposure data have spatial and
temporal dependence.
• Focus on causal inference: bias, measurement
error and confounding,

5. Health Effect Estimation – Health Data
Administrative Databases
• Death and birth certificates
• Medicare and hospital billing records
• Medical records
Advantages
• Individual-level data
• Centralized and cost efficient
Disadvantages
• Missing important individual-level confounders
• Crude geographical information/history

6. Some Research Interests
• Identify susceptible and vulnerable populations.
• Improve exposure assessment methods for
health effect/impact analyses.
• Estimating joint effects of multiple exposures.
• Evaluate environmental regulations/policies
(accountability) and mitigation strategies.

7. Atlanta Emergency Department Visits
• 20-county Atlanta
metropolitan area.
• Billing records of ED visits
between 1993 to 2012.
• ≈ 10 million individual
records

8. Daily Time-Series Analysis
2002 2004 2006 2008 2010
2000300040005000
Daily Emergency Department Visits in Atlanta
Counts
2002 2004 2006 2008 2010
010203040
Daily Maximum Temperature
Calendar Date
Temperature(Celsius)

9. Daily Time-Series Analysis
Daily total ED counts
Daily
Temperature
(e.g. ATL Airport)
Seasonal trend
Long-term trend
Day-of-week
Proxies for:
Diet
Access to care
Population health
log[𝐸 𝑌𝑡 ] =𝛽𝑋𝑡 + Confounders

10. ED Visits and Temperature
Broad Outcome Groups among the Elderly
INTERN GI DIA FEI CIRC RESP REN
AL
INTERN = all internal causes
GI = intestinal infections
DIA = diabetes
FEI = fluid & electrolyte
imbalance
CIRC = all circulatory diseases
RESP = all respiratory
diseases
RENAL = all renal diseases
*Associations of lag 0 TMX and ED visit outcomes based on primary ICD-9 codes
*RRs for TMX changes from 27 oC to 32 oC (25th to 75th percentile)
Winquist A, Grundstein AJ, Chang HH, Hess J Sarnat SE (2016). Environmental Research 147, 314-323.

11. ED Visits and Heat Waves
Heat wave definitions:
• Heat wave period = ≥2 consecutive days with daily metric
beyond the 98th percentiles.
• The first day of a heat wave period treated removed.
• Max/Avg/Min temperature and apparent temperature.
The added burden of sustained high temperature.

12. ED Visit-Heat Wave, Atlanta 1993-2012
Chen T, Sarnat SE, Grundstein AJ, Winquist A, Chang HH (2017). Environmental Health Perspectives. DOI:10.1289/EHP44

13. Heat Wave Statistics
Upper triangle = union days; lower triangle = intersection days
Heat
Wave
Def
MAXT MINT AVGT MAXAT MINAT AVGAT
MAXT 256 133 122 150 143
MINT 67 (26%) 251 259 232 245
AVGT 81 (61%) 104 (41%) 141 154 147
MAXAT 78 (64%) 82 (32%) 91 (65%) 154 139
MINAT 37 (25%) 96 (41%) 65 (42%) 51 (33%) 139
AVGAT 66 (46%) 105 (43%) 94 (64%) 88 (63%) 75 (54%)
Concordance-discordance between heat wave metrics.

14. Preterm Birth-Heat Wave, Atlanta 1994-2006
0.6
1.2
RR(95%ConfidenceInterval)
0.8
1.4
1.0
Daily counts of preterm birth with joint strata of gestational week,
maternal race, and maternal education, and offset by the number of
fetuses-at-risk of preterm birth on each day.
Darrow LA, Strickland MJ, Chang HH (2015) Society for Epidemiologic Research Annual Meeting. Denver, Colorado.

15. Pediatric ED Asthma Visit and Temperature
O’Lenick CR, Winquist A, Chang HH, Kramer MR, Mulholland JA, , Grundstein AJ, Sarnat SE (2017) Environmental Research,
156, 132-144.

16. Spatial Exposure Uncertainty

17. Data Integration for Exposure Assessment
Exposure
Monitoring
Measurement
Satellite
Imagery
Computer
Model
Simulation
Meteorology
Physical
Variables

18. Fine-Scale Temperature Assessment
Daytime Nighttime
1 km spatial resolution over metropolitan Houston for daytime (left; 4:00 pm)
and nighttime (right; 12:00 am), created with the HRLDAS model.
Monaghan AJ, Hu L, Brunsell NA, Barlage M, Wilhelmi OV (2014). Journal of Geophysical Research: Atmospheres, 119(11),
6376-6392.

19. Estimated Average PM2.5 Levels in 2001
(𝜇𝑔/𝑚3
)
1km x 1km spatial grid in 20-county Atlanta

20. Are Finer Exposure Estimates Better?
CMS: central monitoring station
PWA: kriging + population weighted average
FSD: data fusion

21. Health Effect Associations by Poverty
Exposure Metric
% Change in Risk
(95% confidence interval)
ZIP-codes < 8% below
poverty
ZIP-codes > 8% below
poverty
Central site (Daily) 0.30 (-0.52, 1.13) 0.67 (-0.07, 1.42)
Spatial-resolved estimates
without prediction error 0.40 (-0.60, 1.40) 0.90 (0.00, 1.82)
with prediction error 0.38 (-0.61, 1.39) 0.86 (-0.04, 1.78)
Associations between daily ED visits for RESP per 10 μg/m3
increase in previous day PM2.5 ambient concentration, 2001
to 2009.

22. Exposure Measurement Error?
Smoothing  Berkson-type error
Modeling uncertainty  Classical-type error
Georgia birth records 2001-05 Long-term PM2.5 average
Keller J, Chang HH, Strickland MA, Szpiro A (2017). Epidemiology 28(3):338-345

23. Exposure Measurement Error?
Bootstrap
Cohort Correction Trimester Estimate Std. Err 95% CI
Restricted None 1 –1.53 0.82 (–3.14, 0.09)
2 –1.57 0.83 (–3.19, 0.06)
3 –2.37 0.78 (–3.89, –0.83)
Parameter 1 –1.55 1.01 (–3.53, 0.43)
2 –1.56 0.98 (–3.50, 0.37)
3 –2.39 0.92 (–4.20, –0.58)
Statewide None 1 0.281 0.54 (–0.78, 1.34)
2 –0.620 0.55 (–1.69, 0.45)
3 –0.660 0.51 (–1.67, 0.35)
Estimated difference in birth weight (in grams) associatedwith 1
µg/m3 higher ambient average PM2.5 during the specified trimester.
Keller J, Chang HH, Strickland MA, Szpiro A (2017). Epidemiology 28(3):338-345

24. Diarrhea and Environmental Drivers
21 Villages in Ecuador Weekly Diarrhea Incidence
Topographic Data Weather Stations
NSF 1360330 (PI: Remais)

25. Exposure Assessment for Precipitation
0.25 x 0.25 degree grid of the TRMM 3B42 satellite platform.
Deshpande A, Chang HH, Levy K (2017). Health Related Water Microbiology / Water Microbiology Conference.

26. Precipitation and Diarrheal Diseases
0.50
0.75
1.00
1.25
1.50
HRE, Dry AC HRE, Wet AC No HRE, Dry AC
ExpectedCountRatio
Expected Count Ratios of Diarrhea by Environmental Condition
in Rural Parishes at Lag of 7 Days
Associations between:
1. Daily counts of diarrheal disease in Esmeraldas, Ecuador
2. Heavy rain fall events (HRE) by antecedent conditions (AC)

27. Improving Exposure Assessment
• How do combine various data sources?
Monitoring networks
Model simulations
Satellite retrievals
• Analytical challenges
Missing data (spatio-temporal, informative)
Multiple exposures
Large datasets
Exposure prediction error

28. Accountability Research
Henneman LRF, Chang HH, Liao KJ, Lavoue D, Mulholland JA, Russell AG (2017). Air Quality, Atmosphere & Health DOI 10.1007/s11869-017-0463-2

29. Impacts of Emission versus Climate
Stowell JD, Kim YM, Gao Y, Fu JS, Chang HH, Liu Y (2017). Environment International, 108:41-50.

30. Health Effects of Multiple Exposures
X2
Y
X1
X2
Y
X1
YX1 X2
Joint Effects Effect Modification
Mediation

31. Thank you!
NIEHS R01 ES027892, R21 ES022795, R21 ES023763