1. Copyright 2008, The Johns Hopkins University and Francesca Dominici. All rights reserved. Use of these materials permitted only in accordance with license rights granted. Materials provided “AS IS”; no representations or warranties provided. User assumes all responsibility for use, and all liability related thereto, and must independently review all materials for accuracy and efficacy. May contain materials owned by others. User is responsible for obtaining permissions for use from third parties as needed. This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike License. Your use of this material constitutes acceptance of that license and the conditions of use of materials on this site.Creative Commons Attribution-NonCommercial-ShareAlike License

2. How Risky is Breathing? Statistical Methods in Air Pollution Risk Estimation Francesca Dominici Department of Biostatistics Bloomberg School of Public Health Johns Hopkins University

3. From crisis to questions We began with crisis-- the London fog in 1952, and have moved to questions: –Are there adverse effects of today’s air pollution? –How large are these risks?

4. December 5 1952: London's Piccadilly Circus at midday Particulate levels – 3,000 This image has been deleted because JHSPH OpenCourseWare was not able to secure permission for its use.

5. Maureen Scholes, a nurse at the Royal London Hospital in 1952, says the smog penetrated through clothes, blackening undergarments Source: Royal London Hospital Archives and Museum 4,000 deaths the first week 8,000 over next 2 months This image has been deleted because JHSPH OpenCourseWare was not able to secure permission for its use.

6. Designer Smog Masks - London 1950’s Davis When Smoke Ran Like Water (2002) This image has been deleted because JHSPH OpenCourseWare was not able to secure permission for its use.

7. Air pollution and mortality: Then and now London, December, 1952Mortality and PM 10 in Chicago, 2000 This image has been deleted because JHSPH OpenCourseWare was not able to secure permission for its use. "London "Killer" Smog of 1952" from Environmental Health. Available at: http://ocw.jhsph.edu. Copyright © Johns Hopkins Bloomberg School of Public Health. Creative Commons BY-NC-SA. Adapted from Turco, R. P.http://ocw.jhsph.edu

8. Air pollution and health: Fundamental questions Is there a risk at current levels? How can we estimate it? How big is the risk? What causes it? yes The risk is very small but everyone is exposed! ??? By integrating national data sets and developing methods to analyze them

9. Bad air day? Chicago PM 2.5 = 10 This image has been deleted because JHSPH OpenCourseWare was not able to secure permission for its use.

10. Bad air day? Chicago PM 2.5 = 20 This image has been deleted because JHSPH OpenCourseWare was not able to secure permission for its use.

11. Bad air day? Chicago PM 2.5 = 30 Standard setting process in the US is evidence-based This image has been deleted because JHSPH OpenCourseWare was not able to secure permission for its use.

12. National Data Sets

13. National Morbidity Mortality Air Pollution Study Collected data 100 largest cities in the United States –Daily mortality –Daily temperature –Daily level of PM10 Long time series –1987 to 2000

14. The National Medicare Cohort Study, 1999-2005 (MCAPS) Medicare data include: –Billing claims for everyone over 65 enrolled in Medicare (~48 million people), date of service treatment, disease (ICD 9) age, gender, and race place of residence (zip code) Approximately 204 counties linked to the PM 2.5 monitoring network

15. MCAPS study population: 204 counties with populations larger than 200,000 (11.5 million people) This image has been deleted because JHSPH OpenCourseWare was not able to secure permission for its use. Please visit www.biostat.jhsph.edu/MCAPS for maps and other MCAPS informationwww.biostat.jhsph.edu/MCAPS

16. Daily time series of hospitalization rates and PM 2.5 levels in Los Angeles county (1999-2005) This image has been deleted because JHSPH OpenCourseWare was not able to secure permission for its use. Please visit www.biostat.jhsph.edu/MCAPS for maps and other MCAPS informationwww.biostat.jhsph.edu/MCAPS

17. Statistical Ideas

18. 3 Statistical Ideas for Analysis of Observational Studies 1.Adjusting for confounding –Semi-Parametric Regression 2.Combining health risk estimates across counties – Bayesian Hierarchical Models 3.Accounting for the uncertainty in the selection of the statistical model –Model averaging for confounding adjustment

19. Compare day-to-day variations in hospital admission rates with day-to- day variations in pollution levels within the same community Avoid problem of unmeasured differences among populations Key confounders  Seasonal effects of infectious diseases and weather Statistical Methods for multi-site time series studies

20. Statistical Methods Within city: Semi-parametric regressions for estimating associations between day- to-day variations in air pollution and mortality controlling for confounding factors Across cities: Hierarchical Models for estimating: –national-average relative rate –exploring heterogeneity of air pollution effects across the country Dominici Samet Zeger JRSSA 2000

21. Confounding The association between air pollution and mortality is potentially confounded by: –Weather –Other pollutants –Seasonality –Long-term trend

22. 1) Semi-parametric regression model for estimating health risk within a county # of adverse events on day t # of people at risk on day t health risk Time varying confounders: Weather variables seasonality Kelsall Samet Zeger Xu AJE 1997 air pollution series

23. 2)Bayesian hierarchical models for pooling risks across cities County-specific risk estimate County-specific true risk Within-county statistical error Pooled risk Across-county variance of the true risks

24. 3) Do I have the “right” statistical model? Explore the sensitivity of the risk estimates to the statistical model

25. Sensitivity of the national average lag effect of PM 10 on mortality to different statistical models to adjust for confounding (NMMAPS 1987-2000) Peng Dominici Louis JRSSC 2006 Reported estimate Different statistical models to adjust for confounding weakmoderatestrong

26. 3) Do I have the “right” statistical model? XYXY Z1Z1 Z2Z2 Z 1 is a predictor of Y Z 2 is a confounder Regression ModelsWeights based on prediction(BIC) Weights based on ability to adjust for confounding 0.90.0 0.9 0.1 Estimating risks by averaging across statistical models

27. 3) Model averaging for confounding adjustment in observational studies We assign zero weights to models that have optimal prediction properties but that do not include all the potential confounders We identify all the potential confounders by searching for good predictors of the exposure X Theoretical results and simulation studies have showed that this approach outperform existing methods to account for model uncertainty Crainiceanu Dominici Parmigiani Biometrika 2007 Wang Crainiceanu Parmigiani Dominici technical report 2007

28. Biostatistics in Action: The weight of the evidence

29. This image has been deleted because JHSPH OpenCourseWare was not able to secure permission for its use. Full-text available at http://content.nejm.org/cgi/content/abstract/343/24/1742 http://content.nejm.org/cgi/content/abstract/343/24/1742

30. November 17 2004 O3O3 Mortality

31. This image has been deleted because JHSPH OpenCourseWare was not able to secure permission for its use. Full-text available at http://jama.ama-assn.org/cgi/content/full/292/19/2372

32. March 8 2005 PM 2.5 Hospital Admissions

33. This image has been deleted because JHSPH OpenCourseWare was not able to secure permission for its use. Full-text available at http://jama.ama-assn.org/cgi/content/full/295/10/1127

34. The new challenge: Estimating the toxicity of the PM complex mixture

35. New Scientific Questions and Statistical Challenges What are the mechanisms of PM toxicity?  Size?  Chemical components?  Sources? New Methods for estimating health effects of complex mixtures

36. PM 2.5 PM 10 PM 10-2.5 Chemical constituents SizeTotal mass Cl OC SO4 Si K EC NO3 Ca Al Fe Biomass burning Vehicles Crustal Emission sources Bell Dominici Ebisu Zeger Samet EHP 2007

37. % change in CVD hospitalization rate associated with 10 increase in PM 10-2.5 on average across 108 US counties (1999-2005) PM 10-2.5 alone PM 2.5 alone PM 10-2.5 adjusted by PM 2.5 PM 2.5 adjusted by PM 10-2.5 Lag Peng Bell Chang McDermott Zeger Samet Dominici tech report 2007 Lag

38. The policy impact

39. NAAQS: Science has had an Impact From US EPA NAAQS Criteria Document 1996: “Many of the time-series epidemiology studies looking for associations between O3 exposure and daily human mortality have been difficult to interpret because of methodological or statistical weaknesses, including the failure to account for other pollutants and environmental effects.” From US EPA Criteria Document 2006: “While uncertainties remain in some areas, it can be concluded that robust associations have been identified between various measures of daily O3 concentrations and increased risk of mortality.”

40. Assessing the Public Health Impact of the Air Quality Regulations

41. Reproducible research We want to reproduce previous findings –“Did you do what you said you did?” Test assumptions, robustness of findings; check methodology –“Is what you did any good?” Implement and test new methodology –“I can do it better!”

42. Peng Dominici Zeger AJE 2006

43. NMMAPSdata package for R R is a free software environment for statistical analysis and graphics NMMAPSdata package contains the entire updated (1987—2000) NMMAPS database as an add-on module for R Supplemental code available online for reproducing canonical NMMAPS analysis and other analyses iHAPSS: Internet-based Health and Air Pollution Surveillance System –http://www.ihapss.jhsph.edu/http://www.ihapss.jhsph.edu/ Peng Welty R news 2004 Zeger Peng McDermott Dominici Samet HEI 2006

44. A new book to appear this summer… Environmental Epidemiology with R: A Case study in Air Pollution and Health Roger Peng & Francesca Dominici Peng & Dominici

45. Concluding Thoughts The weight of the evidence: –Has an explicit role in the Clean Air Act New NAAQS process New Research underway: especially PM Components and Sources – the cycle begins anew Policy Questions Data Methods Evidence Biostatistics in Action! analyses of observational studies can be used to address other questions beyond air pollution

46. Collaborators in the BSPH Michelle Bell Patrick Breysse Ciprian Crainiceanu Mary Fox Alyson Geyh Aidan McDermott Tom Louis Giovanni Parmigiani Roger Peng Jonathan Samet Ron White Scott Zeger PhD Students Howard Chang Sandy Eckel Sorina Eftim Jennifer Feder Haley Hedlin Yun Lu Chi Wang Yijie Zhou Medicare data users and collaborators in the BSPH and SOM Gerald Anderson Emily Smith Ben Brooke Lia Clattenburg Robert Herbert Peter Pronovost Funding sources EPA: PM Research Center (Samet) NIEHS: Training grant in Environmental Biostatistics (Louis and Dominici) NIEHS R01: Statistical methods in Environmental Epidemiology (Dominici)