Impact of Air Pollution on Public Health: Transportability of Risk Estimates Jonathan M. Samet, MD, MS NERAM V October 16, 2006 Vancouver, B.C. Department of Epidemiology
What is transportability? The idea that risks of air pollution observed in one or more populations can be extended to other populations. AKA: generalizability or external validity
Why transport risk estimates? Local evidence not available as basis for policy formulation. Use external evidence as framework to strengthen interpretation of locally derived evidence. To estimate burden of disease globally
What factors influence transportability? Characteristics of the air pollution mixture in study community(ies). Population characteristics that determine susceptibility to air pollution. Methodologic issues –Characteristics of exposure and outcome data –Data analysis approach –Publication bias
Local vs non-local risk estimates Local estimates –Motivate policy –Facilitate burden estimation –Accountability assessment Non-local estimates –Credibility –Stability and precision –Bound risks
St. George’s data base
Time-series estimates to 2006 Daily all-cause mortality and PM 10 (n=314 ) St. George’s data base, 10/06
Time-series estimates to 2006 Daily cardiovascular mortality and PM 10 (n=177 ) St. George’s data base, 10/06
Time-series estimates to 2006 Respiratory mortality and PM 10 (n=47 ) St. George’s data base, 10/06
All-cause mortality: % change in number of deaths associated with 10 µg/m 3 increase in daily PM 2.5 Source: Anderson HR et al. WHO 2004
What is responsible for heterogeneity? Publication bias? Population characteristics? Methodologic approaches?
What is publication bias? A tendency for the publication process to differentially lead to publication of papers reporting statistically significant findings. –May influence data analysis and selection of findings for publication Can it be identified? –Graphical approaches –Analytic approaches
Ozone for example: a meta-analysis 144 effect estimates from 39 time-series studies Strong statistically significant association identified between ozone and mortality for total deaths and cardiovascular disease Implied relationship between ozone and respiratory disease mortality Large heterogeneity in individual study estimates Some indication of publication bias Bell et al., 2005
% Increase in Daily Mortality for 10 ppb in Daily O 3 Provided a single lag Provided multiple lags Publication bias? Ozone effect estimates and lags: publication bias?
Comparison of ozone meta- analysis and multi-city results Bell et al. 2005
Funnel plot for estimates for respiratory mortality and ozone Source: Anderson HR et al. WHO 2004 Publication Bias Zone
Variation in ozone effect by cause Percent increase in daily total mortality for a 10 ppb increase in daily ozone (95% PI) Total: 0.87% (0.55, 1.18%) CVD: 1.11% (0.68, 1.53%) Respiratory:0.47% (-0.51, 1.47%) Source: Bell et al. Epidemiol 2005
Variation in ozone effect by location Percent increase in daily total mortality for a 10 ppb increase in daily ozone (95% CI) U.S.: 0.84% (0.48, 1.20%) –11 estimates from 9 studies Non-U.S.: 0.92% (0.47, 1.38%) –20 estimates from 14 studies Heterogeneity among estimates Source: Bell et al. Epidemiol 2005
Variation in ozone effect by age Percent increase in daily total mortality for a 10 ppb increase in daily ozone (95% CI) All ages: 0.83% (0.53, 1.12%) 65+ or 64+: 1.27% (0.65, 1.89%) Source: Bell et al. Epidemiol 2005
Ranking of PM 10 estimates for all-cause mortality by annual average levels of PM 10 * *left y-axis: mean PM 10 levels in µg/m 3 ; right y-axis: RR in total mortality of a 10 µg/m 3 increase of PM 10 Source: Anderson HR et al. WHO 2004
Are all meta-analyses the same?
Some solutions Maintained data base and periodic meta- analysis Multi-city analyses Periodic global analyses Also needed: –Unbiased publication processes –Transparent analytic approaches –Bayesian methods for handling local data
National Morbidity Mortality Air Pollution Study 1987—2000
City-specific and regional estimates City, Regional and National Estimates
Sensitivity of the national average estimates of the PM 10 - mortality association to adjustment for seasonality and model choice ( ) Peng, Dominici, Louis JRSS (2006)
Sensitivity of national average estimates to model selection methods National average estimates of the % increase in mortality for a 10 g/m 3 increase in PM 10 Previously reported results appear robust to choice of model selection method
Reproducible Research ( Benefits: Verifying published findings Conducting alternative analyses of the same data Eliminating uninformed criticisms which do not match data Expediting interchange of ideas among investigators
Overview of APHENA Air Pollution and Health: a Combined European And North American Approach (APHENA) The APHENA Group Europe: Touloumi G, Samoli E, Pipikou M, Atkinson R, Le Tertre A, Anderson R, Katsouyanni K US: Dominici F, Peng R, Schwartz J, Zanobetti A, Samet J Canada: Ramsay T, Burnett R, Krewski D. Supported by the Health Effects Institute
Develop a common approach for first-stage analyses of mortality and admissions time-series data and assess sensitivity of findings to critical elements of the model (using simulations and real data). Comparative evaluation of different methods to identify and combine dose-response curves; Comparison of alternative methods for addressing mortality displacement, and eventual application of one or more approaches to the various databases; Development of a data base on potential effect modifiers with exploration of differences in common, core items across the involved countries; Parallel and combined analyses of the air pollution and mortality data, and air pollution and hospitalization data, including exploration of geographic heterogeneity. Objectives:Objectives:
% increase in daily number of deaths (75+ years old), associated with 10 μg/m 3 increase in PM 10 (lags 0 and 1) in 21 European and 15 U.S. cities with daily PM 10 data
HEI’s PAPA-SAN Project
Looking Ahead Need for tracking development of risk estimates through systematic data bases Empiric evaluation of impact of local research estimates is needed Methods development needed for use of local risk estimates in context of regional and global estimates Continuation of APHENA-like approaches?