Bayesian Distributed Lag Models: Estimating Effects of Particulate Matter Air Pollution on Daily Mortality Leah J. Welty, PhD Department of Preventive.

Slides:



Advertisements
Similar presentations
Copula Representation of Joint Risk Driver Distribution
Advertisements

Bayesian inference Jean Daunizeau Wellcome Trust Centre for Neuroimaging 16 / 05 / 2008.
MCMC estimation in MlwiN
Bayesian network classification using spline-approximated KDE Y. Gurwicz, B. Lerner Journal of Pattern Recognition.
Insert Date HereSlide 1 Using Derivative and Integral Information in the Statistical Analysis of Computer Models Gemma Stephenson March 2007.
A. The Basic Principle We consider the multivariate extension of multiple linear regression – modeling the relationship between m responses Y 1,…,Y m and.
Chapter 7 Title and Outline 1 7 Sampling Distributions and Point Estimation of Parameters 7-1 Point Estimation 7-2 Sampling Distributions and the Central.
1 Parametric Sensitivity Analysis For Cancer Survival Models Using Large- Sample Normal Approximations To The Bayesian Posterior Distribution Gordon B.
Jun Zhu Dept. of Comp. Sci. & Tech., Tsinghua University This work was done when I was a visiting researcher at CMU. Joint.
Visual Recognition Tutorial
Copyright 2008, The Johns Hopkins University and Francesca Dominici. All rights reserved. Use of these materials permitted only in accordance with license.
Predictive Automatic Relevance Determination by Expectation Propagation Yuan (Alan) Qi Thomas P. Minka Rosalind W. Picard Zoubin Ghahramani.
Assessing the Health Effects of Air Pollution; Statistical and Computational Challenges Scott L. Zeger on behalf of The Environmental Biostatistics and.
Machine Learning CUNY Graduate Center Lecture 3: Linear Regression.
Prediction and model selection
Model Choice in Time Series Studies of Air Pollution and Health Roger D. Peng, PhD Department of Biostatistics Johns Hopkins Blomberg School of Public.
Distribution Function Estimation in Small Areas for Aquatic Resources Spatial Ensemble Estimates of Temporal Trends in Acid Neutralizing Capacity Mark.
Distribution Function Estimation in Small Areas for Aquatic Resources Spatial Ensemble Estimates of Temporal Trends in Acid Neutralizing Capacity Mark.
Using ranking and DCE data to value health states on the QALY scale using conventional and Bayesian methods Theresa Cain.
Bayesian Analysis for Extreme Events Pao-Shin Chu and Xin Zhao Department of Meteorology School of Ocean & Earth Science & Technology University of Hawaii-
Health Risks of Exposure to Chemical Composition of Fine Particulate Air Pollution Francesca Dominici Yeonseung Chung Michelle Bell Roger Peng Department.
Distribution Function Estimation in Small Areas for Aquatic Resources Spatial Ensemble Estimates of Temporal Trends in Acid Neutralizing Capacity Mark.
Distribution Function Estimation in Small Areas for Aquatic Resources Spatial Ensemble Estimates of Temporal Trends in Acid Neutralizing Capacity Mark.
Peter Congdon, Centre for Statistics and Department of Geography, Queen Mary University of London. 1 Spatial Path Models with Multiple.
Review of Lecture Two Linear Regression Normal Equation
Are the Mortality Effects of PM 10 the Result of Inadequate Modeling of Temperature and Seasonality? Leah Welty EBEG February 2, 2004 Joint work with Scott.
Modeling Menstrual Cycle Length in Pre- and Peri-Menopausal Women Michael Elliott Xiaobi Huang Sioban Harlow University of Michigan School of Public Health.
ECE 8443 – Pattern Recognition LECTURE 06: MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION Objectives: Bias in ML Estimates Bayesian Estimation Example Resources:
Prof. Dr. S. K. Bhattacharjee Department of Statistics University of Rajshahi.
A Beginner’s Guide to Bayesian Modelling Peter England, PhD EMB GIRO 2002.
Estimating parameters in a statistical model Likelihood and Maximum likelihood estimation Bayesian point estimates Maximum a posteriori point.
Term 4, 2005BIO656 Multilevel Models1 Hierarchical Models for Pooling: A Case Study in Air Pollution Epidemiology Francesca Dominici.
Bayesian Analysis and Applications of A Cure Rate Model.
Receptor Occupancy estimation by using Bayesian varying coefficient model Young researcher day 21 September 2007 Astrid Jullion Philippe Lambert François.
Calculation of excess influenza mortality for small geographic regions Al Ozonoff, Jacqueline Ashba, Paola Sebastiani Boston University School of Public.
2006 Summer Epi/Bio Institute1 Module IV: Applications of Multi-level Models to Spatial Epidemiology Instructor: Elizabeth Johnson Lecture Developed: Francesca.
Therapeutic Equivalence & Active Control Clinical Trials Richard Simon, D.Sc. Chief, Biometric Research Branch National Cancer Institute.
HOW HOT IS HOT? Paul Wilkinson Public & Environmental Health Research Unit London School of Hygiene & Tropical Medicine Keppel Street London WC1E 7HT (UK)
Bayesian Multivariate Logistic Regression by Sean O’Brien and David Dunson (Biometrics, 2004 ) Presented by Lihan He ECE, Duke University May 16, 2008.
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.
2005 Hopkins Epi-Biostat Summer Institute1 Module 3: An Example of a Two-stage Model; NMMAPS Study Francesca Dominici Michael Griswold The Johns Hopkins.
Estimation in Marginal Models (GEE and Robust Estimation)
Additional Topics in Prediction Methodology. Introduction Predictive distribution for random variable Y 0 is meant to capture all the information about.
Introduction Outdoor air pollution has a negative effect on health. On days of high air pollution, rates of cardiovascular and respiratory events increase.
Multi-level Models Summer Institute 2005 Francesca Dominici Michael Griswold The Johns Hopkins University Bloomberg School of Public Health.
Lecture 2: Statistical learning primer for biologists
Assessing Estimability of Latent Class Models Using a Bayesian Estimation Approach Elizabeth S. Garrett Scott L. Zeger Johns Hopkins University Departments.
Stochastic Loss Reserving with the Collective Risk Model Glenn Meyers ISO Innovative Analytics Casualty Loss Reserving Seminar September 18, 2008.
G. Cowan Computing and Statistical Data Analysis / Stat 9 1 Computing and Statistical Data Analysis Stat 9: Parameter Estimation, Limits London Postgraduate.
1 Chapter 8: Model Inference and Averaging Presented by Hui Fang.
July Hopkins Epi-Biostat Summer Institute Module II: An Example of a Two- stage Model; NMMAPS Study Francesca Dominici and Scott L. Zeger.
Machine Learning CUNY Graduate Center Lecture 6: Linear Regression II.
1 Part09: Applications of Multi- level Models to Spatial Epidemiology Francesca Dominici & Scott L Zeger.
- 1 - Preliminaries Multivariate normal model (section 3.6, Gelman) –For a multi-parameter vector y, multivariate normal distribution is where  is covariance.
Parameter Estimation. Statistics Probability specified inferred Steam engine pump “prediction” “estimation”
1 Module IV: Applications of Multi-level Models to Spatial Epidemiology Francesca Dominici & Scott L Zeger.
Methodological Considerations in Assessing Effects of Air Pollution on Human Health Rebecca Klemm, Ph.D. Klemm Analysis Group, Inc. American Public Health.
Skill of Generalized Additive Model to Detect PM 2.5 Health Signal in the Presence of Confounding Variables Office of Research and Development Garmisch.
Ch 1. Introduction Pattern Recognition and Machine Learning, C. M. Bishop, Updated by J.-H. Eom (2 nd round revision) Summarized by K.-I.
Markov Chain Monte Carlo in R
Bayesian Semi-Parametric Multiple Shrinkage
Roberto Battiti, Mauro Brunato
Kernel Stick-Breaking Process
STA 216 Generalized Linear Models
More about Posterior Distributions
Comparisons among methods to analyze clustered multivariate biomarker predictors of a single binary outcome Xiaoying Yu, PhD Department of Preventive Medicine.
Module 3: An Example of a Two-stage Model; NMMAPS Study
Task 6 Statistical Approaches
Parametric Methods Berlin Chen, 2005 References:
Applied Statistics and Probability for Engineers
Presentation transcript:

Bayesian Distributed Lag Models: Estimating Effects of Particulate Matter Air Pollution on Daily Mortality Leah J. Welty, PhD Department of Preventive Medicine Northwestern University Joint work with Scott Zeger, Roger Peng, Francesca Dominici Johns Hopkins Department of Biostatistics

Extreme Air Pollution Events London Fog, December 1952 When Smoke Ran Like Water, D. Davis, Perseus Books © 2002

Time Series: Daily Deaths & Daily Air Pollution

Current Air Quality Standards: Adverse Health Effects? Energy Information Administration United States Dept. of Energy United States EPA

Time Series Models for Mortality on Air Pollution log (expected mortality) ~ long term trends + season + weather + pollution + day of week Poisson regression Smooth functions of time, season, and weather GLM/GAM with natural/smoothing splines

Distributed Lag Models (DLMs) These are time series models with lagged exposure variables as predictors. The acute health effects from pollution exposure may take a day or more to develop. The time from exposure to event may vary among individuals. Compared to DLMs, models with single day exposures may under or over estimate the risk of mortality.

Distributed Lag Models Allow response to depend on exposure over several days Effect of yesterdays pollution on todays mortality Total effect = % increase in daily mortality associated w/ 1 unit increase in PM on previous days

Distributed Lag Functions Effect of unit increase in pollution 7 days ago on todays mortality lag i

Estimating Distributed Lag Models Allow response to depend on exposure over several days Difficult to estimate since x s are correlated

Example distributed lag functions Yesterday and day before constrained to have same effect

Constrained Distributed Lag Models Unconstrained Constrained - Step function constraints - Polynomial - Spline -Other How do we set up constraints that are consistent with the true distributed lag function?

Air Pollution Distributed Lag Function: Prior Knowledge Constraint is application of prior knowledge Prior knowledge Acute risk varies smoothly as function of lag Acute risk goes to zero as time from exposure increases Not part of polynomial or spline constraints i

Estimation Problems (Chicago, PM10)

Bayesian Distributed Lag Models: Outline Propose Bayesian Dist Lag Model (BDLM) Approximate posterior distribution Gibbs sampler implementation Compare to other constraints (spline, poly) Apply to Chicago PM & Mortality BDLMs in context of smoothing

1.No knowledge of early lag effects 2.Lag effects must eventually go to zero 3.Lag effects tend to zero smoothly Bayesian Constraints for DLMs Prior on distributed lag coefficients Construct as to reflect i

Constructing Distributed Lag Prior 1.No knowledge of early lag effects 2.Lag effects must eventually go to zero Large Variances Small Variances 3.Lag effects tend to zero smoothly Uncorrelated Correlated

Fast Variances 0 Slow Less Correlation More Simulated Dist Lag Functions from Prior

Bayesian Model If the likelihood for the distributed lag coefficients is normal, then

Bayesian Model: Hyperparameters If the likelihood for the distributed lags is normal, then this mixture is a mixture of normals Posterior distribution for η determined from data (smoothness of DL function estimated from data)

Simulation Study BDLM vs. unconstrained, maximum likelihood polynomial of degree 4 p-spline estimated via GCV p-spline estimated via REML 25 different true distributed lag functions some consistent with prior knowledge some not -- e.g. the dist lag fun a non-zero constant 500 outcome series for each function

Comparing BDLM to Common Methods black = true DL function white = estimated DL function gray = 95% posterior/confidence bands

Comparing BDLM to Common Methods black = truth DL function white = estimated DL function gray = 95% posterior/confidence bands

Comparing BDLM to Common Methods: MSEs* Distributed Lag Function Bayes Poly GCV REML * Expressed as a percent of the MSE for the MLE (unconstrained) Distributed Lag Model TOTAL EFFECT LAG 14

Comparing BDLM to Common Methods BDLM performs consistently well Captures features of DL function Narrower confidence bands When the goal is estimating the total effect BDLM 10-15% better Better estimation at longer lags

Application: Chicago Mortality & PM10 log (expected mortality) ~ long term trends + season + weather + pollution + day of week everything else: trends, season, weather, dow

Distributed Lag for PM10 & Mortality Chicago: Necessary Extensions Two problems 1.Likelihood Poisson No closed form posterior 2.Prior on β? Two approaches 1.Pretend MLEs normal; ignore β uncertainty 2.Gibbs sampler

Gibbs Sampler Rather than estimating the full model all at once: Alternate between: 1. Sampling θ (normal approximate as proposal), using last β in offset: 2. Sampling β (random walk Metropolis, flat prior), using last θ in offset:

Posterior Mean & 95% Posterior Region PM 10 on Mortality Chicago Using last 4000 of 5000 iterations of Gibbs sampler

Joint Posterior for Hyperparameters η PM 10 on Mortality Chicago Using last 4000 of 5000 iterations of Gibbs sampler Fast Variances 0 Slow More Correlation Less

Estimation Method Comparison: Normal Approx vs Gibbs Sampler Using last 4000 of 5000 iterations of Gibbs sampler

Sensitivity of BDLM for Chicago to value of σ and prior on η Altering σAltering η

BDLMs in Context of Smoothing P-spline approach to smoothing distributed lag coefficients: To estimate p-spline, minimize over : Prior on dist lag coefficients penalty matrix: penalty

BDLM in the Context of Smoothing Equivalent results from BDLMs & specific p-spline use reasonably flexible spline basis P-spline penalties related to jumps in 3 rd derivative Difficult to relate to biological or prior knowledge BDLMs transparent method for eliciting P-spline penalties consistent w/ objective function

Conclusions: BDLMs Introduce flexible Bayesian DLM –Incorporates prior knowledge –Degree of smoothness of DLM est from data Simulation Study –smaller MSEs than comparable methods Relates to P-splines for DLMs –BDLM analogous to specific p-spline –Method of eliciting a prior distribution

Conclusions: PM 10 and Mortality Chicago Largest effect lag day 3 10 μg/m 3 increase in PM 10 three days previous associated with 0.17% increase daily mortality Total effect -0.21% (-0.86, 0.41) increase mortality Normal approximation and Gibbs sampler similar results Large # daily deaths (~116), t = 1, … 5114 Less agreement for shorter time series, less normally distributed outcomes

Future Directions: BDLMs Naturally extends to additional hierarchy Peng, Dominici, Welty. Estimating the time course of hospitalization risk associated with air pollution using a Bayesian hierarchical distributed lag model. in press JRSS-C Missing data in exposure time series Many cities air pollution 1/3 or 1/6 days

Bayesian Distributed Lag Models: Leah J. Welty, Roger D. Peng, Scott L. Zeger, and Francesa Dominici, Bayesian Distributed Lag Models: Estimating Effects of Particulate Matter Air Pollution on Daily Mortality Biometrics [Epub ahead of print April 2008] Roger Peng, Francesca Dominici, and Leah J. Welty, A Bayesian hierarchical distributed lag model for estimating the time course of hospitalization risk association with particulate matter air pollution JRSS-C (in press) NMMAPS ( ) Data & Programs: Roger D. Peng, Leah J. Welty, and Aidan McDermott, "The National Morbidity, Mortality, and Air Pollution Study Database in R" (2004). Contact:

Constructing Prior: More Details 3.Lag effects tend to zero smoothly X has desired correlation structure - - -

Comparing BDLM to Common Methods