ACE Workshop at 2007 SER Meeting - Boston, MA
June 19, 2007
Boston Park Plaza Hotel 

June 19, 2007 - Educational Workshops

  • Systematic Review and Meta-Analysis
  • Introduction to Bayesian Modeling of Epidemiologic Data

We apologize, the following workshops have been cancelled.

  • Introduction to Pharmacoepidemiology: Practical Applications and Analytic Methods
  • Draw your assumptions before your conclusions: Graphs for causal inference
  • Use of automated databases in epidemiological research

Introduction to Pharmacoepidemiology: Practical Applications and Analytic Methods *CANCELLED*


  • Full Day Session
  • Morning Session
  • Afternoon Session

Morning session:

  • Cathy Critchlow, PhD, Global Epidemiology, Amgen Inc.
  • Brian Bradbury, DSc, MA, Global Epidemiology, Amgen Inc.

Afternoon session:

  • Sebastian Schneeweiss, MD, PhD, Division of Pharmacoepidemiology and Pharmacoeconomics, Brigham and Women's Hospital

Pharmacoepidemiology is concerned with the study of the use and effect of drugs in populations. Recent controversies surrounding drug safety have refocused interest, both within and outside the pharmaceutical industry, on the role and contributions of observational epidemiologic studies to the processes of drug development, drug approval, and pharmacovigilance. Prior to drug approval, pharmacoepidemiologists conduct studies to assess disease burden; define disease natural history; determine appropriate target populations, case definitions and outcome measures; and evaluate the potential for risks associated with drug treatment. Post-approval, epidemiological studies support the establishment of new indications and evaluate outcomes associated with long-term drug use. However, marketed drugs are prescribed to patients for specific reasons (indications), and the characteristics of persons with indications for a given drug can affect the occurrence of clinical outcomes of interest. Thus, comparisons of clinical outcomes based on drug exposure must address the potential for alternate explanations based on patient selection or “confounding by indication.”

In the morning session of this two-part workshop, we will provide an introductory overview of important concepts in the practice of pharmacoepidemiology. We will present and discuss examples of applications typically encountered by epidemiologists working in the pharmaceutical industry, and discuss how partnerships with academic collaborators facilitate the conduct of epidemiological research within industry. We will follow with a discussion of current issues of particular concern, such as conflict of interest and ethical conduct.

In the afternoon session, we will discuss in more detail analytic methods for the conduct and analysis of pharmacoepidemiology studies. Current examples will be used to illustrate the range of study designs and analytic techniques that are used for observational studies of the utilization, safety and effectiveness of drugs. In particular, techniques specifically addressing the control of confounding by indication in the assessment of clinical outcomes associated with drug exposure will be discussed.

Participants should expect to gain a broad understanding of the practice of pharmacoepidemiology in both the pharmaceutical industry and academic setting, and more specifically, an awareness of the potential contributions of epidemiology to drug development, drug approval, and post-marketing drug safety assessment. Participants will also gain an understanding of specific analytic techniques used to reduce or eliminate confounding by indication.

Systematic Review and Meta-Analysis


  • Afternoon Session


  • Charles Poole
    Department of Epidemiology
    University of North Carolina School of Public Health
    Chapel Hill, NC

This workshop provides a brief introduction to systematic review and meta-analytic methods. Topics include defining the purpose and scope of the review, defining and searching the literature, retrieving reports, extracting results and study characteristics, and contacting authors for additional information. Within-study analyses, when indicated, include sensitivity analysis and categorical regression. Among-study analyses, when indicated, include graphical displays, analyses of funnel plot symmetry, imputation of hypothetically missing results, analysis of overall homogeneity, and stratified and meta-regression analyses of study characteristics. STATA or SAS code is provided for most analyses. Contraindications to summary aggregation, with or without random effects, are stressed. The need for even more caution than usual in interpreting results is underlined. Numerous published examples are presented and discussed.

Introduction to Bayesian Modeling of Epidemiologic Data


  • Morning Session


  • David B. Dunson, Senior Investigator, NIEHS/NIH & Adjunct Professor of
    Statistics, Duke University
  • Amy H. Herring, Associate Professor of Biostatistics, UNC-Chapel Hill
  • Richard F. MacLehose, Postdoctoral Associate, NIEHS/NIH


Bayesian statistical methods have been increasingly used in analyses of epidemiologic data due to some compelling advantages. In simple settings, such as a logistic regression analysis with a large sample size and few predictors, Bayesian methods yield very similar results to maximum likelihood analyses. However, epidemiologic research is frequently confronted with very highly correlated data, massive numbers of predictors or, more generally, with sparse data problems. For example, with improvements in speed and cost of genotyping, genetic epidemiology studies often collect data for single nucleotide polymorphisms (SNPs) at many loci. Even when genetic information is not available, studies may collect information for many different chemical or dietary exposures as well as for multiple lifestyle factors. In the presence of data with very many predictors, some of which likely have missing observations, the advantages of Bayesian methods that include informative priors become clear.

This workshop is designed to provide an introductory overview of Bayesian analyses of epidemiologic data, motivated in particular by the problems of large numbers of predictors, missing data and uncertainty in the regression model relating predictors to a response. The motivation will not be philosophical, but will focus on the practical details involved in a Bayesian analysis, including prior choice, calculation of the posterior and interpretation of the results from a Bayesian analysis. In addition, we will highlight differences with frequentist results in order to help epidemiologists decided when a Bayesian analysis is worth the additional effort in implementation. We will not assume a background in Bayesian statistics, but will assume a basic knowledge of linear and logistic regression modeling.

Some key technical concepts that will be introduced include Markov chain Monte Carlo (MCMC) algorithms, model averaging, and shrinkage estimation through hierarchical modeling. MCMC is the standard approach for implementing Bayesian analyses, with WinBUGS providing a freely available software package and new versions of SAS incorporating procedures for Bayesian analyses of linear, logistic, log-linear and proportional hazards regression models using MCMC. In addition to a basic description of how MCMC works, we provide some guidance in how to use the results from an MCMC analysis for epidemiologic inferences. A second concept, model averaging, is important when there is uncertainty about which regression model to fit. Model averaging can be used to avoid well known bias in stepwise selection, while also producing more realistic uncertainty estimates that reflect error in choosing the model to fit. The concept of shrinkage through hierarchical modeling is critical in addressing problems encountered analyzing highly correlated or high-dimensional predictors. In these settings, maximum likelihood analyses fail, while shrinkage provides an approach for obtaining stable estimates that reflect scientific knowledge about plausible values for the coefficients. These concepts will be illustrated fully through a number of examples, including reproductive, genetic and environmental epidemiology studies. Sample datasets and WinBUGS code will be provided.

Background Reading:

Dunson, D.B. (2001). Practical Advantages of Bayesian Analysis of Epidemiologic Data. American Journal of Epidemiology 153(12):1222-1226.

About the Instructors:

Dr. Dunson is Senior Investigator, Biostatistics Branch, National Institute of
Environmental Health Sciences. Dr. Dunson has over 90 peer-reviewed publications and serves as an Editor of Bayesian Analysis and as Associate Editor of the Journal of the American Statistical Association, Biometrics, and Biostatistics. He has published extensively on Bayesian methods in the statistical literature and also has numerous first- and co-authored publications in reproductive epidemiology journals.

Dr. Herring is Associate Professor, Department of Biostatistics, University of North Carolina at Chapel Hill. A graduate of Harvard University, Dr. Herring has built a statistical research program centered on developing methods for "messy" exposures, including high-dimensional exposures, highly-correlated exposures, and missing exposure data. She is co-investigator of numerous studies in reproductive and environmental epidemiology and has over 40 peer-reviewed publications.

Dr. MacLehose is a postdoctoral researcher in the Biostatistics Branch at the National Institute of Environmental Health Sciences. Dr. MacLehose received a doctorate in epidemiology in 2005 at the University of North Carolina at Chapel Hill and has been working to make new statistical developments more readily available to epidemiologists.

Draw your assumptions before your conclusions: Graphs for causal inference



  • Morning Session


  • Miguel A. Hernán, MD, DrPH
    Department of Epidemiology
    Harvard School of Public Health

Causal directed acyclic graphs (DAGs) can be used to summarize, clarify, and communicate one's qualitative assumptions about the causal structure of a problem. The use of causal DAGs is a natural and simple approach to causal inference from observational data. It is also a rigorous approach that leads to mathematical results that are equivalent to those of counterfactual theory. As a result, causal DAGs are increasingly used in epidemiologic research and teaching.

This workshop will provide a non-technical overview of causal DAGs theory, its relation to counterfactual theory, and its applications to causal inference. It will describe how causal DAGs can be used to propose a systematic classification of biases in observational and randomized studies, including the biases induced by the use of conventional statistical methods for the analysis of longitudinal studies with time-varying exposures.

Use of automated databases in epidemiological research *CANCELLED*


  • Afternoon Session


  • Susan Jick, D.Sc.
    Boston Collaborative Drug Surveillance Program
    11 Muzzey Street
    Lexington, MA 02421

There is increasing interest in and use of, automated databases in epidemiological research. This is particularly true in the area of pharmacoepidemiology. Why are these databases so important for this specialty area? What kinds of automated databases are there? How are they different from one another? When should one use one data resource instead of another, or what factors should one consider when choosing the best database for a particular research question? These are the basic issues to consider before pursuing use of an automated data resource. This workshop will provide information on how to evaluate the many databases available for research. It will go on to describe important limitations and pitfalls in using these resources, and how to avoid or work with them to obtain the most valid and defensible results possible.

Proceed to registration