The December 2010 issue of the Journal of the American Statistical Association features an article by Ross Prentice of the University of Washington and the Fred Hutchinson Cancer Research Center that surveys the state of the art in chronic disease prevention research. The article, “Chronic Disease Prevention Research Methods and Their Reliability, with Illustrations from the Women’s Health Initiative,” is based on Prentice’s 2008 Fisher Lecture at the Joint Statistical Meetings in Denver, Colorado.

Access Free Featured Articles

“Modeling Competing Infectious Pathogens from a Bayesian Perspective: Application to Influenza Studies with Incomplete Laboratory Results” by Yang Yang, M. Elizabeth Halloran, Michael J. Daniels, Ira M. Longini Jr., Donald S. Burke, and Derek A. T. Cummings

“Chronic Disease Prevention Research Methods and Their Reliability, with Illustrations from the Women’s Health Initiative” by Ross L. Prentice

“Estimability and Likelihood Inference for Generalized Linear Mixed Models Using Data Cloning” by Subhash R. Lele, Khurram Nadeem, and Byron Schmuland

The Women’s Health Initiative (WHI), an NIH-sponsored prevention research program addressing causes of death and disability in postmenopausal women in the United States, provides examples. The WHI began in the 1990s and ultimately included more than 100,000 women. The WHI has had an enormous impact. Indeed, plugging “Women’s Health Initiative” into Google produces more than a million hits.

Quite apart from the substantive findings and their importance to women’s health, the WHI—which included both randomized trial and observational study components—also has had a significant effect on statistics. The primary study was a randomized controlled trial evaluating the health impacts of postmenopausal hormone therapy (HT) on a sample of more than 15,000 women.

A novel feature of the WHI design was the inclusion of a companion prospective observational cohort study, involving more than 90,000 postmenopausal women drawn from the same catchment populations, with much commonality in demographic, medical history, and behavioral data collection, as well as outcome data ascertainment procedures.

The observational study component provides a unique opportunity to compare clinical trial and observational study results for interventions that were evaluated in the randomized controlled trial. These comparisons provide insight into the adequacy of observational-study methods for bias reduction and the comparative strengths and weaknesses of these two major study designs more generally.

Theory and Methods

Cloning. First it was Dolly the sheep being cloned and now it is data! The Theory and Methods Section includes an article that proposes data cloning to improve inference. Although you won’t find any mention of somatic cells in “Estimability and Likelihood Inference for Generalized Linear Mixed Models Using Data Cloning” by Subhash Lele, Khurram Nadeem, and Byron Schmuland, you will find a novel, interesting, and useful new method for computing likelihoods for generalized linear mixed models (GLMM) and related hierarchical statistical models.

The authors exploit recent advances in Bayesian computation—Markov Chain Monte Carlo (MCMC) methods in particular—to implement frequentist inference in GLMMs without high-dimensional numerical integration or maximization or differentiation. The approach is Bayesian computation-based and uses well-known MCMC methodology that can be easily implemented in standard software such as WinBUGS.

The key to the authors’ method is that the posterior of the K-times “cloned” data, obtained by repeating the observed data vector K times, concentrates on the original data maximum likelihood estimate. Pretending the cloned data are from K independent experiments, MCMC methods are used to generate random variates from the cloned-data posterior distribution. For large K, the MLE is simply the mean of these random variates. Furthermore, when the parameter space is continuous, the inverse of the Fisher information based on the original data is obtained simply as K times the sample covariance matrix of the random draws.

The paper describes the use of data cloning for GLMMs, provides a graphical procedure for determining an adequate number of clones, shows how to obtain prediction intervals for random effects, and provides a simple graphical procedure to determine estimability of parameters in hierarchical models. The authors’ data-cloning methods are illustrated using examples involving the logistic-normal model for over-dispersed binary data and Poisson-normal models for repeated and spatial count data.

Applications and Case Studies

Winter brings flu season and the December JASA issue features an article that describes methods for studying influenza transmission. “Modeling Competing Infectious Pathogens from a Bayesian Perspective” by Yang Yang, Elizabeth Halloran, Michael Daniels, Ira Longini, Donald Burke, and Derek Cummings presents a competing-risk model that assumes influenza-like illness can result from the influenza virus or other circulating pathogens. A model for infection and transmission of each pathogen/virus is developed.

If specimens are collected and tested to determine the specific cause of each illness, it is possible to study infection and transmission separately for each pathogen. In practice, however, it is generally not feasible to perform laboratory analysis on a specimen from each sample. The cost for doing so can be prohibitive, and even if it is affordable, the logistics may make it impractical. Thus, the nature of the infecting pathogen is unknown for many of the illnesses.

The framework developed by Yang et al. allows for inference about the key infection and transmission process parameters with missing laboratory results as long as the missingness is independent of the type of infection (i.e., the lab test is neither more nor less likely to be missing for any one of the pathogens). It turns out that a moderate amount of missing test results has only a small impact on inference.

A possible difficulty with the approach is that it requires specifying the number of alternative (non-influenza) pathogens circulating in the environment. Fortunately, simulate results show the proposed model works well, even if this number is mis-specified. The authors demonstrate their approach on a study of an educational flu prevention program in Pittsburgh-area elementary schools.

Of course, the above articles are just a sample. The December issue includes Theory and Methods contributions on dimension reduction, model selection, and change-point estimation. The Applications and Case Studies section includes autoregressive mixture models for tracking violent crime intensity in a geographic area and an analysis of factors associated with firms’ decision to withdraw initial stock offerings.