Winter 2007

January 10, 2007
Title:	Statistical Framework for an HIV/AIDS Clinical Trial on Antiretroviral Prevention Strategies
Speaker:	Ying Qing Chen, Ph.D.
Abstract:	Antiretroviral Therapy (ART) developed in the late 1980s has been shown to dramatically reduce the morbidity and mortality of HIV infection through sustained reduction in HIV viral replication. Nevertheless such therapy does not cure HIV infection, and viral resistance is expected to develop in most of the patients on the regimens that are not completely suppressive. Therefore modified ART regimens are usually required to maintain viral suppression. In this informal talk, we will discuss the background and rationale of an ongoing clinical trial on two different ART strategies in HIV/AIDS transmission prevention. A statistical framework will be presented in consideration of trial design, monitoring and data analysis. Challenges and opportunities in methodological research will be also discussed.
January 17, 2007
Title:	Constructing complete networks from egocentric network data: A comparison of parametric and nonparametric approaches
Speaker:	David Lockhart
Abstract:	The structure of the social network which surrounds an individual is a risk factor for a variety of diseases, but in many contexts – such as sexual networks – it may be prohibitively expensive to fully measure the network. Often all that is feasible to obtain is information about sampled individuals and their immediate relations. In this talk I will discuss a variety of methods to reconstruct the complete network out of these egocentric fragments, focusing on the effects of correlation of attributes among partners of the same individual. When strong, such a correlation causes the network to effectively divide up into smaller subnetworks that are not strongly connected to each other. This pattern slows the spread of infectious disease through the network. A parametric approach using exponential random graph models performs well when the correlation across partners is correctly specified, but can provide biased estimates if the correlation is ignored or misspecified. In contrast, a nonparametric approach which fully preserves the observed network fragments can incorporate such correlation without specifying its nature. However, unless the correlation is explicitly and correctly modeled, its strength is attenuated. Thus, they also produce biased network estimates, though less severely biased than misspecified parametric models.
January 24, 2007
Title:	Efficiency of independence estimating equations
Speaker:	Ben French
Abstract:	This seminar could appropriately be titled, “Things you should have learned in 571 (had you been paying attention).” I begin by reviewing the statistical background for independence and generalized estimating equations. I summarize published results on the efficiency of the independence estimator, and highlight situations in which the estimator is inefficient. I conclude by discussing implications of these results for volume-outcome studies, where volume represents both the predictor of interest and the cluster size.
January 31, 2007
Title:	Biocreep in Non-inferiority Clinical Trials
Speaker:	Siobhan Everson-Stewart
Abstract:	In many clinical settings, it may be unethical or inappropriate to perform a placebo-controlled clinical trial of an investigational treatment. A trial may then run to prove that the new treatment is not significantly worse than the standard of care; such trials are called non-inferiority clinical trials. However, if each new accepted treatment is somewhat less effective than its predecessor, there is concern that a completely ineffective product may be approved, a phenomenon known as “biocreep.” A simulation study was performed to examine the conditions under which biocreep may occur; these findings will be summarized.
February 7, 2007
Title:	Harnessing Naturally Randomized Transcription to Infer Regulatory Relationships Among Genes
Speaker:	Lin Chen
Abstract:	A problem of much interest is how to characterize the genome-wide transcriptional regulatory network of an organism based on high-throughput molecular profiling data. Correlation and its extensions have been used to group genes with similar expression patterns, for example, via correlation-based clustering or Bayes networks. However, variation in RNA levels, protein levels, higher order phenotypes, and the environment may all affect each other, implying that correlation among any two variables will usually not imply any particular regulatory or causal effect of one on another. Recombinant inbred line experiments allow one to randomize the genetic content of an organism from two or more genetic backgrounds, thereby producing independent, randomized realizations of DNA content. At the same time, one may perform high-throughput molecular profiling of the recombinant individuals, such as measuring genome-wide expression. We show how to employ this experimental system to infer regulatory relationships of one gene on another at the genome-wide level. Our approach is based on the fact that naturally randomized DNA variation in turn creates a randomization of RNA levels, which can be used to infer causal relationships among the transcriptional levels of genes. We develop a method called TRIGGER and apply the method to an experiment in yeast, showing that the approach yields biologically coherent information, recovering known regulatory pathways. We also estimate a lower bound on the number of transcriptional regulatory relationships among yeast genes in this setting, yielding new insights into the topology of the yeast transcriptional regulatory network.
February 14, 2007
Title:	Missing Data: A Practical Approach
Speaker:	Eric Johnson
Abstract:	While missing data is prevalent in many studies, methods for dealing with it are not in wide use. I feel that this is due to the presentation of the methods: the focus is typically on on statistical derivation rather than an intuitive description of how and when to use each technique. My talk will focus on the practical aspects of various methods, particularly different types of multiple imputation.
February 21, 2007
Title:	Understanding Blood Stem Cell Kinetics via Reversible Jump MCMC
Speaker:	Youyi Fong
Abstract:	An adult produces 100 billion blood cells each day. They mostly come from hematopoietic stem cells (HSC) in the bone marrow. It is of interest to estimate the rates of replication and differentiation of HSC in vivo. Bone marrow transplantation experiments done on hybrid cats produced a dataset which contain information about these rates. We build a hierarchical model to describe the data and apply Bayesian inference technique to solve the estimation problem. The primary challenge comes from the fact that the middle level parameter is a function of time. We have chosen a parameterization that allows us to use reversible jump MCMC to sample from the posterior distribution of this function. We will look at identifiability issue and touch upon goodness of fit testing.
February 28, 2007
Title:	Mendelian Randomization: Applications and Limitations
Speaker:	Carolyn Hutter
Abstract:	Exposure-outcome associations in observational epidemiology studies may be confounded by unknown or unmeasured variables. Mendelian Randomization (MR) has been proposed as a method for deriving unconfounded estimates of exposure-outcome relationships. In the MR approach a genetic variation underlying the exposure of interest is identified. Since Mendel’s second law predicts independent assortment of genes during meiosis, proponents of the MR approach argue that information about genotype-exposure and genotype-outcome associations can be used for inference about the exposure-outcome relationship. More formally, MR uses genotypes as instrumental variables to estimate the “causal” effect of the exposure on disease. My talk will include an introduction to MR, a brief overview of instrumental variable methods, and a discussion of examples from the literature that use MR approaches. I will focus on key assumptions underlying MR for both testing and estimation and will present simulation results demonstrating the impact of violations of these assumptions.
March 7, 2007
Title:	Current Issues in Non-inferiority Trials and Some Ideas for their Resolution (a work in progress)
Speaker:	Katie Davis
Abstract:	Non-inferiority (NI) trials are designed to evaluate whether the effect of a new treatment is unacceptably worse than a standard treatment. Methods for NI trials differ from those for superiority trials, and there are outstanding issues regarding these methods, some of which have resulted in controversy in recent news. In late 2006, five members of Congress called for an investigation of the FDA’s acceptance of non-inferiority trials in deciding on approvals of antibiotics and other drugs. The members cited NI studies used in the approval of the antibiotic Ketek (telithromycin), which was later linked to fatal cases of liver damage. In this case, and many others, the methods used to establish non-inferiority were outdated and unjustified. During my presentation, I will provide an overview of the current methods used to establish non-inferiority, the limitations of these methods, and other concerns regarding the design, conduct, and analysis of NI trials. I will also discuss these issues as they relate to examples of trials reviewed by FDA Advisory Committees in the past 5 years. Finally, I will outline some ideas for development of new methods and applications of existing methods to address some of these problems.