Fall 2004

October 6, 2004
Title:	Issues in the Timing of Data Collection for Estimating Treatment Effects in Randomized Clinical Trials
Speaker:	Amalia Meier, Biostatistics recent graduate
October 13, 2004
Title:	Statistical methods for a 2-sample cost-effectiveness analysis
Speaker:	Phil Dinh, Biostat Ph.D. student
Abstract:	Effectiveness is often an outcome in medical research. These could be the number of cases of diseases avoided, the improvement in quality of life, or the increase if quality-adjusted life years, etc. Health economists are also often interested in costs. Cost-effectiveness analysis is a way to mingle the two. Two measures commonly used in a cost-effectiveness analysis are the incremental cost-effectivenss ratio (ICER) and the net health benefit (NHB). In this talk, I will summarize these two measures, discuss their statistical properties, and provide new methods (based on Edgeworth expansion) for inference. Validity of the new methods will be demonstrated via a simulation study. This is a work in progress under the supervision of Dr. Andrew Zhou.
October 20, 2004
Title:	Alternative Endpoints for Mortality In Studies Of Patients with Atrial Fibrillation: the AFFIRM Study Experience
Speaker:	April Slee, UW Biostatistics graduate student
Abstract:	As the numbers of proven effective therapies have increased, the magnitude of any single clinically meaningful treatment effect has decreased. Since the power of a clinical trial is directly related to this treatment effect, there has been increased interest in surrogate endpoints, i.e., events occurring earlier and more frequently that accurately predict the event of interest, particularly surrogates for mortality. The interest has been particularly intense for trials involving patients with cardiac conditions such as Atrial Fibrillation. The failure of arrhythmia suppression to correlate favorably with survival in the Cardiac Arrhythmia Suppression Trial (CAST) has tempered the acceptance of surrogate endpoints for mortality in subsequent trials of arrhythmia therapies. However, in a post-hoc analysis of the Antiarrhythmics Versus Implantable Defibrillators (AVID) Trial, hospitalization for CHF was shown to be a suitable surrogate for subsequent death. This presentation will explore the performance of hospitalization for cardiovascular reasons as a surrogate for mortality in the Atrial Fibrillation Follow-up Investigation of Rhythm Management (AFFIRM) Study, a randomized clinical trial to determine whither atrial fibrillation is best treated with a heart rate control versus a heart rhythm control strategy. I will discuss how cardiovascular hospitalization meets standard criteria for an appropriate surrogate outcome, and demonstrate the potential improvement in power due to the increased event rate if this surrogate were used in place of mortality.
October 27, 2004
Title:	Working correlation matrix in the GEE model – how complex can it be?
Speaker:	Hyoju Chung, UW Biostatistics graduate student
Abstract:	Since the publication of Liang and Zeger (1986), Generalized Estimating Equations (GEE) have become a popular statistical approach to correlated data. In the GEE model, inter-correlated feature of the data is handled through working correlation matrix. Although the regression parameter is known to be consistent even with misspecified working correlation matrix, the efficiency of the regression parameter depends on the working correlation matrix. So, there still remain theoretical and practical issues regarding choice of working correlation matrix and effects of working correlation matrix on inference of the regression parameter. In this talk, I will briefly summarize Liang and Zeger (1986) – their key ideas, extensions and modifications. Then I will address some issues when modeling and estimating working correlation matrix from the data with moderate-large cluster sizes. Then I will discuss some ongoing researches (including two-index asymptotics and simulation study).
November 3, 2004
Title:	Real life biostatistics: working with the transplanters at FHCRC
Speaker:	Barry Storer, FHCRC
Abstract:	This is not really a talk about statistical methodology, but I’ll describe the work of the Clinical Statistics group at Fred Hutchinson and our interaction with the clinical research program in marrow and stem cell transplant. Transplant is a complex procedure, with complications that may require years of treatment and follow-up. There are many challenges in designing clinical studies related to the transplant process, and in the analysis of the data that arise from transplant studies. Some of these challenges relate to the temporal interrelationships of transplant events (competing risks). I’ll talk about some of the statistical issues that we face as collaborators with clinical researchers – in designing, analyzing, and displaying results from transplant studies.
Title:	Evaluating biomarker for early detection of cancer
Speaker:	Yinge Zheng, FHCRC
Abstract:	Molecular markers can be useful in detecting cancer in early stage, determining prognosis and monitoring disease progression. The development and validation of clinically useful biomarkers, especially from high-dimension genomic and proteomic information, pose great research challenges. In this short talk, I will outline a few statistical issues the researchers at the Early Detection Research Network (EDRN) have been working on, including optimal marker combination with censored outcome, correction for overestimations when the initial candidate pool is large relative to sample size and subsequent investigations are expensive.
November 10, 2004
Title:	Methods for the Analysis of longitudinal data incorporating family genetic information
Speaker:	Grace Ge, Biostat doctoral student
Abstract:	Most of the current methods that analyze longitudinal phenotypic data do not take the family genetic information into consideration. However, many medical measures such as Blood Pressure, glucose level, etc. show the family inheritance pattern which implies that Genetics play an important role in the mechanism of these measures. Although these measures are essential in some complex diseases such as Hypertension, Diabetes, Cardiovascular diseases, the role of Genetics remains unclear. Lander and Botstein 1989 proposed a method based on likelihood, named interval mapping (IM), to find QTL(Quantitative Trait Loci). We’ll briefly review this method. Genetic Analysis Workshop 13 (GAW 13) was held on November 11-14, 2002 in New Orleans, Louisiana. This conference focus on finding the region of the genes influencing common, complex diseases and their risk factors, with an emphasis on use of longitudinal data. Golla et al. 2003 analyze GAW13 longitudinal data by fitting linear regression of SBP on age. Then the slope of the SBP over age were used as the phenotype for QTL linkage analysis. Combining the two methods above: Interval Mapping and considering a QTL for the slope of SBP, we proposed a new model including two QTLs based on Likelihood. Then the possible interesting cases will be discussed.
November 17, 2004
Title:	Segmented Linear Regression for Improved Threshold Estimation in Audiometric Data
Speaker:	Bryan Goldman, Biostatistics graduate student
Abstract:	Segmented linear regression, or changepoint modeling, is essentially a linear spline function in which the location of the cut points is estimated. Although it may be generalized to any number of predictors and any number of cut points, most applications involve the simplest case of a single predictor and single cut point. For example, dose-response relationships have often been modeled this way when it is believed that a minimum dose is required to produce any response, and estimation of this minimum dose, or threshold level, is of interest. In the audiology world, there has been growing interest in threshold estimation using a metric called the distortion-product otoacoustic emission (DPOAE). In this context, the threshold represents the minimum sound level that elicits a response from the ear. However, the methods for estimating this threshold to date have been developed largely ad hoc, and have involved data-driven exclusions, likely leading to biased and inefficient estimation. In this talk I will describe a method for estimating DPOAE thresholds using segmented regression, and compare the properties of this estimator with that derived from previous methods.
December 1, 2004
Title:	Bivariate linkage analysis of quantitative traits using samples of random families : Power, Type I error, parameter estimation of the variance components method
Speaker:	Angel Wan, Biostat graduate student
Abstract:	Linkage methods aim at localizing putative chromosomal locations that influence a particular trait/traits or disease by using marker and pedigree information. In particular, one such semi-parametric method is variance components, which uses a mixed effects linear model to estimate the genetic variance attributable to a putative chromosomal location. In this mixed effects model, the additive components of the trait are specified as the QTL effect, the polygenic effect, the non-genetic sibship effect, and the environmental or residual effect. Traditionally, a single trait analysis is implemented at each time, but under pleiotropy, in which one locus may influence more than one trait, a joint analysis of more than one trait may prove to be more powerful in certain circumstances. Previous studies have focused only on nuclear families or sibships and have considered different component-specific variances; however, these examinations of bivariate linkage analysis in power simulation studies do not investigate other situations or model assumptions that might be violated (Allison, et al 1998; Amos et al 2001). Here the main purpose is to examine thoroughly the simple case of a bivariate linkage analysis in terms of power, Type I error, and model parameter estimation using the variance components method under the following situations: (1) variation of the component-specific variance magnitude (2) variation of the magnitude and the direction of the component;specific correlations (3) non-normal residuals (4) unspecified sibship effect in the analysis (5) family size and structure. The background, methods for the simulation set-up, and preliminary results will be discussed for this thesis talk.
December 8, 2004
Title:	A mutltivariate Neyman-Pearson approach for testing differential expression in microarray experiment
Speaker:	James Dai, UW Biostat student
Abstract:	A common goal of microarray experiments is to identify genes that are differentially expressed between two or more biological conditions. In the past several years, the problem of testing for differential expression has been studied in great effort and many procedures were proposed. A key aspect of existing methods is that traditional frequentist hypothesis testing is first applied in a gene-by-gene basis, then multiple testing problem is taken into account. A truly multivariate approach which targets at maximizing the power of detecting differential expression seems to be lacking. Here we propose a multivariate Neyman-Pearson (MVNP) approach which extends Neyman-Pearson paradigm to a high-dimension setting where thousands of hypothesis are tested. Motivated by maximizing average power for a fixed average type I error, our rule is also optimal for maximizing average power given a fixed false discovery rate (FDR). Instead of using t or F statistic, our estimated scores for differential expression are a weighted average of gene-specific likelihood ratio statistics. It allows us borrows strength across genes without necessarily following Bayesian framework or making extensive model assumptions. In both simulated and real data, our estimated procedure shows marked power improvement with respect to simple t test procedure and the highly used but heuristically motivated SAM procedure. Further exploration of theory and application has been carried out. The method is implemented in the freely distributed EDGE software package. This is ongoing research with Dr. John Storey.