Biostatistics Student Academics Club 2006-2007

SCAD-penalized regression in AFT models 
5/2/07
Huiliang Xie

Abstract: In linear regression, variable selection becomes important when a large number of predictors are present, for the purpose of interpretation and prediction. As a solution to this, penalized regression can perform variable selection and coefficient estimation simultaneously. In survival analysis, the presence of right-censored responses complicates the situation. Theoretical properties of variable selection via penalized partial likelihood was studied for the Cox model in Fan and Li (2002). Our study focuses on the accelerated failure time (AFT) regression with the smoothly clipped absolute deviation (SCAD) penalty, when the responses are possibly censored. The study is framed to allow the number of predictors p_n to go to infinity as the number of observations n goes to infinity.  Under reasonable assumptions the estimator is demonstrated to be consistent in variable selection and asymptotically normal for the nonzero coefficients, as if the subset of significant variables were known beforehand. The majorize-minimize algorithm is adapted to compute the SCAD-penalized least squares estimator. Finite sample behavior of this estimator is studied via simulation and compared with other approaches.

Bayesian Power and Sample Size Calculations for Multi-Center Clinical Trials with Binary Responses
4/25/07
Emine Bayman

Abstract: An important issue is to decide the total sample size, and sample size for each center in a multi-center clinical trial. A Bayesian hierarchical model for binary responses representing success or failure is assumed, with exchangeable random center effects. A concept of Bayesian power is defined as the probability of reaching a specific conclusion for a fixed set of parameter values. An algorithm is described for determining the sample size to give appropriate Bayesian power. It is assumed that there are 2 treatments and N centers. Each center has a pre-specified relative ability to enroll. In addition, the impact of balance and imbalance between sample sizes in each center on power is examined. The algorithm is implemented using WinBUGS and R, and the R2WinBUGS interface.       

Optimal Adaptive Group Sequential Design for Phase II Clinical Trials: A Bayesian Decision Theoretical Approach
4/20/07
Yiyi Chen

Abstract: In designing and conducting a phase II clinical trial, various factors should be taken into consideration, such as the cost of conducting the trial, the severity of adverse event, the utility gain/loss of patients resulting from the study treatment, the potential profit that pharmaceutical companies may derive from effective treatments, the cost of waiting to make a decision, etc.  We propose an adaptive group sequential design using a Bayesian decision theoretic approach based on consideration of the aforementioned factors.  Efficacy and toxicity are monitored simultaneously as bivariate binary outcomes, and trade-off between them is allowed. A carefully contrived clinical trial is presented as an example to illustrate our approach.  

Making PowerPoint-like Presentation Using LaTeX Class Prosper and Editing Statistical Programs Using an Able Editor Emacs Speaks Statistics (ESS)
4/6/07
Suhong Zhang

Abstract: The prosper class permits producing high quality PDF slides for both printing and displaying with a video-projector. It offers many features for lively presentations. It is also easily customizable. You can edit any of the default styles to create your own style -- colors, ornaments, background, font and margins. This talk is meant to give a tutorial on how to produce lively features using Prosper. Emacs Speaks Statistics (ESS) is an Emacs package. It adds two major modes to Emacs: (1) ESS mode for editing statistical languages like SAS, S-PLUS, R and BUGS; and (2) inferior ESS (iESS) mode for interacting with inferior  statistical processes like S-PLUS or R. ESS knows the  syntax and grammar of statistical analysis packages and  provides consistent display and editing features based on  that knowledge. It provides a convenient way of writing and executing code without frequently switching between programs. This talk is meant to share some of the experiences of using ESS as a beginner.

Assessment of Bayesian Estimates of Biomarker's Surrogacy for a Time-to-Event Clinical Endpoint in a Single Trial
3/30/07
Qian (Cicci) Shi

Abstract: Cowles developed the Bayesian joint models (BJM) to estimate two widely accepted biomarker's surrogacy measures for time-to-event clinical endpoint. They are relative effect (RE) and adjusted association (AA) proposed by Buyse et al.. RE is intended to predict the treatment effect on the clinical endpoint based on its surrogate. AA measures the individual level association between the clinical and surrogate endpoints, adjusted for treatment. We extended Cowles's work to a class of Bayesian joint models. The fundamental framework was established. A summarized individual level surrogacy measure – R 2 for multiple biomarkers (RMB) was proposed. Simulation studies under one specific BJM show that AA can be precisely and accurately estimated.  Inclusion of an endpoint with high AA improves the prediction of censored failure time significantly. In contrast, estimation of RE has tremendous variability, and is influenced by the significance of the treatment effects. Alternative measures of RE, multiple surrogate endpoints, and ideas of using a marker with high individual-level surrogacy in a single trial were discussed.

Inference on Association Measure for Bivaraite Survival Data with Hybrid Censoring and Application to an HIV Study
2/16/07
Suhong Zhang

Abstract: Some prior studies suggest that GBV-C, a harmless virus, increases the survival time among HIV patients, while some other studies fail to find the beneficial effect.  All the previous studies compared the Kaplan-Meier survival curves between the GBV-C positive and negative HIV patients.  We now propose to estimate the association between the HIV survival time and the GBV-C persistence time, in the case that HIV survival time is subject to right censoring and the GBV-C persistence time is observed as current status data. We established consistency and asymptotic normality of the proposed estimator and our simulation studies showed its good performance given a reasonable sample size. The application of our method to a Multicenter Aids Cohort Study shows that the GBV-C is associated with increased survival among HIV and GBV-C co-infected individuals.

Analysis of DNA Copy Number Variations Using Penalized Least Absolute Deviations Regression With Fused Lasso
11/29/06
Xiaoli Gao

Abstract: Deletions and amplifications of the human genomic DNA copy number are the cause of numerous diseases such as cancer. Therefore, the detection of DNA copy number variations (CNV) is important in understanding the genetic basis of human diseases. Various techniques and platforms have been developed for genome-wide analysis of DNA copy number, such as array-based comparative genomic hybridization (aCGH), SNP arrays, and high-resolution mapping using high-density tiling oligonucleotide arrays. Since complicated biological and experimental processes are involved in these platforms, data can be contaminated by outliers. Inspired by the robustness property of the LAD regression, we propose a penalized LAD regression with the fused lasso penalty for detecting CNV. This method incorporates the spatial dependence and sparsity of CNV into the analysis and is computationally feasible for high-dimensional array-based data. We evaluate the proposed method using simulation studies and demonstrated it on two real data examples.

Semiparametric Estimation Methods for anel Count Data Using Monotone Polynomial Splines
11/6/06
Minggen Lu

Abstract: We study semiparametric likelihood-based methods for panel count data using monotone polynomial splines with proportional mean model. The generalized Rosen algorithm, proposed by Zhang & Jamshidian (2004), is used to compute the estimators of both and. We show that the proposed spline likelihood-based estimators of are consistent and their rate of convergence can be faster than n1/3. The normality of estimators of  is also established. Simulation studies with moderate samples show that the spline estimators of are more efficient both statistically and computationally than their alternatives proposed in Wellner & Zhang (2005). A real example from a bladder tumor clinical trial is used to illustrate the methods.

Student Academic Club 2005-2006

Lorem ipsum dolor

Lorem ipsum dolor sit amet, consectetaur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.