Winter 2018

Title: A comparison of missing data methods in time-to-event analyses in the presence of time-varying covariates

Abstract: Changes in exposure over time are common in many studies in medicine, e.g., when patients change treatment regimens in response to their current health status. However, it is not always possible to record changes in measurements that may trigger such treatment alterations, giving rise to missing data in potential confounders. A theoretically sound and reasonably accessible method to address missing data is multiple imputation (MI). Implementation of MI in the context of longitudinal studies, particularly with right-censored outcomes, poses challenges, however. Our goal is to evaluate methods for handling missing data induced by incorporating time-varying information. We will characterize the properties of estimates obtained from an extended Cox model with time-varying covariates that makes use of a standard implementation of MI and compare these to commonly applied methods for handling missing data through a series of simulation studies.

About Dr. Montez-Rathgave: Dr. Montez-Rath completed her PhD in Biostatistics from Boston University in 2008 focusing on methods for modeling interaction effects. She has been working as a biostatistician in the Division of Nephrology at Stanford University since 2010 where she collaborates with numerous clinical investigators to study a variety of research questions relevant to kidney disease. Her methodological interests are mainly data-driven and much of her research involves data collected from the USRDS Database, a rich data source describing all end stage renal disease patients in the United States. She is also currently a co-investigator on a PCORI-funded project entitled “The Handling of Missing Data Induced by Time-Varying Covariates in Comparative Effectiveness Research Involving HIV Patients.” Dr. Montez-Rath's presentation today describes some of this work.

Video:
 

Title: Challenges in Statistics in the Semiconductor Industry

Abstract: Statisticians in the semiconductor industry face many challenges in order to assist engineers in creating sellable material. While the main goal is to create as much sellable material as possible, we at times look at large data sets and small data sets in order to increase yield. Increasing yield can be done in many ways, such as reducing the number of particles that are present on the material or implementing improvements that decrease down time for the tools that process material. While money is the biggest factor, we also must consider many other factors depending on the situation, such as physical constraints, detection limits, and sampling plans. This talk looks at these various challenges and the factors that play into them.

About Dr. Beverlin: Lucas Beverlin is a statistician at Intel, where he assists fab engineers with optimizing statistical process control and teaches statistics courses on basic statistics and design of experiments. He graduated from Louisiana State University with master's degrees in mathematics and applied statistics, and from Iowa State University with a Ph.D. in statistics. His research interests include nonlinear modeling, reliability, and machine learning.


Video: