|
|
|
|
Title: |
Determining non-compliance status in a clinical trial through model-based methods |
|
Speaker: |
Katie Gower, Graduate student, Biostatistics |
|
Abstract: |
In a clinical trial, compliance is a measure of how closely a participant adheres to the prescribed treatment that he or she was randomized to receive. Non-compliance can severely impact the integrity of a clinical trial: if participants do not adhere to the treatment regimen, there will be a diminished difference in the treatment effect. Therefore it is important to have some measure of the extent of non-compliance. The Prostate Cancer Prevention Trial (PCPT) is designed to determine if a daily dose of finasteride will assist in preventing prostate cancer. The PCPT employs two measures of compliance: pill counts and a biomarker. For this talk, I will present a test of non-compliance based on a mixed effects model for the biomarker and give an evaluation of the test based on the pill counts. |
|
|
|
|
Title: |
Quality of Afterlife: dropout and death in longitudinal data |
|
Speaker: |
Thomas Lumley, Assistant Professor, Biostatistics |
|
Abstract: |
I will explain why maximum likelihood methods for longitudinal data analysis can work in the presence of missing data, and why this is a bad thing when some subjects die. I will also handwave vaguely about what can be done about this problem. |
|
|
|
|
Title: |
Comparing predictive accuracy of prognostic factors |
|
Speaker: |
Chaya Moskowitz, Graduate student, Biostatistics |
|
Abstract: |
This will be a practice for my general exam. This talk can be tailored to the audience's experience level so as to be accessible to all students. The ability to compare prognostic factors can have many important applications, both in the medical field and beyond. Existing statistical methodology for this purpose, however, particularly in the setting of a prospective study design, is very limited. We propose summarizing predictive accuracy of prognostic factors with two familiar and easily understood quantities, the positive and negative predictive values, and extend their usual definitions for binary outcomes and binary factors to situations where the outcomes are censored failure times and/or the factors are continuous. We recommend quantifying the difference in the predictive accuracy of two factors in terms of relative predictive values and propose a marginal regression framework for estimation of the predictive and relative predictive values that can accommodate clustered data and adjustment for covariates. The methods to be presented provide a unified framework for evaluating and comparing the predictive accuracy of two prognostic factors whether they are binary or continuous predictors of a binary or failure time outcome. |
|
|
|
|
Title: |
Assessing maternal genetic effects using the log-linear approach to case-parent triad data |
|
Speaker: |
Jackie Starr, Graduate student, Epidemiology |
|
Abstract: |
We recently applied for funding
to evaluate the hypothesis that the development of testicular cancer in sons
is associated with maternal genetic variation. We proposed to use a recently
developed log-linear approach to case-parent triad data as our study design
and analytic method (Wilcox, Weinberg, & Lie, 1998; Weinberg, Wilcox
& Lie, 1998). This design/analysis can be used to assess mothers' genetic
associations with disease in their children. It can also be used to assess
cases' own genetic associations with their disease while obviating potential
bias due to ethnicity; or to evaluate whether a genetic locus potentially
relates to a disease through a parent-of-origin effect (e.g. the association between
a trait and variation at an imprinted locus depends on whether that variation
was inherited from one's mother or one's father). I will describe the rationale behind this log-linear approach and present results from simulations performed 1) to estimate statistical power in our study setting; 2) to evaluate performance at large and smaller sample sizes; and 3) to compare this approach to a case-control design. |
|
|
|
|
Title: |
Analysis of Binary Longitudinal Data with Dropout and Death |
|
Speaker: |
Brenda Kurland, Graduate student, Biostatistics |
|
Abstract: |
Parameters for longitudinal binary data regression models are often susceptible to bias when dropout and/or association (correlation) between responses are not accommodated properly. Semiparametric methods (such as GEE) and likelihood-based methods (such as marginalized transition models) for regression models of longitudinal binary data with monotone dropout may be compared with respect to bias and efficiency. Likelihood-based methods have an efficiency advantage over semiparametric methods, and dropout is ignorable if data are missing at random. Therefore, the likelihood-based marginalized transition model (Heagerty, '00) is extended to accommodate nonignorable dropout. The second part of my dissertation research considers longitudinal studies in which both death and dropout occur. Reasons for treating death and dropout separately are examined, and possible analysis methods compared. |
Last Modification: 10 October 2001