11/5/2008: Modeling protein subfamilies; finding the number of mixture components from sequences of generalized Bernoulli random variables

Youyi Fong

Genome sequencing and annotation projects generate a lot of predicted protein sequences. To learn more about the biological functions of these proteins, we investigate subfamily structure of protein families. A protein family is a group of proteins that share similar sequences and functions, but it is generally not a homogeneous group. We propose as model for protein sequences belonging to a protein family, as iid realizations of a mixture of high-dimensional (p=20 to 50) generalized Bernoullis, and identify the order of this mixture model. Several methods are applied to a simulation dataset; we take a close look at the winning approach and discuss the challenges in applying this method to real datasets and possible extensions of the model.

10/29/2008: A Bayesian Method for Cross-Trial Inference in the Non-Inferiority Setting

Siobhan Everson-Stewart

In non-inferiority clinical trials, researchers would often like to show that the new therapy would have been proven superior to placebo, had one been included in the trial, in addition to being non-inferior to the active control. Such proof requires cross-trial inference, which previously required strong, often unrealistic assumptions. For example, most traditional methods require that the constancy assumption holds, i.e., that the effect of any one therapy remains constant from trial to trial. In therapeutic settings with changing patient populations, however, this assumption is often highly suspect. This can lead to the approval of ineffective and even harmful therapies. We present a Bayesian method for cross-trial inference when treatment effect changes as a function of the study population. The performance of this method is examined under a variety of circumstances.

10/22/2008: Statistical performance of group sequential methods for evaluating post-market vaccine and drug safety: A simulation study

Shanshan Zhao

Conducting observational post-marketing vaccine and drug safety surveillance is important for detecting rare adverse events (AEs) not identified pre-licensure. Highly frequent monitoring of vaccine safety has been proposed using continuous sequential monitoring methods, but such frequent testing may result in a loss of power. Interim monitoring using group sequential methods has been widely used to monitor safety in clinical trials, but the performance has not been examined in observational settings or when more frequent surveillance is desired. We conducted a simulation study to compare the power and timeliness to correct detection of an elevated AE risk among vaccinated people. We vary both the testing frequency (daily, weekly, monthly, quarterly) and shape of the stopping boundary (Pocock, O’Brien-Fleming). Results of this evaluation can inform the design of future sequential monitored safety studies.

10/15/2008: Comparison of ROC Regression methods for Ordinal Outcomes

Daryl Morris

The accuracy of a diagnostic test can be summarized in a receiver operating characteristic (ROC) curve, a plot of the true positive (TP) versus false positive (FP) rates associated with varying thresholds c for the test results Y: TP(c) = P[Y > c | disease present] and FP(c) = P[Y > c | disease not present]. ROC curves have become well accepted measures of accuracy in diagnostic medicine. They display the trade-offs possible between increasing true positive and increasing false positive rates as the positivity criterion varies. They describe the inherent capacity of a test for discriminating diseased from non-diseased subjects without linking the test to any specific positivity criterion. ROC curves are particularly useful for comparing diagnostic tests since tests are put on the same scale (even if the test result variables themselves are on entirely different scales) and the scale relates directly to the notion of accuracy.

Methods for estimating and comparing ROC curves have long been available. ROC regression methodology offers the opportunity to investigate how factors such as characteristics of the study subjects or test environment influence test accuracy. For ordinal outcomes, Tosteson and Begg (1988) proposed the use of ordinal regression models to induce regression models for ROC curves. For continuous markers, we have proposed direct modeling of the ROC curves themselves (“Placement Value methods”). We have not previously applied the direct modeling in the case of ordinal outcomes.

In this talk we compare and contrast joint ordinal regression method models (a la Tosteson & Begg) and direct modeling of the ROC curves when applied to ordinal outcomes.

10/08/2008: Sub-clinical disease measures on HIV infected and uninfected participants

Chris Delaney, Post-Doctoral Research Fellow, CHSCC

When comparing measures of Carotid Intimal-Medial Thickness, a measure of sub-clinical cardiovascular disease, between HIV-infected cohorts and unifected cohorts, serious concerns were noted with the possible impact of reader effects. This talk will discuss the magnitude of these effects, the solutions implemented and the estimates of the effect of HIV infection on Carotid Intimal-Medial Thickness.