Abstracts
Pablo Bonangelino, Food and Drug Administration. "Bayesian Designs for Therapeutic Medical Device Trials"
This talk will review the Bayesian designs which have been used in medical device trials that have been submitted to the General and Surgical Devices Branch of FDA's Center for Devices and Radiological Health (CDRH) in recent years. I will begin with an overview of the approach taken by our Branch towards Bayesian statistical methods. I will then discuss examples of Bayesian methods, which include likelihood based methods for incomplete data, Bayesian adaptive designs using predictive probability, and incorporating prior data in a confirmatory study. Regulatory considerations regarding these designs are explored. I will conclude by considering the future for Bayesian techniques in therapeutic device studies.
Vladimir Dragalin, Wyeth. "Adaptive model-based designs in clinical drug development"
The objective of a clinical trial may be either to target the maximum tolerated dose or minimum effective dose, or to find the therapeutic range, or to determine the optimal safe dose to be recommended for confirmation, or to confirm efficacy over control in a Phase III clinical trial. This clinical goal is usually determined by the clinicians from the pharmaceutical industry, practicing physicians, key opinion leaders in the field, and the regulatory agency. Once agreement has been reached on the objective, it is the statistician's responsibility to provide the appropriate design and statistical inferential structure required to achieve that goal. There is a plenty of available designs on statistician's shelf. The greatest challenge is their implementation. We exemplify this in three case studies.
Mani Lakshminarayanan, Pfizer, Inc.
"Handling Count Data in Clinical Trials"
In clinical trials, quantifiable endpoints must be appropriately chosen to ensure that the primary objectives are met based on rigorous statistical analysis. These endpoints may include measurements with continuous variation, dichotomous, event times, ordered and unordered categories and counts. Clinical endpoints based on count data are most common in trials designed to study chronic diseases where the counts are done within an interval in time, usually resulting in a number 0 or 1 (eg, survival). Endpoints such as survival counts are well described in the statistical literature. The type of count data that is primarily considered in this paper deals with count measurements that can take on values higher than 0 or 1. Couple of recently published articles describe such data: number of new enhancing lesions read on MRI scans in multiple sclerosis trials (NEJM 2003; 348: 15-23) and the number of seizures in epilepsy trials (NEJM 2001; 345:311-318). For ease of reporting, these types of count data are often reformulated to a continuous or a discrete form before the primarily analysis is performed and presented in clinical study reports. As a result, the underlying analysis might suffer from loss of information similar to situations where a clinical response is defined (as a dichotomous) variable based on a continuous endpoint. In this presentation, clinical endpoints based on count data are considered as realizations from a Poisson distribution and then use Poisson regression models as a general framework for the underlying analysis. Common approaches for treating such count data are compared including over-dispersed count data. Simulations and examples are used to illustrate the approaches and their comparisons.
Key Words:
Poisson distribution; Over-dispersion; Negative binomial; Random effect; Zero-inflated Poisson models; EM algorithm
David Madigan, Rutgers University. "Lasso Regression: Some Recent Developments."
L1-penalized regression has attracted considerable attention in recent years.
The approach provides simultaneous variable selection and shrinkage and
scales to ultra high-dimensional problems. Recent developments
include the group lasso and the fused lasso. This talk will review these
developments. Motivated by applications in vaccine studies and
early-stage drug safety modeling I will describe some recent efforts
that combine hierarchical partition models with fused and grouped lasso
within a Bayesian framework. (joint work with Suhrid Balakrishnan)
Len Oppenheimer (J&J), Bruce Binkowitz (Merck),Amy Ko (Merck) "Extension Studies."
In todays todays drug development environment there is a heightened need to
obtain more comprehensive risk/benefit information as soon as possible.
Extension studies can provide the much needed long term safety and efficacy data sooner
than traditional multiple independent studies and obtain this data in a cost
effective manner.This talk will define and detail some of the advantages of
extension studies as well as discuss issues related to study design,
hypotheses, analysis, reporting, and interpretation. The concepts discussed will be
illustrated using actual examples from drug development programs.
Jose Pinheiro, Novartis. "Evaluating Adaptive Dose-Ranging Studies and Methods: A Report from the PhRMA Working Group."
Inadequate dose selection following Phase II trials has been indicated
by the FDA and other health authorities as a key factor influencing the
decline in the number of successful submissions over the last decade.
To investigate and propose recommendations on how to address this issue,
PhRMA formed a working group on Adaptive Dose-Ranging Studies (ADRS).
In this innovative class of designs, the number of doses and the allocation
of patients to doses are allowed to change during the study, by incorporating
information in the accruing efficacy and safety data. This presentation will
discuss the main results of an extensive simulation study conducted by the
ADRS working group and its recommendations to improve the efficiency of
dose finding trials.
Hui Quan, Daowen Zhang, Ji Zhang and Laure Devlamynck, Sanofi-Aventis, "Analysis of a Composite Endpoint with Missing Data in Components"
Composite endpoints are often used in clinical trials in order to increase the overall event rates, reduce the sizes of the trials and achieve desired power. For example, in a trial to study the effect of a treatment on the prevention of venous thromboembolic events (VTE) after a major orthopedic surgery of the lower limbs, the primary endpoint is usually a composite endpoint consisting of any Deep Vein Thrombosis (DVT) identified by systematic venography of lower limbs, symptomatic and well documented non-fatal pulmonary embolism, and death from all causes. Just as any endpoints, missing data can occur in the components of the composite endpoint. If a patient has missing data on some of the components but not all the components, this patient may not have complete data but partial data for the composite endpoint. To be consistent with the intention-to-treat principle, the patient should not be discarded from the analysis. In this research, we propose approaches for the analysis of composite endpoint with missing data in components. The main idea is to first derive the probabilities of all possible study outcomes based on the appropriate model and then to combine them to obtain the overall rate for the composite endpoint. Simulations are conducted to compare the approach with several naove methods. A data example is used to illustrate the application of the approach.
William E. Strawderman, Rutgers University. "Combining Biased and Unbiased Estimates in High Dimensions".
Occasionally, there is more than one source of information concerning a parameter vector
of interest. For example, a (multivariate) sample mean may give an unbiased estimator,
while a (biased) regression based estimator may also be available. Often, the regression
based estimator is biased as a result of non-linearity. For example, if there is a regression
of the log of the variable of interest on known covariates, the inverse (exponential) transform
to the variable of interest will introduce bias. We propose a method to combine the
estimators in a way that mimics the behavior of the usual weighted (inversely to variance)
linear combination when both estimators are unbiased, but which improves on the first
unbiased estimator whether both estimators are unbiased or not. We will describe the
method, study its theoretical properties, give an example, and present results of a simulation study.
|
