The two invited papers to be presented at the workshop are:
``Multi-Course Treatment Strategies for Rapidly Fatal Diseases''Full Paper as Postscript or PDF |
|
Peter F. Thall (Department of Biostatistics, Box 447) Hsi-Guang Sung (Department of Biostatistics, Box 447) Elihu H. Estey (Department of Leukemia, Box 428) M.D. Anderson Cancer Center 1515 Holcombe Boulevard Houston, TX 77030, USA |
| Therapy of rapidly fatal diseases often requires multiple courses of treatment. In each course, the treatment may achieve the desired clinical goal, "response," the patient may survive without response, "failure," or the patient may die. When treatment fails in a given course, it is common medical practice to switch to a different treatment for the next course. Most statistical approaches to such settings simply ignore the multi-course structure. They characterize patient outcome as a single binary variable, combine death and failure, and identify only one treatment for each patient. Such approaches waste important information. We provide a statistical framework, based on a family of generalized logistic regression models, that incorporates historical data while accommodating multiple treatment courses, a trinary outcome in each course, and patient prognostic covariates. The framework serves as a basis for both data analysis and outcome--adaptive clinical trial conduct. Rather than focusing on individual treatments, we evaluate multi--course treatment strategies that specify which treatment to give in each course within each prognostic subgroup. We describe a general approach for constructing clinical trial designs that may be tailored to different multi--course settings. For each prognostic subgroup, based on a real-valued function of the covariate-adjusted probabilities of response and death, the design drops inferior treatment strategies during the trial and selects the best strategy at the end. The methodology is illustrated in the context of a randomized two--course, three--treatment acute leukemia trial with two prognostic covariates. The model is first fit to an historical data set to obtain a reasonably informative prior on non-treatment related parameters for use in trial design and conduct. We describe a simulation study of the design under several clinical scenarios. The simulations show that the method can reliably identify treatment--subgroup interactions based on moderate sample sizes. Extensions of the leukemia trial design to more complex multi-course settings are discussed. |
| Discussants: Thomas Louis and J.R. Lockwood, and Steve Goodman |
``Bayesian Methodology in Genomics Research''Full Paper as Postscript#1 or PDF | Jun S. Liu Xiaole Liu Mayetri Gupta Dept. of Statistics, Harvard University Charles E. Lawrence Medical Informatics, Stanford University, Stanford, CA |
| With the completion of genomes of many species and the advances of microarray technologies, we begin to possess a tremendous amount of valuable biological data --- but these ``raw products'' are still far from usable. One of the most challenging problems of this century is to decipher this huge amount of biological information and turn the data into knowledge. The past decade has witnessed a number of successful applications of sophisticated statistical models in computational biology. This article focuses on one of these success stories: using statistical methods to find short repetitive patterns in a set of DNA or protein sequences, a task often referred to as {\it motif discovery}. In particular, we review a few probabilistic models that have recently been shown useful for motif discovery and provide a novel framework based on a Bayesian segmentation model to unify these approaches. We show how to combine the dictionary model with the Gibbs sampler and how a segmentation-based motif sampler can be implemented. A few interesting open problems are also discussed. |
| Discussants: Gary Churchill and Michael Newton |
| The previous five Workshops provided extended presentation and discussion on diverse topics. |
|
|
Panel Discussion on Scientific ReportingStandards for reporting of Bayesian analyses in scientific investigationsLink to paper | Panel: Constantine Gatsonis and Steve Goodman Moderators: Donald Berry, Joel Greenhouse |
|
The enormous methodologic and computational advances in the last two
decades have brought Bayesian statistics into the mainstream of scientific
study design and analysis. However, whereas considerable experience and
consensus exists for the manner in which frequentist analyses should be
conducted and reported, the corresponding consensus regarding Bayesian
analyses is only now emerging. An initiative is underway to develop
guidelines for authors and editors of scientific journals, with an emphasis
on medically-related publications, for reporting of Bayesian analyses. These
guidelines will be accompanied by a companion document that explains the
rationale for each. This initiative is modeled after recent similar
initiatives in the medical literature and will include several steps of
drafting and revising the documents, based on input from a broad spectrum of
the research community, including statisticians, scientific subject-matter
researchers, and journal editors.
The 6th Workshop on Case Studies in Bayesian Statistics provides an ideal forum for in-depth discussion of Bayesian reporting standards by methodologists with special interest and expertise in Bayesian methods. The format will be that of a panel discussion. After brief prepared statements by the organizers and two invited discussants, the floor will be open to workshop participants. Drafts of the material will be made available in this web site approximately two weeks in advance of the workshop. In addition to the exchange of opinions during the panel discussion, workshop participants will have a chance to comment on future drafts of the documents as they are finalized. |
|
|
Recognition of Faces versus Greebles:
| Kert Viele, Robert E. Kass, Michael J. Tarr, Marlene Behrmann and Isabel Gauthier
|
Prosopagnosia is a cognitive deficit in which patients are impaired at
recognizing faces. There is considerable debate as to whether the part
of the brain that is damaged in patients with prosopagnosia is
substantially dedicated to face recognition or whether it is a locus
for identification of similar objects within more general classes.
This debate is an instance of a controversy between advocates of a
modular view of brain function and those who hold that brain activity
should be understood as distributed across a network of locations.
The issue is of critical importance for understanding the relationship
between the brain and behavior and has led to heated exchange in the
last few years. To investigate it, two patients and 30 controls were
presented with face and non-face images in pairs and were asked to
state whether the two items in the pair were the same or
different. This resulted in a total of approximately 20,000 binary
observations. The inference problem is one of comparing multiple
pairs of Binomial outcomes, that is, multiple $2\times 2$ tables: we
may consider ratios of odds ratios, or differences of log-odds ratios;
in probit form the comparisons fit within the framework of the
widely-applied psychophysical technique known as ``signal detection
theory'' (ROC curves) and this is the modeling approach we follow in
the results presented here.
To answer the questions of scientific interest, we here consider the
statistical comparisons in terms of both estimation and testing. In
estimation, for each of nine combinations of subject and stimulus
conditions we may obtain posteriors on the coefficient of a binary
explanatory variable in a probit model, and from these generate the
posteriors on the quantities of interest. Actually, we must do this
for each of four different experimental conditions, thus beginning
with a total of 36 posteriors. These relatively straightforward
analyses might be considered adequate. For example, we can estimate
the magnitude of the deficit of each of the patients in recognizing
faces as compared to non-face objects, and we can provide a posterior
standard deviation for the estimate. In addition, however, we might
attempt to meet the scientific question head on, by providing a
probability that there is no difference in a patient's ability to
discriminate between faces and his ability to discriminate between
non-face objects (in the experimental setting). This requires
selection among alternative models that, in our context, set
particular interaction terms to zero.
Using an ad-hoc procedure we will describe, some of the conclusions we
drew from the latter approach were consistent across probable models,
and with the estimation approach, but a particular issue is whether
the impairment of one of the patients is more severe for faces than
for common non-face objects and the answer to this is more sensitive
to model choice.
Data and a draft paper are available at
http://www.ms.uky.edu/~viele/casestudies2001/prosoreadme.html
| Discussants: Jim Berger, Sid Chib, Andrew Gelman and
Edward George
| Back to the top of the page | Back to the Bayes01 Home
Page | |