| April 7, 2004 | |
| Title: | Nonparametric Confidence Intervals for the One and Two-sample Highly Skewed Data |
| Speaker: | Phil Dinh, graduate student UW Biostatistics |
| Abstract: |
Confidence intervals for the mean of one sample and the
difference in means of two independent samples based on the ordinary-t
statistic suffer deficiencies when the samples come from highly skewed
families and the sample sizes are small to moderate. In this talk, I will
present simulation study evaluating several existing methods and
propose new methods to improve coverage accuracy. The methods examined
include the ordinary-t, the bootstrap-t, the biased-corrected acceleration
(BCa), and three new intervals based on transformation of the t-statistic.
Our study shows that the new transformation intervals and the bootstrap-t
intervals give best coverage accuracy for a variety of skewed
distributions; and that our new transformation intervals have shorter
interval lengths.
|
| April 14, 2004 | |
| Title: | Estimating Causal Effect Using Observational Data: The Intensity-score Approach to Adjusting for Confounding |
| Speaker: | Mary Redman, graduate student UW Biostatistics |
| Abstract: |
Observational studies are characterized by non-random allocation of
treatments to study participants. Measures of the impact of a treatment
on an outcome can be subject to confounding by both measured and
unmeasured factors. Recently, Brumback, Greenland, Redman, et. al.
introduced the intensity-score approach to adjusting for confounding to
estimate the effects of both point-source and time-varying treatments.
The intensity score is a contrast between treatment received and expected
treatment conditional on confounders. This approach relies on proper
specification of a model for the expected treatment as a function of
measured confounders and specification of form of treatment effect (i.e.
presence or lack of effect-measure modification by measured confounders)
to obtain unbiased parameter estimates.
In general, when deriving the intensity score, the true expected treatment is unknown and an estimate needs to be used in its place. In this talk I will give a brief overview of the intensity-score approaches and discuss the impact of using estimates of the expected treatment on the variance of the treatment effect parameter.
|
| April 21, 2004 | |
| Title: | The reversed Berk-Jones statistic |
| Speaker: | Leah Jager, graduate student UW Statistics |
| Abstract: |
In nonparametric testing problems, we often use statistics based on the
empirical distribution function to test whether or not the underlying
distribution of the data is what we think it might be. One classical
example is the Kolmogorov-Smirnov statistic. Berk and Jones introduced
another such statistic which has certain optimality advantages over
Kolmogorov-Smirnov. Here we'll look at a closely related statistic (the
reversed Berk-Jones statistic) and the motivation behind it, as well as
some finite sample characteristics and optimality properties. Consider
yourself warned...
|
| April 28, 2004 | |
| Title: | Image Analysis and Signal Extraction from cDNA Microarrays |
| Speaker: | Tracy Bergemann, graduate student UW Biostatistics |
| Abstract: |
Open discussions about microarray technology invariably lead to concerns
about data quality and objectivity in assessment. The focus of my
dissertation has been to address some of these concerns in a simple and
straightforward manner. The first half of the talk will discuss image
analysis techniques for robust automated detection of microarray spots.
These techniques are packaged into a MATLAB application called
SignalViewer that is freely available at
http://qge.fhcrc.org/signalviewer.
The second half of the talk will cover methods for describing spot quality measures. The goal here was to develop a metric that describes within-spot variability while accounting for spatial correlation of image pixels. This will include a discussion of estimating equations, weighted estimating equations and prediction error. Methods are evaluated on real and simulated data.
|
| May 5, 2004 | |
| Title: | Competing Risk Current Status Data |
| Speaker: | Marloes Maathuis, graduate student UW Statistics |
| Abstract: |
We study competing risk data subject to current status censoring. I will
discuss example data sets from cross sectional studies in which this type
of data arises. We are primarily interested in nonparametric estimation of
the subdistribution functions, i.e. the cumulative probabilities of a
certain failure type.
Our main focus is on the nonparametric maximum likelihood estimator of the subdistribution functions. However, for comparison, we also look at a very simple 'naive estimator'. I will talk about properties of both estimators, such as computational aspects, graph theoretic interpretations, uniqueness, consistency, rate of convergence, and (first steps towards) a limiting distribution.
|
| May 12, 2004 | |
| Title: | Regression Analysis of Longitudinal Data with Subject-specific sampling Times |
| Speaker: | Patra Miksova, graduate student UW Biostatistics |
| Abstract: |
Abstract available at:
http://students.washington.edu/~miksova/research.pdf
|
| May 19, 2004 | |
| Title: | Surveillance of Geographical Cancer Incidence |
| Speaker: | Alan Dabney, graduate student UW Biostatistics |
| Abstract: |
Cancer surveillance involves the systematic examination of
incidence rates over a predefined time period and collection of geographical
regions for localized increases in risk. State departments of health
receive frequent calls from citizens concerned with perceived clusters of
cancer. The vast majority of such alarms turn out to be unfounded. A
surveillance method, in fact, is unlikely to turn up many substantive
clusters. However, in light of the fact that incidence data for all cancers
is collected periodically at local registries (CSS at the Hutch, for
example), it is perhaps a duty of those registries to make an official
effort. I will motivate the use of surveillance with maps of bladder and
lung cancer incidence in Washington census tracts. I will then discuss
existing surveillance methods and propose the use of a Bayesian hierarchical
model.
|
| May 26, 2004 | |
| Speaker: | Bryan Comstock, graduate student UW Biostatistics |
| Abstract: |
Biomarkers have become an increasingly used tool in both the diagnosis and
monitoring of various diseases. Increased levels of serum carbohydrate
antigen 19-9 (CA19-9) have long been known to be associated with pancreatic
cancer. In this commonly fatal form of cancer, CA19-9 levels decrease
immediately following initial treatment but are then expected to rise again
in the following months with the almost sure resurgence of the tumor.
Previous studies have primarily focused on determining whether or not CA19-9
is a worthwhile screening tool for pancreatic and other forms of cancer.
While several studies have also examined the prognostic ability of baseline
CA19-9 on the time of survival, the focus of our methods and analyses here
is to create a longitudinal model for serial CA19-9 measurements. By doing
so, our goal is to take a first step towards examining the prognostic value
of having longitudinal CA19-9 data on patient survival. Using serial
post-surgical CA19-9 data on 262 pancreatic cancer patients, we use a reversible jump MCMC
algorithm to average between longitudinal CA19-9 models in a Bayesian
framework. Before we proceed with a fully Bayesian joint model for CA19-9
and time of survival, the resulting longitudinal model is used as the basis
for a time-varying covariate in a Cox regression model to assess its
potential predictive value.
|
| June 6, 2004 | |
| Title: | Estimation when the outcome of interest is subject to misclassification |
| Speaker: | Pamela Shaw, graduate student UW Biostatistics |
| Abstract: |
In many settings, presence or absence of disease is measured with an
imperfect test. These misclassified outcomes can lead to biased estimates of
covariate effects and survival time. This talk will examine these issues and
methods to address them for binary outcomes and time to event data. For
logistic regression, the EM algorithm can be applied to obtain unbiased
estimates of the parameters. For time to event data, the different
computational issues arise with discrete and continuous time. For discrete
time, methods for both the product limit survival estimate and covariate
effects for the proportional hazards model have been developed. Existing
approaches and some open questions for continuous time data will be
presented. Particular attention will be given to the situation of interval
censored data; these data commonly arise in clinical settings where repeat
testing is performed for the detection of asymptomatic disease.
|
| June 6, 2004 | |
| Title: | Estimation when the outcome of interest is subject to misclassification |
| Speaker: | Pamela Shaw, graduate student UW Biostatistics |
| Abstract: |
In many settings, presence or absence of disease is measured with an
imperfect test. These misclassified outcomes can lead to biased estimates of
covariate effects and survival time. This talk will examine these issues and
methods to address them for binary outcomes and time to event data. For
logistic regression, the EM algorithm can be applied to obtain unbiased
estimates of the parameters. For time to event data, the different
computational issues arise with discrete and continuous time. For discrete
time, methods for both the product limit survival estimate and covariate
effects for the proportional hazards model have been developed. Existing
approaches and some open questions for continuous time data will be
presented. Particular attention will be given to the situation of interval
censored data; these data commonly arise in clinical settings where repeat
testing is performed for the detection of asymptomatic disease.
|