| Room | Time | Schedule |
| NA1-142 | 11:25 — 11:30 | Welcome |
| 11:30 — 12:00 |
Stanley Lemeshow, The Ohio State University Use of Logistic Regression Modeling in Randomized Clinical Trials | |
| 12:00 — 1:15 | Lunch | |
| NA5-68 | 1:15 — 1:45 |
Ralph O'Brien, Cleveland Clinic Foundation Current and Future SAS-Based Tools for Sample-Size Analyses |
| 1:45 — 2:15 |
Nahida Gordon, Case Western Reserve University What Are the Long-Term Relative Benefits and Risks of Adjuvant Therapy for Early Breast Cancer? | |
| 2:15 — 2:35 | Break | |
| 2:35 — 3:30 |
Anastasios Tsiatis, North Carolina State University Efficient Estimation of the Mean of a Time-Lagged Variable Subject to Right Censoring |
In many clinical trials the endpoint of interest may not be available
immediately, but rather evolves over time. Examples of this are
numerous. Survival time is clearly such an example, but also cost of
care, quality adjusted lifetime, or even dichotomous response such as
whether viral load will go below detectable limits after treatment for
AIDS patients are also examples of time-lagged responses. The lag time
may be part of the biological process or due to administrative delays.
Since patient entry is staggered and follow-up is of limited duration,
some of the response variables will be missing or incomplete due to
censoring. For most of these censored time-lagged response endpoints,
standard survival methods for dealing with censoring do not apply
because of induced informative censoring. We will show how the theory
of inverse probability weighting of complete cases developed by Robins
and Rotnitzky can be used to find consistent estimators for the mean
of the time-lagged variable. We will also show how to use additional
information collected during the study to increase efficiency.
The analysis of modern clinical trials may find many useful
techniques that are standard in the analysis of epidemiologic data.
Epidemiologists typically use observational studies to determine the
association between a risk factor and a disease of interest. Because
patients cannot be randomized to exposure groups in epidemiologic
studies, extraneous factors must be controlled for in the analysis of
the data. Failure to appropriately control for these factors can lead
to bias and a corresponding invalidation of results.
In modern clinical trials, the treatment groups are designed to be as
homogeneous as possible. This is done through the randomization of
patients to treatment arms. Alternatively, if there is a known
variable that must be controlled for, randomization can be performed
within strata defined by levels of that variable. Although it is
understood that, in the long run randomization will produce treatment
groups that are well balanced (or homogeneous) with respect to other
extraneous factors, it does not guarantee homogeneity in the short run
(i.e., within a single trial). If, for example, in a randomized
clinical trial of treatment therapies for patients with sepsis, one
group happens to have greater severity of illness than a second group,
this would greatly bias the results towards the group with the lower
severity.
A variable that is related both to the outcome and to the treatment
variable is called a confounder. A variable that has a significant
interaction with the treatment is called an effect modifier.
Epidemiologists have recognized the need to adjust for potential
confounders and to identify effect modifiers. These methods can also
be applied to the analysis of data from clinical trials. Just as
epidemiologists control for imbalances in exposure groups through
stratification, statisticians analyzing clinical trial data must be
able to control for imbalances between treatment arms using
appropriate statistical methodology. This talk presents methods
appropriate in the analysis of clinical trials data that can be used
to control for such imbalances.
We will summarize the current, near future, and long-term future of
SAS-based tools for performing sample-size analyses. UnifyPow
(www.bio.ri.ccf.org/UnifyPow) is a widely-used freeware base-SAS
module/macro that already offers substantial flexibility over an
extensive set of methods. It is due for a major new release in
late-summer 2001. The SAS Institute itself is now developing regular
SAS System procedures that will eventually supersede UnifyPow. We will
summarize these developments and show some examples.
Finite mixtures of distributions have been used
extensively to model heterogeneous data in many fields. The potential
of finite mixture models has become increasingly recognized and used
in the field of survival analysis to analyze failure-time data in a
variety of situations. In particular finite mixtures distributions
provide a way of modeling time to failure in the case of competing
risks.
I propose a mixture distribution for long-term survival after breast
cancer diagnosis and adjuvant treatment. This mixture model expresses
survival, conditioned on the patient's age, as a mixture of two
Gompertz survival distributions each representing one of two mutually
exclusive causes of death: from breast cancer or from other causes
without evidence of a recurrence of breast cancer. Covariate vectors
and their regression parameters are incorporated into the hazard
function of each of the two Gompertz distributions. Cohort frailty,
defined and estimated for patient treatment groups, provides the means
to consider the question posed in the abstract title.
Efficient Estimation of the Mean of a Time-Lagged Variable
Subject to Right Censoring
Anastasios Tsiatis
Professor, Department of Statistics
North Carolina State University
Abstract
Stanley Lemeshow
Center for Biostatistics
The Ohio State University
Ralph O'Brien
Department of Biostatistics and Epidemiology
Cleveland Clinic Foundation
Adjuvant Therapy for Early Breast Cancer?
Nahida Gordon
Department of Epidemiology and Biostatistics, School of Medicine
Case Western Reserve University