Conference Schedule

8:15 – 9:00
Registration and Continental Breakfast

9:00 – 9:10
Introductory Remarks
Boris Iglewicz, Temple University

9:10 – 9:55
Design of Standard Experiments for Nonstandard Conditions
Raymond H. Myers, Virginia Polytechnic Institute and State University

9:55 – 10:20
A Discussion of Left-truncated data with Examples
Brenda W. Gillespie, University of Michigan

10:20 – 10:45
Illustrating the Neyman-Pearson Lemma With a Stopping Rule of Order k
Thomas E. Bradstreet and Milton N. Parnes, Merck Research Laboratories and Temple University

10:45 – 11:15
Break

11:15 – 12:00
Regression Diagnostics based on Cumulative Residuals
L. J. Wei, Harvard University

12:00 – 12:25
Analyzing Safety Data in Clinical Trials
Joseph F. Heyse, Merck Research Laboratories

12:25 – 1:35
Lunch

1:35 – 2:20
Statistical Models for Assessing Familial Aggregation of Disease
Nan Laird, Harvard University

2:20- 2:45
Group Sequential Designs for Jointly Assessing Efficacy and Safety in Phase II Oncologic Drug Trials
Alex Kouassi, Fujisawa

2:45 – 3:30
On the Use of Surrogate Endpoints in Clinical Trials
Colin B. Begg, Sloan Kettering Cancer Center

3:30 – 3:35
Closing Remarks


Abstracts

On the Use of Surrogate Endpoints in Clinical Trials

Colin B. Begg, Sloan Kettering Cancer Center

Surrogate endpoints are frequently used in clinical trials, primarily to permit a trial of feasible duration or cost. Over the past decade considerable attention has been paid to the statistical foundations of studies using surrogate endpoints, with a view to providing mechanisms for assessing the validity of inferences from these trials with respect to the impact of the treatments on the true (unobserved) endpoints. The Prentice Criterion has provided an influential framework for this research. In this talk the validity of the Prentice Criterion itself is examined critically, and an alternative conceptual framework is presented. The related problem of using auxiliary data to augment the inference about a true clinical endpoint with missing or incomplete data will also be discussed.


Illustrating the Neyman-Pearson Lemma With a Stopping Rule of Order k

Thomas E. Bradstreet and Milton N. Parnes, Merck Research Laboratories and Temple University

We present an example based upon a curved exponential family which provides insight into using the Neyman-Pearson (N-P) lemma beyond the usual introductory examples constructed from regular exponential families. Consider generalized inverse binomial sampling where independent strings of i.i.d. Bernoulli trials are each constructed by sampling until k, k > or equal 2, successes in-a-row are observed. Path counts to the points in the sampling plan form Pascal triangles of order (k,s). The N-P most powerful (MP) test is constructed from the two-dimensional statistic which is minimally sufficient for the one-dimensional parameter space, and it partitions a two- dimensional sample space. We remark on some characteristics of the simple likelihood ratio and the two dimensional critical region. Power curve comparisons between the N-P MP test and immediately competing tests clearly illustrate that the N-P test is MP but not UMP. We presented this example in a masters level probability theory and mathematical statistics course sequence for statisticians. Student feedback and teaching points are highlighted.

KEYWORDS: Curved Exponential Family; Pascal Triangle of Order (k,s); Statistical Education; Classroom Example; Homework Problem; Exam Question.


A Discussion of Left-truncated data with Examples

Brenda W. Gillespie, University of Michigan

Left-truncated data are characterized by the property that observations with values below a certain threshold are not observed. The idea was popularized in the statistical literature in the early years of the AIDS epidemic, when HIV cases could not be detected until progression of the disease process allowed them to be identified. Left- truncation often arises with right-censoring, although it may also occur with complete data. Software to handle such data is now available. This talk will highlight some examples of left-truncated data in practice.

In addition, using an example of lifetime recall of physical abuse among women, left- truncation will be used to examine the effect of recall bias among middle-aged women by including only the more recent recall time in the truncation window. Thus, in addition to using these methods for left-truncated (delayed entry) data, they can also be used as an investigative tool for checking recall bias in a lifetime recall context. This concept can be generalized to check for any differences in the survival distribution using subsets of times at risk, such as proximity to the reporting time.


Analyzing Safety Data in Clinical Trials

Joseph F. Heyse, Merck Research Laboratories

Evaluating the safety of candidate drugs and vaccines is an important objective of randomized clinical trials. While much attention has been given to the statistical analysis of clinical efficacy data, approaches to analyzing clinical safety data can offer additional complexities. This presentation will provide an overview of the statistical considerations in analyzing clinical safety data. Two primary scenarios will be discussed. The first is a situation when the primary hypothesis for the trial is based on a specific adverse experience. A second scenario is when data on adverse experiences are collected on patients in the trial without specific prior information to suggest a possible treatment effect on one or more of the adverse experiences collected. The appropriate use of equivalence designs will be discussed, and the multiplicity aspects of both scenarios will be contrasted.


Group Sequential Designs for Jointly Assessing Efficacy and Safety in Phase II Oncologic Drug Trials

Alex Kouassi, Fujisawa

Traditional Phase II anti-cancer drug trials are designed to gather data to help determine whether new treatments are sufficiently active to warrant further investigation. For ethical reasons, these designs allow for early termination when there is enough evidence of either unacceptably low or extremely high therapeutic effect. However, they rarely provide guidelines for safety monitoring even though severe side effects may still turn out to be a major problem after careful Phase I screenings. In this paper, we propose a Phase II design procedure which allows joint efficacy and safety monitoring. The proposed procedure assumes independent discrete primary endpoints. If the efficacy and safety outcomes are correlated, as is often the case for most anti-cancer agents, we show through simulations that the proposed methodology is still reliable in the sense that its associated error probabilities remain somewhat within their pre-specified acceptable bounds. In fact, the stronger the correlation between the outcome variables, the smaller the likelihood of erroneously treating patients with either an inactive or toxic drug while the chance of mistakenly rejecting an otherwise active and safe compound is not significantly inflated.

KEYWORDS: Group Sequential; Clinical trials; Phase II trials; Error probabilities; Likelihood of early termination.


Statistical Models for Assessing Familial Aggregation of Disease

Nan Laird, Harvard University

A first step in exploring the genetic basis of disease is demonstration that disease is clustered, or aggregated, in families. Commonly, case-control designs are used to gather disease outcome data on fist degree relatives of cases and controls. We review some commonly used approached to the analysis, and compare methods based on logistic regression to some likelihood based methods. We discuss extensions of the basis approaches to study the co-aggregation of more than one trait. Applications to lung cancer, and to eating disorders and depression, will be discussed.


Design of Standard Experiments for Nonstandard Conditions

Raymond H. Myers, Virginia Polytechnic Institute and State University

This paper provides discussion of the difficulty in designing efficient experiments when the response is binomial, Poisson or from another member of the exponential family. Illustrations of the use of optimal experimental design for logistic and Poisson regression through the process of apriori “parameter guessing” are illustrated. In addition, standard experimental designs, i.e., factorial and fractional factorials are shown to be extremely robust to the variance heterogeneity and the nonlinear nature of the generalized linear models.


Regression Diagnostics Based on Cumulative Residuals

L. J. Wei, Harvard University

We will discuss a unified approach to regression diagnostics based on the cumulative sums of residuals over certain coordinates (e.g., covariates or fitted values). For a variety of statistical models and data structures, including generalized linear models with independent or dependent observations and proportional hazards models with censored failure times, the distributions of these stochastic processes under the assumed model can be approximated by zero-mean Gaussian processes. Each observed process can then be compared, both graphically and numerically, with a number of realizations from the approximate null distribution by computer simulation. These comparisons enable one to determine objectively whether a trend seen in a residual plot reflects model misspecification or natural variation. The proposed techniques are particularly useful in checking the functional form of a covariate, the link function and the proportional hazards assumption. Several real examples will be used for illustration.

Keywords: Generalized linear models; Goodness of fit; Link function; Longitudinal data; Marginal models; Model checking; Model misspecification; Proportional hazards; Residual plots; Survival data; Transformation.


Registration

This event has passed.


Directions

DOUBLE TREE GUEST SUITES, PLYMOUTH MEETING
640 W. Germantown Pike, Plymouth Meeting PA 19462
(610) 834-8300

From Airport: Take 95 South to 476 North to the last exit #20(Germantown Pike-West). Merge with Germantown Pike and follow for 3 lights. Make a right onto Hickory Rd. at the 3rd light. The hotel is the 3rd building on the left.

New York/ New Jersey Turnpike: Take the New Jersey Turnpike to exit #6, which is PA turnpike. Go west to exit #333- Norristown. Follow signs to Plymouth Rd. Go to the 1st light and make a left. Go to the next light and make a right onto Germantown Pike. Go to the second light and make a right on Hickory Rd. The hotel is the second driveway on the left.

Washington D.C., Wilmington, and Delaware: Take I-95 North to Route 476 North. Take Route 476 to the Germantown Pike West exit #20. Go to the third light, Hickory Rd., and make a right. The hotel is the 2nd driveway on the left.

Route 476: Take 476 to the Germantown Pike West exit #20. Go to the third light, Hickory Rd., and make a right. The hotel is the 2nd driveway on the left.

From downtown Philadelphia: I-76 west Plymouth Meeting exit #331B (Route 476). Take Route 476 north to Germantown Pike exit. Go to the third light, Hickory Rd., and make a right. The hotel is the 2nd driveway on the left.