Sponsoring Section/Society: ASA-Biopharm
Session Slot: 8:30-10:20 Thursday
Estimated Audience Size: 150-200
AudioVisual Request: Two Overheads
Session Title: Permutation Tests in Clinical Trials
Theme Session: No
Applied Session: Yes
Session Organizer: Berger, Vance Food and Drug Administration
Address: CBER Food and Drug Administration 5600 Fishers Lane Rockville MD 20857
Phone: 301-827-3977
Fax: 301-827-3529
Email: bergerv@a1.cber.fda.gov
Session Timing: 110 minutes total (Sorry about format):
Opening Remarks by Chair - 5 minutes First Speaker - 25 minutes Second Speaker - 25 minutes Third Speaker - 25 minutes Fourth Speaker - 25 minutes Discussant - None Floor Discusion - 5 minutes
Session Chair: Berger, Vance Food and Drug Administration
Address: CBER Food and Drug Administration 5600 Fishers Lane Rockville MD 20857
Phone: 301-827-3977
Fax: 301-827-3529
Email: bergerv@a1.cber.fda.gov
1. Issues on Randomization Tests: Applications in Analysis of Ischemic Stoke Lesion Volume in the NINDS t-PA Strok Trial
Lu, Mei, Henry Ford Health System
Address: Department of Biostatistics and Research Epidemiology Henry Ford Health System 3E 1 Ford Place, Detroit, MI 48202
Phone: 313-874-6413
Fax: 313-874-6730
Email: mlu1@biostat.hfh.edu
Tilley, Barbara, Henry Ford Health System
Li, Shuhui, Henry Ford Health System
Abstract: The NINDS Stroke Trial demonstrated the effectiveness of tissue-type plasminogen activators (t-PA) given within three hours of stroke onset for the treatment of acute ischemic stroke. This effectiveness was defined as clinical recovery measured at three months. Assessment of the impact on stroke lesion volumes measured at three months is desirable. We initially proposed an analysis for the effect of t-PA on stroke lesion volume using randomization tests, given the highly skewed data and the distribution-free assumption for randomization tests. However, we encountered difficulties in analyzing th estroke lesion volume. Results from randomization tests are dependent on test statistics calculated from observed data. In this paper, we discuss the aspects of randomization tests and their limitations, illustrate our findings in analyzing the stroke lesion volumes for the effect of t-PA, and compare randomization tests with nonparametric statistics. To study the effect of t-PA on the stroke lesion volumes, we finally proposed a parametric analysis using a generalized estimate equations (GEE) approach on the transformed data and validated the results using randomization tests with a relevant test statistic. This experience with the NINDS t-PA trial data emphasizes that randomization tests are useful for clinical trials but have limitations. The choice of test statistics influences randomization tests, and the choice of test statistics should be appropriate to the given data and hypotheses. Randomization tests are equivalent if test statistics are equivalent.Key Words: Randomization Test; Permutation Test; Exact Test; Nonparametric Test; GEE; Robustness; Statistical Validation.
2. A New Look at Rank Tests in Ordered 2 x k Contingency Tables
Permutt, Thomas, Food and Drug Administration
Address: FDA/CDER 5600 Fishers Lane (HFD-170) Rockville MD 20857
Phone: 301-443-3741
Fax: 301-443-7068
Email: PermuttT@cder.fda.gov
Berger, Vance, Food and Drug Administration
Abstract: The P-P plot allows properties of rank tests in trials with two groups and ordinal outcomes to be visualized. We consider several tests that have been proposed in this application (Smirnov, Wilcoxon, and maximum chi-square). We compare them to the van der Waerden test, well known but not in this setting, and a new, cumulative chi test. We conclude that rank tests are a viable alternative to a priori assignment of parametric scores for outcomes. Both Smirnov-like tests and Wilcoxon-like tests may be useful. However, both the Smirnov and Wilcoxon tests are bad examples of their types in that they give equal weight to some deviations that are highly improbable under the null hypothesis and to others that are much less improbable. The maximum chi-square test is more powerful than the Smirnov test, and the van der Waerden and cumulative chi tests are more powerful than the Wilcoxon test, over a wide region of the parameter space.
3. A Comparison of One-Sided Methods to Identify Significant Individual Outcomes in a Multiple Outcome Setting
Troendle, James, National Institute of Child Health and Development
Address: National Institutes of Health Bld. 6100, Rm. 7B13 Bethesda, MD 20892
Phone: 301-496-6811
Fax: 301-402-2084
Email: jft@helix.nih.gov
Legler, Julie L., National Cancer Intstitute
Abstract: We compare two approaches to the identification of individual significant outcomes when a comarison of two groups involves multiple outcome variables. The approaches are all designed to control the familywise error rate in the strong sense. The first approach is initially to use a global test of the overall hypothesis that the groups are equivalent for all variables, followed by an application of the closed testing algorithm. The global tests considered here are ordinary and generalized least squares, an appoximation to the likelihood ratio test, and an approximation to the most powerful similar test for a simple normal alternative. The second approach is that of stepwise testing, which tests the univariate hypotheses in a specific order with appropriate adjustment to the univariate p-values for multiplicity. The stepwise tests considered are those of Holm and Hochberg, along with step-down and step-up permutation tests. The permutations allow the tests to incorporate the correlation between outcomes without any restrictive assumptions.We illustrate the tests with two examples of birth outcomes: a comparison of cocaine exposed newborns to control newborns on neurobehavioral and physical growth variables, and, in a separate study, a comparison of babies born to diabeticmothers and babies born to non-diabetic mothers on minor malformations. Simulations are used to provide a basis for chosing among the procedures considered.
4. Exact Statistical Inference in Familial Disease Clusters
Zelterman, Dan, Yale University
Address: Division of Biostatistics PO Box 208034 Yale University New Haven, CT 06520
Phone: 203-737-1768
Fax: 203-785-4116
Email: Daniel.Zelterman@Yale.edu
Yu, Chang, Yale University
Claus, Elizabeth B., Yale University
Abstract: In many epidemiologic studies the first indication of an environmental or genetic contribution to the disease is the way in which the diseased cases cluster within the same family units. We assume that all individuals are exchangeable, except for their disease status. This assumption of exchangeability is used to test the initial hypothesis of no familial link with the disease. Conditional on the distribution of the sizes of the various familial units, we obtain the exact probability of observing a given set of disease cluster frequencies. The expected frequencies and other moments are given. Three numerical examples demonstrate these methods. We describe an algorithm for obtaining exact statistical inference by enumerating all possible outcomes consistent with the numbers and sizes of the family units. A parametric model is given to identify useful risk factors for the disease outcome.Keywords: exact inference; genetics; statistical computing; multivariate discrete distributions.
List of speakers who are nonmembers: None