| Speaker: |
John W. Green DuPont Newark, DE | |
| Title: | Statistical Analysis of Ecotox Experiments: Practice, Relevance and Power |
Abstract:
Using the results of a benchmark survey as backdrop, several methods of analyzing continuous data in a concentration-response experiment will be compared for power and appropriateness, including Dunnett's classic multiple comparison procedure and step-down applications of the Jonckheere-Terpstra trend test and a new contrast based approach.
Two underlying principles are recommended: (1) The statistical model should reflect the experimental design and both biological expectation and uncertainty. (2) Demonstration of adequate power of a statistical test should be a requirement for analysis. In general, a step-down approach is recommended, based on a presumed monotone concentration-response relationship, backed by an appropriate check of the presumptive model. In this context, contrasts using normalized ranks and Bartholomew's test for trend or departure from trend are compared.
Extensive simulation results and actual laboratory data are used to demonstrate the power and application of two possible statistical models consistent with these recommendations and these are compared to Dunnett's test. As a rule, step-down tests based on monotone concentration-response models are more powerful than multiple comparison methods which assume no structure. The use of one- or two-sided tests depends on biological appropriateness.
The Jonckheere-Terpstra test is old, but the step-down application of this test is relatively new and not well documented. Both the theoretical basis of this approach and simulation results are given. This approach is consistent with approaches to analysis of binary data. An alternative model is based on Helmert contrasts and was developed by A. C. Tamhane, C. W. Dunnett, J. D. Wetherington and this author. This latter method is especially appropriate when there is agreement on the magnitude effect it is desirable to find.
The variety of conditions of normality, variance homogeneity, multiple variance component designs, lack of monotone concentration-response and presence of massive ties in the response found in ecotox data mean none of these methods is universally appropriate, but uniform logical flow charts exist.
| Date:
| Thursday, March 9, 2000 |
| Time: | Social hour -- 6:00 pm Dinner -- 6:30 pm Speaker -- 7:15 pm |
| Place: | Columbus Inn 2216 Pennsylvania Avenue Wilmington, DE |
| Menu: |
1. Shrimp Diablo with Arabiata sauce Orzo Grilled leeks and prosciutto topped with fresh mozarella and basil Mixed green salad Warm rolls and butter Warm fruit cobbler with cinnamon ice cream Coffee or tea
2. Grilled Spring Vegetable Cavatappi |
| Cost: | $25 for members, $5 for students |
Reservations to Jane Buck by noon Monday, March 6, 2000.
410-885-3117,
buck@count.com
Please be sure to RSVP by this date. We cannot guarantee a meal for late reservations.
Note: Indicate your entree choice with your RSVP.