March 2004 meeting
We will have our March ASA chapter meeting this Thursday
(March 18), 6-8:15 PM in Sid Richardson Hall, Room 3.109 (our usual room
at the LBJ School of Public Affairs).
The LBJ School of Public Affairs at UT, Austin, is in Sid Richardson Hall
adjoining the LBJ Library on Red River Road, just south of Dean Keeton
(26th St.), and offers plenty of free parking. Red River Road is one
block west of IH-35, which it parallels. See map at
http://www.utexas.edu/maps/ for further information.
6 PM - 6:45 PM Refreshments and networking
6:45 - 7:00 PM Announcements and business
7 PM -~8:15 PM "Reflections on Overuse of Randomization in Test and
Selection Procedures: SUPPLEMENTAL STATISTICAL REFINEMENT, A Simple Method of Splitting Points
in Rank Tests"
It is hoped that this meeting during the UT spring break will afford some
a more relaxed meeting date as well as serving as a transition to the
several statistical events in late March and early April.
Reflections on an Overuse of Randomization in Test and Selection
SUPPLEMENTAL STATISTICAL REFINEMENT,
A Simple Method of Splitting Points in Rank Tests
by Ray L. Marr
Randomization to split points in hypothesis testing was proposed by
Neyman and Pearson (1933) as an artifice to extend the applicability of
the Neyman-Pearson Lemma to cases where the test statistic has discrete
atoms that straddle the desired significance level. Although
randomization solves the problem theoretically, it is often
unsatisfactory to resolve a test of hypotheses by an irrelevant
randomization that makes the test results often unrepeatable.
For many rank tests, ties are all too common with small samples and do
not permit finding exact alpha-level tests for many values of alpha. A
potential remedy is to refine the partial ordering on the sample space
induced by the rank test statistic by using one or more supplemental
statistics to break ties in the value of the test statistic and induce a
test with the same asymptotic properties as the original but with smaller
A simple, intuitive, and reasonable method of accomplishing this will be
presented and applied to some common rank tests such as Spearman's
rank-order correlation coefficient, Spearman's footrule, and Wilcoxon's
test, which can be adapted to achieve some exact or more nearly exact
alpha-level tests with small samples.
This talk is intended to be worthwhile for anyone who understands the
basics of hypothesis testing and is familiar with elementary rank tests.
Although the speaker would like to provide the audience with a few
methodological take-aways, he also hopes to stimulate the audience to
re-consider some fundamental ideas.
Ray Marr earned master's degrees (ABD) in mathematics and mathematical
statistics from Indiana University in 1971 and 1980. After working in
the pharmaceutical industry in New York and medical device industry in
Tampa, he returned to school for his PhD in statistics at North Carolina
State under Professor Charles Quesenberry of Q-Chart fame. He has worked
for many years as an internal statistical consultant to the semiconductor
industry at AMD and Sematech in manufacturing, advanced SPC, and design
and analysis of industrial experiments; additionally at AMD, he developed
two hundred pages of mathematical proofs establishing rigorous error
bounds for computer arithmetic implementations of transcendental function
algorithms in the various rounding modes, a subject he is all too happy
to talk about. Ray is an external statistical consultant in Austin and
local chapter president of the ASA.