|
|
Pre-Conference Short Course
June 3, 2008
Pyle Center, Madison, Wisconsin
GRAPHICAL ANALYSIS OF DESIGNED EXPERIMENTS
Dr Veronica Czitrom
It’s a source of frustration to statisticians that designed experiments,
as useful as they are, aren’t used more extensively. This workshop
provides a graphical approach that’s easy to understand even without
statistical training.
The concepts are presented with the help of numerous real-life industrial
examples. The material will be useful to statisticians trying to present
results to customers and students, and to scientists and engineers trying
to understand the results of their experiments.
The first example, stripped down to essentials with one observation per
run, displays the results using a variability graph, introduces main effects
and interactions in an intuitive way, and presents visual t-tests and
advantages of DOEs such as hidden replication and “recycling”
at its best – every observation is used to estimate several effects.
Another example presents multiple observations and sources of variability
(variance components) graphically. A third example presents the results
of a fractional factorial robust designed experiment as an array of pictures
involving nine variables.
Note:participants with laptops can follow the DOE analysis
during the workshop. To do so, download a free trial version of the JMP
statistical data analysis software at www.JMP.com,
and download the DOE data sets from www.StatsTrain.com.
TABLE OF CONTENTS
Introduction
- Example of 23 full factorial with one observation per run
plus center points: main effects, interactions, hidden replication, visual
t-tests, “recycling” to use each observation to estimate several
effects, creating a full factorial with pencil and paper using a tree
diagram, extension to central composite design if there’s curvature.
Advantages of DOE
- Why DOE works – or why you can study several factors simultaneously
- Advantages of DOEs: they require less resources for the amount of information
obtained, provide higher precision, systematically estimate effects
- DOE versus one-factor-at-a-time experiments
Full factorial designs
-Main effects and interactions
-Example of 23 full factorial for a batch process with multiple
observations per run: graphical analysis of raw data and of variance components
(graphical analysis avoids more complex ANOVA due to restriction on randomization);
DOE eliminated manufacturing bottleneck and increased Cpk from unacceptable
0.86 to excellent 2.12
Fractional factorial designs
-Aliasing
-Example of 25-1 fractional factorial robust designed experiment:
responses are yield and pictures, the pictures are presented as an array
involving nine variables; DOE revealed the failure mechanism causing a
drop in yield
INSTRUCTOR
Dr Veronica Czitrom does statistical training and consulting in Asia,
Europe and the United States, primarily in the semiconductor industry.
She worked as an industrial statistician for Bell Laboratories when it
was part of AT&T, for Chartered Semiconductor in Singapore, and as
a professor at a couple of universities. She has published technical articles
in refereed journals and seven books, and was elected Fellow of the American
Statistical Association. She obtained a PhD in Mathematics with concentration
in Statistics from the University of Texas at Austin, as well as a BA
in Physics and an MS in Engineering from the University of California
at Berkeley.
Veronica's Biography can be found via the following link
http://www.statstrain.com/pages/profile.htm
|
|
|