A defense of Taguchi

si...@niobbs1.em.cdc.gov

unread,

Oct 29, 1992, 10:32:00 AM10/29/92

to

April Milliken asks "if anyone can inform me as to why Taguchi methods are
used." In particular, she is concerned that many of the methods "exclude
interactions" and eliminate "potentially important information".

I'm not an expert in Taguchi methods, but I hope I can come to his defense.
Taguchi is in the company of some excellent statisticians like George Box and
Jeff Wu when he proposes designs that have small sample sizes and "exclude"
interactions. In many engineering problems, it is critical to gain an
understanding of the simultaneous impact of many variables. There is
susbstantial empirical and theoretical evidence that this is far superior to
focussing on one factor at a time. In fact, it is far better than doing a
very complete study of two, three, or four factors and all their interactions.

If you have a limited budget, you put your money (or your data) where they
will give you the most information. In many engineering problems, the size
of the main effects tends to be much larger than the size of the
interactions. The importance of the main effects tends also to be much
larger than the importance of the interactions. It is often more efficient,
more economical, and more logical to study 16 factors with no interactions
than to study only 4 of these factors with all possible interactions.

When you find yourself in very complex settings or when you are running the
first experiment in a series of experiments, experiments with many factors
are especially useful. This is counter-intuitive: most people would like to
run a small focussed experiment on a few factors (and all their interactions)
to "get their feet wet". But the engineers know better: get a very general
picture looking at all factors without worrying about interactions and then
run the small focussed experiments based on that picture.

There are times when estimating interactions are necessary, and Taguchi
methods allow for estimation of interactions when needed.

To be fair, Taguchi methods are not always without problems. Technometrics
does an outstanding job of explaining the strengths and weaknesses of Taguchi
methods: take a look at the May 1992 issue, for example.

I like what Taguchi, Box, Wu and the engineers of the world have done. Other
disciplines can learn from their experience and wisdom. I've been promoting
these ideas with the biologists and toxicologists that I work with.

* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
* Steve Simon, Mail Stop C-22 * Delivering on the *
* Chief, Statistics Activity * nation's promise: *
* National Institute for Occupational * safety and health *
* Safety and Health * at work for all *
* 4676 Columbia Parkway * people...through *
* Cincinnati, OH 45226 * prevention. *
* * *
* 513-533-8271 (office) * NIOSH vision *
* 513-533-8510 (FAX) * statement *
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
* SMTP gateway to Internet *
* si...@niobbs1.em.cdc.gov *
* X400 gateway to Internet *
hhscdc!sis1/o=cdc/ou=cinc/ou=nio...@mhs.attmail.com *
* *
* Views expressed here are my own. Mention of a product or *
* company name does not constitute an endorsement by NIOSH. *
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *

(Kernon Gibes)

unread,

Oct 29, 1992, 11:09:51 AM10/29/92

to

I don't have a major problem with what Steven Simon wrote in defense of
Taguchi's methods. I agree that there are situations where one should
do a "screening" (resolution III) design (e.g. Plackett-Burman) first
before running factorials, RSMs, etc. However, *IF* the researchers
have several factors to consider AND have good reason to believe that
certain factors interact, I could not advocate a screening design
without a certain amount of caution. Yes, I'd agree that generally
main effects are larger than interactions, but in resolution III
designs the two-factor interactions are aliased (confounded) with the
main effects. If certain interactions are important, they'll show up as
main effects. If you're doing a series of experiments this may not be
a major problem insofar as the lack of importance of a particular
factor could be discovered further down the line. But researchers
doing screening designs should at least be appraised of this risk.
It's even possible (though perhaps unlikely) that one or more two-factor
interactions are aliased with important main effects in such a manner as
to make the main effect look small. Box, I believe, advocates running
fractional factorials so that one can augment the design with another
fraction if one suspects that some confounding did occur.

Kernon Gibes
Internet: gi...@swirl.monsanto.com

(sheldon haynie)

unread,

Oct 29, 1992, 11:39:24 AM10/29/92

to

in a recent message Steve Simon, says

As a development engineer for ~15 years I have to agree that
empirical results of
simple orthogonal "splits" gives a good initial "feel" for the
response of the system.
These splits are on the "major" variables that are either indicated
by physical models
or are those that are availble in a "black box" process.

We tend to use the "goldilocks" approach to selecting variable
values:
- too small
- too big
- just right

based on exisiting process data, simulation or available setpoints. I
typically
like to select hi/lo points that are likely to produce mean output
shifts of about
1.5 sigma to be "sure" of the actual repetition of the experiment.
Sometimes
that strategy is not selectable due to a lack of knowledge for a new
process.
These split conditions are randomized and the processes run for data
collection.

After initial run(s) depending on time and budget constraints
(sometimes time to do it over,
rarely time to do it "right") we will select the most "interesting"
variables and
design an experiment that focuses on the interactions and sequences.
(many times
the most important factor was the previous conditions :-( )

in the semiconductor industry, you rarely have the time to track all
possible
interactions of a 200+ step process, so some insight and "engineering
horse-sense"
as well as occasional blind MVA seems to be the only path.