|Date: ||Thu, 13 Jun 1996 09:30:17 -0500|
|Reply-To: ||Mitch Gallant <mgallant@PILOT.IH.NAVY.MIL>|
|Sender: ||"SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>|
|From: ||Mitch Gallant <mgallant@PILOT.IH.NAVY.MIL>|
|Subject: ||Seek Advice--4 ingred. mixture expt.|
SAS Users Far and Wide,
I am new to SAS which we purchased to use for designing and analyzing
experiments. A short while ago I posted a more complicated request and was
delighted in the quality of the responses. I have not posted a summary
because (1) I am still digesting that response, and (2) I wish to pursue the
designing step further.
My question for today may seem easy to some, but I wish to have some of my
A coworker wants to propose a study to help another division with a
formulation. The formulation has a specification range, but the group doesn't
understand the properties they might get at different combinations. As far as
I know, there is only one response. I can't recall specifically, but it is a
burning-property response for which the lab reports an average of three or
more tests per sample (the reliability is unknown).
It's a flare composition with a three polymeric component binder and one
metal, e.g., B1, B2, B3, and M1. The constraints are quite tight.
My questions and assumptions:
The costs to prepare a "mix" is VERY expensive, so it's best to get the most
from the least. It is preferred to generate a model and show a countour of
the response for the customer. I am considering two approaches.
Using SAS's XVERT algorithm (McLean-Anderson), there are 39 observations in
the full set. Of course, this is too many. If I am on the right track, there
are six vertices. Following the quidelines of the algorithm, the centroid
should be included and a few check points inside the design space. (Total 10
mixes so far.) Should there be a replicated point or two? Twelve mixes total
is about the reasonable limit from a proposal standpoint...
But can we do it with less...
The constraints are very close, so wouldn't a linear model be good assumption.
Again if I on the right track with this software, a five run design is
generated for the linear model. It has four vertices and one point with the
metal at its low value. The D-efficiency is 86.6. This approach would start
with 5 runs; however, does it build a foundation for well planned follow-on
runs to verify and better define the model? And what about the two-way
Thanks for you your patience and consulation while I am cutting my teeth on
Naval Surface Warfare Center
Indian Head, MD USA