DoE Design For Multi Level Factors?
Six Sigma – iSixSigma › Forums › Old Forums › General › DoE Design For Multi Level Factors?
 This topic has 9 replies, 9 voices, and was last updated 18 years, 11 months ago by Som.

AuthorPosts

May 13, 2003 at 9:00 pm #32232
Brad M.Participant@BradM. Include @BradM. in your post and this person will
be notified via email.I need some input on what type of DoE model I should use. We are conducting a DoE on a solderability defect. Our experiment involves two (2) factors. One (1) factor will be varied over three (3) levels and the other factor will be varied over four (4) levels. What recommendations do you have? Should a screening experiment be done first or is there a design that can provide valid response data with factors varied over multiple/differing levels? Thanks for any feedback you may provide.
0May 14, 2003 at 12:21 am #85853If you have prior knowledge that these factors are significant and that the levels are legit – run the 3 x 4 (12 unique treatment combination) experiment.
0May 14, 2003 at 1:34 pm #85875
Brad M.Participant@BradM. Include @BradM. in your post and this person will
be notified via email.Stan,
Thanks for the reply. What specific DoE design would you recommend?0June 5, 2003 at 5:25 am #86680
J AndellParticipant@JAndell Include @JAndell in your post and this person will
be notified via email.If the two factors are continuously variable in nature, you might want to consider either a simple 2level factorial, or a response surface design (central composite may be the best for a 2factor, but there can be reasons to favor others). If the factor levels are basically fixed (such as with categorical data, or with fixed commercial formulations), then a 3×4 full factor experiment is called for.
0June 5, 2003 at 5:27 am #86681
Mike NellisParticipant@MikeNellis Include @MikeNellis in your post and this person will
be notified via email.Hello Brad M.,
Run the factors at the level combinations below and record your response(s) after each run. Generate your summar data, charts, and graphs. Rerun the test to confirm any conclusions you made from the first experiments.
Hope this helps.
Mike
A B Response2 1 .003 (example)
1 2
2 2
1 1
3 1
1 4
1 3
3 3
2 4
2 3
3 2
3 4
0June 5, 2003 at 7:08 am #86687
Ljubomir LazicParticipant@LjubomirLazic Include @LjubomirLazic in your post and this person will
be notified via email.First of all you should tel us a priory knowledge about outcome?
If Second order model of outcome is satisfactory then you should apply CENTRAL COMPOSITE ROTATABLE DESIGNS (CCRD) for k=2 with 5 replica at NULL factor level, which design plan is:TABLE 2.147 CCRD
Design of experiment
matrixNo of
exper.X1 (factor
level)X2 (factor
level)1
+
+
2
–
+
3
+
–
4
–
–
5
1,414
0
6
+1,414
0
7
0
1,414
8
0
+1,414
9
0
0
10
0
0
11
0
0
12
0
0
13
0
0
So yu should have 13 experiments in order to get second order regression model for outcam with high confidence. The question is :
– ar you restricted with number of experiment ( time, expenses etc.)?
if answer is no then I propose CENTRAL COMPOSITE ROTATABLE DESIGNS (CCRD) for k=2 with 5 replica at NULL factor level.0June 5, 2003 at 8:23 am #86688Maybe this is an alternative:
If in case you are not really sure if there are quadratic effects, then try a 2×2 using the low and high levels of your factors.
Check for lack of fit or validate runs on the midpoints to check for curvature.
If no significant curvature, then the experiment is OK.
Else, augment the design to form a central composite design (facecentered).
This is assuming your factors are continuous.
0June 5, 2003 at 1:46 pm #86699Brad,
I used a General Full Factorial DOE in Minitab for a 2 x 2 x 2 x 3. This should work for what you have too.0June 5, 2003 at 10:14 pm #86719
jediblackbeltParticipant@jediblackbelt Include @jediblackbelt in your post and this person will
be notified via email.Have you looked at a balanced ANOVA? I have done this in the past has a multistep style experiment. Some of the other replies are better off though. I especially like the one where you run a design with a curvature check.
0June 6, 2003 at 2:34 pm #86730I suggest you run a full factorial experiment with 2 replicates, since you have only two factors – if you have resource constrain you can live with one replicate as with 12 data points you are left with 9 df to estimate your error.
0 
AuthorPosts
The forum ‘General’ is closed to new topics and replies.