DSD Screening Replication

[#1]

In trying to try to control a visual quality characteristic, we are nearly ready to run a 4 factor (all are continuous variables, e.g., temperature, feed rate, etc.) DSD at 3 levels per factor.  This will have us run 9 treatments for the DSD.

My question is: Should I run one or more replications on the DSD screening-type DOE?

[#2]

If you have the resources to consider replication of the design then perhaps I would challenge you to brainstorm additional factors above the 4 factors that you are considering. The true power of the DSD's manifests itself when one is considering many more factors than 4. Of course the answer to your question is yes there is an advantage to adding replicates in any design especially if your measurement system my in your case be subjective?

[#3]

LouV,

We've already narrowed down the likely factors, and performed a multi-vari analysis to be sure our production system (i.e., equipment & measurement systems) and setup/operators were not causing variation. 

Now we're down to "incoming material variation" & "process parameters".  And the incoming material isn't varying much.

Per our Cause & Effect diagram (the Ishikawa had 24 possibilities), we were able to get to the four most likely factors.

The DSD method was chosen because the # DSD runs (9) is less than the normal screening # runs (16), and because we can screen for multiple levels in one screening run.

I'm more interested in detecting factor curvature* vs. economical design.

* I'm pretty sure one or more of the factors exhibit curvature.

Don't know if this is very relevant, but I'm using JMP7.

Thanks for the help.

RB

[#4]

I would recommend considering the 6 factor DSD for the 13 runs (and drop the 2 extra factors).

The 9-run DSD can't fit the full RSM model for any subset 3 factors, while the 13-run design is able to. In JMP 12, we now use the 13-run as the default.

Good luck!

[#5]

Ryan,

Thanks for the 6 factor DSD tip, & I'll look into it today.

Are you referencing this paper:

http://www.jmp.com/content/dam/jmp/documents/en/white-papers/definitive-screening.pdf

specifically the designs on page 5?

And by "dropping the 2 extra factors", do you mean dropping the two right-most columns of the m=6 design?

[#6]

You are right on with the "dropping the 2 extra factors".

I would use the m=6 design rather than using Augment design, but I'm partial to the special structure of DSDs.

Lou makes a very good point about considering extra factors. If you use the DSD for 6 factors, you're getting to study 2 extra factors for "free". If there's truly no effect on those extra factors, it's simple enough to remove them from the model, but you may end of finding something you didn't expect.

[#7]

Ryan,

I got the m = 6 DSD screening (13 runs) finished & gave it to our process engineer.  Hopefully he'll be able to start running the treatments soon.  I think I got it copied into this message (see below).

Due to confidentiality considerations I have to rename Factor 3 & 4. And I sorted the table based on the Temp column since it take forever to adjust our furnace's temperature & wait for it to normalize... so technically it probably should be a Blocked Factor.

I will add "Y" data when I get it.

Best regards,

RB

Treat-ment   #	Pattern	Temp.	Feed	Proprietary Factor 3	Proprietary Factor 4	"Y" rating aka undesirable visual quality characteristic
Low	-	540	5	100	0.7
Nom	0	560	10	300	1.3
High	+	580	15	500	1.7
1	-0+-	540	10	500	0.7
2	--0+	540	5	300	1.7
3	-++0	540	15	500	1.3
4	-+-+	540	15	100	1.7
5	----	540	5	100	0.7
6	0+--	560	15	100	0.7
7	0-++	560	5	500	1.7
8	0000	560	10	300	1.3
9	+0-+	580	10	100	1.7
10	++0-	580	15	300	0.7
11	+--0	580	5	100	1.3
12	+-+-	580	5	500	0.7
13	++++	580	15	500	1.7

[#8]

Hope the experiment goes well!

In the future, if blocking for DSDs has been added in JMP 12 which will give you some extra flexibility in the number of blocks you can consider (as well as more balanced block sizes instead of 3 runs for the "center" block).

There have also been improvements to the way we construct DSDs compared to the original paper - I'm not sure if the DSD add-in works in JMP 7.

Cheers,

Ryan

[#9]

Table below has the results data (Original data then two replications).

Obviously the area around run 2 is where future experimentation will be focused.

RUN	Pattern based on m = 6	Temp.	Feed	Factor 3	Factor 4	Visual Rating 1=Best 5=Worst Signal-to-Noise (LIB)
						Y1	Y2	Y3	S/N	Rank
1	-0+-	540	10	500	0.7	2.6	2.1	2.0	-7.039	2
2	--0+	540	5	300	1.7	2.3	2.0	2.0	-6.464	1
3	-++0	540	15	500	1.2	3.6	3.6	2.1	-10.048	4
4	-+-+	540	15	100	1.7	4.8	4.8	4.0	-13.158	13
5	----	540	5	100	0.7	3.5	3.4	2.3	-9.868	3
6	0+--	560	15	100	0.7	4.4	4.5	4.0	-12.680	12
7	0-++	560	5	500	1.7	3.5	3.3	3.0	-10.299	5
8	0000	560	10	300	1.2	3.8	3.2	3.6	-10.985	7
9	+0-+	580	10	100	1.7	4.0	4.1	3.7	-11.903	11
10	++0-	580	15	300	0.7	3.7	3.4	3.8	-11.216	9
11	+--0	580	5	100	1.2	3.8	3.6	3.8	-11.445	10
12	+-+-	580	5	500	0.7	3.9	3.7	2.5	-10.688	6
13	++++	580	15	500	1.7	4.1	4.0	2.5	-11.146	8

NOTE: Color shading in Y1, Y2, & Y3 only used to indicate better (green), moderate (yellow-orange), and bad (red) visual-quality rating.

Rating method was to arranged all Y1, Y2 & Y3 samples side-by-side for comparison/evaluation purposes.  We had to do this type of rating method because no one make a reasonably affordable gage (giving quantitative results) for this product's visual characteristic we're trying to reduce.

RB

PS Thanks to all who provided help on this (LouV & Ryan Lekivetz).

PPS. Now I have ammo to justify getting JMP12... woohoo!

[#10 · Accepted solution]

RB,

I would give a try to (-,-,+) for Temp = 540,Feed = 5 and Factor 3 = 500. Factor 4 does not seem to impact response so choose the economical/ergonomical setting for that one.