Subscribe Bookmark RSS Feed

Retrospectively determine matrix design?

bwilliamson

Community Trekker

Joined:

Jul 27, 2015

How best to try to retrospectively determine what factors an experimental matrix was designed to evaluate?

I have a block of data that were dumped off on me that presumably came from a fractional factorial type experimental matrix. I have the responses and I have the levels of the various parameters that were varied for the matrix. What I don't have is what terms were targeted in the original design.

Any help is appreciated.

Thanks!

1 ACCEPTED SOLUTION

Accepted Solutions
Solution

If you load your design into JMP, you can go to DOE->Evaluate Design, and pick your 6 columns for factors. It pre-loads the main effects for your model, with all the 2-factor interactions as alias terms (ie., not in the model, but potentially active). Examining the Color Map on Correlations, I can see it's a regular resolution IV fractional factorial with 3 center runs. To see which interactions are aliased you can hover over the off-diagonal entries that are red.

Cheers,

Ryan

11340_pastedImage_0.png

7 REPLIES
txnelson

Super User

Joined:

Jun 22, 2012

Assuming that you received the data in the JMP data table created from the JMP DOE, you can then click on the DOE Dialog script in the Tables Panel, and it will return you back to the design dialog.

11328_pastedImage_0.png

11329_pastedImage_1.png

Jim
bwilliamson

Community Trekker

Joined:

Jul 27, 2015

Unfortunately it was sent to me in Excel.

I don't know what software package was used and the organization that the data came from isn't very responsive so I was hoping I could piece it together.

Thanks for the reply!

ian_jmp

Staff

Joined:

Jun 23, 2011

Stating the obvious perhaps, but a factor at n levels allows terms in the model up to (n-1) in that factor. So looking at the number of unique values in each column describing how a factor was changed should help. You could also try using 'Columns > Recode' to make the design easier to read using integers for levels. If you post that here, somebody might recognise it.

But, by definition, a fractional factorial design always involves some confounding relative to the full factorial (which allows all terms to be unambiguously estimated). So, even if you 'did' know the logic behind the design you have, it will still support multiple models which you have to somehow disambiguate between (often using sparsity of effect or heredity arguments). Even without this knowledge you may still be able to build a plausible model, though how 'plausible' will depend on what you actually want to use the model for.

bwilliamson

Community Trekker

Joined:

Jul 27, 2015

Below is the pattern of factors that was used. We have 6 factors at essentially two levels with three center points sprinkled in. I don't see any replicates, save the three center point runs. This design shouldn't be capable of detecting curvature. I was mostly interested in trying to pry out the interactions that are resolvable and knowing the confounding pattern...

Thanks for your help!

Factor AFactor BFactor CFactor DFactor EFactor F
-1-111-11
-1111-1-1
000000
1-11-1-1-1
111111
1-1111-1
-11-1111
000000
1-1-1-111
-1-11-111
11-1-11-1
111-1-11
11-11-1-1
1-1-11-11
-1-1-111-1
-11-1-1-11
-111-11-1
-1-1-1-1-1-1
000000
Solution

If you load your design into JMP, you can go to DOE->Evaluate Design, and pick your 6 columns for factors. It pre-loads the main effects for your model, with all the 2-factor interactions as alias terms (ie., not in the model, but potentially active). Examining the Color Map on Correlations, I can see it's a regular resolution IV fractional factorial with 3 center runs. To see which interactions are aliased you can hover over the off-diagonal entries that are red.

Cheers,

Ryan

11340_pastedImage_0.png

ian_jmp

Staff

Joined:

Jun 23, 2011

Nice!

bwilliamson

Community Trekker

Joined:

Jul 27, 2015

Beautiful.

I didn't know that could be done.

Thank you!