264087_en-US

liliconcom

CCD facial Design with blocked factors

Hi,

I have trouble to replicate a design a customer had made with Statistica. My two goals are to once replicate and second to find a probably better or similar good design with less runs.

 

The setup is the following:

• They do a CCD facial design with one replicate, the 35th run is one setting they want to test in each lot, as this is an extreme result they had got in pre-experimentation and is only in the first two lots
• Over the run of the experiments they will have three starting materials (Lots):
  • the first lot has 10 Liter aka max of 10 runs,
  • the second lot has a 10 Liter aka 10 runs max
  • the third can have more (in their design 15 runs) 
  • they cannot be randomly changed they are one after each other
  • they want to understand the lot by lot variation, therefore they have it as blocking factor in their design
• They want to do one complete replicate for the design
  • so two Operators do the same experiments 
  • want to randomly vary them within the Lot block
  • currently they just alternate the runs between operators as they cannot do this in their tool
• Three continuous factors
• (want to relate the factors to several responses)

 

Attached a JMP table of an imported Excel file. It is anomized due to confidentiallity.

 

What I have done so far:

• I tried the custom designer, but failed with restriction of different block sizes for first two lots 
• I tried to do a CCD orthogonal blocks design, on face, with 3CP and one replicate aka 34 runs at all (see CCD 3CP 1repl.jmp), however I could not include the Operator (needs to be manually added).

 

==> Do you have any idea on how to use custom designer to generate a similar design?

==> Do you know of any other design approach what would be helpful in such a requirement?

 

Thanks,

Martin

 

Re: CCD facial Design with blocked factors

As far as I know, specification of different block sizes is not an out of the box feature in the Custom Designers. I'd be surprised if this wasn't built in an ad-hoc manner and not directly available in Statistica. If it is, I'd be curious to know how it is implemented. @bradleyjones  did something like this a while back for split plot designs for a journal article but used a script to create them.

Re: CCD facial Design with blocked factors

Hi Martin,

Does it help if you start with a table for the 35 runs by Lot factor? (10 rows Lot = 1, 10 rows Lot = 2, 15 rows Lot =3)

Then use this a covariate factor in Custom Design.

Regards,

Phil

Re: CCD facial Design with blocked factors

If you do this, I'd remove Lot from any term in the model when you build the design. Once the design is created, change lot to Nominal and add it as a Random effect to the Model dialog.

Re: CCD facial Design with blocked factors

1. With covariate I will loose the ability for replicates. but duplicating the created design should work as well, right

 

2. In addition the run order will be random throughout the covariate variable (which makes sense of course, but does not help as the lots starting material is defined. If the 1st run in lot 1 is e.g. all high values (1, 1, 1) then the second run should be the same just done by Operator 2.). Can I just ignore the run number (aka keep the checkbox uncheckedbefore making the data table)?

 

3. I used once a Lot covariate with 35 rows as described by Phil and set up a 35run experiment. Then I tried a covariate table with 5/5/6= 16 rows and created a 16run Design, dublicated it in the data table to have same experiments for both Operator 1 an 2. 

I won't get to a lower #runs as 11 or 22 if duplicated.

Can I then simply compare the designs using e.g. evaluate design? I am unsure if we then assume something which cannot be hold. 

 

4. Your comment on the nominal Lot to be used as repeated measure is not clear to me? We do not jsut measure the experiment by two Operators, but conduct the experiment twice, one time with Operator one, one time with Operator two. Or did I get it wrong and you mean it should be treated as blocking variable and therefore the random effect?

 

Thanks  a lot for clarifications. Attached my three designs.

 

Re: CCD facial Design with blocked factors

1. If by replication you mean the entire design (as opposed to a single run or set of runs), then yes.
2. I thought this would be the case, too, but the designs I've created keep Lots in the same order as the covariate table. Not sure if this is relatively new (works this way in 14, too) or something about Lot.
3. Why not add Operator as a two level Categorical Factor? For your first design, don't include the Lot*X2, Lot*X3 and Lot*X4 cross terms. You want to treat Lot as a (random) block to estimate the variability due to the Lot to Lot differences. You will also need to alter the model that gets saved to the design table (see below).
4. Are you treating Operator as an additional blocking factor or trying to estimate the differences between the operators? If you want two blocking factors, it gets a bit more complicated. Blocking factors are design components where the interest is in estimating variability not mean differences. This is done by treating the factor as a random effect. In your attached design, you need to apply the Random attribute to Lot (and remove any cross terms it's involved with). With a bit of care, setting up a two blocking factor design isn't difficult, but requires you specify the correct model for both design and analysis.

 

Re: CCD facial Design with blocked factors

First of all, thanks Don!

1. yes replication of the entire design was what I thought about, as they want to have the similar runs done by two operators.

2. ok

3. They do some process characterization and thought having some idea of the Operator effect could help to gain additional insight if there is something wrong with the way they handle the experiment. Then they said, that - though they do it this way - at the end they take all the data independent of the operator. With that information, the operator factor might not be necessary at all. (Overall the Operator to operator variation should be checked upfront in a measurement study if possible.) 

If they want to understand if there is an operator effect due to e.g. different shifts I like the two level cat. factor approach more than replicating the design. Thanks.

About the model, yes, I handled the other three designs that way. That was a good hint:)

4. I think they were not really clear about how to treat operator. they asked to use it as blocking factor, then they don't lock at it in the analysis, and finally after we saw some different behaviour in the operators behaviour of another design they thought about looking at the variation. I believe I will recommend them to use the model approach with the random effect for Lot, and make a design with the categorical Operator factor.  then they also might be able to gain appropriate insight with less than the 35 runs.

 

 

4.  

Re: CCD facial Design with blocked factors

Don, one last question: 

If I use Operator as blocking factor. How would I need to adjust the model:

- main effects for all, interaction/quadratic effects just for the three continuous factors?

- and in the model (data table) I need set "Lot & Random", and then? Nesting Lot and Operator? 

 

Thanks,

Martin

Re: CCD facial Design with blocked factors

Martin,

If Operator is another blocking factor, it would not be crossed with any of the other experimental factors, but may possibly be crossed with Lot. It would also be random. The model would look something like this (order is not important):

 

A, B, C, A*B, A*C, B*C, A*A, B*B, C*C, Lot&Random, Operator&Random, Lot*Operator&Random

 

The Lot*Operator interaction is quantifying the variability due to different operators using different lots. It's not something I'd expect to contribute much to variability (if anything). If it were something like Tool*Operator then, it might be worth looking at (e.g., different operators my be better at using particular tools, so you might expect more variability).

Don