Adding candidate runs in mixture design

[#1]

HI, Is it possible to include candidate runs in mixture design. When i tried to add 4 runs that i wanted to include in the design. It is considering them as another factors. So, may someone explain how this works?

[#2]

Yes. If you make a separate table with your 4 candidate runs, then use the 'Select Covariate Factors' button, these candidate runs will be considered/included (depending on the check boxes at the bottom) in your design.

[#3]

Thank you  @Jed_Campbell . So my question is, Why does JMP treat them as different factors?. As far as i understand, by candidate runs we mean to add some custom values of the factors that we would like to see as part of the design. Please correct me if i am wrong . And in my case they were treated as only covariate factors and not mixture factors. Hope my question is clear for you. 

thanks in advance.

[#4]

Hi @Mathej01,

Some additional information about the option proposed earlier :

The candidate set approach, described by @Jed_Campbell, may be useful if you have a lot of data points fully covering your (constrained) experimental space. If you start from 4 points only and follow the option above, you might end up with an error or small design not very adequate for your objective(s).

I would consider using these 4 initial runs and build the design around it using the "Augment Design" platform. This way, the factors will be kept as initially described, not as different factors (and same for the response(s)). You might have to specify some column properties for your mixture factors before launching the Augment platform : properties "Mixture", "Design Role" and "Factor Changes" (if not already present in your initial table).

Example here with 4 initial runs (red) augmented in a Scheffe Cubic design with 8 additional runs (in blue) : 

You can have a look at both approaches with the toy dataset I have created with 4 initial runs, and see what looks best for you in your use case.

Hope this answer will help you,

[#5]

Hi 

Thank you @Victor_G . I believe your answer could be one approach to solve my issue. But i think i already made a mistake in understanding the concept of having candidate runs. I think the candidate runs i wanted to include the runs that already fall within the limits of the factors (mixture factors) that i specified. So my idea was that, i wanted to have some of my current formulations to be included as runs . And then make a design of total 15 runs to have prediction model. Does any of this make any sense ? or in this case it wont make any sense to add candidate runs?

And if its possible, how would JMP make a design . And will this limit the efficiency of the design?

Hope my question was clear.

thanks in advance.

[#6 · Accepted solution]

Hi @Mathej01,

It sounds like you want to have "forced runs" in your design more than using a Candidate set approach, which requires a table with all possible experimental combinations runs in which the Custom Design platform may pick the most informative runs regarding the model you want to investigate.

The Augment Design platform may be a suitable choice for your use case. I totally understand your choice to include formulations already done, and the Augment Design platform is completely able to handle the information from these initial runs and provide new runs that help complete the information needed by the model you specify.

In the toy dataset that I provide, you can test your use case by augmenting the 4 initial runs into a design (checking the "Group new runs into separate block" may be a good idea to evaluate any variability change between the initial and augmented set of experiments), specify the model you want and the number of runs in total (for example 15 runs including the 4 initial runs). In the example provided, you can specify a Scheffe Cubic model (with the blocking enabled) and have a design proposal with 15 runs, so it does make sense.

The efficiency of the design is linked to various paremeters, such as the number of runs you can afford, information brought by your initial runs (repartition of the points in your experimental space), complexity of the model you assume...

So there might be no definitive answer to your question, the best thing to do is to try different scenario (design generated by augmenting initial runs or design generated from scratch) and compare the designs obtained through the "Compare Design" platform.

You will have a better overview of the potential benefits of augmenting a design, since the 4 initial runs are already done, so that means you may compare a 15-runs augmented mixture design to a 11-runs mixture design from scratch (if you can't afford more runs), and benefit from this additional information to have better precision in terms estimation, better predictive performances and/or the possibility to assume a more complex model.

Hope this answer helps you,