Beginner Question

[#1]

Hi community, 

this is my first post ever in a forum and also the first time I'm using DoE for experiments. I have watched the DoE Video from JMP and have a question or rather look for confirmation: 

I have an experiment with two continuous factors (e.g. range of pHs) and one categorical factor (e.g. type of acid). I want to use RSM but apparently, it won't work when a categorical factor is included. Hence, I have decided to go for a "Custom Design" and included all the factors. Now the system is saying "Categorical factors cannot appear in polynomial model terms. Only the quadratic terms for continuous factors are being added to the model.".

Would a more senior DoE user say this is a correct approach? And is there anything I need to keep in mind? Alternatively, I was thinking of doing it like all the others and assume the categorical factor (e.g. type of acid) as a continuous factor. But this just doesn't make sense in my point of view. Thank you for any help.

[#2]

The much older RSM designs only accommodated continuous factors. The modern Custom Design can handle any kind of factor, so you can have continuous and categorical factors in the same experiment. You can also include terms for interaction effects between continuous and categorical factors, and non-linear effects of continuous factors.

The warning is just letting you know that Acid*Acid will not appear in the list of model terms.

[#3]

Hi Mark and all others who replied, 

thank you very much for explaining. I still have one question. Recently, I read a paper (S. Sadhukhan, U. Sarkar /Energy Conversion and Management 118 (2016) 450–458) where the aim was to increase the glycerol purity to a maximum, and they did following (which confused me a bit): 

They used a CCD with three continuous factors: 

1.) Type of Acid 

2.) pH 

3.) Amount of adsorbent

What is confusing me is exactly what @statman said: factor 1) Type of acid is not a continuous factor. Actually, it should be a categorical variable. Hence, it is actually a pseudo RSM as @P_Bartell said, since they converted an actually categorical factor into a continuous factor, correct? (Q1)

Is this approach valid from a scientific point of view? (Q2) I have seen at least 4-5 other authors doing exactly the same.

According to my understanding, you cannot do this, as there is no continuity between the type of acids (it is a category! e.g. sulphric, phosphoric, acetic etc...), so no curvature. I'm doing something similar and would like to know: would a correct approach be defining three different variables which are continuous e.g.

1) pH - Saponification - conti

2) pH - Acidification - conti

3) Salting out (=ratio of solvent:aqueous layer) - conti

and use a fourth categorical factor as "type of base"? (Q3) If yes, this would mean that the model does not take into account the relationship between factor 4 and factors 1-3, correct? (Q4) But this would also be not logical since the influence of the type of base is highly relevant for factors 1-3 as well. Your replies are very much appreciated.

[#4]

A1: They are using the older RSM design and analysis, so they had to treat the type of acid as a continuous factor. You do not need to do the same thing. You are correct that it is a categorical factor and should be modeled as such.

A2: It is absolutely valid.

A3: Yes, your definition of the factors makes sense.

A4: You can model interactions between X4 and the first three factors. Just cross the factors in the model. For example, X1*X4 would estimate and test the interaction effect between them.

[#5]

Dear Mark, 

many thanks for your replay. Last question: Why does the Run 9 have a saponification of 8.99. Is there any logical explanation to this? Thank you in advance!!

[#6]

Custom Design is based on the 'coordinate exchange algorithm' and an 'optimality criterion.' It is not based on a combinatoric method. It starts with a completely random design. That is to say, the factor levels in the initial design are uniformly randomly distributed. Not a good design! But the algorithm determines better coordinates for one run, changes them, and iterates like this until it converges on the best value for the criterion. This guided search is not guaranteed to find the globally optimum design, so the process is repeated many times.

This approach sometimes leads to designs that are not intuitive. It can also lead to a different design if you click Make Design again. This result is the case when there is more than one design that are equally optimal by the criterion.

In this case, a value close to 9 will provide more informative data with regard to the model and the criterion than another value. On the other hand, you may change it. For example, it might be more practical to use 9 than it is to achieve 8.99. The design won't break or fail to perform if you change a value like this one.

[#7]

Ok, now I'm a bit confused. Does this automatically mean that I would be better of using the JMP-CCD with the base as continuous factors to determine the process which yields the highest glycerol purity, or is my design better? 

[#8 · Accepted solution]

I would not recommend the CCD when you have a custom design that is optimally tailored to your experiment.

[#9]

Dear Mark, 

many thanks for your useful hints and one question appeared during the course of the work. If it is not possible to go to a pH of 14 with one of the runs, is it possible to change specific experimental runs and change the pH from 14 to for example 12 within the table? 

[#10]

@Artaxerxes90 It's important to add an '@' sign to the beginning of a name in a reply so as to 'tag' that person in the Community. Then if they have notifications turned on they will get an email notification from the system and in turn hopefully respond. I can tell you that if the 'table' you are referring to is a JMP data table generated by any of the JMP DOE platforms, then it's fully editable. Obviously it changes the specific treatment combination within the design so you may want to create a completely NEW data table with the second design and then go through the Evaluate Design platform comparing both designs to see the impact of the change...mostly noting correlation of parameter estimates and if any effects are no longer estimable would be two things to check for sure. Without knowing the original design, my strong suspicion, is that making this change to ONE and only one run won't have serious implications...but that's just my suspicion.