Solved: Re: When creating a DoE, is the modelling type or design role or both considered...

DaniellaS · Jun 8, 2023 2:05 PM

Our experiment has a factor which is set to "low" / "middle" / "high" - these levels will be adjusted by a mix of settings

Data Type is set to Character

Modeling Type is set to Ordinal

Design role is set to Categorical

Value Order is set to "low" / "middle" / "high"

Does JMP create the design with that information, that low is lower than middle, etc.?

Or will it be the same design as with three categorical factors which have no order or relation?

The alternative is to rename "low" / "middle" / "high" into "1" / "2" / "3" with

Data Type set to Numeric

Modeling Type set to Ordinal

Design role set to Discrete Numeric

For our experiment, we need less recommended runs.

Which way is the correct way or how do we decide which one to use?

Thanks in advance for any help!

Daniella

Dan_Obermiller · Aug 19, 2021 11:52 AM

If a factor is a character string, JMP will treat the factor as discrete. That means that the levels are just names and there is no implied order (ordinal and nominal are treated the same -- remember that in the Custom Designer JMP does not ask for the modeling type). The disadvantage for a truly categorical factor is that interpolation cannot be used to help estimate the parameters at the various levels. You MUST take data at each level, and taking fewer data points at a level will decrease the estimation of the parameter for that level.

If a factor is made as discrete numeric, JMP will create the factor as continuous. The key difference is that in order to achieve all of the discrete levels, JMP will add additional "if possible" terms to the model to force the extra levels. For your example, it would add an extra squared term for your factor to ensure that the "2" level will appear in the design. This approach MIGHT lead to the middle level not being chosen as often as if the character string were used. Why? Because if the factor truly is continuous, then interpolation will can help us in the parameter estimation. I know a bit more about the middle of the factor range because I will have data collected at the low and high settings.

Generally, if a factor can be treated as continuous, I would recommend making it continuous for the design. But if that is the case I would then question: why are you needing three levels? Mathematically the only reason for the third level is to estimate curvature. If that is the case, then the model should have a squared term in it. So there is lots of discussion to be had around the factors and which way they should be treated, but I think most of those answers are based on subject matter knowledge.

Ultimately, create the design both ways. Which has better properties that meet your needs?

Dan Obermiller

View solution in original post

Dan_Obermiller · Aug 19, 2021 11:52 AM

If a factor is a character string, JMP will treat the factor as discrete. That means that the levels are just names and there is no implied order (ordinal and nominal are treated the same -- remember that in the Custom Designer JMP does not ask for the modeling type). The disadvantage for a truly categorical factor is that interpolation cannot be used to help estimate the parameters at the various levels. You MUST take data at each level, and taking fewer data points at a level will decrease the estimation of the parameter for that level.

If a factor is made as discrete numeric, JMP will create the factor as continuous. The key difference is that in order to achieve all of the discrete levels, JMP will add additional "if possible" terms to the model to force the extra levels. For your example, it would add an extra squared term for your factor to ensure that the "2" level will appear in the design. This approach MIGHT lead to the middle level not being chosen as often as if the character string were used. Why? Because if the factor truly is continuous, then interpolation will can help us in the parameter estimation. I know a bit more about the middle of the factor range because I will have data collected at the low and high settings.

Generally, if a factor can be treated as continuous, I would recommend making it continuous for the design. But if that is the case I would then question: why are you needing three levels? Mathematically the only reason for the third level is to estimate curvature. If that is the case, then the model should have a squared term in it. So there is lots of discussion to be had around the factors and which way they should be treated, but I think most of those answers are based on subject matter knowledge.

Ultimately, create the design both ways. Which has better properties that meet your needs?

Dan Obermiller

When creating a DoE, is the modelling type or design role or both considered?

Re: When creating a DoE, is the modelling type or design role or both considered?

Re: When creating a DoE, is the modelling type or design role or both considered?