nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Curved Prediction Lines instead of Linear

JoshE87

I am very new to JMP, as this question will likely let you know. I designed a DoE in the hopes that I could predict the outcome of future experiments. After using "Fit Model" and basically keeping everything how it defaulted (Standard Least Squares and Effect Screening), I am seeing only straight lines when I go to the prediction section, as shown below.

I know based on other experiments I've performed that some of these relationships should be curved lines (polynomial, logarithmic, exponential...what have you). I'd like to see these prediction lines looking more like the below image from the JMP website:

Please let me know how to get better predictions from my data! Thanks very much!

Accepted answers

Dan_Obermiller

Well, yes, 3 levels does define a curve, but you have to consider the combinations of all of the variables. Look at this plot that shows all four variables together (a plot of B versus A for each combination of C and D). There is never a spot where you have three points in the same quadrant that would allow estimation of a quadratic.

Graph Builder.png

If you were to remove a factor, say D, you MIGHT be able to get some quadratic effects, but you would still not be able to estimate interactions.

Graph Builder2.png

Essentially, the Taguchi design that you are looking at is called a resolution III design. It is intended to fit main effects only, not even interactions between the factors. You ALWAYS need to choose the design that will meet the objectives of the experimentation. If you want to estimate quadratic effects you will need a different design type.

Remember that JMP's custom designer will help you with this. Rather than choosing a named design (like Taguchi), specify your factors and the model that you wish to estimate. JMP will then choose the design that allows you to estimate your chosen model.

All comments

Dan_Obermiller

To completely answer your question could fill an entire book as several books on design of experiments have been written to help answer your question.

But, to try and be brief, your design will allow you to fit a specific type of model (the design and the model go together). If you desire to see curves, then you need to have a model that exhibits that behavior. This typically means that your model needs to have quadratic or squared terms. Your choice of design needs to provide the right experimental runs that allow estimation of such terms.

JoshE87

Thanks very much for the explanation.

I was under the impression that having a 3 stage continuous parameter set would allow for curved fits? I used a Taguchi method design of the following type:

Shouldn't that allow for curved fits since there are three numerical levels to each continuous parameter? Thanks again!

Dan_Obermiller

Graph Builder.png

If you were to remove a factor, say D, you MIGHT be able to get some quadratic effects, but you would still not be able to estimate interactions.

Graph Builder2.png

JoshE87

Thanks, this was very helpful!

P_Bartell

To supplement @Dan_Obermiller 's replies, an efficient place to begin building BOTH your basic DOE skills AND 'How to in JMP' is through completion of select SAS training classes that cover the DOE domain. Two to get started might be:

https://support.sas.com/edu/schedules.html?ctry=us&crs=JMDOE

https://support.sas.com/edu/schedules.html?ctry=us&crs=JDRS

And if you only have resources (time, money, etc.) as @Dan_Obermiller hints at, I'd complete the first entry above. With JMP and it's Custom Design course...it's the way to go nowadays for efficiency, flexibility, and adapability of many, many types of practical problems to the DOE problem solving approach.