Prediction profiler from DOE model does not show anything.

Question

Hi all,

I am rather new with JMP 15 and I would like your input regarding a DOE model issue I am having. I have seen some related topic, but not exactly to the same extend. ( Prediction Profiler  )

I have generated a DOE custom design with 6 factors (4 continuous and 2 discrete numeric) and RSM model. The 4 continuous have one center point, while the 2 discrete numeric are only the extremes value. My data table is fully filled with data. I have attached an example with randomized values. I have followed the minimum number of 24 runs. I have set to "if necessary" some effects I knew would not have any influence.

My problem is now that when I run the model and fit the data, both the "Effect summary" and the "the prediction profiler" are empty. I have also a singularity detail.

I can get the effect summary to work if I remove the model effects listed in the singularity detail. But the model is poorly fitting and I do not want to remove these effect from my analysis as I know they are important.

Is there a way to still create a prediction profiler with a mix of continuous and discrete numeric data model ? Does any of you has experience with running a reliable model with these type of data available ?

print_screen_model.PNG

CD_RSM_6_Parameters_24_runs_random_data.jmp

Mark_Bailey · Answer

Point out that the problem arose because you changed some terms' estimability in the model to 'if possible and then selected the minimum number of runs. That path means you cannot estimate all of the terms in the model. You might consider removing terms from the default RSM model if you believe they are unnecessary. You can remove them now when you open the Fit Model launch dialog. You can save this change from the red triangle menu in the dialog so you don't have to remove them again if you want to run the model again.

statman · Answer

The reason for the issue is you have entered the wrong model into JMP.  You should have the model in mind before you run your experiment.  The experiment is a way of evaluating your model.  I would not advise using stepwise regression on existing DOE data.  Stepwise is an additive model building approach, and I recommend you use a subtractive approach (start with a saturated model and remove unimportant terms from there).  I have added your JMP table with the script Fit Least Squares with the correct model (I only have one response , but you can add the others).

CD_RSM_6_Parameters_24_runs_random_data.jmp

Victor_G · Answer

Hi @Chantal,

Yes, you have a singularity in your model, as the intercept is a linear combination of some of the terms, so it can't be estimated independently, creating the error in the prediction profiler you mentioned.

One possible option to create your responses' model is to launch the platform "Fit Model" from the script "Model" in the datatable provided, but switch the model personality from "Standard Least Squares" to "Stepwise Regression". This way, you'll be able for each response to test which terms are significant, and get more reliable models. Click on CTRL+"Go" to launch stepwise regression on all responses, and on CTRL+"Run Model" to open a "Fit Group" platform containing each model for the responses. Once the stepwise regression is done, you can still remove manually terms in each response's model if needed from the generated "Fit Response/Model" platform.

See script "Stepwise analysis" in your datatable to see the results for each of the 4 responses and the prediction profiler.

For more infos about Stepwise regression and how to use it, you can search here : Stepwise Regression Models (jmp.com)

I hope this first answer will help you,

CD_RSM_6_Parameters_24_runs_random_data.jmp