Screening AnalysisI will be submitting the below background and four questions to JMP technical support.
I just finished teaching a classroom course on Design of Experiments to some P&G employees. We had some people on JMP 12 and some on JMP 13. It seems that JMP 13 makes it much more difficult to get the analysis out of Fit Model, and instead really wants you to use the new Fit Two-Level Screening platform. In JMP 12 you could get an analysis out of Fit Model or out of the old Screening platform.
First question: Why is it much harder to find screening analysis in JMP 13? It's under Analyze > Specialized Modeling > Specialized DOE Models (rather than simply Analyze > Screening). Fortunately it's easier if the design was created by a DOE platform to just click green-triangle for Screening.
Second question: Why has JMP stopped giving Summary of Fit and Analysis of Variance for screening designs by default in Fit Model?
Third question: When there are no error DF, JMP doesn't give any estimates or p-values by default any more. Why? I know you can get the analysis for no error DF left from Fit Model in JMP 13 by red-triangle > Estimated > Sorted Estimates. But the more intuitive place to look for me would be red-triangle > Effect Screening > Scaled Estimates, and this doesn't work.
Fourth question: There is a difference in p-values between Fit Model and Fit Two-Level Screening (but same estimates, t-ratios). Why? See Image 1 and Image 2 below. Data set also attached.
Image 1 - Fit Model outputImage 1 - Fit Model output
Image 2 - Fit Two-Level Screening outputImage 2 - Fit Two-Level Screening output
Re: Screening AnalysisTechnical Support's response (paraphrased, tracking number 7612015122):
1. Analyze menu was rearranged in JMP 13 and the new placement of the Screening Platform under multiple submenus is as intended.
2. The Effect Screening emphasis of Fit Model overrides any JMP Platform Preferences and does not show the Summary of Fit and ANOVA tables unless explicitly requested by Red Triangle or JSL. This is as intended by the developer since someone who is screening shouldn't be worried about getting a good model fit but rather about identifying the correct effects.
3. Even when using Effect Screening in Fit Model, Sorted Parameter Estimates are no longer automatically triggered. If it's not intuitive to turn that option on in the red triangle (or Preferences or JSL), then the Screening platform may be a better choice.
4. Screening platform calculates the p-values by simulation, while the Fit Model platform does not. You can force Screening to give the Fit Model results although this is not recommended.
My reactions (not sent on to support):
1. Fine. I don't like the platform being positioned this deep but I understand why it was done.
2. I think this is an adjustment in thinking that I could make.
3. Given number 4, this is probably the correct behavior.
4. Is there a white paper or anything else that explains why the simulation is better? And if it is better, why is Fit Model still calculating it the old way?
