I am a new user of JMP. I am using Response Surface (BOX-B) to optimize my formula. I have three factors, each factor has three levels. I run 14 tests in total (including two central points).
My problem now is the overall significance from ANOVA (F-test) is higher than my significance level (0.05), does this mean the model doesn't fit although the lack-of-fit p-value is higher than my significance level? What should I do with this?
Thanks a lot!
You might try removing some of the least significant higher order terms like X3*X3 and X1*X2. You could also change the fitting personality in Fit Model dialog to Stepwise and use it as an aid to model selection.
Thanks for sharing your JMP results - things otherwise look good.
Agree with Mark's assessment. Looks good and probably improve by removing the insignificant terms.
The additional thing I would share after many years is the "more the merrier" with respect to screening factors prior to any RSM. With the recently introduced Definitive Screening Designs, for instance one can efficiently screen 10 factors in 25 runs. I am only sharing this ideas because when asked "what would you do if I had to do it all again", I would say that I wished that I could have screened more factors before potentially sub-optimizing a few.
Based on what you are showing it looks like you are making mixtures of of the rubber, peroxide and co-agent. If this is truly what you are doing then there are some other things to take into consideration with your model along with what @markbailey was suggesting.
Thanks, Bill. Yes, I am blending rubber with polymers. Could you please point out what else I need to take into consideration? Thanks.
Just a couple of things you might consider.
Because you are working with mixtures or mixture components all of you mixture components will be completely correlated or said another way are not/cannot be independent of one another so with that in mind you are going to have aliasing/confounding with your model which will lead to a higher ANOVA (F-Test) value. This is more or less expected behavior with mixtures.
Based on your domain knowledge, are there combinations of your mixture components that you know will not work or will give you bad results? If so, you can add constraints to your design so that the combinations of concern will not be part of your design.
One other fit option you could try is a Scheffe Cubic to see if you get a better fit. Be careful though because sometimes a better Rsquare doesn't mean a better predicitve model. Use your Effect Summary here as well to remove higher order terms that are not important to the model. Compare your models and see which model is best.