I have an extremely high R^2 (Pred) of 96% and R^2 (adj) of 99%. Using Minitab's response optimizer I generate a confidence interval and prediction interval for specific settings. Yet when I get to the shop floor to test these settings my results are generally out of the interval.
The intervals for specific settings are generally .002 wide which is extremely tight for machining processes.
I am assuming the operator is having a significant effect on the parts.
Have I over fitted the model? Taking out terms reduces the R^2 (pred) but it widens the intervals which may be more relevant.
Adam,
As you are using MINITAB, you can first check VIF. If some ones are much greater than 1 (10, 20, ...) you have non independence problems. The "4 in one" graph or residuals can give you also informations. I don't know in what domain you are working, but in mine (chemistry), a R²(pred) of 96% is "too nice". Roughly, it means that the "experiment + measurement" error is very low. Is it your case?
Without data, for me, it's difficult to give you more informations.
If you can send me the data (if they are confidential, call independent variables x1, x2, ... and dependent ones y1, y2, ...), perhaps I'll can tell you more.
Pierre,
Thank you for the reply. All VIF's were 1.00. As far as domain, I'm modeling a machining process in the aerospace industry. I have attached the data file. TIR 1 is the response variable.
Interesting you say error being too low. I was under the assumption that minimizing your error was the hallmark of a properly performed experiment.
I appreciate any insight you may have.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis