DOE - Fit Model and Profiler

Question

I am using Fit Model and Profiler to help me identify which factors are important and predict the response.

After Fit Model -> Stepwise analysis, only 2 out of 5 factors are important. I try and put all 5 factors into Fit Model -> Standard Least Squares & Effect Screening, the Profiler gave me different prediction. That make sense to me because now JMP fit a different model. But I know the first model with only 2 factors are more accurate. If I try to maximize my response, how do I find out the setting of the other 3 factors that showed not significant to my response?

louv · Answer

Agree with Kevin,

Assuming you have enough degrees of freedom to fit the model, I would start there and pull terms out that are not adding value. You used stepwise but did not delineate which stopping rule that you employed. I also read into your question that you were trying to ascertain what settings the other terms that were removed should be set at. In that case include those main effect terms in your model and the Profiler will then indicate their impact on the response.

Kevin_Anderson · Answer

When you say you "know the first model with only 2 factors are more accurate", what do you mean?  How do you know?  Because Stepwise said so??

A web search on "problems with stepwise regression" will quickly detail many technical reason to distrust Stepwise.  Not just JMP's...any software...

My favorite quote about Stepwise procedures:  "Personally, I would no more let an automatic routine select my model than I would let some best fit procedure pack my suitcase." Ronan Conroy.

Part of the problem is that Stepwise sets the parameter estimate of the Unchosen factors to zero, so the settings of those factors don't even matter in the Stepwise model.  Is that legitimate in your application?  This can cause significant bias.  Why not use the full model?  If you're unhappy with the full model for some legitimate reason, I would recommend using some penalized regression technique like the lasso or elastic net in the Generalized Regression platform.  At least those methods constrain the parameters and deal with multicollinearity (unlike Stepwise).