Lack of Fit report provided based on dataset instead of model

[Kangwon]

What inspired this wish list request? 

When I create a 2X2X2 factorial design with 2 replicates (16 total runs), and run the reduced model with factor effects A, C, and AC, I expect to get the lack of fit test outcomes because in this case, I have 8 pure error degrees of freedom and remaining 4 degrees of freedom can be used for lack of fit test. But when I really run the model, JMP doesn’t activate the lack of fit report.

This is because JMP completely ignores a column B and recognize the model with A, C, and AC as a full model in two factors A and C. But it is truly not a full model in terms of three factors A, B, and C.

What is the improvement you would like to see? 

I suggest let JMP keep the information of factors involved in the data (not the model) and use it to provide a LoF report.

Why is this idea important? 

For example, suppose a user built an initial model with all possible terms in 3 factors factorial design. By investigating the effect summary window, the user can remove some unnecessary terms one by one. Whenever a term is removed, the user may want to confirm the elimination by checking the LoF report. However, currently, if all the terms involving a specific factor are eliminated, LoF is not provided.

This is probably not only restricted to 2-level experimental data; the same thing might happen with any general datasets, I think.

2x2x2 Factorial.jmp

[SarahGilyard]

Thank you for posting Kangwon. The platform is behaving as required based on the nature of how LOF is computed. The "full model" is defined in reference to the terms that are specified in the model effects dialogue box. When you specify the model as having effects A;C and AC, you have specified a full model. The model is still being fit using all the data (ie: all 16 rows of data are being used). In your attached design, there are only factorial design points (replicated), with no center points. If you included replicated center points in the design you would generate LOF that could test for quadratic curvature even when a full model is fit. With only replicated factorial points and no center points, you will only generate LOF if a reduced model is fit and it will be testing for whether interactions are needed. I hope this explanation was helpful. Thanks!