Center Points and Screening Designs

[#1]

Hi members, 

I am trying to do a custom main effects screening design using JMP 16. I have five factors and one of the factors is categorical with 4 levels. The custom design generates a 16 run experiment and I was wondering if it's necessary to add center points to the study? I am only interested in determining the factors that matter and which of the 4 categorical variables I should use for an RSM design. Any help will be greatly appreciated. 

[#2 · Accepted solution]

Necessary? Strictly speaking...no, if your primary goal is estimating influential main effects and main effects only. Especially for a screening design where you expect and plan to conduct follow up RSM model based experimentation.

The only reason I might consider adding center points is you are looking for some sort of internal to the experiment independent estimation of pure error wrt to the experimental space. But to me this begs another question...since ultimately you'll be picking one or more levels of the categorical factor...wouldn't you want to estimate pure error at EVERY level for the categorical factor to see if that 'noise' is constant across the categorical factor? If the answer is 'yes'...then you'll need to replicate the center points across all levels of the categorical factor, which will balloon the number of runs in your design pretty quickly...where a preponderance of runs are now in search of pure error and not significant main effects. Long way of saying...I'd leave the center points out...

Others may feel differently?

[#3]

Thank you very much @P_Bartell I will go ahead and do a screening experiment without the centre points. 

[#4]

Center points are about continuous factors (or discrete numeric factors). You say you have 5 categorical factors. If I interpreted your description correctly, then you only need to be concerned about main effects and interaction effects.

[#5]

I have one categorical factor with 4 levels and 4 other continuous variables at two levels each. 

[#6]

Thank you very much @Mark_Bailey 

[#7]

Working with 4 levels of a categorical factor in a screening experiment with only 16 runs seems like it might over simplify your system.  If there is any chance that there is an interaction between the categorical factor and the continuous factors, it might be a good idea to add the interaction term to your model.  This makes your screening experiment a little bigger; however, a few extra runs for each level of the categorical variable would increase your ability to detect differences between them. 

As an example, let's say that only 3 of the continuous variables are important, and you select one level of the categorical variable, then an additional 8 runs would be necessary to add the interactions and quadratic effects for the remaining 3 continuous variables. (DOE>AUGMENT)

If the experimental units aren't difficult or expensive, more runs is better anyway.

 

[#8]

Thank you very much @Byron_JMP 

I have to limit the number of runs I can do due to cost and I also suspect that two of the factors may not be significant. I opted for the custom design's D-optimal design that gave me the 16 runs, hopefully, it can help me identify one of the categorical variables and the important factors for an RSM study.