How I can specify in the model that my factor values cannot have decimal numbers (ie. 0.5)?

Question

Hello!

Two of the factors (in my Custom Design) are parameters that I have set up in an instrument. But the instrument doesn't let you use decimal units because it must be a whole number. The instrument lets the user change the value of these factors in increments of 1 unit (ie. 21, 22, 23...).
I got a nice solution from @Victor_G (https://community.jmp.com/t5/forums/postpage/board-id/discussions), but I was curious to know of others do it in a different way?
Thank you in advance!

Victor_G · Accepted Answer

Hi @P_Bartell  and @ADouyon,

Here is the answer I gave (and hoping the link will correctly work): https://community.jmp.com/t5/Discussions/Adding-a-complex-constrain-to-a-Custom-Design-with-4-factors/m-p/556004/highlight/true#M76999

I agree with you @P_Bartell, having a discrete numeric factor would be my default choice. But since the equipment can still have a lot of different settings, I was considering also possible to have a numeric factor and change the column format to get 0 decimals.
This won't prevent JMP to find optimal response values with decimals, but rounding the result to the closest whole number could do the trick.

I would say the optimal solution depends on the increment size (here 1) vs. factor range, and also on the type of design/required runs (a design requiring runs with only min and max levels for this factor will only have whole number, so there should be no problem).

What do you think ?

P_Bartell · Accepted Answer

The link worked fine to go to your original solution...where indeed you mention the discrete numeric factor idea. If ultimately the experimenter is headed to optimization land (vs. screening...if screening stop reading now), I'm not sure I'm a fan of modifying the data table values to be whole integers because I'm not sure how the Prediction Profiler optimization routines will work for searching for optimal levels of these modified factors...it still might suggest mid range factor values for optimal solutions? I'd have to play with the Prediction Profiler desirability functions, constraining factor values, and the optimization routines to see how they would behave.

Bottom line: I think all would work as we would expect if the experimenter would just go with the discrete numeric factor idea.