DOE for Binomial response

Question

Hey all.

Do you have a standard response or recommendation for customers who wish to model a binary response using DOE? I'm not talking about deterministic systems like computer simulations -- just something analogous to the typical DOE with a continuous response, except the end analysis would be logistic regression instead of a standard linear model.

One idea I've heard is to replicate a traditional or custom design a number of times, but I can't find any references describing this situation, in JMP documentation or elsewhere.

I believe that Brad Jones was discussing this idea at kickoff this year when demoing the new simulation feature for JMP 13?

Cheers,

Jordan

Ryan_Lekivetz · Answer

As far as I'm aware, the simulate response feature is in all JMP (ie., you can generate a random binomial response column), but the simulator itself to simulate the p-values (or whatever else you choose) in Fit Model is Pro only.

DonMcCormack · Answer

Last I heard the simulator for binary DOEs will be a JMP Pro only feature. Is that still the case?

Ryan_Lekivetz · Answer

Hi Jordan,

Using a custom design is certainly a possibility, and what you'll usually find is that running only one trial per run from a DOE doesn't give much power to detect effects.

To get an idea of power/quality of estimates, you can't use the power from DOE, but in JMP 13 you can do a lot with the new simulator feature. Attached is one example from Brady.

To make it work, you need a formula column with a random piece that will be updated each time (the Y column in this example).

We're working on a new Simulate Responses in DOE that ties nicely into the simulate feature, in that it looks similar to what was there before, but creates the formula column needed for the new simulator feature to work on other platforms. The DOE Simulate Responses lets you specify a binomial distribution to generate the data as one of the options.

Cheers,

Ryan

Logistic Regression Design N5.jmp