Is there a way to force JMP to keep a model prediction (for instance, a "reaction time") positive (i.e., above zero)? I'm wondering if that constraint would help build a more realistic model for my data set.
As far as I know, there is no capability in JMP to fit a linear model and force the predictions to be positive. However, the general question you ask is not really a JMP question but a modeling question.
I believe that one solution here is to fit a non-linear model to your data. If you choose the right non-linear model, the predictions will be forced to be positive. Another solution might be to transform your data somehow, fit the model, and then un-transform (again, if you do this properly you can force the predictions to be positive). What the right non-linear model is, and/or what the right transformation is, depends on your data.
I could provide some more info that might be helpful. I am modeling within a DoE, so with the Fit Model approach (factorial to degree 2, with 7 factors in the mix), I get a complex, non linear prediction formula. The mathematical model has a good fit (r-sq close to 1), but will predict a non-realistic result given certain tweaking of the factors from within the profiler.....for instance it will predict a time of negative 45 hours. I am doing other experiments to confirm (or not) and refine the model as we go. But I'd like to keep JMP modeling within the frame of reality (positive numbers for time and masses of ingredients) if possible.
Ah, terminology. When a statistician (such as me) speaks about a non-linear model, that term non-linear is guaranteed to confuse anyone who isn't a statistician. Sorry about that.
Your quadratic model (factorial to degree 2) is considered a "linear" model in this terminology, even though it has squared terms and interactions and the result isn't a straight line.
What I meant by non-linear was something like an exponential, or a piece-wise fit, or anything other than a polynomial, which is what JMP fits in Fit Model.
The above isn't really relevant to solving the problem, but it might be in the future. So, where does that leave us? I think you need to decide if you need a good fit near a response value of zero, where you are having trouble, or a good fit elsewhere, or both. Is the goal of this modeling to predict near a response of zero? Also, when you use the profiler and get a time of –45, this could be indicating you are trying to predict in an infeasible area, or an area where the model doesn't apply. Lastly, I remain concerned that you claim you get a r-sq close to 1 (what does "close" mean?) and yet the profiler is giving you negative predictions. Are you dragging the profiler sliders beyond the range of the x data?