Using Prediction Profiling to Maximize Model Proficiency - Part 1
Created:
| Last Modified:
The JMP Prediction Profiler is an option available from over 20 places in JMP to use DOE, observational and formula data to visualize and explore models and make decisions.
See how to:
- Interpret the profile trace that shows each combination of X and Y variables and the black lines show how the predicted Y changes as X changes
- Understand how the X axis shows factor, input variable, X column, independent variable, term and process parameter
- Understand how the Y axis shows response, output variable, Y column, dependent variable and outcome
- Examine the 95% Confidence Interval for Predicted Values
- For continuous X variables you see a blue line and shaded area
- For categorical X variables you see height of error bar
- Examine current values of X as a vertical dotted line and red text on x axis
- Examine predicted Y for current values of X as the horizontal dotted line and red text on Y axis
- Examine 95% CI for predicted Y as the blue text on Y axis
- ID important model factors that are important (steepness)
- ID effects are significant (main, quadratic, interaction)
Optimize Desirabilities and ID combination of factor settings that will optimize the response
- Define desirabilities
- Remember settings
- Lock factor settings (reset factor settings
- Handle multiple responses
- Simultaneously optimize responses
- Handle response that is more important than the others
Save predicted values
- Save to JMP Public
Questions answered by Andrea @andreacoombs1 and Mike @MikeD_Anderson during the live webinar:
Q: What does decrease in desirability indicate? Does it mean that it's harder for us to meet the conditions specified in the profiler?
A: A decrease in desirability means that the current settings for the factors do not perfectly meet the desired outcomes set by the user. That perfect condition may not actually be possible, so desirability is more of a qualitative measurement.
Q: If I maximize desirability for %FPY or another response that is in %age, how do I limit the maximize desirability without going over 100% or under 0%? Would it just be match target?
A: You would need to use a transform on y or change the distribution that the model is using. An applicable transform on y might be Logit Percent or Logistic Percent.
Q: Is JMP making an estimate for the standard deviation for the simulation factors based on the data? I know you can type in your SD from production data.
A: Yes, that’s the starting point and you can tweak that manually, as well.
Q: Can you explain the 'Assess variable importance' option under the red triangle, and the different alternatives under this option?
A: Assess variable importance, at a high level, does exactly what it says - help you determine important variables. The details are documented here.
Q: Could I use desirability if I want to push the variable closer to LSL? For example, in manufacturing we want to push product specs closer to min spec to gain more volume which is more profitable.
A: You’d set the desirability to “minimize” and then set the lower limit in the desirability to match your spec limit.
Q: Can you copy a predicted formula from one file to another?
A: Yes, using the Formula Depot in JMP Pro. You can use a right click option on a formula column under Edit and then select Copy Column Properties, and from there go into another data table and right click and select Paste Column Properties. (this will bring over the column formula as well as any other properties in the Column).
Q: What is an interaction?
A: An interaction would be the combined effect of two factors. Mathematically it would show up as “Factor A * Factor B” as a term in the model. In words, an interaction describes to what degree Factor A depends on the level of Factor B in relation to the Y response, or alternatively and equivalently, to what degree does Factor B depend on the level of Factor A in relation to Y.
Q: Can you constrain factors to a certain range instead of locking them?
A: Yes, you can Add Linear Constraints. If your data is from a designed experiment where you defined the constraints for your design, the constraints from the DOE will automatically be applied.
Q: Sometimes you get yield over 100% with normal distribution. How can this be avoided so that the Prediction Profiler does not give you a non-physical value?
A: The JMP Pro Generalized Regression modeling capability avoids that by allowing you to change the distribution of your response that would bound it to between 0 and 1, for example.
Q: If someone wanted to push a variable closer to the lower spec limit, how would they do that using variables.
A: You can drag the upper and lower spec limits from within the Prediction Profiler or manually change by Alt/Click on PC within the Prediction Profiler Desirability area, get a pop-up and manually change values and desirability there.
Resources
- Controlling Extrapolation in the Prediction Profiler in JMP Pro Discovery Summit 2021 presentation
- Prediction Profiler documentation
