nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Prediction Interval in Prediction Profiler

Florian_Vogt

Hi all,

a customer keeps asking for prediction intervals in the prediction profiler. At the moment, this is not natively possible and there already is a wishlist item for it ( Use of prediction intervals / distribution in prediction profiler and simulator ).

They have been using this in Design Expert for years already and see this as a critical functionality for their work with JMP.

Does anybody know of a workaround to accomplish this maybe?

Thanks,

Florian

Accepted answers

jiancao

I got the answer from Adam Morris of JMP Tech Support. The slight difference that I see between the prediction interval shown from the Prediction Profiler in Graph and the saved Indiv Confidence Interval directly from Fit Model is due to using the normal distribution versus t-distribution—the Normal quantile of 1.96 versus t quantile of 2.02 in my example. As sample size increases, the difference will go away. See the tech support track for details.

So, the work-around steps are:

1A. Save the Prediction Formula and StdErr Formula after your model is fit: Pred Formula height and PredSE height as in my example of regressing height on weight;

1B. Modify the PredSE height formula by adding the MSE to account for the variation in the response, e.g., 9.1816. Keep the same column name since PredSE height formula needs to be paired with Pred Formular height;

2. Run the Profiler from Graph with both the prediction formula and new SE formula. What's shown is the prediction interval for a single future observation using the normal quantile. prediction profiler.png

The data with the saved scripts and formula columns, Prediction Intervals.JMP, can be down loaded from here

All comments

malcolm_moore1

Would the simulator be an acceptable workaround:

Maximise desirability
Turn on simulator
Keep all the factors fixed at the "best" settings
Add random noise to responses
Simulate to table (using default of 10000)
Use Distribution platform on simulated data table - the 95% confidence interval of this simulated data would be equivalent to the prediction interval

Florian_Vogt

Thanks @malcolm_moore1.

The suggestion is good and working, however they wish to have the PI across the entire response surface. So the workaround for this could then be to repeat the process for different settings of course.

Any ideas how this could be done?

Phil_Kay

I would think that it would be possible to create a formula column for the prediction interval. You can then use that along with the saved prediction formula column in the Profiler platform from the Graph Menu.

I can't remember how you calculate a prediction interval so I would start by looking at this white paper. It is old but I am sure it is all still valid and I found it really useful many years ago when I needed to understand prediction intervals.

This thread should help. Also this blog post from one of our JMP Partners.

jiancao

I thought I found a work-round but not yet :-). Here is what I submitted to the Support for clarification. https://sirius.na.sas.com/Sirius/ShowTrack.aspx?trknum=7613396563. Stay tuned.

As you man be already aware that JMP Fit Model does let you save the prediction intervals to the data table (i.e., Indiv Confidence Interval, and Indiv Confidence Limit Formula-- they (values) are the same, )

jiancao

So, the work-around steps are:

1A. Save the Prediction Formula and StdErr Formula after your model is fit: Pred Formula height and PredSE height as in my example of regressing height on weight;

The data with the saved scripts and formula columns, Prediction Intervals.JMP, can be down loaded from here

jiancao

JMP 17 Update:

The minor difference noted above due to using the t-quantile vs the normal quantile has been resolved; JMP 17 uses the t-quantile in Graph' Profiler as is used in Indiv Confidence Interval