turn on suggestions

Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

- JMP User Community
- :
- Discussions
- :
- Discussions
- :
- Negative numbers in prediction profiler when zero is limit

Topic Options

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Jul 21, 2014 6:05 AM
(3615 views)

I have attached an 8 mixture custom DoE in which 3 ingredients are constant. When I run the model, the prediction profiler predicts the maximum Y1 & Y2 within reason of the actual maximum. But when I minimize the desirabilities, I get negative values. How is it getting these numbers and how would I prevent any predictions falling well below the actual low response values?

1 ACCEPTED SOLUTION

Accepted Solutions

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Jul 21, 2014 6:27 AM
(6122 views)

Bwanders,

The numbers for Y in the Prediction Profiler are coming from the prediction formula for your model. You can see the estimated model by looking at the Parameter Estimates. In your particular case, because there is no replication, the model is a perfect fit for the data. However, when you are trying to minimize the Y response, the prediction profiler searches the design space and finds a location with a negative predicted Y.

Linear regression assumes a continuous response, so a negative value is just fine as far as linear regression is concerned. If you wish to restrict the predictions to only positive results, a different modeling technique should be used that can enforce that restriction. However, the real culprit here is that the combination that provides that negative Y prediction is an extrapolation. The location that minimizes Y is the first set of conditions in your table except it is X1 Type = B instead of A. You never ran that combination. When you have a low level of X1, switching to X1 Type = B will lower the response (see the Prediction Profiler). In this case, low enough to give a negative value. The trends are what was shown in the data you did collect, but leads to an extrapolation because not all combinations were collected. My guess is that if you were to actually run that condition you would get a very low Y value.

Dan Obermiller

1 REPLY

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Jul 21, 2014 6:27 AM
(6123 views)

Bwanders,

The numbers for Y in the Prediction Profiler are coming from the prediction formula for your model. You can see the estimated model by looking at the Parameter Estimates. In your particular case, because there is no replication, the model is a perfect fit for the data. However, when you are trying to minimize the Y response, the prediction profiler searches the design space and finds a location with a negative predicted Y.

Linear regression assumes a continuous response, so a negative value is just fine as far as linear regression is concerned. If you wish to restrict the predictions to only positive results, a different modeling technique should be used that can enforce that restriction. However, the real culprit here is that the combination that provides that negative Y prediction is an extrapolation. The location that minimizes Y is the first set of conditions in your table except it is X1 Type = B instead of A. You never ran that combination. When you have a low level of X1, switching to X1 Type = B will lower the response (see the Prediction Profiler). In this case, low enough to give a negative value. The trends are what was shown in the data you did collect, but leads to an extrapolation because not all combinations were collected. My guess is that if you were to actually run that condition you would get a very low Y value.

Dan Obermiller