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
- :
- How do I make a prediction model based on my old experiment data?

- 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

Aug 26, 2015 1:54 PM
(1565 views)

I am new to JMP software ( version 11.2.0).

I have a question on prediction model. Before having JMP, I ran 2 experiments shown below. ISOD is the response, and temperature & aging time are both continuous. **Using these 2 tables, can I make a prediction model for ISOD yield in temperature range 100 to 180 & aging time 1 to 20 together? For an example, what will ISOD be at like 150 C with 15 days of aging time?** I was not able to do it. I know that I can make matrix with DOE for prediction model, but that means I have to do experiments again.

experiment 1

temperature (C) | aging time (days) | ISOD |

100 | 2 | 14 |

140 | 2 | 79 |

180 | 2 | 67 |

experiment 2

temperature (C) | aging time (days) | ISOD |

100 | 1 | 0.3 |

100 | 10 | 75 |

100 | 20 | 99 |

Thank you for your time, and I would appreciate your help.

6 REPLIES

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

You can do what you desire knowing the context of the data. It is happenstance data and not data originating from a controlled designed experiment which would provide additional information around any interactions.

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

Thank you, Lou.

I really appreciate your help. I see what you did. You just added T & aging time into factors and ran standard least squares/minimal report. I have 3 additional questions on the report.

1. I watched one tutorial video, and a guys mentions that Prob >F should be less than 0.01 as a rule of thumb. Mine is shown as 0.0938. Do I play around with "model effect" and "degree" in "fit model" window to decrease this value?

2. My response ISOD is percentage and it should be in range of 0-100. In prediction profile for 150 C and 15 days, I can understand ISOD being 120 % as 100 %, but is there a way to set so that prediction does not go over 100?

3. According to this prediction model, I get ISOD = 106 at 100 C and 20 days. My actual data has ISOD = 99 at that condition. Is there a way to refine it?

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

YIKES!!!

I would be very dubious about predicting any results outside the experimental space. You have only one data point at 150+ deg. C and then only 2 days aging. You also have only one data point with more than 15 days aging and that is at only at 100 deg. C . Therefore, predicitng at 150 deg C and 15 days aging would be very precarious. My advice: Perform more experiments designed around the parameters within which you wish to be able to predict with any confidence/accuracy.

Steve

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

Aug 27, 2015 5:37 AM
(1315 views)
| Posted in reply to message from Steven_Moore 08/27/2015 08:13 AM

Hi smoore2,

You are right. The error in prediction model is probably too big to be useful. I need more data points to have more accurate prediction model... .

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

Your original experiment looks like an OFAT...one in which One-Factor-at-A-Time is varied. It's not a very efficient way to generate information.

Like Lou Valente and smoore2 recommend, you might consider running a designed experiment. "More data points" is not all you need to form a more accurate prediction model.

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

A couple more ideas:

Instead of creating a totally new designed experiment or repeating the same experimental runs, you can add experiments to your current data using the "Augment" feature. With your current data table open, select **DOE > Augment Design** from the menu bar. Enter ISOD as the Y and the other two variables as X's.In the Augment Design window, click on the Augment button. For a standard model that is useful for prediction (interactions and curvature effects), click on the RSM button. The new designed experiment will consist of your 6 experimental runs plus several more runs to complete.

Regarding predicted values above 100: one could do a Logitpct transform on the response data. I would want to know more about the nature of the response before transforming it, but this would guarantee that the model predictions stay between 0 and 100. In the model dialog, put ISOD in the Y role, click on it to select it, and then click on the Transform red triangle and select LogitPct.

Howard