I want to find significant factors in the process of spherical crystallization. I did Definitive Screening Design with 6 factors on three levels, with four additional runs (2 fake factors). I am measuring several different responses.
I analysed the data with Fit Definitive Screening and got the results which make sense. When I went on and made a model, using Standard Least Squares, I noticed, that the Parameter Estimates are different in a model made with Standard Least Squares and in Fit Definitive Screening . p- values are the same for some factors and very close for others. It worries me also, that in one factor (V ratio) the Estimate got the minus sign (See the photo below) . That changes the model a lot. (+ sign is more logical, based on the understanding of the process). As I know, the Parameter Estimates should be the same. How can I get the same Parameter Estimates?
When I analysed my other responses, I also got different Parameter Estimates, but the sign (+/-) stayed the same. However, some results in Standard Least Squares report in Prediction Profiler were very unlogical (I got the negative number for the mean diameter of a particle). That makes me think that there is something wrong with the Standard Least Squares report, hence I also can't trust the data in the Actual by Predicted Plot, etc. Is there any better way to evaluate the model fitting?
I also noticed, that in the Standard Least Squares report, the Prediction Profiler has Yield on the Y axis, despite the fact, that I entered Root Square Yield as a response. On the axis of Actual by Predicted Plot, Root Square Yield is used, as it should. How can I fix that problem?
Did you design the experiment for Yield and then create a new column for Root Square Yield? The table scripts won't capture that change so you might be comparing 'apples to oranges.'
Also, did you use the Model script in the data table to get Fit Least Squares or did you click Run Model at the bottom of the Fit Definitive Screening platform? Again, it might make a difference. The fact that you observe Yield, not Root Square Yield, that makes me wonder.
On the other hand, did you enter Yield at the Y variable and the apply a transformation in the Fit Model dialog box? That action would also provide the untransformed response in the prediction profiler.
I know I am wrong but I am just trying to find where the discrepancy comes from.
No, I transformed Yield in the window, where you have to select response and factors, prior to Fit Definitive Screening.
I then clicked Make Model and leave everything as it was (Square Root Yield is selected as a Response and factors, which were find active in Fit Definitive Screening as factors). If I just click run model I get the same result.
I want to emphasize, that I get different Parameter Estimates also when I analyse untransformed response. However , Parameter Estimates in the Standard Least Squares report are the same I get, if I analyse the same data with Stepwise regression. If I compare the result I get from transformed data, all three Parameter Estimates from different analysis are different.
I am aware that we are talking about different analysis, but Dr. Bradley Jones states in the video http://video.sas.com/detail/videos/analytics-in-action/video/4492688979001/a-new-analytical-approach..., that the Parameter Estimates should be the same. That is also true for the example data in JMP for Fit Definitive Screening (Extraction 3). (In this example, Parameter Estimates are the same with all three different analysis, which are mentioned above).
I am familiar with the published results (journals, JMP Help, and Dr. Jones) and I have reproduced them without any problem. The same is true of other experiments, which I cannot share here. There has never been a discrepancy.
You have clearly stated your actions and shared your results, which I cannot resolve. I suggest that you submit this case to JMP Technical Support (firstname.lastname@example.org) and then share with their findings here. You should be able to use Fit Definitive Screening with confidence.
Remember that the Fit Definitive Screening and Fit Least Squares are very different analyses. The former is a two-stage analysis where as the latter is one stage. The former combines the results from the two stages where as the second obtains all the results simultaneously.
I solved the problem. The problem was, that I was analysing data sheet, where I entered the minimum value of factor as -1, a middle value as 0 and a maximum value as 1. That made no difference in the results from Fit Definitive Screening, but apparently a big difference with Standard Least Squares and Stepwise Regression. When I entered "real" values in the data sheet, the results from all three analyses were the same, as expected.
Thank you for your help.