Hello,
While using Loglinear Variance model, there is an inconsistency in the model for the Variance model. I put up an example below to explain the situation.
I build 2 models using Loglinear Variance:
Model 1 : Response A. Factors A, B, C.
Model 2 : Response A , B. Factors A, B, C.
Note: Response A and Factor A,B,C are the same data for both Model 1 and 2.
Comparing Model 1 to Model B , we focus on the Mean Model and Variance model for Response A.
The Mean Model for Response A is exactly the same between Model 1 vs Model 2 (the green background).
However, the Variance (or Stdev) model for Response A between Model 1 vs Model 2 (the red background) is totally different.
The comparison is also done with the same Factor settings and same scaling.
It seems that by just adding another response to the model (in this case we add Response B to the model with Response A) the Variance model become different. Response A and B are independent and should not affect each other.
Is this situation with Loglinear Variance considered normal and can be trusted?
Thanks to advise.
B.r,
Chris
Hi @ChrisLooi,
I tried to reproduce the situation you have with the JMP sample dataset InjectionMolding.jmp and creating a new response column and two models, one with response Shrinkage and one with Shrinkage and Shuffle[Shrinkage] responses, with the same terms in both models.
I see no differences in the profiler when creating a model with one or two responses, using Loglinear Variance :
Do you have the same terms in the Variance Effects panel from Fit Model window ?
Could you share an anonymized dataset where we could reproduce the problem ?
Which version of JMP do you use ?
Hi @ChrisLooi,
One preliminary remark : even if your responses may be individually measured, they are not rigourously independent, as a strong correlation can be seen between the two responses :
Then, I just had a quick look at the datatable provided and noticed that you have a lot of missing value for response B.
Should some rows be identicals, or do you have a different granularity/detail level in your data between response A and B ?
I think the missing values in B may cause the problem you encounter.
Here is the profiler with the two responses "as it is" (same as the one you showed previously) :
However, when doing automated data imputation technique on column "response B" to avoid missing values (and using this imputed response B as the second response), the profiler show better behaviour for response A and B (and profiler of response A is now identical to the profiler when the response A is the only response modelized) :
You can also have Profilers with the same behavior by hiding and excluding rows with missing values for B, to only keep for the modeling the rows having complete informations for both responses A & B :
So there may be a correction to do on your datatable, to fill up the missing values, impute them or exclude rows with incomplete information between responses A and B. The difference of granularity in your data (and resulting missing values) might be responsible for the change you see in profilers, as you imply the same model for both responses, but the limited amount of information in column B does disturb the variance model (as it needs more observation for each terms in the model than the mean model : Loglinear Variance Models)
You can also fit the two responses independantly (launch the platform for one response only) with a different model (probably a smaller model for response B since you have less observations available for the variance model).
Attached is the datatable and scripts used,
