Testing for poolability of all factors and factor combinations” in stability extrapolation as per ICH. How to do it in JMP?

[#1]

As per ICHQ1E guideline “section B.3.2.2 Testing for poolability of all factors and factor combinations” all factor combination can be used to test poolability. if other factors like Pack Size and Orientation need to include in poolability and Model selection, how to do it in JMP? Do we have provision to do this analysis in JMP with consideration of other factors?

[#2 · Accepted solution]

I think that you can directly extend the example with batch poolability. Use tests for the factors and their interactions with time. If they are not significant, then you can pool the data over all of the factor levels. If the interaction is significant, then you must leave that factor and interaction in the model. You can use the Prediction Profiler to determine expiry after selecting the final model.

Use Analyze > Fit Model. Select the data columns for time and factors, click Macros, and select Factorial to Degree. Click Run. Remove terms that you decide are not significant, starting with interactions first. Do not break model hierarchy.

[#3]

Thanks Mark for the solution. yes I am  totally agree to the point and was first choice for model selection in JMP the way you describe it. Please let suggest how to conduct ANCOVA with multifactor in JMP. Similarly with Multiple factors there will be  more number of possible combinations for comparing slope and intercept significance between batches. I am referring to the below mentioned article(link). Unfortunately I don't have hard copy of this article. https://www.yumpu.com/en/document/read/7434354/shelf-life-estimation-for-multifactor-stability-studi... 

Please suggest. Hope this article will help explaining query i have.

Regards.

rsomankar

[#4 · Accepted solution]

I will use the Big Class example that is installed with JMP. I want to model the linear relationship between :weight and :height. There are also the :age and :sex factors. I enter these terms in the model to test their effect I cross :age and :sex with :height to produce the interaction terms. These terms test if the :height effect depends on the :age or :sex.

Here are the initial results:

We would then follow the Q1A and Q1E guidance and test higher order terms first. We eliminate the least significant term (i.e., highest p-value) first and proceed one step at a time. Then we might arrive at a model like this one for interpretation and prediction:

Such a result is evidence that the data across :sex and :age groups may be pooled, but there is a difference across :age groups.