Stepwise regression with heredity restriction

[#1]

Dear all, 

I want to automate a model selection process via scripting using the Fit model platform utilizing stepwise regression for a stability analysis (I also know the stability platform in JMP but for "reasons" I want to use Fit model).  

My terms are Storage time, Batch ID and Storage Time*Batch ID. Using Stepwise regression the idea is now to leave terms within the model with an p-value <0.25. An initial script looked (in parts) like this:

Y( :response ),
	Effects(
		:"Storage Time (2-8°C Analysis)"n, :Batch ID, :"Storage Time (2-8°C Analysis)"n * :Batch ID,
	),
	Center Polynomials( 0 ),
	Personality( "Stepwise" ),
	Run(
		Stopping Rule( "P-value Threshold" ),
		Prob to Leave( 0.25 ),
		Direction( "forward" ),
		Rules( "Whole Effects" )
	)

My problem is now, that by using the Rule "Whole effects", terms are excluded following not the hierarchical principle, meaning that my model contained after using stepwise regr. e.g. The interaction and time effect, but not the batch effect. See also:

I know that there is also the combine or restrict rule (to have hierarchy), however by applying these rules my term Batch ID (multi-level) is splitted into different groups which is not my desire. I really want just to have the whole effect Batch ID if it is significant (or its interaction with time). 
So is there a way to combine the "whole Effect" term display with the hierarchical principle for model selection? A script would also work for me.
I hope I could clarify my question.

Thanks in advance for any solution or remarks! 

Best Johann

[#2]

I believe that you cannot combine rules. You can select and use only one rule at a time.

You have a simple case of linear instability with the possibility of differences between lots. You might avoid the limitations and overhead of Stepwise. The model selection might be easier using a direct approach based on Fit Least Squares instead of Stepwise since you have only three terms. Test for poolability (model selection) if you are following ICH Q1E and Q1A.

1. Fit the full model (intercept, batch, time, interaction) with Fit Least Squares.
2. Examine the p-values for the interaction.
3. If the interaction is significant, then estimate the expiry for each batch and choose the shortest life.
4. Fit the reduced model without the interaction term if it is not significant and use a common slope.
5. If the batch is significant using the second model, then estimate the expiry for each batch and choose the shortest life.
6. Fit the reduced model without the batch effect if it is not significant in the second fit and use a common intercept.
7. If time is significant, determine the expiry using the reduced model.

Fit Least Squares supports inverse prediction, too, so you can immediately determine expiry for any of the three poolability cases.

[#3]

Hi Mark,

thank you very much for your fast reply. The ICHQ1E based procedure you mentioned is familiar to me as well as how to perform the analysis in Fit Model. The point is that I want to automate the model selection process (exclude terms with p-value >0.25 following the hierachical principle) for our QC-department. Do you think that there is a way to script this? 

 I don't want to use the stability platform in JMP as this platform uses for different intercept/different slope the unpooled MSE. However ICHQ1E is requesting the pooled MSE for the analysis, which is indeed also used by utilizing the Fit Model platform (also with different intercept/different slope model). And utilizing either pooled or unpooled MSE could have significant impact on expiry date calculation.  As far as I know JMP 17 has now a checkbox to preset the usage of the pooled MSE in the stability platform. Unfortunately neither me nor our QC department have this version. That's why I'm looking for an automated "work-around" via script using the fit model platform. 

Kind Regards

Johann

[#4]

---

@JohannD wrote:

As far as I know JMP 17 has now a checkbox to preset the usage of the pooled MSE in the stability platform. Unfortunately neither me nor our QC department have this version.

---

JMP 17 is a complimentary upgrade included in most licenses. Contact your organization's JMP software administrator and encourage them to get JMP 17. They were recently sent information about how to get it.

If you don't know who your JMP software administrator is you can contact Tech Support (support@jmp.com) and they can help put you in contact.